Mechanistic Modelling in Pig and Poultry Production

MECHANISTIC MODELLING
IN PIG AND POULTRY

PRODUCTION
This page intentionally left blank
MECHANISTIC MODELLING
IN PIG AND POULTRY
PRODUCTION
Edited by
R. Gous
University of KwaZulu-Natal
Pietermaritzburg
South Africa
T. Morris
University of Reading
Reading
UK
and
C. Fisher
Consultant
Midlothian
Scotland
CABI is a trading name of CAB International
CABI Head Office CABI North American Office

Nosworthy Way 875 Massachusetts Avenue
Wallingford 7th Floor
Oxfordshire OX10 8DE Cambridge, MA 02139
UK USA
Tel: +44 (0)1491 832111 Tel: +1 617 395 4056

Fax: +44 (0)1491 833508 Fax: +1 617 354 6875
E-mail: cabi@cabi.org E-mail: cabi-nao@cabi.org
Website: www.cabi.org
© CAB International 2006. All rights reserved. No part of this publication

may be reproduced in any form or by any means, electronically,
mechanically, by photocopying, recording or otherwise, without the prior
permission of the copyright owners.
A catalogue record for this book is available from the British Library,
London, UK.
A catalogue record for this book is available from the Library of Congress,
Washington, DC.
ISBN-10: 1–84593–070–3
ISBN-13: 978–1-84593–070–7
Typeset by Columns Design Ltd, Reading

Printed and bound in the UK by Cromwell Press, Trowbridge
Contents
List of Contributors vii

Preface ix
Acknowledgements xi
1 An Introduction to Modelling in the Animal Sciences 1
T.R. Morris
2 Scientific Progress and Mathematical Modelling: Different 6
Approaches to Modelling Animal Systems
J. France and J. Dijkstra
3 Basic Concepts Describing Animal Growth and Feed Intake 22
N.S. Ferguson
4 The Effects of Social Stressors on the Performance of 54
Growing Pigs
I.J. Wellock, G.C. Emmans and I. Kyriazakis
5 Modelling Populations for Purposes of Optimization 76
R.M. Gous and E.T. Berhe
6 Advancements in Empirical Models for Prediction and 97
Prescription
W.B. Roush
7 The Problem of Predicting the Partitioning of Scarce 117
Resources during Sickness and Health in Pigs
I. Kyriazakis and F.B. Sandberg
8 Nutrient Flow Models, Energy Transactions and Energy 143
Feed Systems
J. van Milgen
v
vi Contents
9 Evaluating Animal Genotypes through Model Inversion 163

A.B. Doeschl-Wilson, P.W. Knap and B.P. Kinghorn
10 Considerations for Representing Micro-environmental 188
Conditions in Simulation Models for Broiler Chickens
O.A. Blanco and R.M. Gous
11 Using Physiological Models to Define Environmental 209
Control Strategies
M.A. Mitchell
12 Modelling Egg Production in Laying Hens 229
S.A. Johnston and R.M. Gous
13 Comparison of Pig Growth Models – the Genetic Point of View 260
P. Luiting and P.W. Knap
14 Mechanistic Modelling at the Metabolic Level: a Model of 282
Metabolism in the Sow as an Example
J.P. McNamara
15 The Place of Models in the New Technologies of 305
Production Systems
D.M. Green and D.J. Parsons
Index 325
Contributors
E.T. Berhe, Animal and Poultry Science, School of Agricultural Sciences and
Agribusiness, University of KwaZulu-Natal, Private Bag X01, Scottsville 3209,
South Africa.
O.A. Blanco, Animal and Poultry Science, School of Agricultural Sciences and
South Africa.
J. Dijkstra, Animal Nutrition Group, Wageningen Institute of Animal Sciences,
Wageningen University, Marijkeweg 40, 6709 PG Wageningen, The
Netherlands.
A.B. Doeschl-Wilson, Sygen International, Scottish Agricultural College, Bush
Estates, Penicuik, Edinburgh, EH26 0PH, UK.
G.C. Emmans, Animal Nutrition and Health Department, Scottish Agricultural
College, West Mains Road, Edinburgh, EH9 3JG, UK.
N.S. Ferguson, Maple Leaf Foods Agresearch, 150 Research Lane, Guelph,
Ontario, Canada, N1G 4T2.
J. France, Centre for Nutrition Modelling, Department of Animal and Poultry
Science, University of Guelph, Guelph, Ontario, Canada, N1G 2W1.
D.M. Green, University of Oxford, Department of Zoology, South Parks Road,
Oxford, OX1 3PS, UK.
R.M. Gous, Animal and Poultry Science, School of Agricultural Sciences and
South Africa.
S.A. Johnston, Animal and Poultry Science, School of Agricultural Sciences and
South Africa.
B.P. Kinghorn, Sygen International, Scottish Agricultural College, Bush Estates,
Penicuik, Edinburgh, EH26 0PH, UK.
P.W. Knap, PIC International Group, Ratsteich 31, D-24837 Schleswig, Germany;
and Sygen International, Scottish Agricultural College, Bush Estates, Penicuik,
Edinburgh, EH26 0PH, UK.
vii
viii Contributors
I. Kyriazakis, Animal Nutrition and Health Department, Scottish Agricultural

P. Luiting, PIC International Group, Ratsteich 31, D-24837 Schleswig, Germany.
J.P. McNamara, Department of Animal Sciences, Washington State University, PO
Box 646351, Pullman WA 99164-6351, USA.
M.A. Mitchell, Roslin Institute, Roslin, Midlothian, EH25 9PS, UK.
T.R. Morris, School of Agriculture, Policy and Development, The University of
Reading, New Agriculture Building, PO Box 237, Reading, RG6 6AR, UK.
D.J. Parsons, Cranfield University, Silsoe, Bedford, MK45 4HS, UK.
W.B. Roush, USDA-ARS Poultry Research Unit, Mississippi State, MS 39762,
USA.
F.B. Sandberg, Animal Nutrition and Health Department, Scottish Agricultural
J. van Milgen, INRA – UMR SENAH, Domaine de la Prise, 35590 Saint-Gilles,
France.
I.J. Wellock, Animal Nutrition and Health Department, Scottish Agricultural
Preface
This volume records the proceedings of a conference held in South Africa

in April 2005 with the title ‘Recent Advances in Pig and Poultry Modelling’.
The Conference, organized by the University of KwaZulu-Natal and the
South African Branch of the World’s Poultry Science Association, brought
together scientists from several countries and from different modelling
traditions to share their ideas and recent developments. The papers
published here create a permanent record of these deliberations.
The Conference was held at the Ithala Game Reserve in KwaZulu-Natal
from 13–16 April 2005 and was attended by 65 delegates. Ithala offered an
unusual and stimulating location for the meeting and scientific sessions were
interspersed with the viewing of wild game and the exploration of a very
beautiful wild bush. A team of staff and students from the Department of
Animal and Poultry Science, University of KwaZulu-Natal, Pietermaritzburg,
led by Professor Rob Gous, ensured that the requirements for a successful
meeting were fully met. Several companies supported the Conference by
sponsorship as shown under Acknowledgements.
The meeting had three main aims: to provide a discussion and record
of recent developments in the mechanistic modelling of pig and poultry
production systems; to provide a written record of these discussions; and
to mark the contribution of retiring Professor Trevor Morris of the
University of Reading to the field of animal modelling and systems.
Mechanistic modelling of animal systems has already provided a great
deal of understanding of the underlying principles of pig and poultry
nutrition and production. This is an ongoing process and a review at this
time is particularly appropriate. Notable amongst these papers is the
consideration of new components of the animal production process, such
as social stressors and disease. Also the understanding of some new
systems, such as the physiological control of egg production in hens, has
benefited greatly from the development of a modelling approach. The
integration of modelling into the wider aspects of animal production
ix
x Preface
systems is an area that is developing quite rapidly at this time. On the more
general questions of animal modelling, many different philosophies and
approaches are viable and the reader will find a reflection of this diversity
in these pages. The day-to-day application of modelling in management
decision-making is still some way off but the progress towards this ideal is
reflected in this book.
Professor Trevor Morris led the way in the application of quantitative
methods in poultry production and nutrition. Mainly using statistical
methods but always with a clear view of the underlying mechanisms, for
many years he showed the benefits of combining the results of different
experiments into a set of simple and applicable quantitative rules. Whilst
this is some way from modern, computer-based mechanistic modelling,
these earlier ideas undoubtedly showed the way and, in particular, trained
and motivated many students who later became modellers. Professor
Morris is now retired and this volume is warmly dedicated to him in
recognition and appreciation of his work and contribution. It will be an
appropriate testament to his work if this volume encourages some young
animal scientists to see how the application of modelling techniques and
ideas can enhance their own work.
R.M. Gous
Acknowledgements
Financial support for this symposium is gratefully acknowledged from:
Aviagen
Degussa
EFG Software (Natal)
Elsevier Publishing
Maple Leaf Foods
University of KwaZulu-Natal
World’s Poultry Science Association
xi
This page intentionally left blank
1 An Introduction to Modelling in
the Animal Sciences
T.R. MORRIS
School of Agriculture, Policy and Development, The University of Reading,
New Agriculture Building, PO Box 237, Reading, RG6 6AR, UK
trevor@trevormorris.wanadoo.co.uk
Scientists and engineers use models to represent parts of what they regard
as the real world; to help them to convey to others an understanding of the
way in which things work and, sometimes, to help them to make
predictions about the consequences of alternative courses of action.
Some models are pictorial and none the less useful for that. We could
draw a diagram of the digestive system of a pig that would be helpful in
explaining how the animal converts its food into components that can be
absorbed, leaving a residue to be excreted. Note that the diagram does not
need to look much like the real guts of a pig and, indeed, a colour
photograph of an alimentary tract, although more ‘accurate’ than a simple
line drawing, is actually less suitable for our purpose. A single photograph
would not reveal the teeth, the salivary glands, the bile duct, the pancreas
and the hepatic portal vein. For similar reasons, an accurate scale map of
the London underground railway system, showing all its tracks, stations and
platforms might be valuable to a maintenance engineer but is not helpful to
the visitor trying to find his or her way around London. The map which
does appear on the walls of London underground stations is a much
simplified diagram, showing the relationships between stations on a given
line (but not their real geographical locations) and the interconnections
between lines, which are distinguished by the use of colour codes. This
model is the work of an electrical engineer who knew how to draw a circuit
diagram and applied his skill to the problem of making the planning of a
journey as simple as possible. What these examples tell us is that models
have to be sufficient for their purpose and, therefore, that we must define
the purpose carefully before setting out to construct a model.
A diagram of a pig’s digestive system and a map of an underground
system are examples of models which represent something in the real
world, but they do not involve equations. Wordsworth (1798) wrote that
‘poetry is emotion recollected in tranquillity’ and we might argue that
© CAB International 2006. Mechanistic Modelling in Pig and Poultry Production

(eds. R. Gous, T. Morris and C. Fisher) 1
2 T.R. Morris
‘science is observation encapsulated in equations’. The aim of most

scientific research is (or should be) to reduce a mass of observational detail
to equations which have powerful predictive value. Galileo could (and did),
like others before him, construct models of the solar system which were
helpful when explaining the relative motion of the sun and its planets; but
Newton explained why the planets moved in elliptical orbits around the
sun and provided equations capable of predicting their future positions.
The hypothesis, later derived from Einstein’s work, that the ultimate
future of the sun itself (not just our sun – any sun) is limited by natural
processes (as distinct from divine judgement), does not detract from the
value of Newton’s equations for any planetary predictions that we mortals
might wish to make. The proposition that light travels in straight lines has
been shown to be false by quantum theory, but it still works on the small
scale of our solar system and will do well enough for landing a man on the
moon (and bringing him home again). The validity of any particular model
can only be judged in relation to the purpose it is intended to serve.
The early history of the application of models in animal agriculture
was well reviewed by Baldwin and Hanigan (1990). Two strands became
interwoven with very fruitful results. On the one hand there were the early
models of energy and protein utilization due to Kellner (1909) and Armsby
(1917), later codified by Brody (1945). On the other hand, there was the
work of economists who sought predictive models for systems and, in our
particular case, for agricultural systems. These economic models were
usually designed to maximize profit in a particular system or to minimize
costs. The use of linear programming to maximize profit within a farm
business, and the application of the same linear programming metho-
dology to minimizing the ingredient cost of a diet for a particular set of nu-
trient specifications and raw materials, were early examples of economic
modelling applied to agriculture.
Fawcett, an economist, wrote in 1973 that ‘much effort is now being
applied to simulation techniques for the purpose of designing optimal
processes, but simulation is no substitute for mathematical analysis’.
Whittemore and Fawcett (1974, 1976) were subsequently among the first to
combine the thinking of economists and animal nutritionists to produce, in
their case, a simulation model for the growing pig. Meanwhile, Morris
(1968) had quantified the response of laying hens to changes in dietary
energy concentration and had shown how this could lead to optimizing
energy level in relation to any particular set of ingredient prices (Morris,
1969). Fisher and Wilson (1974) did the same thing for the energy content of
broiler diets. These last were examples of quantitative models, derived
empirically and not based upon any direct understanding of the mechanisms
at work.
A model of the laying hen’s response to amino acid intake (Fisher et al.,
1973), on the other hand, was explicitly derived by considering the uti-
lization of the limiting amino acid for the synthesis of egg and body
proteins. Here, theory and experimental data fitted neatly together.
McDonald and Morris (1985) subsequently argued that, since the theory
Introduction to Modelling in Animal Sciences 3
could be shown to fit all the available experimental data for lysine,

methionine, tryptophan, isoleucine and valine, it should be adopted as the
preferred method of calculating responses to these and other amino acids,
rather than undertaking more experiments.
The potential of good models to obviate the need for further
experiments seems often to have been overlooked. Workers will often plan
an experiment to answer a particular question about the response of one
current genotype to nutrient or environmental inputs, without stopping to
consider how this might help to predict the response of future genotypes
(and tomorrow is part of the future) to a somewhat different array of input
variables. Worse, they may use the results of such trials to list ‘requirements’
for current genotypes (e.g. Chiba, 1999; Leeson and Summers, 1999). This
ignores the important proposition that ‘nutrient requirements’ cannot be
defined for groups of chickens or pigs for three reasons. First, the response
of any group of animals to increasing inputs of any limiting variable is
curvilinear (Morris, 1983): this means that an optimum input can be
determined, but it should not be labelled a ‘requirement’. Secondly, the
response curve will shift with changes in the potential output of the group
of animals being considered. Thirdly, the position of the optimum on the
curve will shift with changes in the cost of the input or the value of the
output. Thus it is possible to produce equations defining curves which
represent the mean response of groups of animals (with defined potential)
to various inputs, but this cannot lead to a calculation of the optimum dose
until prices have also been defined. The marriage of economic thinking and
nutritional knowledge, which took place some 30 years ago, is indissoluble.
The scope of this publication has been limited to mechanistic
modelling, a term which seems to have been introduced by Thornley and
France in 1984. By this we mean models which are quantitative and which
aim to represent the underlying mechanisms that produce end results.
This is in contrast to other quantitative models, which use equations
derived from observations in the real world, but not necessarily
representing any understanding of the causal mechanisms at work. For
example, we might construct an epidemiological model showing how the
risk of heart attack in a given population is affected by factors such as
smoking, excess body weight or physical exercise. This would be a good
description of the historical data analysed and might be a valuable
predictor of future risk but does not require any understanding of how
these risk factors actually alter the frequency of heart attacks in a
population. Conversely, a team of doctors might establish quite convincing
explanations of the causal connection between excess energy intake,
atherosclerosis and myocardial infarction, but would not, from that evidence,
be in a position to predict how many heart attacks would be avoided by a
defined reduction in body mass index.
Causal, or mechanistic, models are much to be preferred when we can
find them. This is because they are much more likely to be robust and to
apply to situations outside the range of conditions actually tested. You can
travel the universe with an equation such as E = mc2 and expect it to be
4 T.R. Morris
obeyed everywhere. However, there are many cases where we are not yet
in a position to build a causal model from fundamental components and
we must therefore, for the time being, rely on empiricism in those cases.
Two examples will illustrate this point. We know that changes in photo-
period have marked effects on the rate of sexual development in pullets
and we can trace changes in the concentration of gonadotrophic hormones
in the blood, following an increase or decrease in photoperiod. We are
confident that the brain responds to photoperiod by altering the flow of a
releasing hormone from the hypothalamus to the pituitary gland, which in
turn adjusts the flow of gonadotrophins to the ovary. However, this
knowledge does not put us in a position to write equations representing
the effect of a stated series of photoperiods during rearing upon the age at
which the pullet will lay her first egg. We can, however, provide quite
robust empirical equations to predict age at first egg for any specified
pattern of photoperiod applied during rearing (Lewis et al., 2002, 2003). It
is perhaps going too far to say that this particular gap between theory and
empiricism will never be bridged, but it seems unlikely that a full
mechanistic model would have any better predictive capability than the
present empirical one; and so it may be that further effort to quantify the
mechanism is not justified. A second example is the response to added
copper in a diet. A small amount of copper (about 6 mg/kg diet) is needed
for metabolic purposes, but this level is supplied by almost all natural diets.
If copper sulphate is added to raise copper levels to 50–100 mg/kg, growth
rate of pigs and baby chicks is enhanced. This effect is attributed to
modification of the gut flora in a manner that is beneficial to the host
animal. At higher concentrations, copper begins to be toxic and growth is
depressed. All these effects, copper deficiency at very low levels, copper as
a growth promoter in the medium range and copper toxicity at high doses,
can be demonstrated and quantified by appropriate trials, but this does not
lead to a theory connecting the cause to the effects in a mechanistic model.
On the other hand, we could produce a well researched response curve,
which could be reliably used to predict the effect of adding copper
sulphate to pig and poultry diets.
There are many more examples available where we have no immediate
prospect of being able to develop a good mechanistic model, even though
we believe that we understand a good deal about the mechanisms involved.
Empirical modelling is therefore not to be despised if it is the best available
tool for solving a particular problem. What is not acceptable is the use of
empirical modelling where others have already developed a mechanistic
model capable of resolving the particular question being approached
empirically. I believe that the modelling of growth and development in
pigs and poultry has now reached a sufficiently advanced state that
empirical models in this area can no longer be justified.
Introduction to Modelling in Animal Sciences 5
References
Armsby, H.P. (1917) Nutrition of Farm Animals. Macmillan, New York.
Baldwin, R.L. and Hanigan, M.D. (1990) Biological and physical systems: animal
sciences. In: Jones, J.G.W. and Street, P.R. (eds) Systems Theory Applied to
Agriculture and the Food Chain. Elsevier Science, Barking, UK, pp. 1–21.
Brody, S. (1945) Bioenergetics and Growth. Hafner, New York.
Chiba, L.I. (1999) Feeding systems for pigs. In: Theodoru, M.K. and France, J.
(eds) Feeding Systems and Feed Evaluation Models. CAB International,
Wallingford, UK, pp. 181–209.
Fawcett, R.H. (1973) Towards a dynamic production function. Journal of
Agricultural Economics 20, 543–549.
Fisher, C., Morris, T.R. and Jennings, R.C. (1973) A model for the description and
prediction of the responses of laying hens to amino acid intake. British Poultry
Science 14, 469–484.
Fisher, C. and Wilson, B.M. (1974) Response to dietary energy concentration by
growing chickens. In: Morris, T.R. and Freeman, B.M. (eds) Energy
Requirements of Poultry. Constable, Edinburgh, UK, pp. 151–184.
Kellner, O. (1909) The Scientific Feeding of Animals. Translated by Goodwin, W.
Duckworth, London.
Leeson, S. and Summers, J.D. (1999) Feeding systems for poultry. In: Theodoru,
M.K. and France, J. (eds) Feeding Systems and Feed Evaluation Models. CAB
International, Wallingford, UK, pp. 211–237.
Lewis, P.D., Morris, T.R. and Perry, G.C. (2002) A model for predicting the age at
sexual maturity for growing pullets of layer strains given a single change in
photoperiod. Journal of Agricultural Science, Cambridge 138, 441–448.
Lewis, P.D., Morris, T.R. and Perry, G.C. (2003) Effect of two opposing changes in
photoperiod upon age at first egg in layer-hybrid pullets. Journal of Agricultural
Science, Cambridge 140, 373–379.
McDonald, M.W. and Morris, T.R. (1985) Quantitative review of amino acid intakes
for young laying pullets. British Poultry Science 26, 253–264.
Morris, T.R. (1968) The effect of dietary energy level on the voluntary calorie
intake of laying birds. British Poultry Science 9, 285–295.
Morris, T.R. (1969) Nutrient density and the laying hen. In: Swan, H. and Lewis,
D. (eds) Proceedings of the Third Nutrition Conference for Feed Manufacturers.
University of Nottingham, Nottingham, UK, pp. 103–114.
Morris, T.R. (1983) The interpretation of response data from animal feeding trials.
In: Haresign, W. (ed.) Recent Advances in Animal Nutrition – 1983. Butterworths,
London, pp. 2–23.
Thornley, J.H.M. and France, J. (1984) Role of modelling in animal research and
extension work. In: Baldwin, R.L. and Bywater, A.C. (eds) Modelling Ruminant
Nutrition, Digestion and Metabolism; Proceedings of Second International Workshop.
University of California Press, Davis, California, pp. 4–9.
Whittemore, C.T. and Fawcett, R.H. (1974) Model responses of the growing pig to
the dietary intake of energy and protein. Animal Production 19, 221–231.
Whittemore, C.T. and Fawcett, R.H. (1976) Theoretical aspects of a flexible model
to simulate protein and lipid growth in pigs. Animal Production 22, 87–96.
Wordsworth, W. (1798) Preface. In: Wordsworth, W. and Coleridge, S.T. (eds)
Lyrical Ballads. Longmans, London.
2 Scientific Progress and
Mathematical Modelling:
Different Approaches to
Modelling Animal Systems
J. FRANCE1 AND J. DIJKSTRA2
1Centre for Nutrition Modelling, Department of Animal and Poultry Science,
University of Guelph, Guelph, Ontario, Canada, N1G 2W1;
2Animal Nutrition Group, Wageningen Institute of Animal Sciences,
Wageningen University, Marijkeweg 40, 6709 PG Wageningen, The

Netherlands
jfrance@uoguelph.ca
Introduction
A general understanding of science influences the scientific questions that
are asked, the choice of problems for scientific investigation and also how
these are attacked. A more widespread understanding of this topic might
enable a greater contribution to be made for the same effort. This chapter
attempts to describe what science is, how it progresses, the role and practice
of mathematical modelling and different approaches to modelling animal
systems. This is done with particular reference to animal science, with
examples from poultry and pigs. It represents, of course, a personal view.
Nature and Progress of Science

The zoologist E.O. Wilson has stated that science is ‘the reconstruction of
complexity by an expanding synthesis of freshly demonstrated laws’
(Wilson, 1978). This contrasts with the view of the famous engineer and
physicist Ernst Mach that science is a minimal problem (Mach, 1942).
Mach’s widely-accepted principle is cogently stated by the biologist and
geneticist J.B.S. Haldane who wrote ‘in scientific thought we adopt the
simplest theory which will explain all the facts under consideration and
enable us to predict new facts of the same kind’ (Haldane, 1927). However,
the philosopher Karl Popper, who has much to say on the nature of science

6 (eds. R. Gous, T. Morris and C. Fisher)
Scientific Progress and Mathematical Modelling 7
(Popper, 1968), stresses the importance of predictive ability rather than

simplicity or complexity. All this may be distilled into the statement that
science is about the correspondence of our ideas with the real world:
Ideas ↔ Real World (2.1)
In Eqn 2.1 ideas means such things as concepts, hypotheses or theories,
and the real world means the world contacted through our senses, extended
or not by instrumentation. Ideas are connected to the real world by means
of experiments. Often theoretical prediction (deduced from a scheme of
ideas or model) is compared with experimental data. If the experiments
(the interactions with the real world) are quantitative and numbers are
measured, the ideas should similarly be expressed numerically in order to
make a proper connection. To express ideas quantitatively, it is necessary
to use mathematics.
Most practising scientists share the views of Thomas Kuhn (1963) that
scientific progress is largely evolutionary, in contrast to those of Karl Popper
(1968) that science progresses entirely by a series of revolutions or
catastrophes. Kuhn argues that most scientists are conservative, seeking to
apply accepted methods and theories to new problems. When inconsistencies
in experimental data build up and some new paradigm is offered, a scientific
revolution occurs and more and more scientists abandon the old paradigm
in favour of the new one. Indeed, there is little doubt that from time to time
revolutions do occur, usually by the replacement of a theory by a more
embracing alternative theory. In trying to understand a particular
phenomenon, current theory must be taken as the starting point. An attempt
is made at connecting the corpus of current scientific knowledge to the
problem which concerns some aspect of the real world. This attempt to make
a connection will usually fail at the desired level of precision. However, after
perhaps repeating the experiment and modifying or extending the theory,
some success may be achieved. The scientist will then be better able to make
predictions and will feel he has arrived at a better understanding of the
problem. Going round this cycle (Fig. 2.1) again and again, the concepts and
ideas become more articulated and more precise, and are matched to nature
at more points and with more precision. It is stressed that movement is
always in the direction of increasing precision. At some point in the cycle it
will become necessary to use mathematics or mathematical modelling for
formulating the ideas and for making the connection between theory and
experiment. Thus, a set of mathematical equations or a model can be viewed
simply as an idea, a hypothesis or a relationship expressed in mathematics.
Role and Practice of Mathematical Modelling

Modelling is a central and integral part of the scientific method. As
phrased eloquently by Arturo Rosenbluth and Norbert Weiner,
the intention and result of a scientific inquiry is to obtain an understanding and
control of some part of the universe. No substantial part of the universe is so
8 J. France and J. Dijkstra
Fig. 2.1. Nature and progress of science.
simple that it can be grasped and controlled without abstraction. Abstraction

consists in replacing the part of the universe under consideration by a model of
similar but simpler structure. Models, formal or intellectual on the one hand, or
material on the other, are thus a central necessity of scientific procedure.
(Rosenbluth and Weiner, 1945)
Models therefore provide us with representations that we can use.
They provide a means of applying knowledge and a means of expressing
theory and advancing understanding (i.e. operational models and research
models). They are simplifications not duplications of reality. To quote from
an editorial that appeared in the Journal of the American Medical Association,
a model like a map cannot show everything. If it did it would not be a model
but a duplicate. Thus the classic definition of art as the purgation of superfluities
also applies to models and the model-makers problem is to distinguish
between the superfluous and the essential.
(Anon, 1960)
This is, of course, an affirmation of Occam’s Razor, that entities are not to
be multiplied beyond necessity.
To appreciate fully the role of mathematical modelling in the biological
sciences, it is necessary to consider the nature and implications of
organizational hierarchy (levels of organization) and to review the types of
models that may be constructed.
Organizational hierarchy
Biology, including pig and poultry science, is notable for its many
organizational levels. It is the different levels of organization that give rise
to the rich diversity of the biological world. For animal science, a typical
scheme for the hierarchy of organizational levels is shown in Table 2.1.

This scheme can be continued in both directions and, for ease of
exposition, the different levels are labelled …, i + 1, i, i – 1, …. Any level of
the scheme can be viewed as a system, composed of subsystems lying at a
lower level, or as a subsystem of higher level systems. Such a hierarchical
system has some important properties:
1. Each level has its own concepts and language. For example, the terms of
animal production such as plane of nutrition and liveweight gain have little
meaning at the cellular level.
2. Each level is an integration of items from lower levels. The response of
the system at level i can be related to the response at lower levels by a
reductionist scheme. Thus, a description at level i – 1 can provide a
mechanism for responses at level i.
3. Successful operation of a given level requires lower levels to function
properly, but not vice versa. For example, a microorganism can be
extracted from the caecum of a pig and grown in culture in a laboratory, so
that it is independent of the integrity of the caecum and the animal, but
the caecum (and hence the animal) relies on the proper functioning of its
microbes to function fully itself.
This organizational hierarchy helps to explain three categories of model:
teleonomic models which look upwards to higher levels, empirical models
which examine a single level, and mechanistic models which look downwards,
considering processes at a level in relation to those at lower levels. A more
detailed classification of models is given in Thornley and France (2006).
Teleonomic modelling
Teleonomic models (see Monod, 1975, for a discussion of teleonomy) are

applicable to apparently goal-directed behaviour, and are formulated
explicitly in terms of goals. They usually refer responses at level i to the
constraints provided by level i + 1. It is the higher level constraints that,
via evolutionary pressures, can select combinations of the lower level
mechanisms, which may lead to apparently goal-directed behaviour at level
i. Currently, teleonomic modelling plays only a minor role in biological
Table 2.1. Levels of organization.

Level Description of level
i+3 Collection of organisms (herd, flock)

i+2 Organism (animal)
i+1 Organ
i Tissue
i–1 Cell
i–2 Organelle
i–3 Macromolecule
modelling, though this role might expand. It has hardly been applied to
problems in animal physiology though it has found some application in
plant and crop modelling (Thornley and Johnson, 1989).
Empirical modelling
Empirical models are models in which experimental data are used directly to
quantify relationships, and are based at a single level (e.g. the whole animal)
in the organizational hierarchy discussed above. Empirical modelling is
concerned with using models to describe data by accounting for inherent
variation in the data. Thus, an empirical model sets out principally to
describe, and is based on observation and experiment and not necessarily on
any preconceived biological theory. The approach derives from the
philosophy of empiricism and adheres to the methodology of statistics.
Empirical models are often curve-fitting exercises. As an example,
consider modelling voluntary feed intake in a growing pig. An empirical
approach to this problem would be to take a data set and fit a linear
multiple regression equation, possibly relating intake to liveweight,
liveweight gain and some measure of diet quality.
We note that level i behaviour (intake) is described in terms of level i
attributes (liveweight, liveweight gain and diet quality). As this type of
model is principally concerned with prediction, direct biological meaning
usually cannot be ascribed to the equation parameters and the model
suggests little about the mechanisms of voluntary feed intake. If the model
fits the data well, the equation could be extremely useful though it is
specific to the particular conditions under which the data were obtained,
and so the range of its predictive ability is limited.
Mechanistic modelling
Mechanistic models are process-based and seek to understand causation. A

mechanistic model is constructed by looking at the structure of the system
under investigation, dividing it into its key components, and analysing the
behaviour of the whole system in terms of its individual components and
their interactions with one another. For example, a simplified mechanistic
description of intake and nutrient utilization for our growing pig might
contain five components, namely two body pools (protein and fat), two
blood plasma pools (amino acids and other carbon metabolites) and a
digestive pool (gut fill), and include interactions such as protein and fat
turnover, gluconeogenesis from amino acids and nutrient absorption. Thus
the mechanistic modeller attempts to construct a description of the system
at level i in terms of the components and their associated processes at level
i – l (and possibly lower), in order to gain an understanding at level i in
terms of these component processes. Indeed, it is the connections that
inter-relate the components that make a model mechanistic. Mechanistic

modelling follows the traditional philosophy and reductionist method of
the physical and chemical sciences.
Model evaluation
Model evaluation is not a wholly objective process. Models can be

perceived as hypotheses expressed in mathematics and should therefore be
subject to the usual process of hypothesis evaluation. To quote Popper,
these conjectures are controlled by criticism; by attempted refutations, which
include several critical tests. They may survive these tests, but they can never
be positively justified … by bringing out our mistakes it makes us understand
the difficulties of the problem we are trying to solve.
(Popper, 1969)
A working scientific hypothesis must therefore be subjected to criticism and
evaluation in an attempt to refute it. In the Popperian sense, the term
validation must be assumed to mean a failed attempt at falsification, since
models cannot be proved valid, but only invalid. Validation is thus best
avoided.
Following Popper’s analysis, the predictions of a model should be
compared with as many observations as possible. However, there is often a
lack of suitable data to compare predictions with observations, because the
available data were used to estimate model parameters and hence cannot be
used to evaluate the model independently, or because the entities simply
have not or cannot be measured experimentally. We refute the opinion of
some referees and editors that a model is valuable if and only if its
predictions are fully accurate. The evaluation of research models depends
on an appraisal of the total effort, within which mathematical modelling
serves to provide a framework for integrating knowledge and formulating
hypotheses. For applied models, evaluation involves comparison of the
results of the new model and of existing models, in a defined environment
(the champion-challenger approach). In all cases, the objectives of a
modelling exercise should be examined to assess their legitimacy and to
what extent they have been fulfilled.
Mathematical Approaches
At this point in our discussion, it is important to give a correct picture of the
nature of mathematics. Mathematics is often seen as a kind of tool, as the
handmaiden of science and technology. This view fails to acknowledge or
reflect the potential role of mathematics in science and technology as an
integral part of the basic logic underlying the previewing and developmental
imagination which drives these vital disciplines. The use of the word tool to
describe mathematics is, we submit, pejorative. Tools operate on materials in a
coercive way by cutting, piercing, smashing, etc. Mathematics is used in a
completely non-coercive way, by appealing to reason, by enabling us to see the

world more clearly, by enabling us to understand things that we previously
failed to understand.
Mathematics itself is an umbrella term covering a rich and diverse
discipline. It has several distinct branches, e.g. statistics (methods of
obtaining and analysing quantitative data based on probability theory),
operational research (methods for the study of complex decision-making
problems concerned with best utilization of limited resources) and applied
mathematics (concerned with the study of the physical world and including
e.g. mechanics, thermodynamics, theory of electricity and magnetism). The
mathematical spectrum is illustrated in Fig. 2.2.
Statistics has had a major influence on research in animal science and
in applied biology generally, and is well understood by biologists. This is
hardly surprising given that many of the techniques for the design and
analysis of experiments were pioneered in the 1920s to deal with variability
in agricultural field experiments and surveys caused by factors beyond the
control of investigators such as the weather and site differences. Other
pertinent branches of mathematics, such as applied mathematics and
operational research, are less well understood. In the rest of this chapter,
we explore a key paradigm from each of these three branches, viz. the
regression and the linear programming (LP) paradigms from statistics and
operational research, respectively, and the rate:state formalism of applied
mathematics (biomathematics).
Regression paradigm
Linear multiple regression models pervade applied biology. The mathe-

matical paradigm assumes there is one stochastic variable Y and q deter-
ministic variables X1, X2, …, Xq, and that E(Y | X1, X2, …, Xq), the expected
value of Y given X1, X2, …, Xq, is linearly dependent on X1, X2, …, Xq:
E(Y X1 , X2 , ..., X q ) = β0 + β1 X1 + β2 X2 + ... + β q X q , (2.2)

and the variance V(Y | X1, X2, …, Xq) is constant:
V(Y X1 , X2 , ..., X q ) = σ 2 .
Y is known as the dependent variable, X1, X2, …, Xq as the independent

variables, and the equation:
Y = β0 + β1 X1 + β2 X2 + ... + β q X q ,
Statistics Operations Applied Numerical Pure

research mathematics analysis mathematics
← Empirical modelling →← Mechanistic modelling
Fig. 2.2. Mathematical spectrum.

as the regression equation. The parameters 1 ,2 , ..., q are the partial
regression coefficients.
It is convenient to write Eqn 2.2 in the form:
˜ + ( X − x ) + ( X − x ) + ... + ( X − x ),
E(Y X1 , X2 , ..., X q ) = 0 1 1 1 2 2 2 q q q
where the –x i ’s are computed from the n observations (y1, x11, x21, …, xq1),
n
(y2, x12, x22, …, xq2), …, (yn, x1n, x2n, …, xqn) as, e.g. x1 = ∑ x1 j n. The sum
j=1
of squares of the yj’s from their expectations is therefore:
∑ [ yj ],
n 2
S(˜0 , 1 , ..., q ) = ˜ − ( x − x ) − ( x − x ) − ... − ( x − x )
− 0 1 1j 1 2 2j 2 q qj q
j =1
˜
and the least squares estimates of the parameters β0 , β1 , β2 , ..., β q are the
solutions of the normal equations:
∂S ∂S ∂S ∂S
= = = ... = = 0.
˜
∂0 ∂1 ∂2 ∂ q
Parameter β0 can be determined knowing β̃0. A linear multiple regression
model is linear in the parameters 1 , 2 , ..., q. A non-linear model that
can be transformed into a form which is linear in the parameters (e.g. by
taking natural logarithms) is said to be intrinsically linear. Draper and
Smith (1998) is recommended reading on regression methods.
Many of the models applied in pig and poultry science are systems of
linked regression equations, e.g. current feed evaluation systems.
LP paradigm
An LP problem has three quantitative aspects: an objective; alternative

courses of action for achieving the objective; and resource or other
restrictions. These must be expressed in mathematical terms so that the
solution can be calculated. The mathematical paradigm is:
q
min Z = ∑ c j X j , [objective]
j =1
q
∑ aij X j ≥ or ≤ bi ; i = 1,2, ..., m , [constraints]

j=1
X j ≥ 0; j = 1, 2, ..., q, [non-negativity constraints]
where Z is the objective function and the Xj’s are decision variables. The cj’s,
aij’s and bi’s (bi 0) are generally referred to as costs, technological
coefficients and right-hand-side values, respectively. The paradigm is

generally solved using a simplex algorithm (see Thornley and France, 2006).
This formalism is much less restrictive than it first appears. For
example, maximization of an objective function is equivalent to minimizing
the negative of that function; an equality constraint can be replaced by
entering it as both a and a constraint; and any real variable can be
expressed as the difference between two positive variables. Also, there are
various extensions of this paradigm that allow, e.g. examination of the way
the optimal solution changes as one or more of the coefficients varies
(parametric programming); nonlinear functions of single variables to be
accommodated (separable programming); decision variables to take integer
values (integer programming); the objective of an activity or enterprise to
be expressed in terms of targets or goals rather than in terms of optimizing
a single criterion (goal programming); and multiple objective functions to
be considered (compromise programming). Further description of these
techniques can be found in Thornley and France (2006).
Typical applications in pig and poultry production include:
formulating feed compounds and least-cost rations; allocating stock to
feeding pens; and deciding on the amounts of fertilizer to apply to land.
Rate:state formalism
Differential equations are central to the sciences and act as the cornerstone of
applied mathematics. It is often claimed that Sir Isaac Newton’s great
discovery was that they provide the key to the ‘system of the world’. They
arise within biology in the construction of dynamic, deterministic, mechanistic
models. There is a mathematically standard way of representing such models
called the rate:state formalism. The system under investigation is defined at
time t by q state variables: X1, X2, …, Xq. These variables represent properties
or attributes of the system, such as visceral protein mass, quantity of substrate,
etc. The model then comprises q first order differential equations which
describe how the state variables change with time:
dX i
= f i ( X1 , X2 , ..., X q; S); i = 1, 2, ..., q, (2.3)
dt
where S denotes a set of parameters, and the function fi gives the rate of
change of the state variable Xi.
The function fi comprises terms which represent the rates of processes
(with dimensions of state variable per unit time), and these rates can be
calculated from the values of the state variables alone, with of course the
values of any parameters and constants. In this type of mathematical
modelling, the differential equations are constructed by direct application of
scientific law based on the Cartesian doctrine of causal determinism (e.g. the
law of mass conservation, the first law of thermodynamics) or by application
of a continuity equation derived from more fundamental scientific laws. The
rate:state formalism is not as restrictive as first appears because any higher-

order differential equation can be replaced by, and a partial differential
equation approximated by, a series of first-order differential equations.
If the system under investigation is in steady state, solution to Eqn 2.3 is
obtained by setting the differential terms to zero and manipulating
algebraically to give an expression for each of the components and processes
of interest. Radioisotope data, for example, are usually resolved in this way,
and indeed, many of the time-independent formulae presented in the animal
science literature are derived likewise. However, in order to generate the
dynamic behaviour of any model, the rate:state equations must be integrated.
For the simple cases, analytical solutions are usually obtained. Such
models are widely applied in digestion studies to interpret time-course data
from marker and in vitro experiments, where the functional form of the
solution is fitted to the data using a curve-fitting procedure. This enables
biological measures such as mean retention time and extent of digestion in
the gastro-intestinal tract to be calculated from the estimated parameters.
For the more complex cases, only numerical solutions to the rate:state
equations can be obtained. This can be conveniently achieved by using one
of the many computer software packages available for tackling such
problems. Such models are used to simulate complex digestive and
metabolic systems. They are normally used as tactical research tools to
evaluate current understanding for adequacy and, when current
understanding is inadequate, help identify critical experiments. Thus, they
play a useful role in hypothesis evaluation and in the identification of areas
where knowledge is lacking, leading to less ad hoc experimentation. Also, a
mechanistic simulation model is likely to be more suitable for extrapolation
than an empirical model, as its biological content is generally far richer.
Recent examples of this type of model include the simulation of nutrient
partitioning in growing pigs to predict anatomical body composition
(Halas et al., 2004) and the simulation of calcium and phosphorus flows in
layers to evaluate feeding strategies aimed at reducing P excretion to the
environment in poultry manure (Dijkstra et al., 2006).
Sometimes it is convenient to express a differential equation as an
integral equation; for example Eqn 2.3 may be written:
t
X i = X i (0) + ∫ fi ( X1, X2 , ..., Xq; S)dt; i = 1, 2, ..., q,
0
where Xi(0) denotes the initial (zero time) value of Xi. Integral equations
arise, not only as the converse of differential equations, but also in their
own right. For example, the response of a system sometimes depends not
just on the state of the system per se but also on the form of the input.
Input P and output U might then be related by the convolution (or
Faltung) integral:
t
∫0
U ( t) = P( x)W ( t − x)dx = P( t) * W ( t),
where x is a dummy variable ranging over the time interval zero to the
present time t during which the input has occurred, and W is a weighting
function which weights past values of the input to the present value of the
output. The symbol * denotes the convolution operator. Integral equations
are much less common in biology than differential equations though they
occur as convolution integrals in areas such as tracer kinetics. Further
discussion of these issues can be found in Thornley and France (2006).
Application of the rate:state formalism is illustrated with reference to
coccidiosis, an intestinal disease in chickens caused by protozoan parasites
of the genus Eimeria. The life cycle of E. tenella, a typical species that
invades the caecum, is depicted in Fig. 2.3.
Fig. 2.3. Life cycle of E. tenella: (a) sporulated oocyst; (b) sporozoite being liberated from
oocyst and sporocyst; (c) sporozoite; (d) trophozoite parasitizing an epithelial cell; (e) early
schizont; (f) mature first-generation schizont; (g) first-generation merozoite parasitizing another
epithelial cell; (h, i) second-generation schizonts; (j) rupture of second-generation schizont; (k)
second-generation merozoite may parasitize other epithelial cells (l) for a third asexual cycle, or
may parasitize an epithelial cell (m) to become a female gametocyte (q); merozoite parasitizing
an epithelial cell (n) and becoming a male gametocyte (o); (p) liberated microgametes unite with
macrogamete (r), which develops into oocyst (s) and is liberated in the faeces by host (t),
sporulation (u) of oocyst occurs in outside environment (source: Reid, 1984).
The main features of this complex life cycle are: (i) the exogenous
development of newly excreted oocysts in the litter to become infectious
sporulated oocysts; and (ii) programming of the parasite to undergo
controlled replication within the intestinal mucosa, with a time delay
between each stage. Endogenous development occurs from the eight
sporozoites released from each oocyst which then undergo a maximum of
three cycles of sexual divison (schizogony) with known multiplication rates.
Each of the three generations of schizonts contains a different but
relatively constant number of protozoan forms known as merozoites.
Merozoites released from third generation schizonts give rise to the sexual
phases of the cycle, forming either male microgametes or female
macrogametes. Fertilization of the macrogametes results in zygotes which,
after the development of a protective wall, are released as unsporulated
oocysts. A much simplified version of this life cycle is shown in Fig. 2.4.
The system as represented is defined at time t by five state variables: X1,
X2, …, X5. These variables represent the number of oocysts per bird in the
litter (X1), the number of oocysts inside a single bird (X2), the number of
sporozoites inside a bird (X3), the number of schizonts inside a bird (X4),
and the number of zygotes inside a bird (X5). The model then comprises
five first order differential equations given by Eqn 2.3 with q = 5. The solid
arrows between boxes represent flows (per unit time) between the different
stages of the life cycle included in the model. Time delays are incorporated
to allow for stages not explicitly or inadequately represented. The model
can be solved using an appropriate set of parameter values (S) to give values
of the state variables over time, and to simulate the effects of intervention
strategies such as the use of vaccine oocysts in the feed (Fig. 2.4).
A second application of the formalism is demonstrated by considering
the synthesis of milk fat and lactose, two of the principal constituents of
Oocysts Ingested
Zygotes, Time
in litter, oocysts,
X5 X1 X2
Time
delay Vaccine
oocysts
Total Merozoites Sporozoites,

schizonts, X3
X4 Time delay
Fig. 2.4. Simplified representation of the life cycle of E. tenella for use as a model (from
Parry et al., 1992).
milk, in the mammary gland of the lactating sow. The biochemical

pathways involved are shown in Fig. 2.5.
Triacylglycerols comprise over 97% of the lipids in milk. Biosynthesis
of fatty acid precursors occurs in the mitochondria, that of fatty acids,
glycerol, and other related intermediates in the cytosol, and that of
triacylglycerol in or near the endoplasmic reticulum. The primary pathway
for fatty acid synthesis is glucose through glycolysis to pyruvate, followed
by oxidative decarboxylation to form acetyl CoA. Acetyl CoA, together with
oxaloacetate (OAA), may be further oxidized to CO2 in the citric acid cycle
(Fig. 2.5). Lactose is a disaccharide composed of one molecule of glucose
and one of galactose. The synthetic pathway is glucose-1-P to uridine
diphosphate (UDP)-glucose to UDP-galactose, then UDP-galactose plus
glucose to lactose (Fig. 2.5).
A highly simplified version of these pathways is shown in Fig. 2.6. The
system as represented is defined at time t by five state variables: X1, X2, …,
X5. These variables (in mols) represent the precursors fatty acids (X1) and
glucose (X2), the intermediate fatty acyl CoA (X3), and the products milk fat
Fig. 2.5. Biochemical pathways involved in the synthesis of milk fat and lactose in the
mammary gland of the lactating sow (source: J.P. Cant, personal communication).
(X4) and lactose (X5). The model then comprises five first order differential
equations given by Eqn 2.3 with q = 5, as for the simplified representation
of the life cycle of E. tenella (Fig. 2.4). The solid arrows between boxes
represent flows (mols per unit time) between the state variables of the
model. An acetyl CoA transaction is incorporated to generate ATP for
metabolic transactions via its oxidation. The model can be used to simulate
a complete lactation and to help develop practical feeding strategies.
These two applications illustrate the power of differentials and the
rate:state formalism in providing quantitative, dynamic descriptions of
biological life cycles and biochemical pathways, which are central to pig
and poultry science.
Conclusions
The first step in the application of scientific precepts to a problem is to

identify objectives. Next, appropriate information is collated to generate
theories and hypotheses which are subsequently tested against observations
(Fig. 2.1). Mathematical models, particularly process-based ones, provide a
useful means of integrating knowledge and formulating hypotheses. Thus
mathematical modelling is an integral part of a research programme, with
the experimental and modelling objectives highly inter-related.
The mathematical expression of hypotheses in models forms a central
role in a research programme. Kuhn (1963) stressed the importance of
research performed by scientists within a scientific discipline, which slowly
but steadily increases knowledge, and the more rapid progress which from
time to time is achieved by efforts of scientists in a true interdisciplinary
manner. Progress in modelling depends on a variety of approaches and
ideas. Thus, while further refinements of models may provide knowledge
Fatty acids, Glucose,

X1 X2
Fatty acyl CoA,

X3
Oxidation
Milk fat, Milk lactose,

X4 X5
Fig. 2.6. Simplified milk fat and lactose synthesis for use as a model (adapted from
Pettigrew et al., 1992).
that is of value in its own right, that value is greatly enhanced if these
refinements can be related to the interaction between observations
resulting from experiments and from simulations. Modelling increases the
efficiency and effectiveness of experiments with animals and enhances
progress in understanding and controlling pig and poultry production.
Biological research, if it is to remain truly relevant, must be
undertaken at several levels of generality, e.g. cell, tissue or organ, whole
organism, population. There is much more to biology than just molecular
science. Hopefully, the molecular chauvinism that seems to have
dominated biological research thinking (and hence funding) for much of
the last quarter century is finally at an end. This chapter has identified
different modelling approaches, i.e. teleonomic, empirical and mechanistic
modelling, and different mathematical paradigms drawn from different
branches of mathematics. No approach or paradigm is advocated as being
universally superior; no one has a monopoly on wisdom. It is noteworthy
and pleasing that papers on pig and poultry modelling were read at the
present workshop and also formed a significant part of a recently held 5-
yearly farm animal modellers workshop (Kebreab et al., 2006). It is, after
all, a truism that those modelling pig nutrition have things to learn from
their counterparts working, for example, in poultry nutrition, and vice
versa. Thus scientific pluralism, not just across animal species but also
across levels of generality and types of modelling, should be a pillar for
future development of the activity of pig and poultry modelling.
Peering into a crystal ball and attempting to foretell what lies ahead is
usually a futile task. To quote Baldwin (2000): ‘previewing the future is an
equivocal process’. We think it sufficient to conclude by saying that a future
focus for pig and poultry modelling based on scientific pluralism, with
emphasis on solving biological problems rather than applying mathematical
techniques, offers a fruitful way ahead.
Acknowledgement
We thank Dr John Thornley for many useful discussions on this topic over
a number of years.
References
Anon (1960) Working models in medicine. Journal of the American Medical Association
174, 407–408.
Baldwin, R.L. (2000) Introduction: history and future of modelling nutrient
utilization in farm animals. In: McNamara, J.P., France, J. and Beever, D.E.
(eds) Modelling Nutrient Utilization in Farm Animals. CABI Publishing,
Dijkstra, J., Kebreab, E., Kwakkel, R.P. and France, J. (2006) Development of a
dynamic model of Ca and P flows in layers. In: Kebreab, E., Dijkstra, J., Gerrits,
W.J.J., Bannink, A. and France, J. (eds) Nutrient Digestion and Utilization in Farm
Animals: Modelling Approaches. CABI Publishing, Wallingford, UK, pp. 192–210.
Draper, N.R. and Smith, H. (1998) Applied Regression Analysis, 3rd edn. Wiley, New
York.
Haldane, J.B.S. (1927) Science and theology as art forms. In: Possible Worlds. Chatto
& Windus, London.
Halas, V., Dijkstra, J., Babinszky, L., Verstegen, M.W.A. and Gerrits, W.J.J. (2004)
Modelling of nutrient partitioning in growing pigs to predict their anatomical
body composition. 1. Model description. British Journal of Nutrition 92,
707–723.
Kebreab, E., Dijkstra, J., Gerrits, W.J.J., Bannink, A. and France, J. (eds) (2006)
Nutrient Digestion and Utilization in Farm Animals: Modelling Approaches. CAB
International, Wallingford, UK, viii+480 pp.
Kuhn, T.S. (1963) The Structure of Scientific Revolutions. University Press, Chicago,
Illinois.
Mach, E. (1942) The Science of Mechanics, 9th edn. Open Court, LaSalle, Illinois.
Monod, J. (1975) Chance and Necessity. Collin metabolism of lactating sows. Journal of
Animal Science 70, 3742–3761.
Popper, K.R. (1968) The Logic of Scientific Discovery. Hutchinson, London.
Popper, K.R. (1969) Conjecture and Refutations. The Growth of Scientific Knowledge, 3rd
edn. Routledge & Kegan Paul, London.
Reid, W.M. (1984) Coccidiosis. In: Hofstad, M.S. (ed.) Diseases of Poultry, 7th edn.
Iowa State University Press, Ames, Iowa, pp. 784–846.
Rosenbluth, A. and Weiner, N. (1945) The role of models in science. Philosophical
Science 12, 316–321.
Thornley, J.H.M. and France, J. (2006) Mathematical Models in Agriculture, 2nd edn.
CAB International, Wallingford, UK, 886 pp.
Thornley, J.H.M. and Johnson, I.R. (1989) Plant and Crop Modelling. Oxford
University Press, Oxford, UK.
Wilson, E.O. (1978) On Human Nature. Harvard University Press, Cambridge,
Massachusetts.
3 Basic Concepts Describing
Animal Growth and Feed Intake
N.S. FERGUSON
Maple Leaf Foods Agresearch, 150 Research Lane, Guelph, Ontario,
Canada, N1G 4T2
fergusns@mapleleaf.ca
Introduction
The simulation of animal growth potentially provides a way of predicting

animal performance and the subsequent effects on the production of pork,
over a wide range of conditions with an accuracy that would otherwise be
impossible to accomplish. In addition, limiting factors within the pork
production system can be identified, nutrient requirements predicted,
meat quantity and quality estimated, more effective financial and
management decisions made, and the consequences of genetic selection
predicted. Fundamental to any model predicting animal growth and
voluntary feed intake, is the theory describing how the animal grows and
how it interacts with its environment. The accuracy of defining these
biological responses depends on the nature of the theory and how
inclusive and/or exclusive it is of our understanding of animal growth in
general. The basic theory proposed in this chapter has been well defined
and described in the literature, as well as successfully implemented in a
number of modelling applications (Ferguson et al., 1994; Emmans and
Kyriazakis, 1999; Wellock et al., 2003a,b). Essentially, it is driven by an
adequate description of: (i) an animal in some state of being; (ii) the
environment in which the animal exists; (iii) the type and quantity of feed
given; and (iv) the health status. The combination of these components
provides the framework for predicting growth responses to a wide variety
of production scenarios and the numerous commercial applications
thereafter (Fig. 3.1).
The following are the key assumptions and premises of the proposed
theory:

Basic Concepts Describing Animal Growth and Feed Intake 23
Animal
Maintenance
Growth Gut
Capacity
Requirements C
R Feed O
E N
S S
O Physical T
U R
Environment A
R
C I
E N
S Social T
Desired Intake Environment Constrained
Potential Growth S
Intake and Growth
Actual Intake and Growth
Nutrient Nutritional
Carcass Manure
Requirements Management
Grading
Economics LCF/Optimize Feed Budgets N and P
Fig. 3.1. Framework of the processes involved in modelling growth and feed intake
and the subsequent commercial application (after Emmans and Oldham, 1988).
1. The animal will always attempt to achieve its potential rate of growth
which is defined by its current state and genetic potential;
2. The amount of feed eaten will be the lesser of what the diet can offer to
achieve potential growth and the capacity of the gut, within the constraint
of maintaining heat balance;
3. Health status and stocking density are possible constraints on potential
growth;
4. Predicted responses are of the average individual.
With an accurate description of the genotype, the potential growth rate of
the animal may be predicted. When the nutritional and environmental
inputs are inadequate the animal will fail to achieve its potential growth; the
extent to which it is constrained will be defined by a set of rules governing
the partitioning of nutrients according to the most limiting factor (e.g. amino
acid, energy, disease challenge, gut capacity, maintaining heat balance, etc.).
The corollary to this suggests that the nutrient and environmental inputs
required to achieve potential growth can be determined. Based on this
approach, an adequate description of the genotype, the feed, health status,
physical and social environment are required.
24 N.S. Ferguson
Animal Description
There is a plethora of mathematical functions describing the pattern of

potential growth through various phases of life (Gompertz, 1825;
Robertson, 1923; Brody, 1945; Von Bertalanffy, 1957; Parks, 1982; Black et
al., 1986; Bridges et al., 1986; France et al., 1996). However, not all are
appropriate, nor do they all meet the criteria for the framework proposed
in this chapter. Wellock et al. (2004) examined these numerous functions
and concluded that the Gompertz function is a ‘suitable descriptor of
potential growth’ because of its simplicity, accuracy and ease of application.
According to this function, growth will reach a peak at approximately
0.368 of the animal’s mature weight and will then decline to zero at
maturity. But a description of the potential growth of an animal must also
deal with the systematic changes occurring in both chemical and physical
composition of the body. The detailed theory of how these parameters
interact to determine the daily rate of growth has previously been
documented by Ferguson et al. (1994), Emmans and Kyriazakis (1999) and
Wellock et al. (2003a) and therefore only an overview will be presented in
this chapter. In summary, the Gompertz function is used to determine the
potential protein growth rate from which the growth of the remaining
chemical components of the body (lipid, moisture and ash) can be
determined. This is achieved using allometric relationships between
protein and lipid, moisture and ash (Emmans and Fisher, 1986; Moughan
et al., 1990). Based on this approach, the inherent characteristics required
to describe the animal (genotype) are:
1. The rate of maturing (B);
2. The mature body protein weight (Pm);
3. The inherent fatness or lipid:protein ratio at maturity (LPRm); and
4. The allometric coefficients defining the relationships between protein
and water (the water:protein ratio, WPRm), and protein and ash
(ash:protein ratio, APRm) at maturity. According to Emmans and
Kyriazakis (1995) these are constant for a number of breeds.
An important assumption here, is that the rate of maturing is similar across
all four chemical components (protein, lipid, water and ash). Ferguson and
Kyriazis (2003) provided evidence to corroborate this assumption.
Genotypes, therefore, will differ in a number of respects that affect their
potential growth curves, including mature protein size, mature
composition (fat, moisture and ash) and the rates of maturing.
Potential protein growth
Body protein weight over time is determined from the function:

Pt = Pm x e–e ln(–lnuo) – (Bxt) (kg/day),
where Pt = body protein weight at time t (kg)

Pm = mature body protein weight (kg)
uo = degree of maturity at birth (Pt0/Pm)
B = rate of maturing constant (day1)
t = age (days);
while the rate of potential protein growth (pPD) is defined by:
pPD = B Pt ln (Pm/Pt) (g/day),
with maximum pPD (pPDmax), determined as:
pPDmax = B 1/e Pm (g/day).
The equations above indicate that the rate at which an animal grows
will depend almost entirely on its current state and two inherent
characteristics, B and Pm (Taylor, 1980). Potential protein deposition (pPD)
will only be realized if the animal is able to ingest sufficient quantities of
energy and the first limiting amino acid, and if the environment is
sufficiently cool to allow the animal to lose the subsequent heat produced.
Otherwise, actual protein deposition rate (PD) will be lower than pPD.
Examples of estimated constants for different sexes and strains of pigs
derived from the literature and experiments are shown in Table 3.1.
Fat growth
It has been widely acknowledged that an animal has an inherent potential

rate of protein growth, as defined by its maximum rate of growth under
Table 3.1. Growth parameters from various literature sources.

Literature source and pig type B Pm LPRm WPRm APRm
(/day) (kg) (kg/kg) (kg/kg) (kg/kg)
Ferguson and Gous (1993)

LW Landrace – entire males
(South Africa) 0.0107 38.7 2.60 3.30 0.21
Ferguson and Kyriazis (2003)
LW Landrace – entire males
(South Africa) 0.0114 40.0 1.80 3.17 0.22
Ferguson (2004, unpublished results)
Commercial Duroc – mixed sexes
(Canada) 0.0142–0.0156 33.0–34.0 2.0–2.3 3.40 0.21
Knap (2000b)
Commercial – mixed sexes (UK) 0.009–0.0170 31.0 1.4–4.7
Kyriazakis et al. (1990)
LW Landrace – entire males (UK) 0.0150 35.0 2.50 3.05 0.19
Wellock et al. (2003b)
LW Landrace Pietrain – mixed sexes
(France) 0.0175 35.0 2.50 3.05 0.19
26 N.S. Ferguson
ideal conditions (Webster, 1993; Schinckel and de Lange, 1996; Moughan,

1999; Schinckel, 1999; Knap et al., 2003). However, lipid growth does not
appear to have a potential limit, because of its dependence on nutrition, but
rather a desired (preferred) rate of fat deposition. Emmans (1981) alluded
to the concept of a ‘desired lipid growth’ to quantify the relationship
between protein and fat growth and voluntary feed intake. In subsequent
papers by Kyriazakis and Emmans (1992a,b, 1999) and Ferguson and
Theeruth (2002), there is evidence indicating that pigs that are fatter than
‘normal’, will attempt to correct this deviation once the limiting condition
has been removed. Normal, in this case, is defined as the body fat content of
pigs, with a similar body protein content, grown under ideal dietary and
environmental conditions. This desired or preferred fatness is best
described in relation to body protein in the form of a lipid:protein ratio at
maturity (LPRm) and, during growth, an allometric coefficient relating lipid
content to protein (bl) (Emmans and Kyriazakis, 1999). The desired body
fatness (dLt), at a point in time, is therefore defined as:
dLt = LPRm Pm (Pt/Pm)bl (g/day),
where bl = 1.46 LPRm0.23 (after Emmans, 1997).
The preferred rate of fat deposition (dLD) can therefore be described as:
dLD = pPD LPRm bl (Pt/Pm)bl-1 (g/day).
At any age, dLD can be predicted merely as a function of the current
protein weight of the animal. However, the actual rate of fat deposition
(LD) and body fat content (Lt) will be dependent on other nutritional and
environmental factors, including the quantity and quality of food
consumed, the protein:energy ratio in the diet, environmental conditions,
and the state of the animal. With an estimate of dLt, it becomes possible to
determine any compensatory growth responses. Any differences between
Lt and dLt will result in larger or lower daily fat gains. For example,
feeding a poor quality diet (e.g. low protein:energy) will result in excess fat
deposition while restricting feed intake will be associated with a leaner
animal. Compensatory responses are determined by adding the desired fat
growth (dLD) and the difference between actual body fat (Lt) and
preferred body fat content (dLt):
LD = dLD + (dLt – Lt)/1000 (g/day).
If the animal is fatter than desired then LD on the following day will be less,
in order to compensate for the extra fat deposited the previous day
(Kyriazakis et al., 1991; Ferguson and Theeruth, 2002). Provided it is possible
the animal will deposit less fat on the following day. Similarly, if the animal is
leaner than expected, for a given protein content, then LD would be higher
than dLD. An important corollary to the concept of maintaining a desired
level of fatness is that at all times the animal can utilize body fat reserves, to a
greater or lesser extent, to supplement dietary ME, when the need arises.
The use of body fat reserves is limited to periods when dLD is less than or
equal to 0. It is therefore possible to obtain significant protein growth rates

at the expense of fat growth, which would not be possible if a minimum lipid
to protein ratio was used (Moughan et al., 1987; Pomar et al., 1991).
Moisture and ash growth
As the allometric coefficient describing the relationship between body

protein and body moisture content is not unity, the method of determining
moisture deposition (WD) is similar in approach to that of determining fat
growth viz.
WD = PD WPRm bw (Pt/Pm)bw-1 (g/day),
where bw = 0.855 (range from 0.83 to 0.90)
after Emmans and Kyriazakis (1995) and Moughan et al. (1990).
The relative proportion of ash varies little between sexes and strains,
with the rate of ash growth (AD) proceeding at a constant proportion of
protein growth (Moughan et al., 1990; Ferguson and Kyriazis, 2003)
between 0.19 and 0.22 such that:
AD = 0.20 PD (g/day).
Live weight gains
Empty body weight gains (EBWTg) for each day will be calculated from
the sum of the four components, after all other constraints or conditions
have been met, including environmental
EBWTg = PD + LD + AD + WD (g/day).
the empty body weight gains are added to the empty body weight at the
start of the day to give the empty body weight at the end of the day.
To translate empty into total body weights, gut fill has to be considered.
Gut fill is determined from the equation of Whittemore (1998):
Gut fill = 1.05 + 0.05 (0.008 crude fibre 1000 – 40) (g/day),
where crude fibre (CF) is a dietary input value, the daily live weight gains
(ADG) are calculated as:
ADG = EBWTg Gut fill (g/day).
Body weight at any given time (BWTt ) becomes:
BWTt = BWT(t-1) + ADG/1000 (kg/day).
Similarly for each body component, their weights equal the sum of their
starting weight and growth rate for each day. Whether the animal is able to
achieve its potential growth rate each day is dependent on the feed being
offered, the health status and on the environment in which it is housed.
28 N.S. Ferguson
Prediction of Voluntary Feed Intake
The principle behind predicting voluntary feed intake assumes that an

animal will eat what it needs to grow to its potential, within the constraints
of gut volume, health, social stresses and environmental temperature (Fig.
3.1). The basic concept was first proposed by Emmans (1981) and has
subsequently been incorporated into a number of simulation models that
predict voluntary feed intake in growing pigs (Ferguson et al., 1994; Knap,
1999; Wellock et al., 2003a). For a more detailed explanation of the
principles, refer to Kyriazakis and Emmans (1999).
Desired feed intake
The basic premise on which the prediction of voluntary feed intake is

based is that a pig will attempt to consume an amount of feed daily that
will satisfy its requirements for both energy and protein. Unlike the more
popular assumption that animals ‘eat for energy’ (Schinckel and de Lange,
1996), the theory of desired feed intake considers the possibility that
animals may ‘eat for protein’ (Ferguson et al., 2000a,b). Therefore, the
desired feed intake will be the quantity of the diet needed to satisfy the
requirement for the most limiting of either energy or an amino acid, under
non-limiting circumstances.
Energy most limiting

To determine energy requirements for growth and maintenance, use is
made of the ‘Effective Energy’ system proposed by Emmans (1994, 1997).
The effective energy required (EER) by the animal is described as follows:
EER = Em + 50 PD + 56 LD (kJ/day),
where Em = Maintenance energy requirement (kJ/day)
= (1.63 Pt Pm0.27) 1000.
The definition of Em is not the same as the ‘typical’ classification that
equates maintenance energy requirements with fasting heat production
because it removes the effect of the energy lost from the synthesis and
excretion of nitrogen in the urine during fasting. Maintenance
requirements also need to be adjusted for activity and health status. This
will be discussed in more detail later: suffice it to say that activity and
disease can increase Em by as much as 0.15 and 0.20, respectively. In
addition, disease can also reduce the level of activity.
The effective energy content (EEC) of a feed may be described as the
amount of energy available for maintenance and tissue deposition after
deducting energy losses resulting from digestion and defecation. The effect
of fermentation heat losses is considered negligible. The EEC is calculated
as follows:
EEC = MEc – 4.67 dCP – 3.8 IOM + k 12 dCL (kJ/day),

where ME = Metabolizable Energy of the feed (kJ/g)
MEc = ME corrected for zero nitrogen retention (MEc = ME –
5.63 dCP)
IOM = Indigestible organic matter or indigestible carbohydrate
component (g/kg)
dCP = Digestible crude protein content of the feed (g/kg)
k = Proportion of dietary fat retained as body fat (assumed
between 0.3 and 0.8)
dCL = digestible crude lipid content of the feed (g/kg).
The feed intake that will allow the potential energy requirements to be met
in a thermoneutral environment with no social deviances (dFIe), is
calculated as:
dFIe = EER/EEC (g/day).
Protein (amino acid ) most limiting

If protein, or more precisely an amino acid, is the first limiting nutrient in the
feed then the desired feed intake will be based on the protein (amino acid)
requirement and the concentration of dietary protein (available amino acid).
Similar to energy requirements, protein requirement is the sum of ideal
maintenance and potential protein growth, and their respective efficiencies of
utilization. Recent publications have provided evidence to justify the
separation of maintenance protein requirement into its various constituents
and the inclusion of the different amino acid profiles (Table 3.2), to account
for more specific protein losses, including protein turnover, gut and
integument losses (Moughan, 1999; Boisen et al., 2000; Whittemore et al.,
2001c; Green and Whittemore, 2003; de Lange, 2004). In addition, if amino
acid requirements are to be expressed in terms of ‘standardized’ ileal
digestible values then, by definition, the endogenous losses from the gut
unassociated with the feed must be included in the requirements. Moughan
(1999), Boison et al. (2000) and Green and Whittemore (2003) provide a
detailed exposition of the determination of maintenance protein
requirements. In summary, the individual ideal maintenance constituents are:
1. Protein that is deaminated and not reused, which is estimated to be 0.06
of total protein turnover. This is divided into two components, protein
turnover associated with maintaining (PLm) and retaining (PLnm) protein
tissue. Both these processes involve different amino acid profiles (Table
3.2) and therefore need to be determined independently.
PLm = 8 Pt/Pm0.27 (g/day)
PLnm = 0.06 (PD/0.23 x Pm/(Pm–Pt)) – PLm (g/day).
30 N.S. Ferguson
2. Integument protein losses (PLsh) which, according to Moughan (1999),

may be quantified as:
PLsh = 0.105 BWT0.75 (g/day).
3. Total protein endogenous gut loss (PLel) includes losses from both
secretions (PLsec) and physical effects (PLpe) of the feed on the intestinal
lining. The latter gastrointestinal lining losses are a result of the passage of
food and the lack of reabsorption of these amino acids. The inclusion of
both sources of gut loss is important when the supply of amino acids is
expressed in standardized ileal digestible terms (AmiPig, 2000). Intestinal
disease is likely to increase PLel significantly, but the exact extent of the
effect is unknown. If the cost of disease on maintenance protein is to be
considered, it is proposed that PLel could be increased by as much as 0.20,
depending on the severity of the disease. According to Moughan (1999),
Whittemore et al. (2001b) and de Lange (2004):
PLsec = 0.57 BWT0.75 (g/day)
PLpe = 15 g per kg DM intake (g/day)
or PLpe = 1.19 BWT0.75 (g/day)
PLel = (PLsec + PLpe) (1 + DiseaseEffect) (g/day).
4. The ideal protein requirement for maintenance (Pmaint) is the sum of
the four constituents and the efficiency of utilization of absorbed protein
for maintenance (em). Under conditions of normal health em is assumed to
be 0.95, such that:
Pmaint = (PLm + PLnm + PLsh + PLel)/em (g/day).
With these metabolic functions placing demands on different amino acids, it
is necessary to have a specific profile for each amino acid. Table 3.2 provides
suggested amino acid profiles for the various functional requirements.
Boisen et al. (2000) indicated that maintenance protein losses are high
in methionine, cysteine and threonine and therefore requirements for
these amino acids must be increased accordingly. Table 3.2 shows the
significantly higher M+C, Thr and Trp coefficients for PLm, relative to
Table 3.2. Profiles of the amino acid coefficients (expressed as mg/g protein) used for the various
constituents of protein requirements (after Green and Whittemore, 2003 and de Lange, 2004).
Lys M+C Thr Trp Ile Leu P+T His Val
Protein Deposition (PD) 70 37 38 10 35 75 63 30 45

Maint Turnover (PLm) 65 75 90 17 49 45 79 21 44
Non Maint Turnover (PLnm) 66 32 38 10 34 75 63 28 45
Skin and Hair Loss (PLsh) 43 55 32 9 24 50 47 12 36
Endogenous gut losses (PLel) 54 54 46 23 36 60 96 21 63
coefficients for PD. A consequence of these differences in requirements for

individual amino acids, is that there is no longer a constant proportion
between each other and, therefore, relating the requirements of the amino
acids to lysine is no longer appropriate nor correct (de Lange et al., 2001;
de Lange, 2004). For example, Thr:Lys ratio can be 0.64 in pigs < 20 kg
but increase to 0.70 in finisher pigs (>100 kg).
The ideal protein requirement for growth (Pg) is a function of pPD
and the efficiency of ideal protein utilization (eg). A certain amount of
inefficiency does exist when dietary available amino acids are converted
into body tissue, with the result that ideal protein requirements need to be
adjusted before being stated as actual requirements. Efficiency of protein
utilization according to Kyriazakis and Emmans (1992b) is a function of
dietary energy and digestible protein such that eg = 0.0112 ME/dCP
with a maximum of 0.814. However, there is evidence to suggest that the
maximum is closer to 0.85 (Ferguson and Gous, 1997; Green and
Whittemore, 2003). An important assumption with using this function is
that eg is constant across sexes, strains and breeds of pigs and amino acids
(Kyriazakis et al., 1992b). Pg can therefore be calculated as:
Pg = pPD/eg (g/day).
Total ideal protein requirement (Preq) is the sum of maintenance and
growth:
Preq = Pmaint + Pg (g/day).
To determine the amount of feed required to satisfy potential growth, the
available ideal protein content of the feed has to be known. This is
determined by multiplying the digestible protein content of the diet (dCP)
by the value of the protein relative to an ideal balance (BV). The BV value
is the ratio of the proportion of the most limiting amino acid in digestible
dietary protein over the proportion of the same amino acid in ideal body
protein. Therefore, the desired feed intake to satisfy ideal protein
requirements (dFIp) is:
dFIp = Preq / (dCP BV) (g/day).
Constrained feed intake
A potentially limiting factor, preventing an animal attaining its potential

rate of growth is the interaction between the capacity of the gut and the
bulk density of the diet. There is a limit to the volume of food an animal
can ingest, which is determined partly by the size of the animal and partly
by the indigestible components of the diet. For example, a young pig fed a
high fibre diet will have a lower feed intake than an older pig. This limited
intake capacity is referred to as the constrained feed intake. Using this
concept of a bulk constraint (Tsaras et al., 1998) is a more rational
approach to determining dietary constraints than that of imposing fixed
32 N.S. Ferguson
maximum feed intake limits based on animal size (Black et al., 1986)
irrespective of the actual dietary constituents. Kyriazakis and Emmans
(1995) and Whittemore et al. (2001a, 2003) have proposed the use of water
holding capacity as a means of estimating the bulk constraint of a diet.
Unfortunately estimates of water holding capacity for diets are not readily
available and therefore an alternative approach is proposed using the IOM
content (indigestible component) of the diet to predict the bulk density of
the diet (BLKDN):
BLKDN = 0.36 + (0.857 + form) IOM,
where IOM = OM – digOM (kg/kg)
OM = organic matter (1-ash)
DigOM = digestible organic matter (Noblet et al., 2004)
form = physical form of the feed where with pellets: form = 0.0;
crumbles: form = 0.01; mash: form = 0.02.
The maximum feed intake on any one day or constrained feed intake (cFI)
is determined by the following equation, which also incorporates a
rudimentary adjustment for breed differences in appetite capacity:
cFI = ((26.78 + (171.34 Pt) + (2.3316 Pt2) )/ BULKDN)
AdjustBlk (g/day),
where AdjustBlk = appetite factor depending on genotype.
The equation considers both the size effect of the animal and the
indigestible component of the diet, in determining a constrained intake. In
this case, the contribution of size is a quadratic function of protein weight.
Actual feed intake
The desired feed intake (dFI) of the pig in a thermal neutral environment
would be the larger of dFIe and dFIp while the actual daily feed intake
(aFI) would be the lesser of dFI and cFI. For a perfectly balanced diet,
dFIe would equal dFIp. The decision-making process to determine aFI is
illustrated in Fig. 3.2.
Consequences of feed intake on growth
There are three possible pathways to consider, each with their own
consequences on protein and fat growth assuming the physical
environment is not limiting.
1. When energy is most limiting, such that aFI = dFIe

This is the simplest case where the animal consumes enough energy and
protein to satisfy both pPD and dLD. High protein diets will reduce the
EEC of the diet.
Pig description Diet details
Potential growth Gut capacity

Desired fatness
Feed bulk
dFIe dFIp
greater
dFI lesser cFI
Actual FI
Fig. 3.2. Decision-making process to determine actual daily feed intake.
2. When protein is most limiting, such that aFI = dFIp

With protein (or more specifically) an amino acid limiting, the excess
energy consumed above that for pPD and maintenance will be deposited as
extra fat (Ferguson and Theeruth, 2002). The additional fat deposited will
result in the pig being fatter than its preferred level of fatness. The
following day the animal will attempt to compensate by depositing less fat
in order to return to its desired state. It could only achieve this if the
constraining factor, which had caused it to deposit more fat, were removed.
In this case, it would mean increasing the concentration of the limiting
amino acid. Associated with the increase in fat deposition is the amount of
heat produced by the animal. If the heat produced is greater than that
which could be lost to the environment then additional constraints are
placed on growth rate and feed intake. This will be discussed in more
detail in the environmental section below.
3. When intake is constrained by gut capacity, such that aFI = cFI

Voluntary intake is likely to be constrained by the gut capacity in young
pigs (< 50 kg, Whittemore et al., 2003) and/or when the nutrient density of
the diet is low (high bulk diet). In cases such as these, the question becomes
which nutrient becomes the most limiting? This is an important
consideration as it determines whether the animal can attain pPD or not. If
energy was the most limiting then the animal may still be able to consume
sufficient protein to sustain pPD but not dLD. Biological processes dictate
34 N.S. Ferguson
that energy should be allocated first to maintenance functions and then to

protein and lipid growth. If there is insufficient energy for pPD then PD
will be reduced to:
PD = eg ((cFI BV dCP) – Pmaint ) (g/day).
Finally, any remaining energy will be deposited as fat (LD). This amount
will be less than desired resulting in a reduction from the animal’s desired
level of body fatness, such that:
LD = [(cFI EEC) – (50 PD)] / 56 (g/day).
If protein is more limiting than energy, then dietary protein is allocated
according to priority with maintenance first and then the remainder to
growth. The limited amount of dietary protein available for growth will
mean that PD is less than pPD and the energy that would have been used for
pPD (pPD-PD) is deposited as fat. The amount of fat deposited will depend
on the constrained feed intake; it may be higher or lower than dLD.
Compensatory responses
The work by Kyriazakis and Emmans (1991), de Greef (1992), Tsaras et al.
(1998) and Ferguson and Theeruth (2002), provides substantial
experimental support for the previously discussed compensatory growth
theory. These authors found that when young pigs were made fatter, by
eating a poor quality diet (low CP:ME ratio), and then were placed on a
high protein diet, they deposited fat at a much slower rate, than pigs that
were leaner. These data corroborate the idea of a preferred level of fatness
and the desire of the pig to maintain this level (Fig. 3.3).
The implication of this compensation is that pigs will adjust their
voluntary intake to maintain pPD but at the expense of LD once the
causative factor has been removed. It is therefore possible for negative LD
to occur simultaneously with a positive PD, provided there is enough fat to
lose. A minimum amount of body fat is essential for sustaining life, and is
often expressed in terms of a minimum body lipid:protein ratio (L:Pmin).
Wellock et al. (2003a) assumed a L:Pmin ratio of 0.1 but this was not
substantiated. In practice, it is highly improbable that pPD will continue
unabated when the ratio of body lipid:protein decreases below 0.3, even in
young pigs who already have a low L:P ratio (Stamataris et al., 1991).
Physical Environment
Crucial to the prediction of voluntary feed intake and deposition of protein

and fat tissue is the influence of the surroundings of the animal. There are
a number of physical factors that affect the amount of heat the animal can
produce and lose from its body. The interactions between the environment,
animal and diet are regulated by how much heat the animal can lose to its
environment. Therefore, to include the physical environmental effects on
feed intake it is necessary to compare the daily heat production (THP) with
the maximum (THLmax) and minimum (THLmin) daily heat loss limits. If
THP falls within these upper and lower bounds then growth and feed
intake are unaffected by the thermal environment. However, should THP
extend beyond these boundaries, there will be changes in predicted feed
intake and body composition (Fig. 3.4).
Any external stimuli that can affect the temperature the animal actually
‘feels’ at its level will influence the rate of heat loss from the animal
(Whittemore, 1983). This includes factors such as ventilation, the type of floor
material and the insulation of the house. Therefore, ambient temperature
needs to be adjusted accordingly, to reflect an ‘effective’ temperature (Te).
Calculation of THLmin and THLmax
Total heat loss (THL) is the sum of the non-evaporative (or sensible, SHL)
and evaporative heat loss (EHL) components. Therefore, to determine
THLmax and THLmin the minimum and maximum amounts of SHL and
EHL, respectively, have to be determined.
Evaporative heat loss (EHL)

Evaporative heat loss is minimal (EHLmin) and is constant for a particular
live weight at low temperatures (Black et al., 1986, 1999). The maximum
EHL is normally constant, and several times greater than EHLmin, but at
high temperatures the animal will wet its skin and therefore a greater
amount of EHL occurs (Fig. 3.5) (Knap, 2000a).
Desired
Lipid weight (kg)
Total Heat Loss
Improve diet
THLmax
Change diet } Bounds for total

heat loss
THLmin
Restrict Remove
intake restriction
Temperature
Protein weight (kg)
Fig. 3.3. The compensatory responses in fat Fig. 3.4. The relationship between
growth when animals are made either fatter temperature and total heat loss, and the
or leaner than their preferred or desired level maximum (THLmax) and minimum (THLmin)
of fatness. bounds within which growth and feed intake
are unaffected.
36 N.S. Ferguson
The following descriptions of EHLmin and EHLmax are modified

functions from Bruce and Clarke (1979), Black et al. (1986) and Knap
(2000a). Minimum EHL is defined as:
EHLmin = (8+0.07 BWT) (0.09 BWT0.67) (Watts/day).
The maximum EHL is more complicated as it includes an additional
amount of heat loss from evaporation of wet skin. Without losing any heat
from wet skin, the maximum EHL (EHLhot) is determined as:
EHLhot = (12+100 BWT-0.33) humidityfactor 0.09 BWT0.67
(Watts/day),
where humidity factor = 1.36 – (Waterair/35.9)
Waterair = water content of the air calculated from relative humidity (g/kg).
The extra heat lost from the pig wetting its skin is calculated as:
EHLwet = ((45.4 Vel0.6) BWT-0.13 (Abs-Waterair)) 0.35 0.09
BWT0.67 (Watts/day),
where Vel = air speed (m/s)
Abs = water content of the air at 100% Relative Humidity (g/kg).
Maximum EHL can be calculated as:
EHLmax = EHLhot + EHLwet (Watts/day).
Sensible or non-evaporative heat loss (SHL)

Sensible heat loss (SHL) depends on the temperature gradient between the
environmental temperature and the surface of the pig. It is dominant
under cold conditions and diminishes at a constant rate with increasing
temperature. To incorporate the effects of behavioural and physiological
changes, such as huddling, vasoconstriction and vasodilation, associated
with environmental stimuli, the model determines a maximum (SHLmax)
and a minimum (SHLmin). The general theory of how SHL changes with
temperature is illustrated in Fig. 3.6.
SHLmax
Sensible heat loss (SHL)
Evaporative heat loss (EHL)
SHLmin
Above thermoneutral temperature
EHLmax
EHLmin
Temperature Temperature
Fig. 3.5. Relationship between evaporative Fig. 3.6. Relationship between sensible
heat loss and temperature. (non-evaporative) heat loss and temperature.
The amount of heat lost through SHL depends on the slope of the line
in Fig. 3.6 (HLslope), the temperature difference between the animal and its
immediate surroundings, and the surface area from which the heat is lost.
Therefore, SHL can generally be defined as:
SHL = HLslope (Body Temperature – Te) BWT0.67 (kJ/day),
where HLslope = 48
Te = effective temperature.
A minimum amount of heat can be stored by allowing body temperature to
increase from 38 to 40.5 when temperatures exceed the upper bound of the
thermoneutral range. If the animal is hot or cold then certain anatomical
and behavioural changes occur, resulting in the core body temperature
either rising (40.5°C, SHLmax) or remaining constant at 38°C (SHLmin),
respectively. The SHL component contributes very little towards heat
production at high temperatures, because sensible heat loss depends on the
difference in temperature between the environment and the surface of the
pig (Mount, 1975). At an effective temperature of 40.5°C, SHL will be zero.
THLmax and THLmin

Maximum total heat loss is the sum of EHLmax and SHLmax while THLmin
comprises EHLmin and SHLmin.
Comparison of THL with THP
The final stage in determining the effects of the thermal environment on

growth and feed intake is to compare the heat produced (THP) by the pig
with THLmax and THLmin. The THP is calculated as the difference between
the energy consumed and that retained for protein and fat deposition:
THP = (aFI ME) – (23.8 PD) – (39.6 LD) (kJ/day).
Comparing THL with THP determines whether the animal is too hot, cold
or thermoneutral, and enables the appropriate voluntary intake and
growth responses to be calculated.
Responses to environmental constraints

1. THP > THLMAX. When the amount of heat the animal produces is greater
than the maximum that can be lost to the environment then the pig is
effectively ‘hot’ and, therefore, will attempt to reduce THP, such that THP
= THLmax. There are three ways of doing this, depending on the
difference between heat loss and heat produced. First, activity levels
decline and therefore maintenance energy requirements will decrease.
According to Knap (2000a) this reduction can be as much as 7.5%.
Secondly, LD can increase to improve the efficiency of energy utilization
and therefore reduce the heat burden. Thirdly, and most often, feed intake
declines to maintain the energy balance:
38 N.S. Ferguson
aFI = dFIe – (THP – THLmax)/ME (g/day).

The impact this has on PD and LD depends on whether there is still
sufficient protein consumed to meet pPD, given that PD is determined by:
PD = eg ( (aFI BV dCP) – Pmaint) (g/day).
Fat deposition (LD) may increase or decrease depending on the severity of
the reduction in aFI, and is calculated from the difference between energy
intake and the energy retained for PD and lost as heat.
LD = {(aFI x ME) – THP – (23.8xPD)} / 39.6 (g/day).

In most cases, LD will decline rather than increase (Close, 1989).
2. THP < THLMIN. If the amount of heat lost to the environment is greater than
the amount produced then the animal is cold and extra heat (cold
thermogenesis) will be required to maintain body temperature and ensure
THP = THLmin. The energy difference between THLmin and THP, under
thermoneutral conditions, will cause maintenance requirements to increase
and therefore feed intake will increase by:
ExtraFI = (THLmin – THP)/ME (g/day).
The only constraint on ExtraFI, is the bulk constraint, cFI. If (ExtraFI + aFI)
> cFI then feed intake will decline to cFI, and PD and LD will be adjusted
accordingly, as previously discussed under Constrained Feed Intake.
The lower and upper critical temperatures, which define the
thermoneutral boundary, as well as the ideal or comfortable temperature,
can be determined from THP and the various minimum and maximum
heat loss components.
Social Environment
Under commercial growing conditions the environment within which the

pig exists is far from ideal, and changes over the growing period. Factors
contributing to this changing environment include disease challenges,
decreasing air quality, reduced feeder space, high stocking density and
other social stresses. Of these factors only the influences on performance of
stocking density, and to a lesser extent disease, have been quantified
(Kornegay and Notter, 1984; Hyun et al., 1998; Black et al., 1999; Morgan
et al., 1999; Knap, 2000a).
Stocking density
Recent evidence suggests that stress associated with high stocking density
results in a reduction in protein growth, irrespective of feed intake, such that
the amount of feed consumed is driven by the lower PD requirement and not
vice versa (Chapple, 1993; Baker and Johnson, 1999; Morgan et al., 1999;
Matteri et al., 2000; Ferguson et al., 2001). Based on these findings and within
the current context, it is reasonable to assume that stocking density exerts its
effect by reducing protein growth through a lower rate of maturing. When
the amount of space per pig declines below a minimum value then the rate of
maturity will be reduced accordingly. The function used to implement this
effect is based on the surface area of the pig and the amount of space available
(kg0.67/m2) (StDen). The adjusted rate of maturity (Badj) is calculated as:
Badj = B (1 – ((StDen –25) /50)) (/day),
where StDen = (pigs/pen BWT0.67) / pen size (kg0.67/m2).
Disease or health and well-being status
Disease affects the growth performance and feed intake of growing pigs,
the response depending on the source and severity of infection (Baker and
Johnson, 1999; Greiner et al., 2000; Escobar et al., 2002). The possible
effect that health status has on growth under commercial conditions is
illustrated in Fig. 3.7. Although animals may show no signs of clinical
disease, the sub-clinical disease challenge can reduce performance.
The main effects of sub-clinical disease appear to be an increased
maintenance requirement, reduced nutrient digestibility, reduced protein
growth and feed intake (Black et al., 1999; Knap, 2000a). How these are
mediated is not clear. Within the proposed modelling framework, the
question becomes what factors, be they animal or dietary, are likely to be
affected by disease or health status, and how are they adjusted in a simple
Optimal
8 days longer to market
130 Good
At 110 kg Poor
110
Live weight (kg)
90
18 days longer to market
70
50
30
10
50 75 100 125 150 175 200 225
Age (days)
Fig. 3.7. Effects of sub-clinical disease challenge on changes in live weight over time. The
health status of the pig is indicated by Optimal, Good and Poor.
40 N.S. Ferguson
but meaningful way. Logic would dictate that of the animal parameters,
mature protein weight is unlikely to be affected, except under a severe
disease challenge, but the rate at which it achieves this will be adversely
affected. With little evidence to support or disprove this theory, it is
proposed that a proportional reduction in B, relative to the degree of
disease challenge or health status, be adopted to account for the reduction
in protein deposition and subsequent feed intake. In addition, maintenance
protein and energy requirements will be increased, while activity levels will
be reduced.
Reduction in rate of maturity

The model uses a health profile to adjust B. This profile represents the
health status, well-being and barn conditions post weaning (Fig. 3.8) and
can be modified depending on the health status of the animal. Based on
retrofitting data, typical values for high health pigs grown under
commercial conditions are in the range 0.96–1.00, while for diseased pigs
the value could be as low as 0.80.
The health coefficient from the profile is used to reduce B and
therefore PD, LD and feed intake accordingly. The poorer the health
status the more severe the reduction in PD, LD, live weight and feed
intake. Although simplistic in design, and arguably scientifically naïve, the
application of this approach under commercial conditions has shown to
improve significantly the accuracy of prediction (Fig. 3.9).
Increase in maintenance requirements

Black et al. (1999) suggest that maintenance energy be increased by as
much as 0.30 and protein deposition decreased by 0.10 as a consequence
of disease. With insufficient data to justify a 30% increase in maintenance
energy, a more conservative 20% increase is proposed. Although there are
1.2
1.0
Health coefficients
0.8
0.6
0.4
0.2
0
0 5 10 15 20 25 30 35 40
Weeks after weaning
Fig. 3.8. Health profile illustrating the changes in health and well-being status post weaning.
10.0
7.5
Deviations from actual weight (%)

5.0
2.5
0.0
–2.5
–5.0
–7.5
–10.0
21 35 49 63 77 91 105 119 133 147 164
Age (days)
No health adjustment With health adjustment 95% Confidence limits
Fig. 3.9. Comparison of the deviations from actual live weight over time between predicted
results, which include and exclude health status adjustments.
no data to justify similar effects on maintenance protein, for the sake of

brevity, the same increase is applied, such that:
Maint_adj = 1/Health_status,
where Maint_Adj <= 1.20 and >= 1.0.
Maintenance protein (particularly endogenous protein losses) and energy
requirements are increased by multiplying Pmaint and Em by the
adjustment factor. Simultaneously, activity levels are reduced between 1.25
(healthy) and 1.00 (sick), depending on the health status. A reduction in
activity will reduce maintenance energy requirements but not enough to
compensate for the increase due to disease. A consequence of the increased
maintenance requirement, reduced B value and a lower feed intake will be
a reduction in PD and LD.
Commercial example
Comparative data from a number of different commercial production units
within Canada were used. The main cause of the differences in
performance between units was the health and well-being status of the
pigs. The data were separated according to the performances of the best
(High) and worst (Low) health status units, and were compared against the
model predictions. To simulate the differences in health conditions, the
health status coefficients of the Low health units were assumed to be 0.9
42 N.S. Ferguson
times the High health units over the live weight range of 6 to 108 kg. The
growth rate and feed intake results are shown in Figs 3.10a and b.
The similarity between actual versus predicted suggests that the
proposed approach does not produce unrealistic results and is sufficiently
sensitive to differentiate between animals differing in health status.
However, it would be inappropriate and too simplistic to assume that the
process is valid as a means of quantifying the complex effects of specific
health challenges. Nevertheless, the results indicate that it is a reasonable
attempt to incorporate the adverse effects of reduced health status on
growth and feed intake.
Testing Model Theory
Critical to any model is the need to test the underlying theories that drive
or control the modelling process. This is not an easy exercise, as an
(a)
900
800
Growth rate (g/day)
700
600
500
400
300
High health Low health High health Low health
(6–29 kg) (6–29 kg) (26–108 kg) (26–108 kg)
(b) 2300
2000
Feed Intake (g/day)
1700
1400
1100
800
500
High health Low health High health Low health
(6–29 kg) (6–29 kg) (26–108 kg) (26–108 kg)
Fig. 3.10. Comparison of (a) growth rates and (b) feed intakes, between Actual ( ) and
Predicted (䊐), for High versus Low health status producers.
accurate prediction under one set of circumstances does not mean that it is
‘valid’ (Black, 1995). However, confidence in the validity of the model may
be gained when accurate predictions, for a number of diverse
circumstances, are consistently obtained. Unfortunately, there is no specific
way in which models can be validated other than by comparing the
similarities in the model predictions with experimental outcomes. The
three key components of the proposed model that need to be tested are
the growth and intake responses when energy is limiting, when protein
(amino acid) is limiting and when ambient temperatures change.
Responses when energy is limiting
Nursery pigs
To evaluate the effect of Effective Energy content in young nursery pigs,
data from a recent trial (Ferguson, 2005 unpublished results) were used.
Five energy levels, ranging from 13.00 to 13.82 MJ/kg, were fed to pigs
between 12 and 25 kg live weight. The animal definition parameters used
were: B = 0.0142, Pm = 33.0kg, LPRm = 2.30, WPRm = 3.40, APRm =
0.20. The results of the comparison between actual and predicted growth
rates and feed intakes are shown in Figs 3.11a and b. Given that young
pigs (< 30 kg) cannot eat sufficient to satisfy their high relative potential
growth irrespective of the nutrient density of the diet (Whittemore et al.,
2001a), this comparison, in many respects, evaluates the ability of the
model to define the gut capacity and the effects of this constraint on
predicted performance. From Figs 3.11a and b, the results would suggest
that the method proposed in this chapter is sufficiently robust to predict
feed intake and performance within the 95% confidence interval of actual
results.
The higher predicted feed intake value at 13.00 MJ/kg was the result of
the reduced efficiency of protein utilization associated with a low
ME:digestible CP ratio (65 MJ/kg). As the minimum ratio, below which the
efficiency of protein utilization will decline, is 73 MJ/kg (Kyriazakis and
Emmans, 1992b) there will an increased demand for dietary protein to
compensate for the reduced efficiency of utilization. The increased
requirement for protein has resulted in an increase in voluntary feed
intake. Therefore, the slightly higher predicted ADFI at 13.00 MJ/kg would
suggest either an overestimation of the gut capacity or that the equation
predicting efficiency of protein utilization is underestimating the efficiency
in low ME:dCP diets. However, as the overestimation was within the 95%
confidence limits, there is no justification for rejecting the status quo.
Grower-Finisher pigs
The data from Campbell et al. (1985) were used to compare the responses
to energy intake in grower pigs between 48 and 90 kg live weight. The
parameters used to describe the castrate genotype were: B = 0.0125, Pm
44 N.S. Ferguson
(a)
10
Deviation from Actual ADG values (%)

8
6
4
2
0
–2
–4
–6
–8
–10
13.00 13.23 13.48 13.72 13.82
Effective energy content (MJ/kg)
(b)
10
Deviation from Actual ADFI values (%)
8
6
4
2
0
–2
–4
–6
–8
–10
13.00 13.23 13.48 13.72 13.82
Effective energy content (MJ/kg)
Fig. 3.11. Comparison of the deviations in (a) growth rates (ADG) and (b) feed intakes
(ADFI), between Actual and Predicted in response to Effective energy in nursery pigs. (—)
95% Confidence interval.
= 28.0 kg, LPRm = 4.0, WPRm = 3.00, APRm = 0.20. The lysine:energy
ratio was reported to be in excess of that required for maximum protein
deposition on ad libitum feed intake, therefore energy was always the most
limiting nutrient. A summary of the differences between the predicted
outcomes from the model and actual data is shown in Figs 3.12a,b
and c.
The most noticeable differences (> 1 standard deviation) were
confined to PD and LD at very low energy intakes (55% of ad libitum
intakes) where the model overestimated PD and underestimated LD.
These differences are due, in part, to the difficulty in establishing the
relationship between protein and fat deposition when energy is severely
limiting (restricted intakes) and secondly, in the inadequate description of
the environmental conditions given in the scientific paper. Without a
(a)
1000
800
ADG (g/day)
600
400
200
20 25 30 35 40 45
DE Intake (MJ/day)
Fig. 3.12a. Comparison of actual (䊏) versus predicted (—) responses in growth rate (ADG)
to digestible energy intake (DE Intake), in pigs grown from 48 to 90 kg live weight. Data from
Campbell et al. (1985). Bars represent 1 standard deviation.
(b)
4.2
4.0
3.8
3.6
Feed:Gain
3.4
3.2
3.0
2.8
2.6
2.4
20 25 30 35 40 45
DE Intake (MJ/day)
Fig. 3.12b. Comparison of actual (䊏) versus predicted (—) responses in feed:gain ratio to
digestible energy intake (DE Intake), in pigs grown from 48 to 90 kg live weight. Data from
Campbell et al. (1985). Bars represent 1 standard deviation.
proper measure of the housing conditions, a number of assumptions had

to be made (e.g. air temperature, insulation, relative humidity) and
therefore any one of these may not have held true. The differences in
protein, fat, water and ash content of the empty body at 90 kg live weight
in pigs on ad libitum intake, were +5.3, 14.1, +8.0 and 3.2%, relative to
experimental results.
46 N.S. Ferguson
(c)
400
350
300
PD and LD (g/d)
250
200
150
100
50
0
20 25 30 35 40 45
DE Intake (MJ/d)
Fig. 3.12c. Comparison of actual versus predicted responses in protein deposition (PD) (䊏, —)
and lipid deposition (LD) (䉱, ....) to digestible energy intake (DE Intake), in pigs grown from 48
to 90 kg live weight. Data from Campbell et al. (1985). Bars represent 1 standard deviation.
Responses when protein is limiting
This evaluation was conducted to measure the effectiveness of the model

when crude protein or an amino acid is first limiting. The first test
compared the outcome of the model with two experiments (A and B)
published by Gatel et al. (1992). Experiment A investigated the response in
weaned piglets (between 8 and 26 kg) to increasing protein and amino acid
concentrations, with the amino acid:protein ratio remaining constant for
each of the six treatments. Experiment B investigated the growth responses
to increasing amino acid levels but keeping protein constant; hence an
increasing amino acid:protein ratio. The animal description parameters
used were: B = 0.0137, Pm = 35 kg, LPRm = 3.0, WPRm = 3.02, APRm =
0.20. The second test compared differences in total lysine content as well as
differences in amino acid:energy ratios, using data from Kyriazakis et al.
(1990) for pigs grown between 12 and 30 kg. The animal description
parameters used were: B = 0.0135, Pm = 48 kg, LPRm = 3.0, WPRm =
3.4, APRm = 0.20. Although this experiment was primarily concerned with
the effects of choice feeding on performance in young pigs, it does provide
responses to single feeding systems that differ in their dietary crude protein
(total lysine) content only. The results of the comparison are summarized in
Figs 3.13 a, b and c. The results clearly show the similarity between
predicted and actual responses to increasing dietary amino acids. Very few
of the predicted values were outside of the range of 1 standard deviation
above or below the mean. This allows for a greater degree of confidence in
the model, and support for the underlying theory of growth and feed
intake regulation, especially when protein and/or an amino acid is first
limiting, as is often the case in the young growing pig.
(a)
800
700
ADG (g/day)
600
500
400
300
0.5 0.7 0.9 1.1 1.3 1.5 1.7
Total lysine content (%)
(b)
1100
1000
Feed Intake (g/day)
900
800
700
600
0.5 0.7 0.9 1.1 1.3 1.5 1.7
(c)
2.25
2.00
FCR (g/g)
1.75
1.50
1.25
1.00
0.5 0.7 0.9 1.1 1.3 1.5 1.7
Fig. 3.13. (a) Comparison of responses in growth rates (ADG) to total lysine content in the diet.
(b) Comparison of responses in feed intake to total lysine content in the diet. (c) Comparison of
responses in feed:gain (FCR) to total lysine content in the diet. Gatel Expt A, Actual (䊏),
Predicted (—); Gatel Expt B, Actual (䉱), Predicted (....); Kyriazakis Actual (䊉), Predicted (– –).
Bars represent 1 standard deviation.
48 N.S. Ferguson
Responses to ambient temperature
To test the effects of the thermal environment, the data from Rinaldo and
Le Dividich (1991) were selected. This study examined the performance of
growing pigs between 10 and 30 kg live weight when kept in one of four
different temperatures, 12, 18.5, 25 and 31.5°C. The animal description
parameters used were: B = 0.0150, Pm = 35 kg, LPRm = 3.0, WPRm =
3.3, APRm = 0.20. The differences between actual and predicted are
summarized in Fig. 3.14a and b. The only significant difference between
the model predicted and the reported results was in daily growth rates
(ADG) at 12°C, which in turn affected the feed:gain ratio. The very low
ADG observed by Rinaldo and Le Dividich (1991) is contrary to what is
found in most published literature on growth responses at low
temperatures, where ADG remains constant at low temperatures (Nienaber
et al., 1987; Ferguson and Gous, 1997; Wellock et al., 2003b). No
explanation is given for the reduced growth rate reported in the
publication by Rinaldo and Le Dividich (1991).
Conclusions
The theory of growth and feed intake described in this chapter is based on
the proposition that an animal eats to grow to its potential, and if anything
prevents it attaining this growth, then it will grow according to what it has
eaten. Foremost, therefore, is the need to predict desired feed intake and
how the subsequent interactions with the animal and its environment
influence the actual voluntary feed intake. Energy that is available for
maintenance and productive purposes, after removing the heat increment
(b)
(a) 3
ADG and feed intake (g/day)
1400
1200 2.5
Feed:gain
1000 2
800
1.5
600
400 1
8 12 16 20 24 28 32 8 12 16 20 24 28 32
Ambient temperature (°C) Ambient temperature (°C)
Fig. 3.14a. Comparison of actual versus Fig. 3.14b. Comparison of actual versus
predicted responses in growth rate (ADG) predicted responses in feed:gain (䊏, —) to
(䉱, ....) and feed intake (䊏, —) to ambient ambient temperature, in pigs grown from 10
temperature, in pigs grown from 10 to 30 kg to 30 kg live weight. Data from Rinaldo and
live weight. Data from Rinaldo and Le Dividich Le Dividich (1991). Bars represent 1
(1991). Bars represent 1 standard deviation. standard deviation.
of eating and digestion, is compared with the requirements to predict

voluntary food intake. In addition, the response of the animal is to the first
limiting nutrient rather than energy alone. Changes in the chemical
composition of the animal are based on nutrient and physical and social
environmental interactions as well as the current physiological state of the
animal. This includes the effects of health status and pig space. Live weight
changes are determined first, by predicting potential protein growth using
the Gompertz growth function, and secondly, by calculating the remaining
components of the body through established allometric relationships with
body protein content. However, the exact change in body composition is
governed by the desire to maintain an inherent body state, to which the
animal will attempt to return wherever possible. The constraining factors
most likely to cause any deviations from normal growth will be: the gut
capacity, the physical and social environment, and the first limiting
nutrient. When the model described here was used to predict results of
previously published data, the results produced were, with few exceptions,
within the 95% confidence limits of the published data, suggesting that
both the theory and the logic of the model appear to be valid. Where there
were incidences of departure from reality, they were associated with
extreme conditions, which do not normally occur in commercial practice.
References
AmiPig (2000) Ileal standardized digestibility of amino acids in feedstuffs for pigs.
AFZ, Ajinomoto Eurolysine, Aventis Animal Nutrition, INRA and ITCF.
Association Francaise de Zootechnie, Paris.
Baker, D.H. and Johnson, R.W. (1999) Disease, stress, cytokines and amino acid
needs of pigs. Pig News and Information 20, 123N–124N.
Black, J.L. (1995) The testing and evaluation of models. In: Moughan, P.J.,
Verstegen, M.W.A. and Visser-Reyneveld, M.I. (eds) Modelling Growth in the Pig.
EAAP publication no.78, Wageningen Pers, Netherlands, pp. 23–31.
Black, J.L., Campbell, R.G., Williams, I.H., James, K.J. and Davies, G.T. (1986)
Simulation of energy and amino acid utilisation in the pig. Research and
Development in Agriculture 3, 121–145.
Black, J.L., Bray, H.J. and Giles, L.R. (1999) The thermal and infectious
environment. In: Kyriazakis, I. (ed.) A Quantitative Biology of the Pig. CABI
Publishing, Wallingford, UK, pp. 71–97.
Boisen, S., Hvelplund, T. and Weisbjerg, M.R. (2000) Ideal amino acid profiles as a
basis for feed protein evaluation. Livestock Production Science 64, 239–251.
Bridges, T.C., Turner, L.W., Smith, E.M., Stahly, T.S. and Louwer, O.J. (1986) A
mathematical procedure for estimating animal growth and body composition.
Transactions of the American Association of Agricultural Engineers 29, 1342–1347.
Brody, S. (1945) Bioenergetics and Growth. Hafner Publishing Company, New York.
Bruce, J.M. and Clark, J.J. (1979) Models of heat production and critical
temperature for growing pigs. Animal Production 28, 353–369.
Campbell R.G., Taverner, M.R. and Curic, D.M. (1985) The effects of sex and
energy intake between 48 and 90 kg live weight on protein deposition in
growing pigs. Animal Production 40, 497–503.
50 N.S. Ferguson
Chapple, R.P. (1993) Effect of stocking arrangement on pig performance. In:

Batterham, E.S. (ed.) Manipulating Pig Production IV. Australian Pig Science
Association, Victoria, pp. 87–97.
Close, W.H. (1989) The influence of the thermal environment on the voluntary
feed intake of pigs. In: Forbes, J.M., Varley, M.A. and Lawrence, T.L.J. (eds)
The Voluntary Food Intake of Pigs. Animal Production Occasional Publication No.
13, London, pp. 87–96.
de Greef, K.H. (1992) Prediction of production. Nutrition induced tissue
partitioning in growing pigs. PhD thesis. University of Wageningen,
Netherlands.
de Lange, C.F.M. (2004) A systems approach to optimizing phosphorus and
nitrogen utilization in the growing pig. Proceedings of 25th Western Nutrition
Conference, Saskatoon, Canada, pp. 226–246.
de Lange, C.F.M., Gillis, A.M. and Simpson, G.J. (2001) Influence of threonine
intake on whole-body protein deposition and threonine utilization in growing
pigs fed purified diets. Journal of Animal Science 79, 3087–3095.
Emmans, G.C. (1981) A model of the growth and feed intake of ad libitum fed
animals, particularly poultry. In: Hillyer, G.M., Wittemore, C.T. and Gunn,
R.G. (eds) Computers in Animal Production. Animal Production Occasional
Publication No. 5, London, pp. 103–110.
Emmans, G.C. (1994) Effective energy: a concept of energy utilization applied
across species. British Journal of Nutrition 71, 801–821.
Emmans, G.C. (1997) A method to predict the feed intake of domestic animals
from birth to maturity as a function of time. Journal of Theoretical Biology 186,
189–199.
Emmans, G.C. and Fisher, C. (1986) Problems in nutritional theory. In: Boorman,
K.N. and Fisher, C. (eds) Nutrient Requirements of Poultry and Nutritional Research.
Poultry Science Symposium No. 19, Butterworths, London, pp. 9–39.
Emmans, G.C. and Kyriazakis, I. (1995) A general method for predicting the
weight of water in the empty bodies of pigs. Animal Science 61, 103–108.
Emmans, G.C. and Kyriazakis, I. (1999) Growth and body composition. In:
Kyriazakis, I. (ed.) A Quantitative Biology of the Pig. CABI Publishing,
Emmans, G.C. and Oldham, J.D. (1988) Modelling of growth and nutrition in
different species. In: Karver, S. and van Arendonk, J.A.M. (eds) Modelling of
Livestock Production Systems. Kluwer Academic, Dordrecht, Netherlands,
pp. 13–21.
Escobar, J., Van Alstine, W.G., Baker, D.H. and Johnson, R.W. (2002) Growth
performance and whole-body composition of pigs experimentally infected with
Mycoplasma hyopneumoniae. Journal of Animal Science 80, 384–391.
Ferguson, N.S. and Gous, R.M. (1993) Evaluation of pig genotypes. 2. Testing
experimental procedure. Animal Production 56, 245–249.
Ferguson, N.S. and Gous, R.M. (1997) The influence of heat production on
voluntary food intake in growing pigs given protein-deficient diets. Animal
Science 64, 365–378.
Ferguson, N.S. and Kyriazis, S.T. (2003) Evaluation of growth parameters of six
commercial crossbred pig genotypes 1. Under commercial housing conditions
in individual pens. South African Journal of Animal Science 33, 11–20.
Ferguson, N.S. and Theeruth, B.K. (2002) Protein and lipid deposition rates in
growing pigs following a period of excess fattening. South African Journal of
Ferguson, N.S., Gous, R.M. and Emmans, G.C. (1994) Preferred components for
the construction of a new simulation model of growth, feed intake and nutrient
requirements of growing pigs. South African Journal of Animal Science 24, 10–17.
Ferguson, N.S., Arnold, G.A., Lavers, G. and Gous, R.M. (2000a) The response of
growing pigs to amino acids as influenced by environmental temperature. 1
Threonine. Animal Science 70, 287–297.
Ferguson, N.S., Arnold, G.A., Lavers, G. and Gous, R.M. (2000b) The response of
growing pigs to amino acids as influenced by environmental temperature. 2
Lysine. Animal Science 70, 299–306.
Ferguson, N.S., Lavers, G. and Gous, R.M. (2001) The effect of stocking density on
the responses of growing pigs to dietary lysine. Animal Science 73, 459–469.
France, J., Dijkstra, J., Thornley, J.H.M. and Dhanoa, M.S. (1996) A simple but
flexible growth function. Growth, Development and Aging 60, 71–83.
Gatel, F., Buron, G. and Fekete, J. (1992) Total amino acid requirements of weaned
piglets 8 to 25 kg live weight given diets based on wheat and soya-bean meal
fortified with free amino acids. Animal Production 54, 281–287.
Gompertz, B. (1825) On the nature of the function expressive of the law of human
morality and on a new method of determining the value of life contingencies.
Philosophical Transactions of the Royal Society, 513–585.
Green, D.M. and Whittemore, C. T. (2003) Architecture of a harmonized model of
the growing pig for the determination of dietary net energy and protein
requirements and of excretions into the environment. (IMS Pig). Animal Science
77, 113–130.
Greiner, L.L., Stahly, T.S. and Stabel, T.J. (2000) Quantitative relationship of
systemic virus concentration on growth and immune response in pigs. Journal
of Animal Science 78, 2690–2695.
Hyun, Y., Ellis, M., Riskowski, G. and Johnson, R.W. (1998) Growth performance
of pigs subjected to multiple concurrent environmental stressors. Journal of
Knap, P.W. (1999) Simulation of growth in pigs: evaluation of a model to relate
thermoregulation to body protein and lipid content and deposition. Animal
Production 68, 655–679.
Knap, P.W. (2000a) Variation in maintenance requirements of growing pigs in
relation to body composition. A simulation study. PhD thesis. University of
Wageningen, Netherlands.
Knap, P.W. (2000b) Time trends of Gompertz growth parameters in “meat-type”
pigs. Animal Science 70, 39–49.
Knap, P.W., Roehe, R., Kolstad, K., Pomar, C. and Luiting, P. (2003)
Characterization of pig genotypes for growth modeling. Journal of Animal
Science 81, (E. Suppl 2) E187–195.
Kornegay, E.T. and Notter, D.R. (1984) Effects of floor space and number of pigs
per pen on performance. Pig News and Information 5, 23N–33N.
Kyriazakis, I. and Emmans, G.C. (1991) Diet selection in pigs: Dietary choices made
by growing pigs following a period of underfeeding with protein. Animal
Kyriazakis, I. and Emmans, G.C. (1992a) The effects of varying protein and energy
intakes on the growth and body composition of pigs. 1. The effects of energy
intake at constant, high protein intake. Animal Production 68, 603–613.
Kyriazakis, I. and Emmans, G.C. (1992b) The effects of varying protein and energy
intakes on the growth and body composition of pigs. 2. The effects of varying
both energy and protein intake. Animal Production 68, 615–625.
52 N.S. Ferguson
Kyriazakis, I. and Emmans, G.C. (1995) The voluntary feed intake of pigs given
feeds based on wheatbran, dried citrus pulp and grass meal in relation to
measurements of feed bulk. British Journal of Nutrition 73, 191–207.
Kyriazakis, I. and Emmans, G.C. (1999) Voluntary food intake and diet selection.
In: Kyriazakis, I. (ed.) A Quantitative Biology of the Pig. CABI Publishing,
Kyriazakis, I., Emmans, G.C. and Whittemore, C.T. (1990) Diet selection in pigs:
Choices made by growing pigs given foods of different protein concentrations.
Animal Production 51, 189–199.
Kyriazakis, I., Stamataris, C., Emmans, G.C. and Whittemore, C.T. (1991) The
effects of food protein on the performance of pigs previously given food with
low or moderate protein contents. Animal Production 52, 165–173.
Matteri, R.L., Carroll, J.A. and Dyer, C.J. (2000) Neuroendocrine responses to
stress. In: Moberg, G.P. and Mench J.A. (eds) The Biology of Animal Stress. Basic
Principles and Implications for Animal Welfare. CABI Publishing, Wallingford, UK,
pp. 43–76.
Morgan, C.A., Nielsen, B.L., Lawrence, A.B. and Mendl, M.T. (1999) Describing
the social environment and its effects on food intake and growth. In:
Kyriazakis, I. (ed.) A Quantitative Biology of the Pig. CABI Publishing,
Moughan, P.J. (1999) Protein metabolism in the growing pig. In: Kyriazakis, I. (ed.)
A Quantitative Biology of the Pig. CABI Publishing, Wallingford, UK,
pp. 299–331.
Moughan, P.J., Smith, W.C. and Pearson, G. (1987) Description and validation of a
model simulating growth in the pig (20–90kg liveweight). New Zealand Journal
of Agricultural Research 30, 481–489.
Moughan, P.J., Smith, W.C. and Stevens, E.V.J. (1990) Allometric growth of
chemical body components and several organs in the pig (20–90 kg live
weight). New Zealand Journal of Agricultural Research 33, 77–83.
Mount, L.E. (1975) The assessment of thermal environment in relation to pig
production. Livestock Production Science 2, 381–392.
Nienaber, J.A., Hahn, G.L. and Yen, J.T. (1987) Thermal environmental effects on
growing-finishing swine. 1. Growth, food intake and heat production.
Transactions of the American Association of Agricultural Engineers 30, 1772–1775.
Noblet, J., Sève, B. and Jondreville, C. (2004) Nutritional values for pigs. In:
Sauvant, D., Perez, J.M. and Tran, G. (eds) Tables of Composition and Nutritional
Value of Feed Materials. Wageningen Academic Publishers, Wageningen,
Netherlands, pp. 25–37.
Parks, J.R. (1982) A Theory of Feeding and Growth of Animals. Springer, Berlin.
Pomar, C., Harris, D.L. and Minvielle, F. (1991) Computer simulation model of
swine production systems: I. Modeling the growth of young pigs. Journal of
Rinaldo, D. and Le Dividich, J. (1991) Assessment of optimal temperature for
performance and chemical body composition of growing pigs. Livestock
Production Science 29, 61–75.
Robertson, T.B. (1923) The Chemical Basis of Growth and Senescence. Monographs of
Experimental Biology. J.B. Lipincott, Philadelphia.
Schinckel, A.P. (1999) Describing the pig. In: Kyriazakis, I. (ed.) A Quantitative
Biology of the Pig. CABI Publishing, Wallingford, UK.
Schinckel, A.P. and de Lange, C.F.M. (1996) Characterization of growth parameters
needed as inputs for pig growth models. Journal of Animal Science 74, 2021–2036.
Stamataris, C., Kyriazakis, I. and Emmans, G.C. (1991) The performance and body
composition of young pigs following a period of growth retardation by food
restriction. Animal Production 53, 373–381.
Taylor, St.C.S. (1980) Genetic size-scaling rules in animal growth. Animal Production
30, 161–165.
Tsaras, L.N., Kyriazakis, I. and Emmans, G.C. (1998) The prediction of the
voluntary food intake of pigs on poor quality foods. Animal Science 66,
713–723.
Von Bertalanffy, L. (1957) Quantitative laws for metabolism and growth. Quarterly
Reviews of Biology 32, 217–231.
Webster, A.J.F. (1993) Energy partitioning, tissue growth and appetite control.
Proceedings of the Nutrition Society 52, 69–76.
Wellock, I.J., Emmans, G.C. and Kyriazakis, I. (2003a) Modelling the effects of
thermal environment and dietary composition on pig performance: model
logic and concepts. Animal Science 77, 255–266.
Wellock, I.J., Emmans, G.C. and Kyriazakis, I. (2003b) Modelling the effects of
testing and evaluation. Animal Science 77, 267–276.
Wellock, I.J., Emmans, G.C. and Kyriazakis, I. (2004) Describing and predicting
potential growth in the pig. Animal Science 78, 379–388.
Whittemore, C.T. (1983) Development of recommended energy and protein
allowances for growing pigs. Agricultural Systems 11, 159–186.
Whittemore, C.T. (1998) The Science and Practice of Pig Production, 2nd edn.
Longman Scientific & Technical, Essex, UK.
Whittemore, C.T., Green, D.M. and Knap, P.W. (2001a) Technical review of the
energy and protein requirements of growing pigs: food intake. Animal Science
73, 3–17.
Whittemore, C.T., Green, D.M. and Knap, P.W. (2001b) Technical review of the
energy and protein requirements of growing pigs: energy. Animal Science 73,
199–215.
Whittemore, C.T., Green, D.M. and Knap, P.W. (2001c) Technical review of the
energy and protein requirements of growing pigs: protein. Animal Science 73,
363–374.
Whittemore, C.T., Emmans, G.C. and Kyriazakis, I. (2003) The relationship
between live weight and the intake of bulky food in pigs. Animal Science 76,
89–100.
4 The Effects of Social Stressors on
the Performance of Growing Pigs
I.J. WELLOCK, G.C. EMMANS AND I. KYRIAZAKIS
Animal Nutrition and Health Department, Scottish Agricultural College,
West Mains Road, Edinburgh, EH9 3JG, UK
Ian.Wellock@sac.ac.uk
Introduction
Different published approaches to pig simulation modelling have advanced
our understanding of pig performance under a wide range of
environmental conditions. They range from the first relatively simple
attempt to model pig growth by Whittemore and Fawcett (1974, 1976),
where predictions were based upon empirical equations, to more recent
and elaborate attempts such as those made by Black et al. (1986),
Whittemore and Green (2002) and Pomar et al. (2003). These latter models
contain various combinations of empirical and mechanistic equations
usually with an underlying biological basis. Attempts to predict feed intake,
although still not universal, are more frequent in recent modelling attempts
(e.g. Black et al., 1986; Bridges et al., 1992; Ferguson et al., 1994) and more
factors have been considered and introduced as model inputs. Stressors in
the physical environment, such as ambient temperature, humidity, air
velocity and floor type have been comprehensively modelled (e.g. Bruce
and Clark, 1979; Black et al., 1986; Wellock et al., 2003a,b) allowing
predictions of performance under varying conditions to be made. Factors
which may act as social stressors, which include group size (N), space
allowance (SPA, m2/BW0.67), feeder space allowance (FSA, feeders/pig), and
mixing on the other hand, have been largely ignored. This is mainly due to
a lack of quantitative data on which to build models and a lack of
understanding of how such stressors affect performance. Effects of the
infectious environment are yet to be included in a systematic way.
The objective of this chapter is to describe how the effects of social
stressors on the performance of growing pigs can be quantified and to
show how these relationships, including variation between genotypes in
their ability to cope (AB), can be incorporated into a more general pig
growth model. The consequences of introducing individual variation into
the model are investigated and the difficulties of estimating parameter

Effects of Social Stressors on Performance of Growing Pigs 55
values, with particular regard to AB, are described. Finally, the practical
implications of AB in relation to production, welfare and genetic selection
are discussed along with potential future model developments.
Modelling the Effects of Social Stressors

The influence of social stress on pig performance, although undeniable, is
frequently underestimated and, in pig growth modelling, generally
ignored. Black (2002) noted that ‘current pig models do not predict well
the effects of stress encountered by pigs reared in commercial
environments’. Only the pig growth model of NRC (1998) includes a social
stressor effect on performance, with SPA directly affecting dietary energy
intake. The adjustment is calculated from one of three equations according
to body weight (BW) and is added to dietary energy intake. However, this
is considered to be ‘a crude estimate [which] should be used with caution’
particularly at the ‘lower end of the three weight classes’ (NRC, 1998).
Kornegay and Notter (1984) developed linear regression equations
relating average daily feed intake (ADFI, kg), average daily gain (ADG, kg)
and feed conversion ratio (FCR, kg/kg), to SPA and N for pigs in three
weight ranges. More recently Turner et al. (2003) did the same, using the
same three weight ranges (Table 4.1). This latter exercise was over a much
larger data set and included group sizes of up to 120 as opposed to a
maximum of 33 in the analysis of Kornegay and Notter (1984). Whilst these
equations give some insight into how N and SPA may affect performance,
they are difficult to interpret and implement and fail to predict interactions
between the type of pig and the environment in which it is kept. For
example, the equations of Kornegay and Notter (1984) and Turner et al.
(2003) predict an ADG of zero when group size reaches 338 and 1363,
respectively, in growing pigs. This seems unrealistic, since pigs in groups of
Table 4.1. Equations from Kornegay and Notter (1984) and Turner et al., (2003)
relating group size (N) to average daily gain (ADG, kg), average daily feed intake
(ADFI, kg) and feed conversion ratio (FCR).
Kornegay and Notter (1984) Turner et al. (2003)
Weaner period (7.6 to 21.1 kg)
ADG = 0.4178 – 0.0037N (R2 = 0.97) ADG = 0.416 – 0.00036N (R2 = 0.97)
ADFI = 0.8317 – 0.092N (R2 = 0.97) ADFI = 0.618 – 0.00051N (R2 = 0.98)
FCR = 1.9535 – 0.0051N (R2 = 0.94) FCR = 1.650 + 0.00004N (R2 = 0.96)
Grower period (26.6 to 53.5 kg)
ADG = 0.6407 – 0.0019N (R2 = 0.43) ADG = 0.654 – 0.00048N (R2 = 0.90)
ADFI = 1.5950 – 0.0025N (R2 = 0.87) ADFI = 1.790 – 0.00005N (R2 = 0.90)
FCR = 2.4974 + 0.0037N (R2 = 0.94) FCR = 2.750 + 0.00179N (R2 = 0.97)
Finisher period (44.1 to 92.3 kg)
ADG = 0.7497 – 0.0012N (R2 = 0.82) ADG = 0.715 – 0.00009N (R2 = 0.99)
ADFI = 2.3748 + 0.0032N (R2 = 0.92) ADFI = 2.340 + 0.00033N (R2 = 0.84)
FCR = 3.2182 + 0.0060N (R2 = 0.72) FCR = 3.329 + 0.00104N (R2 = 0.97)
56 I.J. Wellock et al.
up to 2000 are now kept in profitable pig production enterprises, and is a

consequence of the range of data used in the empirical analysis.
Developing a Model
To develop a mechanistic model for predicting the effects of social stressors
on pig performance it is necessary to do the following three things: (i)
determine the mechanism by which social stressors affect performance, this
is needed to integrate social stressor effects into a model in a mechanistic
way; (ii) quantify the effects of the individual social stressors; and (iii)
integrate these social stressor effects into an overall growth model. These
three steps are discussed below in turn.
How social stressors affect performance
Integrating the effects of social stressors in the form of mechanistic

equations into a growth model poses the problem of describing how social
stressors affect pig performance. Unlike physical environmental stressors
(e.g. thermal environment) which affect pig performance in an ordered way
via known mechanisms (e.g. Bruce and Clark, 1979) the way in which social
stressors act is not clear. Amongst the possible mechanisms are: (i) a direct
reduction in appetite (Matteri et al., 2000); (ii) a reduction in the capacity to
deposit protein and attain potential growth (Chapple, 1993); and (iii)
increased metabolic demands diverting resources from the growth process
(Elsasser et al., 2000) resulting in a reduction in the efficiency of feed use.
Chapple (1993) used the AUSPIG simulation model developed by
Black et al. (1986) to investigate how changes induced by social stressors
observed in experiments may come about in order to try and elucidate the
mechanism of how social stressors lead to a depression in performance. He
found that a reduction in intake alone, i.e. a direct reduction in appetite,
could not explain the experimental observations. The effect of reducing
intake was predicted to result in leaner pigs with a lower backfat thickness,
but observations showed that an increase in group size resulted in an
increase in P2 backfat depth. A reduction in the pig’s ability to deposit
body protein was required in order to simulate the experimental result.
Consequently, Chapple suggested that the stressors associated with rearing
pigs in groups are mediated through biochemical growth factors that
down-regulate lean tissue growth, resulting in a reduction in feed intake. It
has been suggested that physiological factors such as growth hormone
(MacRae and Lobley, 1991), plasma cortisol (von Borell et al., 1992),
insulin-like growth factor or cytokines (Chapple, 1993) may be responsible.
Experiments where an increased protein supply to crowded pigs did not
overcome their decreased performance relative to non-crowded pigs
(Edmonds et al., 1998; Ferguson et al., 2001) also support the mechanism
suggested by Chapple (1993). For example, Edmonds et al. (1998) found that
crowded pigs with lower feed intakes required a substantially lower quantity
of amino acids than their uncrowded counterparts when expressed as a
percentage of the diet. If the mechanism was a direct reduction in appetite

or increased metabolic demands it would be expected that the crowded pigs
would have a requirement equal to or greater than the uncrowded pigs.
Although not simulated by Chapple (1993), a reallocation of resources is
unlikely to be the mechanism responsible as, providing no other constraint is
present, it is assumed that an animal will eat to meet its requirement for the
first limiting resource (Emmans and Fisher, 1986). Thus it is expected that
any increase in resource demands would be met by compensatory feed intake.
Consequently, it is suggested that a down regulation in lean tissue growth,
which may also be described as a decrease in the animal’s ability to attain its
potential, is the most suitable mechanism to incorporate into a model to
describe the detrimental effects of social stressors on pig performance.
Representing the effects of individual social stressors: functional form and

parameter estimation
The relationship between social stressors and performance is described

here by conceptual equations derived on biological grounds that are then
parameterized using experimental data. Rather than predicting values for
the model output variables, such as ADFI and ADG by an empirical
adjustment (e.g. Kornegay and Notter, 1984; Turner et al., 2003), the
approach used here integrates the chosen functional forms into a general
growth model as mechanistic equations. By this means any interactions
that exist between the type of pig and its environment can be explored
and, at least in principle, predicted.
In order to test the chosen functional forms for their relevance, and to
enable realistic quantitative values to be assigned to the parameters,
experimental data must be used (Table 4.2). To avoid the various problems
of using a strictly empirical approach, several measures were taken,
including the following: (i) using only experiments where all variables
other than the one of interest were controlled for, to try and avoid the
confounding effect of other variables; (ii) using more biologically relevant
methods where possible, for example, by relating space allowance to BW
rather than simply area per pig; (iii) taking differences in live weight into
account. Relative daily gain (R) was used as the measure of performance
rather than daily gain, thus eliminating the need for separate equations
according to BW and allowing a greater amount of information to be
included in the analysis; (iv) accounting for differences in the potential of
pigs used; (v) accounting for differences between the experiments and
giving appropriate weighting for the number of replications in each
experiment; and (vi) checking that the equations used are sensible when
extrapolated over the full range of interest.
Space allowance
Decreasing the space allowance available to a group of pigs depresses

intake and growth (e.g. Edwards et al., 1988; Hamilton et al., 2003). This
Table 4.2. Parameter values for the conceptual equations relating the major social stressors
to pig performance estimated from experimental data in the literature (reproduced from
Wellock et al., 2003c, Table 1).
Equationa Parameter 1 Parameter 2 Parameter 3
1 RSPA = b1 + g1 ln (SPA) b1 = 168.49 g1 = 21.48
(9.62)b (2.65)b –
2 RN = b2 – g2 ln (N) b2 = 100c g2 = 3.6971
(0.69)b –
3 EN = (x1.(N-1)).EMaint x1 = 0.0075 – –
5 FRmax = (g3.BW1) / (1000.WHC) g3 = 2.85 – –
7 RMix = b3 – g4.BW – ((g5.BW).ln.(t)) b3 = 100c g4 = 0.6 g5 = 0.18
8 EMix = (x2 – (x3.ln.(t))).EMaint x2 = 1.15 x3 = 0.050 –
aR
SPA, RN and RMix represent the relative daily gain as a percentage of maximal performance
in relation to space allowance (SPA, m2/BW0.67), group size (N) and mixing (Mix),
respectively. EMaint, EN and EMix represent the energy expenditure (MJ/d) due to maintenance,
group size and mixing, respectively. FRmax is the maximal feeding rate (kg/min) and WHC is
the water holding capacity (kg/kg) of the food used as a measure of its bulk.
b Values in parenthesis are standard errors.
c Denotes fixed parameter values.
may be due to an increase in the frequency of antagonistic encounters that

may to an extent depend on breed. It appears that there is a critical value,
(SPAcrit, m2/BW0.67) below which performance becomes depressed. It is
assumed that above SPAcrit, there is no effect on performance and that
when SPA reaches another crucial value (SPAmin, m2/BW0.67), growth is no
longer able to occur (see Fig. 4.1). SPAmin is reached when the pen is fully
occupied and sets the minimum value for SPA within the model, 0.019
m2/BW0.67 (Petherick, 1983). When SPAmin < SPA < SPAcrit, relative daily
gain, RSPA, in relation to that recorded at a SPA > SPAcrit, is calculated as:
RSPA = b1 + g1.f.(SPA) (4.1)
b1 and g1 determine SPAcrit and the extent of performance depression
below SPAcrit. They are affected by genotype as discussed below. SPA =
Area/BWq where Area is m2/pig and q is the body weight scalar assumed to
be 0.67 (Petherick, 1983). As there appears to be no clear biological
expectation for the shape of the functional form, f, between SPAcrit and
SPAmin, the logarithmic relationship was chosen (Table 4.2) after inspection
of the experimental data. The value assigned to SPAcrit was 0.039
m2/BW0.67. To account for the greater space requirements of pigs housed
on solid floors, SPA in Eqn 4.1, is decreased by 25% in agreement with
Whittemore (1998), when pigs are housed on solid floors.
Group size
There appears to be an effect of grouping per se as individually housed pigs

have been widely shown to outperform their group-housed counterparts
(e.g. Gonyou et al., 1992) and most experiments report a decrease in
1
SPAmin
RSPA
SPAcrit
0
SPA, m2/BW0.67
Fig. 4.1. Effect of space allowance on the relative daily gain (RSPA) in relation to that recorded
at SPA > SPAcrit. SPAcrit represents the point below which space becomes limiting resulting in
a depression in performance and SPAmin represents the minimum amount of space required
for a given animal. Both SPAcrit and SPAmin are affected by body weight and genotype.
performance as N increases (e.g. Wolter et al., 2001; Hyun and Ellis, 2001).
Increasing N by a fixed number has a greater influence on smaller groups
than larger ones, because the social hierarchy of small groups is disrupted
to a greater degree than that of large groups that appear to lack social
structure (Arey and Edwards, 1998; Turner et al., 2001). A logarithmic
form is used to represent the relationship (Fig. 4.2):
RN = b2 – g2.ln.(N) (4.2)
RN is the daily gain as a proportion of that of a singly housed counterpart.
The value of the constant b2 is set to 100 and the value of g2 is assumed to
differ between breeds (see below). Calculated parameter values are given
in Table 4.2.
Although the evidence for greater activity in larger groups is equivocal,
the literature does indicate such a trend (e.g. Petherick et al., 1989; Turner
et al., 2002). Consequently, energy expenditure is increased as N increases.
This increase, EN, is calculated as a proportion of maintenance energy
requirements (Emaint, MJ/day) and included in the calculation of daily
energy requirements. It is assumed that EN will not increase indefinitely
and so a proposed maximum is set at Nm. When N < Nm:
EN = (x1.(N1)).Emaint (4.3)
The value of x1 will differ between genotypes as discussed later. When
NNm, Nm replaces N in Eqn 4.3. To account for a twofold increase in
activity as N increases to Nm, a value of 0.0075 was assigned to x1. A value
of 21 was assigned to Nm.
Feeder space allowance
Intake is reduced when the number of feeder spaces available to a group

of pigs falls below a critical value (FSAcrit, feeder spaces/pig), and continues
to decrease as FSA decreases further (e.g. Nielsen et al., 1995). FSAcrit is
RN
Group size (N)
Fig. 4.2. Effect of increasing group size (N) on relative daily gain (RN) in relation to that of a
singly housed counterpart.
reached when all of the pigs in the group can no longer satisfy their
desired feed intake (FId, kg/day), due to increased pig competition and is
dependent upon N, FId, maximum feeding rate (FRmax kg/min), and the
number of minutes in the day, 1440.
FSAcrit = (FId/(1440.FRmax)) N (4.4)
FRmax depends upon aspects of mouth capacity, feed composition and
method of feed presentation. It can be calculated as:
FRmax = (g3.BW1.0)/(1000.WHC) (4.5)
where WHC (kg/kg) is the water holding capacity of the feed, used as a
measure of feed bulk. It is assumed that g3 is unaffected by genotype.
When FSA < FSAcrit, then the constrained feed intake (FIc, kg/pig), is
calculated as:
FIc = (1440.FSA.FRmax)/N (4.6)
If troughs are used FSA is calculated as the number of pigs able to feed
simultaneously so that FSA = trough width/ j.BWk. The values assumed for
j and k are 0.064 and 0.33, respectively (Petherick, 1983) and these
estimate the width of the pig at the shoulders.
Mixing
Generally results indicate that mixing is a transient stressor and that, given
sufficient time, there are no noticeable effects of mixing on performance in
the longer run (Spoolder et al., 2000; Heetkamp et al., 2002). There is
however, an initial decline in performance immediately after mixing most
likely due to the increased frequency of antagonistic encounters (D’Eath,
2002) associated with establishing a new stable social structure. Mixing
depresses performance to a greater extent in larger animals due to the
increased ferocity of their fighting (Spoolder et al., 2000) before returning
to normal values. The mixing effect is described as:
RMix = b3 – ((g4.BW) – ((g5.BW).ln(t))) (4.7)

RMix is the performance relative to that of a non-mixed pig. The value of
the constant b3 is set to 100. The values of the parameters g4 and g5 are
likely to change with genotype (see below); t is the time in days after
mixing occurs. At some value of t, RMix will be estimated to be 100. From
then on performance is normal and no longer affected by the past mixing
(Fig. 4.3). Values were chosen (Table 4.2) so that mixing decreased
performance by an average of approximately 25% in a 70 kg pig in the first
week after mixing and had an effect that lasted for 2 to 3 weeks (Tan et al.,
1991; Stookey and Gonyou, 1994).
Mixing also increases the energy expenditure of pigs due to increased
levels of aggression, especially in the first few days after mixing (Heetkamp
et al., 1995). This increase in energy expenditure due to mixing, Emix
MJ/day, which decreases over time as activity levels return to normal, is
added to the daily energy requirements.
Emix = (x2 – (x3.ln(t))).Emaint (4.8)
The values of the scalars x2 and x3 were chosen to represent an increase in
Emaint by a maximum of 15%, following EN, and to have an effect that lasts
for 2 to 3 weeks. These values are expected to change with genotype.
Genetic differences
It is envisaged that there is genetic variation between breeds in their ability to

cope with social stressors (Beilharz and Cox, 1967; Grandin, 1994; Schinckel
et al., 2003). These differences are accounted for by introducing a parameter,
AB, to describe the pigs ‘ability to cope’ when exposed to social stressors. This
adjusts both the intensity of a stressor at which the animal becomes stressed,
e.g. SPAcrit, and the extent to which stress reduces performance (see Fig. 4.4)
and increases energy expenditure (activity) at a given stressor intensity. It is
assumed in the model that these two factors are correlated.
tmix
ADG (kg/day)
ADGdep
Non-mix
Mix
Time (days)
Fig. 4.3. Effect of mixing on average daily gain (ADG). The extent of depression in ADG
(ADGdep) and the time taken (tmix) to return to non-stressed levels of ADG is determined by
both body weight and genotype.
SPAcrit
105
100
95
RSPA
90
85
AB = 10
80
AB = 5
75 AB = 1
70
0.2 0.4 0.6 0.8 1.0
Area (m2/pig)
Fig. 4.4. The effect of genotypic differences in ability to cope with social stressors (AB) on
relative daily gain (RSPA) at differing space allowances. SPAcrit represents the points at which
space becomes limiting.
Increasing AB from 1 to 10 represents an increasing ability to cope and

modifies the effect of each stressor by multiplying the estimated parameters
shown in Table 4.2 by appropriate scaling factors (Table 4.3). Because there
is currently no established genetic variation in pig’s ability to cope with
stressors, values for the scaling factors were estimated to represent
deviations of approximately 1% from the mean, AB = 5, per unit change in
AB. For example, AB values of 1 and 10 predict an approximate departure
from the mean of 5 and +5%, respectively, at a given stressor intensity.
SPAcrit falls from 0.042 to 0.031 m2/BW0.67 as AB decreases from 10 to 1. It
is expected that modern, ‘leaner’, genotypes will have lower values of AB
than traditional, ‘fatter’, genotypes (Grandin, 1994; Schinckel et al., 2003).
Incorporating Effects of Social Stressors into a More General Pig

Growth Model
Information required
The model framework used as the starting point is that described and
tested by Wellock et al. (2003a,b). Information is needed about the pig, its
diet and the social and physical environments in which it is kept. No
additional inputs are required to describe either the thermal or dietary
environment. In addition to the three genetic parameters used to predict
potential growth [protein weight at maturity (Pm, kg), the lipid to protein
ratio at maturity (Lm/Pm, kg/kg) and a growth rate parameter (B, d1)], the
parameter, AB, discussed above, is required to describe the pig’s ability to
cope with social stressors. Additional inputs to describe the social
environment are pen area, the number of pigs per pen (N), the number of
individual feeders or the trough length, and the occurrence or not of
mixing. Up to two mixing events are allowed during a run and the weight
at which each mixing occurs is required. The model can be run either to a
final BW (BWf, kg) or for a given period (t, days).
Table 4.3. Scaling factors (z) for the appropriate parameters to account for variation in ability
to copea (AB) with social stressors (adapted from Wellock et al., 2003c).
Space allowance Group size Mixing
zb1 = 1.2 – 0.04AB zg2 = 1.5 – 0.1AB zg4 = 1.2 – 0.04AB
zg1 = 1.5 – 0.1AB zx1 = 1.1 – 0.02AB zg5 = 0.977 + 0.066AB
zx2 = 1.025 – 0.005AB
zx3 = 0.9 + 0.02AB
aAn AB value of 5 represents the mean and therefore all parameters are multiplied
by a scaling factor of 1 when AB = 5. Values were chosen to represent deviations of
approximately 1% from the mean performance per unit change in AB.
Integrating the mechanism and social stressor equations
It has been shown in studies with pigs (Hyun et al., 1998a,b) that the effects
of stressors, at intensities expected under commercial conditions, are likely
to be additive rather than multiplicative, antagonistic or synergistic. It is
therefore assumed that, within the bounds of the model, the effects of
multiple stressors on maximum relative daily gain of the stressed animal,
Rs, are additive and are predicted by summing the effects of the individual
stressors.
Rs = Rp.((100 – ((100 – RSPA) + (100 – RN) + (100 – Rmix))) / 100) (4.9)
Here Rp = ADGp / BW and is the pig’s potential relative daily gain. ADGp is
dependent upon the genotype and the current state of the pig. It is
assumed in the model that social stressors lead to a down-regulation in
lean tissue growth, i.e. a decrease in the animal’s ability to attain its
potential. This is equivalent to lowering the growth rate parameter, B,
which in turn leads to a decrease in the ADGp that the pig is able to
achieve. Consequently, Rs is calculated on a daily basis and used to
calculate the modified growth rate parameter, Bs = Rs B, from which
the maximum daily gain of the stressed animal is predicted, ADGs,
replacing ADGp in the model.
As it is assumed that animals eat to attain their potential, a decrease in
ADGp necessarily leads to a decrease in FId. Consequently, the desired feed
intake of the stressed animal is predicted directly from the animal’s
depressed growth potential. Predictions of FId, actual intake and gain are
then made taking account of any changes in energy requirements due to
increases in activity, EN and Emix, and possible constraints on intake due to
limiting FSA, feed composition and the climatic environment.
Introducing Between-animal Variation into the Model

One of the problems and perhaps main limitations of using a model that
represents a single average pig, is the assumption that all pigs within a
population are the same. In reality of course this is not true. A
consequence of between-animal variation is that there may be differences
between the response of the average individual and the mean response of
the population, which is an average of all individuals (Fisher et al., 1973;
Emmans and Fisher, 1986). These differences may prove important when
predicting nutrient requirements (Fisher et al., 1973; Curnow, 1986),
optimizing pig production systems (Pomar et al., 2003) and devising animal
breeding strategies (Knap, 1995).
In order to account for differences between individuals in a group,
between-animal variation was introduced into the model. In addition to
accounting for differences in growth potential, as in the stochastic pig
growth models of Knap (2000) and Pomar et al. (2003), variation in initial
state as described by initial BW (BW0) and ability to cope when exposed to
stressors were also included. Variation in growth potential was generated
by creating variation around the population means of each of the genetic
parameters describing growth, Pm, Lm/Pm and B (Ferguson et al., 1997).
Individual variation in BW0 is generated from the assigned genotype
mean, (mBW0, kg) and standard deviation (sBW0, kg) using the simulated
genetic parameters of the individual to correlate BW0 with potential
growth. By this means individuals in the group with the greatest potential
will tend to have the highest BW0 as would be expected from non-limiting
growth. For further details see Wellock et al. (2004).
It is expected that within a population or group the social
environment (i.e. position within the social hierarchy) affects an
individual’s ability to cope (Muir and Schinckel, 2002) and that pigs
classified as dominant tend to outperform their subordinates (e.g. Hessing
et al., 1994; D’Eath, 2002). There is also evidence that social dominance is
positively correlated to BW in pigs (Drickamer et al., 1999; D’Eath, 2002).
Taken together these results suggest that the larger pigs within a group
tend to be dominant and better able to cope when conditions are sub-
optimal, i.e. when pigs are exposed to social stressors. Consequently, it is
assumed in the model that there is a positive correlation between BW0 and
AB. Individual values for AB (ABi) are generated around the assigned
genotype mean (mAB) and standard deviation (sAB) of AB, whilst being
positively correlated to BW0.
ABi = mAB + b4.((BW0i / mBW0).(sAB.(mBW0 / sBW0))) ± residuali (4.10)
The parameter b4 determines the degree of correlation between BW0 and
AB and is set equal to one. The residuali is drawn at random taking
account of sAB. Within a population, AB is not directly correlated to
leanness. However, leaner animals will tend to have lower AB values due to
the positive correlations between Lm/Pm and BW0 and BW0 and AB.
Model Simulations: Practical Implications of AB in Relation to

Production, Welfare and Genetic Selection
Effect of growth potential and ability to cope on pig performance
The effects of differing abilities to cope when exposed to social stressors

were explored, using the model, for pigs with two levels of potential
performance, ‘good’ and ‘poor’. Simulations for pigs grown from 80 to 100
kg and either mixed or not on day 1 are shown in Table 4.4. As expected,
mixing led to a decrease in performance, with pigs having the poorest
ability to cope (AB = 1) displaying the largest decrease. For example, good
genotype pigs with AB values of 10, 5 and 1 were predicted to show a
decrease in ADG of 9, 20 and 35%, respectively, compared to the non-
mixed counterparts. Interestingly, poor genotype pigs with high (AB = 10)
and average (AB = 5) abilities to cope were both predicted to outperform
the good genotype pigs with a low ability to cope (AB = 1), reaching 100
kg 6 and 3 days earlier, respectively. This implies that pigs with the highest
potential for growth do not always outperform others and that an animal’s
response to stressors, i.e. its ability to cope, may be as important as its
growth potential, particularly when raised under commercial conditions.
Average individual versus mean population response
Ferguson et al. (1997) stated that ‘there is a marked difference in the

response of the average individual in the population and the mean of the
population’. Pomar et al. (2003) demonstrated clear differences between
the average individual and the mean population response for the rate of
protein retention in response to increasing dietary protein intake.
However, from the model simulations shown in Fig. 4.5 it is clear that
differences between the average pig and the mean population responses
should not always be expected and will depend partly upon the stressors to
which the pigs are exposed. Where all individuals become adversely
affected at the same stressor intensity, e.g. being housed in a group as
opposed to individually or being mixed or not, then no differences
between the average individual and mean population response is predicted
(Fig. 4.5a). This is because all individuals are either affected or not,
although this may be to varying extents. If however the intensity at which
Table 4.4. The effects of pig potential (ADGp) and ability to cope (AB) when exposed to
stressors on the time taken (t) to reach 100 kg, average daily gain (ADG), average daily feed
intake (ADFI) and feed conversion ratio (FCR) from a starting weight of 80 kg. The effect of
mixing pigs on day 1 of the simulation is also shown for the AB = 10 genotype.
ADGp AB Mix t (days) ADG (kg/day) ADFI (kg/day) FCR (kg/kg)
Gooda 10 No 22 0.93 2.51 1.58
10 Yes 24 0.84 2.37 2.82
5 Yes 28 0.74 2.22 3.00
1 Yes 34 0.60 2.01 3.35
Poorb 10 No 26 0.80 2.66 3.33
10 Yes 28 0.74 2.52 3.41
5 Yes 31 0.65 2.34 3.60
1 Yes 38 0.53 2.09 3.94
a Genetic growth parameters: B = 0.016 day1, Pm = 32 kg, Lm/Pm = 1.2 kg/kg.
b Genetic growth parameters: B = 0.011 day1, Pm = 30 kg, Lm/Pm = 2.0 kg/kg.
(a) (b)
1.7 2.50
1.6 2.48
ADFI (kg/day)
ADFI (kg/day)
1.5
2.46
1.4
2.44
1.3
1.2 Average individual 2.42 Average individual

Population mean Population mean
1.1 2.40
0 20 40 60 80 100 0.75 0.80 0.85 0.90 0.95 1.00
Group size Pen area (m2/pig)
Fig. 4.5. Predicted effect of environmental stressors on the average daily feed intake (ADFI)
response of the average individual and population mean; (a) the effect of increasing group
size on the ADFI of pigs from 20 (± 2 kg) to 50 kg; (b) the effect of decreasing space
allowance on the ADFI of 100 kg (± 10 kg) pigs over a simulation period of one day.
the stressor becomes limiting is able to differ between individuals, e.g.

critical SPA (SPAcrit, m2/BW0.67) and upper critical temperature, differences
between the average individual and mean population response are
expected (Fig. 4.5b).
The linear-plateau response of the average individual to decreasing SPA
is a direct outcome of the assumption used in the model. The curvilinear-
plateau response of the population however can be explained by individual
differences in SPAcrit, generated from between-animal variation in BW and
AB. The plateau is predicted to occur when SPA > SPAcrit for all pigs in the
population and the curvilinear transition phase occurs when only a
proportion of the population is constrained, i.e. SPA < SPAcrit for only some
individuals. As the intensity of the stressor increases, the proportion of the
population that is constrained also increases until all individuals are
affected. At a fixed SPA the proportion of pigs limited will increase with
increasing population variance and this will result in a greater degree of
curvature. This was demonstrated by Pomar et al. (2003) for average daily
rate of protein retention in response to increasing protein intake.
Variation in initial state and ability to cope with social stressors
The model predicts that variation in the growth response of a population

is determined to a greater extent by variation in AB and BW0 than by
variation in growth potential, when pigs were exposed to social stressors.
This is demonstrated in Table 4.5 and Fig. 4.6.
Table 4.5 shows the effect of variation in growth potential, BW0 and
AB on the performance of 500 pigs raised from 60 kg to a given BWf of
100 kg when raised under typical commercial conditions. Fig. 4.6 is from
the same simulations and shows how the distribution in the time taken to
Table 4.5. Effect of variation in initial BW (sBW0, kg) and ability to cope (sAB) on the
variation in the time taken (t, days) to reach 100 kg from a mean BW0 of 60 kg. P and L are
the protein and lipid content, respectively. Mixing occurred at 75 kg and pigs were given a
space allowance of 0.7 m2/pig throughout (adapted from Wellock et al., 2004).
sBW0a sAB t (days) ADG (kg/day) ADFI (kg/day) P (kg) L (kg)
0.00 0.00 50.7 (3.93)b 0.801 (0.060)b 2.17 (0.104)b 17.78 (0.352)b 15.59 (1.534)b
2.07 0.00 50.8 (5.48) 0.804 (0.057) 2.17 (0.107) 17.78 (0.333) 15.59 (1.445)
4.19 0.00 51.5 (7.64) 0.802 (0.056) 2.17 (0.119) 17.78 (0.357) 15.59 (1.592)
5.74 0.00 51.5 (9.42) 0.801 (0.052) 2.16 (0.121) 17.78 (0.358) 15.58 (1.544)
7.87 0.00 52.2 (11.59) 0.801 (0.053) 2.16 (0.131) 17.77 (0.346) 15.63 (1.500)
10.04 0.00 52.2 (14.23) 0.795 (0.054) 2.15 (0.132) 17.78 (0.338) 15.60 (1.506)
12.32 0.00 54.1 (17.42) 0.792 (0.054) 2.14 (0.148) 17.78 (0.352) 15368 (1.548)
0.00 0.47 50.8 (4.06) 0.799 (0.061) 2.17 (0.105) 17.76 (0.350) 15.70 (1.548)
0.00 0.94 50.8 (4.16) 0.798 (0.063) 2.17 (0.108) 17.79 (0.359) 15.54 (1.580)
0.00 1.45 50.8 (4.71) 0.799 (0.072) 2.17 (0.117) 17.79 (0.342) 15.61 (1.495)
0.00 1.93 50.8 (5.14) 0.799 (0.079) 2.17 (0.128) 17.77 (0.349) 15.64 (1.579)
0.00 2.53 50.8 (5.52) 0.802 (0.087) 2.17 (0.140) 17.79 (0.354) 15.53 (1.576)
5.77 1.40 52.1 (11.48) 0.801 (0.071) 2.17 (0.161) 17.77 (0.331) 15.63 (1.435)
12.18 2.42 54.2 (21.05) 0.799 (0.080) 2.15 (0.222) 17.79 (0.338) 15.46 (1.479)
a Simulated values.
b Result of variation in growth potential only.
reach BWf changes as variation in BW0 and AB increases. Consequently, it

is suggested that the pig’s potential for growth might be less important
than the pig’s response to stressors when pigs are reared in commercial
environments. This is because improving the ability of pigs to cope would
allow a greater proportion of their potential to be attained and may be a
better way of improving pig performance and enterprise profitability than
increasing potential per se. Schinckel et al. (2003) also noted that ‘the pig’s
genetic potential for protein accretion and feed intake are less important
than the pig’s response to encountered stressors’ and for these reasons
suggested that ‘farm genetic population specific growth and feed intake
parameters are required’.
If, as suggested, AB and lean growth rate are adversely correlated
(Grandin, 1994; Torrey et al., 2001; Schinckel et al., 2003), then there may
be negative implications regarding the welfare of pigs selected for lean
growth. This is because selection for improved lean growth rate would
indirectly lead to selection for poorer ability to cope in the population. Fig.
4.7 shows the correlation between Lm/Pm and AB simulated by the model
for a population of 500 pigs with population mean (± SD) values of 1.2
(0.18) and 5 (1) for Lm/Pm and AB, respectively. Since AB depends in part
upon the structure of the group, then group selection may be necessary in
order to improve the ability of animals to cope when exposed to social
stressors. Griffing (1966) found that individual selection could result in a
negative response of the population mean. The experiments of Muir and
120 (a) 120 (b)

100 100
Frequency
Frequency
80 80
60 60
40 40
20 20
0 0
20 30 40 50 60 70 80 90 100 20 30 40 50 60 70 80 90 100
Time (days) Time (days)
120 (c) 120 (d)

100 100
Frequency
Frequency
80 80
60 60
40 40
20 20
0 0
20 30 40 50 60 70 80 90 100 20 30 40 50 60 70 80 90 100
Time (days) Time (days)
Fig. 4.6. The effect of variation in initial BW (sBW0) and ability to cope with social stressors
(sAB) on the time taken to reach 100 kg from an initial mean BW of 60 kg (N = 500); (a) sAB
= 0, sBW0 = 0, i.e. variation in growth potential only; (b) sAB = 1.5, sBW0 = 0; (c) sAB = 0,
sBW0 = 6 kg, and (d) sAB = 6, sBW0 = 1.5 kg. Mixing occurred at 75 kg and pigs were given
a space allowance of 0.7 m2/pig throughout.
Schinckel (2002) with quail and Muir (1996) and Muir and Craig (1998)
with poultry also demonstrated that selection for desirable associate effects
within a group may be a means to select animals which are better adapted
to their rearing environment. Any genetic correlation between AB and the
growth parameters that can be evaluated could be included in the model by
incorporating the co-variation between the identified parameters and AB.
Quantifying the variation in AB may improve the rate of breeding for
a better ability to cope, as the amount of heritable variation determines the
degree of selection pressure able to be applied. If a parameter such as AB
was included in a selection index then individual pigs with both the
greatest growth potential and best ability to cope could be selected for. For
example, animal ‘a’ shown in Fig. 4.7 may be a better breeding prospect
than animal ‘b’ as it has the desirable properties of having a low Lm/Pm
value and high AB unlike animal ‘b’, which has a high Lm/Pm and low AB.
This would result in benefits for both welfare and production. If increased
growth rate and ability to cope are antagonistic, then trying to increase pig
performance achieved under excellent conditions, i.e. improving potential
8
a
7
AB 4
b
1
0.0 0.5 1.0 1.5 2.0 2.5
Lm/Pm (kg/kg)
Fig. 4.7. The simulated correlation between leanness, as represented by the lipid to protein
ratio at maturity (Lm/Pm), and ability to cope when exposed to environmental stressors (AB).
Five hundred individuals were simulated with a population mean (± SD) of 5 (1) for AB and 1.2
(0.18) Lm/Pm. Animal ‘a’ may be a better breeding prospect than animal ‘b’ as it has the
desirable properties of having a low Lm/Pm value and high AB, unlike animal ‘b’ which has a
high Lm/Pm and low AB.
alone, may not prove to be the best selection strategy. It is likely that
improvements in the growth potential of the animals and in the
environment, particularly better biosecurity and vaccination, are required
in addition to improving pigs’ ability to cope.
Future Model Developments
Estimation of model parameter values with particular reference to AB
Currently there are no means of assigning estimates to the AB parameter

introduced into the model developed here to describe the ability of pigs to
cope when exposed to social stressors. However, assuming that there is a
measurable phenotypic difference between types of pigs and individuals
within a population, it is thought that genetic characterization is possible.
The work of de Greef et al. (2003) and Kanis et al. (2002) supports this.
They described and evaluated a conceptual framework for breeding for
improved welfare in pigs and showed that it is possible to select for abilities
to cope with stressors such as environmental temperature.
To satisfactorily test whether the introduction of AB is useful and to

quantify it by experimentation is likely to require an elaborate experiment
with a large number of pigs of different breeds, strains and sexes exposed
to a large number of treatments. This is unlikely to be carried out.
Nevertheless, it is thought that more modest, small scale, experiments may
allow first tentative estimates of both the genotype mean and between-
animal variation in AB to be made. Animal scientists have long been
designing experiments exposing pigs of different breeds and sexes to a
number of differing social stressors. These have included studies
manipulating group size (e.g. Wolter et al., 2001), space allowance (e.g.
Hyun et al., 1998a), feeder space allowance (e.g. Nielsen et al., 1995) and
mixing (e.g. Stookey and Gonyou, 1994). However, it has been scientists
interested mainly in the behaviour of pigs that have usually conducted
these experiments. As a consequence performance measures have often
been neglected or not suitably reported. For instance, no experiments
could be found in the literature where individual performance of mixed
pigs had been presented. The few ‘mixing’ experiments which reported
any performance information (e.g. Hessing et al., 1994; D’Eath, 2002) did
so only for the group. Simply including measures of performance in
conjunction with the usual behavioural measures would allow progress to
be made. For example, recording individual pig feed intake and gain on a
daily basis for the duration of a ‘normal’ mixing experiment would give an
indication of the effect the stresses of mixing have on individual
performance. Linking the expected decreases in intake and gain due to
mixing with the BW and position of the individual within the dominance
hierarchy would also allow an initial test of the assumption used in the
model: that bigger pigs within the population are the ones that cope best
when socially stressed.
Comparing the variation in performance observed in experimental
data with the variation predicted by the model will also allow an initial
estimate of the variation in AB to be made. This inverted modelling
technique was the method used by Ferguson et al. (1997) when predicting
the variation in B*, Lm/Pm and Pm. However in order for this to be done
successfully a measure of the heritability of AB is also required. It is also
important to know if any correlations exist between AB and any of the
other genetic parameters, particularly leanness described by Lm/Pm. If so,
this will affect the nature and description of the variation of the correlated
parameters (Ferguson et al., 1997) and would need to be accounted for in
the model. This of course relies on the simplistic assumption that
individuals react in the same way to all types of social stressors. However if
this is incorrect, the introduction of further parameters, in addition to AB,
will be required for a sufficient description of ability to cope when exposed
to social stressors.
Modelling the effects of infectious stressors
It was assumed in the model that all animals were in good health and free
from exposure to infectious stressors throughout. Any response to
infectious stressors, such as an increase in resource requirements to cope
with the consequences of infection, acquire and express an immune
response, a change in the efficiency of energy utilization or a voluntary
reduction in feed intake (anorexia), all of which would result in a decrease
in performance, were ignored. In reality of course, pigs are continuously
exposed to many different kinds and intensities of infectious stressors.
These include pathogens and other harmful environmental components
that may trigger tissue injury or further infection, such as bites and
scratches from other individuals in the same pen.
The incorporation of infectious stressors into simulation models is an
important next step in the attempt to predict commercial pig performance
accurately. To include the effects of infectious stressors into a model in a
systematic way it is necessary first to do a number of things. The metabolic
load imposed by infectious stressors, i.e. increased nutrient requirements,
and the extent to which performance is decreased need to be quantified.
How animals allocate resources when exposed to infectious stressors, e.g.
cope with a pathogen challenge, needs investigating and the biological
mechanism responsible for the decrease in performance needs elucidating.
Two possible mechanisms may lead to the decrease in pig performance
observed when pigs are exposed to disease. These are either a decrease in
the pigs’ ability to attain their potential, as suggested by Schinckel et al.
(2003), or a direct decrease in appetite as suggested by Kyriazakis (2003).
There is also likely to be between-animal variation in immune response
and resilience, i.e. differences in the ability of individual pigs to cope and
perform during exposure to pathogens (see Kyriazakis and Sandberg,
Chapter 7, this volume). This should be accounted for in any future
modelling attempt, along with any potential interactions between stress
and disease susceptibility when such information becomes available.
Conclusion
Despite their importance, few attempts have been made to quantify the
effects of social stressors on pig performance and incorporate these effects
into a pig growth simulation model. Here we describe how the effects of
the major social stressors, i.e. group size, space allowance, mixing and
feeder space allowance, can be described by conceptual equations based on
the biology of the animal, quantified and incorporated into a more general
pig growth model. The adapted model allows the performance of both
individuals and populations of growing pigs differing in initial state,
growth potential and ability to cope with social stress when raised under
given dietary, physical and social environmental conditions to be explored
and, at least in principle, predicted.
Among the main outcomes of the model simulations are the following:
(i) allowance for population variance is crucial in making decisions as there
may be differences between the response of the population and the
average individual; (ii) improving management to minimize stressors and
decrease variation in initial state is an important factor in decreasing the
heterogeneity of a group, particularly in commercial production systems
where payment is based upon uniformity; and (iii) if growth rate and
ability to cope when exposed to social stressors are antagonistic, trying to
improve pig performance by increasing growth potential alone may not be
the best selection strategy.
Acknowledgements
This work was supported by the Biotechnology and Biological Sciences

Research Council of the United Kingdom and the Scottish Executive,
Environment and Rural Affairs Department.
References
Arey, D.S. and Edwards, S.A. (1998) Factors influencing aggression between sows
after mixing and the consequences for welfare and production. Livestock
Production Science 56, 61–70.
Beilharz, R.G. and Cox, D.F. (1967) Social dominance in swine. Animal Behaviour
15, 117–122.
Black, J.L. (2002) Pig growth models future. Journal of Animal Science 80 (Suppl. 1),
174 (Abstr.).
Bridges, T.C., Turner, L.W., Stahly, T.S., Usry, J.L. and Loewer, O.J. (1992)
Modelling the physiological growth of swine. Part I: Model logic and growth
concepts. Transactions of the American Association of Agricultural Engineers 35,
1019–1028.
Chapple, R.P. (1993) Effect of stocking arrangement on pig performance. In:
Batterham, E.S. (ed.) Manipulating Pig Production IV. Australian Pig Science
Association, Victoria, pp. 87–97.
Curnow, R.N. (1986) The statistical approach to nutrient requirements. In: Fisher,
C. and Boorman, K.N. (eds) Nutrient Requirements of Poultry and Nutritional
Research. Butterworths, London, pp. 79–89.
D’Eath, R.B. (2002) Individual aggressiveness measured in a resident-intruder test
predicts the persistence of aggressive behaviour and weight gain of young pigs
after mixing. Applied Animal Behaviour Science 77, 267–283.
de Greef, K.H., Kanis, E., Van der Belt, H., Groen, A.F. and Schakel, J. (2003)
Breeding for improved welfare of pigs, a conceptual framework. Proceedings of
the 7th World Congress on Genetics Applied in Livestock Production 32, 39–42.
Drickamer, L.C., Arthur, R.D. and Rosenthal, T.L. (1999) Predictors of social
dominance and aggression in gilts. Applied Animal Behaviour Science 63,
121–129.
Edmonds, M.S., Arenston, B.E. and Mente, G.A. (1998) Effect of protein levels and
space allocation on performance of growing-finishing pigs. Journal of Animal
Science 76, 814–821.
Edwards, S.A., Armsby, A.W. and Spechter, H.H. (1988) Effects of floor area
allowance on performance of growing pigs kept on fully slatted floors. Animal
Elsasser, T.H., Klasing, K.C., Filipov, N. and Thompson, F. (2000) Metabolic
consequences of stress. In: Moberg, G.P. and Mench, J.A. (eds) The Biology of
Animal Stress. Basic Principles and Implications for Animal Welfare. CABI
Publishing, Wallingford, UK, pp. 77–110.
Emmans, G.C. and Fisher, C. (1986) Problems in nutritional theory. In: Fisher, C.
and Boorman, K.N. (eds) Nutrient Requirements of Poultry and Nutritional
Ferguson, N.S., Gous, R.M. and Emmans, G.C. (1997) Predicting the effects of
animal variation on growth and feed intake in growing pigs using simulation
modelling. Animal Science 64, 513–522.
Ferguson, N.S., Lavers, G. and Gous, R.M. (2001) The effect of stocking density on
the response of growing pigs to dietary lysine. Animal Science 73, 459–469.
Fisher, C., Morris, T.R. and Jennings, R.G. (1973) A model for the description and
prediction of the response of laying hens to amino acid intake. British Poultry
Science 14, 469–484.
Gonyou, H.W., Chapple, R.P. and Frank, G.R. (1992) Productivity, time budgets
and social aspects of eating in pigs penned in groups of five or individually.
Applied Animal Behaviour Science 34, 291–301.
Grandin, T. (1994) Solving livestock handling problems. Veterinary Medicine 89,
989–998.
Griffing, B. (1966) Selection in reference to biological groups I. Individual and
group selection applied to populations of unordered groups. Australian Journal
of Biological Science 20, 127–139.
Hamilton, D.N., Ellis, M., Wolter, B.F., Schinckel, A.P. and Wilson, E.R. (2003) The
growth performance of the progeny of two swine sire lines reared under
different floor space allowances. Journal of Animal Science 81, 1126–1135.
Heetkamp, M.J.W., Schrama, J.W., De Jong, L., Swinkels, J.W.G.M., Schouten,
W.G.P. and Bosch, M.W. (1995) Energy metabolism in young pigs affected by
mixing. Journal of Animal Science 73, 3562–3569.
Heetkamp, M.J.W., Schrama, J.W., Schouten, W.G.P. and Swinkels, J.W.G.M. (2002)
Energy metabolism in young pigs as affected by establishment of new groups
prior to transport. Journal of Animal Physiology 86, 144–152.
Hessing, M.J.C., Schouten, W.G.P., Wiepkema, P.R. and Tielen, M.J.M. (1994)
Implications of individual behavioural characteristics on performance in pigs.
Livestock Production Science 40, 187–196.
Hyun, Y. and Ellis, M. (2001) Effect of group size and feeder type on growth
performance and feeding patterns in growing pigs. Journal of Animal Science 79,
803–810.
Hyun, Y., Ellis, M., Riskowski, G. and Johnson, R.W. (1998a) Growth performance
of pigs subjected to multiple concurrent environmental stressors. Journal of

Hyun, Y., Ellis, M. and Johnson, R.W. (1998b) Effects of feeder type, space
allowance, and mixing on the growth performance and feed intake pattern of
growing pigs. Journal of Animal Science 76, 2771–2778.
Kanis, E., de Greef, K.H., Van der Belt, H., Groen, A.F. and Schakel, J. (2002)
Breeding for improved welfare of pigs, is it technically achievable in practice?
Proceedings of the 7th World Congress on Genetics Applied in Livestock Production 32,
43–46 (Abstr.).
Knap, P.W. (1995) Aspects of stochasticity: variation between animals. In: Moughan,
P.J., Verstegen, M.W.A. and Visser-Reyneveld, M.I. (eds) Modelling Growth in the
Pig. EAAP Publication No. 78. Wageningen Pers, Netherlands, pp. 165–172.
Knap, P.W. (2000) Stochastic simulation of growth in pigs: relations between body
composition and maintenance requirements as mediated through turn-over
and thermoregulation. Animal Science 71, 11–30.
Kornegay, E.T. and Notter, D.R. (1984) Effects of floor space and number of pigs
per pen on performance. Pig News and Information 5, 23–33.
Kyriazakis, I. (2003) The control and prediction of food intake in sickness and in
health. In: Wiseman, J., Varley, M.A. and Kemp, B. (eds) Perspectives in Pig
Science. Nottingham University Press, UK, pp. 381–403.
MacRae, J.C. and Lobley, G.E. (1991) Physiological and metabolic implications of
conventional and novel methods for the manipulation of growth and
production. Livestock Production Science 27, 43–59.
Matteri, R.L., Carroll, J.A. and Dyer, C.J. (2000) Neuroendcrine responses to stress.
In: Moberg, C.P. and Mench, J.A. (eds) The Biology of Animal Stress. Basic
Principles and Implications for Animal Welfare. CABI Publishing, Wallingford, UK,
pp. 43–76.
Muir, W.M. (1996) Group selection for adaptation to multiple-hen cages: Selection
program and direct responses. Poultry Science 75, 447–458.
Muir, W.M. and Craig, J.V. (1998) Improving animal well-being through genetic
selection. Poultry Science 77, 1781–1788.
Muir, W.M. and Schinckel, A. (2002) Incorporation of competitive effects in
breeding programs to improve productivity and animal well being. Proceedings
of the 7th World Congress on Genetics Applied in Livestock Production 32, 35–38.
Nielsen, B.L., Lawrence, A.B. and Whittemore, C.T. (1995) Effect of group size on
feeding behaviour, social behaviour and performance of growing pigs. Animal
NRC (National Research Council) (1998) Nutrient Requirements of Swine, 10th edn.
National Academy Press, Washington, DC.
Petherick, J.C. (1983) A biological basis for the design of space in livestock housing.
In: Baxter, S.H., Baxter, M.R. and MacCormack, J.A.C. (eds) Farm Animal
Housing and Welfare. Martinus Nuhoff, Lancaster, UK, pp. 103–120.
Petherick, J.C., Beattie, A.W. and Bodero, D.A.V. (1989) The effect of group size on
the performance of growing pigs. Animal Production 49, 497–502.
Pomar, C., Knap, P.W., Kyriazakis, I. and Emmans, G.C. (2003) Modeling
stochasticity: dealing with populations rather than individual pigs. Journal of
Animal Science 81, E178–E186.
Schinckel, A.P., Spurlock, M.E., Richert, R.T. and Weber, T.E. (2003) Modelling
health status-environmental effects on swine growth. EAAP Annual meeting,
Rome.
Spoolder, H.A.M., Edwards, S.A. and Corning, S. (2000) Aggression among
finished pigs following mixing in kennelled and unkennelled accommodation.

Stookey, J.M. and Gonyou, H.W. (1994) The effect of regrouping on behavioural
and production parameters in finishing swine. Journal of Animal Science 72,
2804–2811.
Tan, S.S.L., Shackleton, D.M. and Beames, R.M. (1991) The effect of mixing
unfamiliar individuals on the growth and production of finishing pigs. Animal
Torrey, S., Pajor, E.A., Weaver, S., Kuhklers, D. and Stewart, T.S. (2001) Effect of
genetic selection for loin-eye area on behaviour after mixing in Landrace pigs.
In: Garner, J.P., Mench, J.A. and Heekin, S.P. (eds) Proceedings of the 35th
International Conference of the ISAE. Centre for Animal Welfare, California, p. 98
(Abstr.).
Turner, S.P., Horgan, G.W. and Edwards, S.A. (2001) Effect of social group size on
aggressive behaviour between unacquainted domestic pigs. Applied Animal
Behaviour Science 74, 203–215.
Turner, S.P., Dahlgren, M., Arey, D.S. and Edwards, S.A. (2002) Effect of social
group size and initial liveweight on feeder space requirement of growing pigs
fed ad libitum. Animal Science 75, 75–83.
Turner, S.P., Allcroft, D.J. and Edwards, S.A. (2003) Housing pigs in large social
groups: A review of implications for performance and other economic traits.
von Borell, E., Biensen, N.J. and Gatnau, R. (1992) Effects of stress hormones on
pig performance and health. In: Swine Research Report. Iowa State University,
USA, pp. 77–78.
Wellock, I.J., Emmans, G.C. and Kyriazakis, I. (2003a) Modelling the effects of the
Wellock, I.J., Emmans, G.C. and Kyriazakis, I. (2003b) Modelling the effects of the
Wellock, I.J., Emmans, G.C. and Kyriazakis, I. (2003c) Predicting the consequences
of social stressors on pig food intake and performance. Journal of Animal Science
81, 2995–3007.
Wellock, I.J., Emmans, G.C. and Kyriazakis, I. (2004) Modeling the effects of
stressors on the performance of populations of pigs. Journal of Animal Science
82, 2442–2450.
Whittemore, C.T. (1998) The Science and Practice of Pig Production. Longman
Scientific and Technical, Essex, UK.
Whittemore, C.T. and Green, D.M. (2002) The description of the rate of protein
and lipid growth in pigs in relation to live weight. Journal of Agricultural Science,
Cambridge 138, 415–423.
Wolter, B.F., Ellis, M., Curtis, S.E., Augspurger, N.R., Hamilton, D.N., Parr, E.N.
and Webb, D.M. (2001) Effect of group size on pig performance in a wean-to-
finish production system. Journal of Animal Science 79, 1067–1073.
5 Modelling Populations for
Purposes of Optimization
R.M. GOUS AND E.T. BERHE
Animal and Poultry Science, School of Agricultural Sciences and
Agribusiness, University of KwaZulu-Natal, Private Bag X01,
Scottsville 3209, South Africa.
gous@ukzn.ac.za
Introduction
Broiler and pig growth models have improved considerably since

Whittemore and Fawcett (1976) developed the Edinburgh Model Pig, yet
in most cases such models still work at the level of the individual rather
than at that of a population. In so doing, the genetic variation inherent in
a population, the variation that exists from one end, or side, of a
production unit (e.g. a broiler house) to the other, and variation in the
composition of the feed offered to broilers or pigs is ignored, the
assumption being that the response of the average individual, housed in an
average environment and fed an average feed, will be sufficiently similar to
that of the average response of the population from which it comes.
Although this is a naïve approach, there is great merit in ensuring that the
simulation of food intake and growth of the individual is sensible and
accurate before expanding the simulation to that of a population. If the
simulation of the growth of an individual meets the above criteria, and has
been modelled mechanistically, then it is not an insurmountable problem
to expand this to a population model. Models of individuals may be
adequate for an understanding of the theory of growth and food intake, as
well as for ‘what-if ’ scenario planning. However, for purposes of
optimization, it is imperative to account for the variation inherent in the
system if a realistic assessment of the population response is to be
simulated.
According to Knap (1995) stochastic simulation can be defined as
producing simulation outputs that reflect not only the expected population
means of the traits of interest but also their expected dispersion, as a result
of deliberately introducing variation in a number of basic parameters of
the simulation model. He outlined the reasons for considering variation
between animals in growth models when simulating different systems.

Modelling Populations for Purposes of Optimization 77
1. The profitability of the systems may be affected to a large extent by the

amount of variation in the production traits.
2. The change from one system to another may have small effects on
average levels but large effects on variation.
3. Differences between systems can be discovered more readily when
variation is made visible.
4. In order to study the relationships between traits, covariance should be
created, which requires variation (Emmans and Fisher, 1986).
This chapter deals with the sources of variation that should be considered
when modelling the growth of a population of broilers for purposes of
optimization, and presents some comparisons of simulation and
optimization outputs when these were conducted at the individual and the
population levels.
Optimizing the Feeding Programme for Broilers
The optimum feeding programme for a broiler producer is that which

results in the highest profit for the enterprise, e.g. maximum margin/m2
per annum or margin over feeding cost. Determining the optimum
nutrient density in the feeds used, the optimum concentrations of amino
acids relative to energy in each feed, and the optimum length of time that
each feed should be fed, are therefore both nutritional and economic
decisions.
The information required for optimization consists of feed costs at
different levels of nutrient provision, a description of all the relevant
animal responses, both fixed and variable costs affecting the production
system and details of revenue. The complexity of the information required
would depend on the level of organization at which the optimization is to
be made. If profit of the broiler grower is to be maximized at the farm
gate, then responses in liveability, growth and feed conversion ratio will
probably suffice. However, and more realistically, a wider view will often be
required, and the effect of broiler nutrition on slaughterhouse variables
(eviscerated yield, rejects, etc.) and further processing (carcass
composition) will need to be defined. Mack et al. (2000) emphasized the
importance of broiler companies considering all aspects of the production
cycle when making nutritional decisions.
Feed costs for any nutritional specification are readily calculated by
linear programming. This will take account of feed ingredient availability,
analysis and costs. Processing and transport costs may be added. Broiler
production costs are complex but will usually be specified by each
company.
So the only persistent problem in optimization lies, as ever, in the
definition of animal response. Consider some of the procedures that would
be needed when optimizing a feeding programme. It would be necessary
to determine the potential growth rate and potential fatness of the birds to
78 R.M. Gous and E.T. Berhe
be fed; the distribution of potential growth rates (greater when mixed

sexes are used); the environmental conditions provided, and the cost of
altering the prevailing conditions; the costs of a range of feeds differing in
nutrient density at each energy-to-protein ratio; and the cost of mixing and
then transporting these feeds to the production site (which would place an
upper limit on the number of feeds that could be considered in any
production cycle). Consider then that the birds can adjust their intake of a
given food to an extent, this being limited by the environmental
conditions; that the effect of feeding a relatively low quality food initially
can be compensated for at a later stage if the conditions are such that the
bird can either consume more food later, or draw on lipid reserves,
thereby exhibiting an improved feed conversion efficiency. Consider, too,
that the amount of lipid in the gain is of importance to some producers,
but not to others; and that the length of the production cycle can be
altered considerably by the use of different feeding programmes. It
becomes clear that it is naive to imagine that any one, or even any series, of
experiments could begin to address the question of defining the optimum
feeding programme. Only with the use of an accurate simulation model
could such an optimization be contemplated.
An optimization tool for broiler production would need to combine
three types of computer program, namely, a feed formulation program, a
growth model and an optimization procedure. The flow of information
for such a procedure, shown in Fig. 5.1, bears similarities to the
continuous quality improvement model of Deming (1986), which consists
of four repetitive steps (Plan, Do, Check, Act), this continuous feedback
loop being designed to assist managers to identify and then reduce or
eliminate sources of variation that cause products to deviate from
customer requirements. The process in this case is straightforward: the
optimizer defines nutritional constraints for practical broiler feeds, which
are passed to the feed formulation program where the least-cost feed that
meets these constraints is determined. The characteristics of this
formulated feed are then passed, as input, to the broiler growth model.
The performance expected from this feed when given to a defined flock
of broilers in a given environment is predicted by the model, and this
predicted performance is then passed to the optimizer to complete the
cycle. The next cycle starts with the optimizer modifying the feed
specifications, moving, according to some in-built rules, to an optimum
point. The objective function to be maximized or minimized can be
defined in terms of any output from the broiler growth model, but
realistically would be an economic index of some sort. Examples are
margin over feeding cost, margin per m2 per year, or maximum breast
meat yield at an age or weight.
The system used in the EFG broiler nutrition optimiser (EFG Software
Natal, 1995), (which optimizes three aspects of a commercial broiler feeding
programme, namely, the amino acid contents in each feed, given a feeding
schedule, the nutrient density of each feed in the schedule and the optimum
feeding schedule given feeds of a fixed composition) considers all of the
Broiler Nutrition Optimizer
Feed Feed Broiler

formulator composition model
Feed Predicted
specification performance
Optimizer
Fig. 5.1. Flow of information in optimizing the feeding programme of a broiler chicken.
criteria mentioned above. The broiler growth model allows for multiple
harvesting from one flock and calculates revenues from any mixture of
whole-bird sales or processing. Typical economic variables are included,
although these may be readily customized to fit with individual enterprises.
Growth models developed in different laboratories may be included in
optimization schemes of the sort described above. The key to this approach
clearly lies in the ability of the broiler growth model to reflect accurately
the performance expected under commercial conditions. However, the
performance of the average individual in the flock, subjected to average
conditions in the broiler house and being fed the average food formulated
by the nutritionist, without considering the variation inherent in genotype,
environment and food, cannot accurately reflect the variation that would
result in commercial conditions. For this, a stochastic model of growth and
food intake is required.
Individuals versus Populations
Because most nutrition experiments in research and all experiments in

commercial poultry production are conducted on groups of birds, models
of the response of individuals are not entirely appropriate for application
in the poultry industry. For this reason some models have been developed
at the level of the group, but a problem with this approach is that the
relationships between inputs and outputs for groups are curves, whereas if
one assumes that marginal efficiencies for nutrient utilization are constant,
then such relationships are not curves (Emmans and Fisher, 1986). The
Reading Model (Fisher et al., 1973) demonstrated convincingly that the
response of an individual (laying hen) differs markedly from that of the
population from which the individual is drawn. Whereas the response of
each hen to an increasing supply of an amino acid can be assumed to be
linear, up to a point where a plateau of output (the genetic potential) is
reached, the population response is a continuous, asymptotic curve with
no abrupt threshold. This curve results from determining the mean
response of a group of individuals at a time. In a population of growing
broilers, as with laying hens, there exist differences in the potential output
(growth rate) of each animal at a time, but because each bird is growing
and food intake is increasing, variation in growth rate is also introduced
within each animal over time (Gous and Morris, 1985), making the need
for a population model of broiler growth even more important (or critical)
than in the case of laying hens, which are in a relatively steady state. A
population model should be built by simulating the responses of many
individuals, and not directly as a group, so that the accepted assumptions
about marginal efficiencies are not compromised.
But it is not only the variation between individuals in their response to
a given feed or environment that controls the variation in the response of a
population of birds to a feeding programme in a given environment.
Variation also exists in the environmental conditions to which the birds are
subjected, as well as the composition of the feed used. Each of these
sources of variation will be addressed in the following review, but emphasis
will be placed on the effect of variation at the level of the genotype on the
response, as this aspect has been used in numerous exercises by the
authors to illustrate the difference in response between an individual and a
population of broilers.
Sources of Variation in a Population of Broilers
Variation in the genotype
When modelling the growth and food intake of a growing broiler it is

sensible to have some idea of the potential growth rate of the bird, the idea
being that the bird has a purpose, namely, to grow at its potential
(Emmans, 1987). Given this goal, the model can then calculate the desired
amount of the given food the bird needs to consume each day to grow at
its potential, and will predict the consequences for food intake, growth and
carcass composition if this intake is constrained by feed bulk or an
environment that is too hot to allow the bird to lose sufficient heat to that
environment.
Describing an individual
Potential feather-free body protein growth may be adequately described by

the three parameters of the Gompertz growth curve, namely, the starting
body protein weight (BP0), the rate of maturing (B) and the mature
protein weight (BPm) (Emmans, 1987). The lipid content of the mature
body is used to define the gross chemical composition of the body at any
intermediate weight. Well-defined allometric relationships between
protein, lipid, ash and water enable the growth of the physical components
of the body to be simulated. However, recent developments in broiler
genotypes in which, for example, breast meat yield has been increased
markedly by selection, mean that these allometric relationships are unlikely
to be universally applicable to all genotypes, and therefore should be

redefined for genotypes whose physical composition has been modified by
selection.
The rate of feather growth must also be defined, and a Gompertz
growth curve may also be used for this purpose. The rate of maturing of
feathers differs from that of body protein growth, but may be seen as a
multiple of the latter. Feather growth differs markedly between genotypes
and sexes, especially in feather-sexable strains. Feathering rate multipliers
(FR) are used in the EFG Broiler Growth Model to calculate the rates of
feathering of male and female broilers, these differing for feather-sexable
and non-feather-sexable strains.
In order to recreate the different responses to dietary protein that are
seen between some commercial strains a further genetic parameter is
defined, which controls the maximum lipid in the gain ratio (MLG) in the
short run. In essence, this parameter reflects the ability of the genotype to
fatten when presented with a feed in which one or more amino acids or
other nutrients are marginally limiting, which causes the bird to
overconsume energy in an attempt to obtain sufficient of the limiting
nutrient. Birds with a low propensity to fatten would be disadvantaged in
such a case, as they would be unable to store the excess energy as body
lipid, the resultant food intake therefore being lower than desired, and
weight gain being below the potential.
The parameters described above are seen as genetic characteristics of
individuals in the flock, many of which have been measured in commercial
strains of broiler (Hancock et al., 1995; Gous et al., 1996, 1999). Realistic
responses of individuals, whose genotypes have been described by allocating
appropriate values to these parameters, may then be predicted using a
simulation model such as that described by Emmans (1987). Such a model
has been developed (EFG Software Natal), which simulates the growth of a
single bird, taking account of genetic parameters, diet composition and
feeding programme, the environment, stocking density and other factors
that may affect the outcome of production decisions in practice. Body
protein content is used to define the current state or condition of the
animal, which is then used to quantify the remaining body constituents and
their respective growth rates (Taylor, 1980). Body protein is the driving
variable in the model, the assumed goal of the broiler being to grow at its
potential body protein growth rate whenever possible (Emmans, 1981).
Food intake, growth, body composition and yield, and a variety of
production indices are calculated in each simulation. The model also carries
out basic economic calculations to guide commercial decisions. This model
was used to perform the simulations that have been used in this chapter.
Describing a population
The parameters describing the genotype of an individual would be

expected to vary normally between individuals in a population. It is this
variation that is used by geneticists to select the more favoured individuals

within a population, thereby moving the mean performance of the strain
in a desirable direction. The extent to which each parameter varies and is
correlated with other parameters, how these values may be predicted, and
the consequences of such variation on the mean performance of the
population will be dealt with below.
Variation in the environment within a broiler house
Controlling the physical microenvironment in a broiler house is an

important element in optimizing the production process (Reece and Lott,
1982; Mitchell, 1985; Parmar et al., 1992; Aerts et al., 2000). However,
depending on the system used for such control, variation in the
environment within a poultry building may be considerable, impacting
significantly on the performance of broilers housed in different areas of
the house. Many factors contribute to this variation such as unadjusted
inlet openings, unsealed cracks, dirty fan shutters, loose curtains, the
amount of time the fans run, the total length of the timer cycle and the
position of thermo-sensors (Al Homidan et al., 1997, 1998). The behaviour
of the broilers themselves will also cause variations in temperature along
the length of the house, as a result of mass migrations and clustering in
some areas (Wathes and Clark, 1981), which will affect thermoregulation.
Such variation in the environment is likely to be positively correlated with
the variation in the weight of the end product of the production process,
raising the question whether the optimum feeds or feeding programme
should be adjusted to take account of this variation.
The vertical temperature profile in a poultry house is affected by many
factors, including heat generated by heaters, the flock and solar radiation,
microbial fermentation in the litter, heat fluxes between poultry house air
and the soil, walls and roofs, due to temperature gradients, moisture loss
from the litter and natural convection around broilers (Van Beek and
Beeking, 1995; Boshouwers et al., 1996). However, of greater importance
in a tunnel-ventilated house is the variation that exists in the horizontal
temperature profile at bird level. Winter and summer conditions alter this
profile, with differences in temperature along the length of the house
being considerably greater in winter than in summer. In winter, chicks
often migrate towards the air inlets, thereby increasing the stocking density
in that area, and the temperature differential in the early stages of growth
can be as high as 7–9°C (Xin et al., 1994) and remains at least 3.5°C cooler
at the far end, with 7% higher relative humidity, by the end of the
production cycle. In summer, because of the higher ventilation rates used,
Xin et al. (1994) found little difference in temperature at either end of the
house. Many factors interact to influence this gradient, so it would be
expected that the temperature and humidity gradient would differ
markedly between houses and between seasons.
Poor air quality, due to environmental contaminants such as carbon
monoxide, carbon dioxide, ammonia and dust, reduces performance,

increases the potential for respiratory disease and may increase mortality
(Weaver and Meijerhof, 1991). Conditions at the exhaust end of a tunnel-
ventilated broiler house may be far more hostile than at the inlet end,
especially in winter, resulting in considerable variation in the performance
of the flock.
The effects of variation in temperature, humidity and air quality in the
broiler house on the range of body weights present in the house at the end
of the production cycle cannot be estimated empirically because of the
expected interactions between these environmental factors and potential
body protein growth rate, feathering rate and the chemical and physical
composition of the feed, all of which are expected to vary also. These
interactions can only be assessed with the aid of mechanistic models.
Variation in the nutrient content of the feed
Variation in the nutrient content of feeds offered to broilers in a

commercial operation is brought about through three main sources:
variation in ingredient composition, mixer inefficiency (including weighing
errors) and separation after mixing and during transportation.
Sources of variation in the physical and chemical characteristics of
grains used in poultry feeds include variety, seasonal effects, growth sites
(Metayer et al., 1993), crop treatment and grain fumigants, post-harvest
storage conditions and period of storage and processing (Dale, 1996;
Hughes and Choct, 1999), rainfall and environmental temperature
patterns during the period of grain maturation, genetic effects, level of
fertilizer usage (Metayer et al., 1993; Hughes and Choct, 1999) and
inclusion rate (Senkoylu and Dale, 1999). This type of variation may be
dealt with in the formulation process through non-linear stochastic
programming (Roush et al., 1996), but the consequences of variation
introduced in this way, and through mixer inefficiency, must be modelled
differently from the variation brought about through separation, which is
intuitively less random than the first two types of variation.
Performance is adversely affected when variation in nutrient content in
feeds is increased (Duncan, 1988; McCoy et al., 1994). Some of the more
sophisticated feed formulation programmes used in the feed industry,
which attempt to account for variation in nutrient content in the
ingredients used in the feed, generally favour the use of ingredients that
exhibit the least amount of variation, thereby potentially reducing the
variation in performance of the broilers being fed such feeds. However,
whereas natural variation in the composition of a feed has zero cost, a
considerable cost is incurred when attempting to reduce variation, as these
non-linear programmes demonstrate, so there is an economic limit to the
extent to which variation in nutrient content should be reduced.
Fawcett et al. (1992) attempted to determine this limit by predicting the
response of broilers, in margin/m2, to a range of dietary lysine and ME
contents with the use of the ‘Poultry Growth Model’ of Emmans (1981),
and then integrating the bivariate probability distribution for different
degrees of variation in nutrient content over the response surface to
determine the value of reducing the variance in the feed. This approach,
although interesting, does not simulate the day-to-day variation in feed
nutrient content to which broilers are subjected in practice, where nutrient
composition varies continuously throughout the rearing period. A more
useful technique would be to predict the response of each bird in the
population to random variations in feed quality on each day of the
growing period, thereby obtaining a more realistic assessment of the effect
of variation in feed quality. This would require a stochastic approach to the
definition of the composition of the feed offered to the broilers, as well as
to the broilers themselves.
Such an approach is a useful tool for determining the upper limit to
the value of variance reduction, but does not address the issue of further
systematic variation in nutrient content brought about by separation
during road, rail, auger, blower or chain transportation. Such variation
tends to separate the fine from the coarse particles, particularly when the
feed is conveyed in a mash form, or when the quality of pellets is poor,
resulting in a high proportion of fines. This problem does not appear to be
addressed in the literature, so the variation introduced in this way is
unknown, but could be considerable. This variation is likely to be
systematic, in that separation would take place along the length of the
feeder lines, with fines being left behind as the coarser particles move to
the end of the line. If the fines consisted predominantly of major and
minor minerals and vitamins, bone development could be seen to worsen
at the far end of a broiler house compared with the end at which the feed
is introduced. The effect of such nutrient separation could be modelled in
the same way as are the systematic changes that take place in the
environment within a broiler house, the birds along the length of the
house being subjected to changes in temperature, humidity and air quality.
But it is not only the chemical separation of the food that takes place that
influences performance along the length of a broiler house; the physical
nature of the feed may also be affected.
Pelleted feeds are known to improve feed conversion efficiency when
compared with mash feeds (Jensen et al., 1962), this being due to the
shorter period of time spent by the broilers consuming pellets, resulting in
lower energy expenditure. The physical nature of the feed therefore has
an effect on performance and, because of the abrasive nature of some
feeding equipment used in broiler houses, the amount of time spent
eating, and hence the energy expended in eating, may well increase as the
distance of the feed trough from the hopper increases.
If the performance of broilers subjected to variation in the chemical
and physical nature of the feed offered to them could be accurately
simulated, a cost benefit analysis could be conducted to determine to what
extent the reduction in such variability is worth pursuing. Only with the
aid of stochastic programming is such an analysis possible.
Generating a Population of Broilers
As argued above, when predicting the outcome of an experiment in a

population of broilers it is more constructive and sound to simulate the
performance of each individual in the population and average these
responses, rather than attempting to predict the response of the
population as a whole. For this approach to work, some theory of the
structure of a population must be developed. The values allocated to each
of the genetic parameters that describe each individual in the population
must be seen as being stochastic, i.e. they are assumed to be normally
distributed in the population, and thus can be described by allocating a
mean and standard deviation to each parameter. The population is then
described by correlated distributions. Correlations between parameters are
easily dealt with when generating a population, but the values assigned to
such correlations are not well researched. Of the six genetic characteristics
used to describe the potential growth rate of a broiler, the correlations
between B (a parameter defining the decline in logarithmic growth rate)
and Pm (the weight of protein in the animal at maturity), and between B
and feathering rate, are potentially the most important to be considered.
The negative correlation between B and Pm (Brody, 1945) may be dealt
with by the use of a scaled rate parameter, B*= B Pm 0.27 that is uncorrelated
with Pm (Taylor, 1980; Emmans and Fisher, 1986). The variation in B and
Pm may be appreciable within a population, with suggested coefficients of
variation (CV) of between 0.06 and 0.10, whereas that of B* may be much
lower, at between 0.02 and 0.04 (Emmans and Fisher, 1986). The CV of the
parameter LPRm (the lipid:protein ratio in the body at maturity) has been
predicted to be around 0.04 (Emmans and Oldham, 1988).
The genetic correlation of total feather score with body weight measured
by Singh and Trehan (2002) ranged between 0.179 and 0.444, while that
between the increase in feather density score from 4 to 6 weeks of age and
body weight was negative, varying from 0.189 to 0.323. Whereas these
genetic correlations are relatively high, the negative phenotypic correlation
is considerably greater than this, as can be demonstrated by simulation. The
heritability (h2) estimates of rate of feathering of broilers measured by Singh
and Trehan (2002) at 10 days varied between 0.231 and 0.580, and the h2 of
feather density scores was even higher (0.279–0.925), implying that
geneticists have the potential to alter these characteristics relatively easily.
This may prove to be a relatively simple method of overcoming the effects of
heat stress as broiler genotypes are selected for ever-faster growth rates, as
suggested by Cahaner et al. (2003), given that feeding programmes are
ineffective in overcoming this stress.
Methods of Generating the Individuals Making Up the Population
The individuals making up a simulated population need to be generated in

such a way that the mean of each of the parameters generated is close to
the required mean, and the distribution of values about the mean reflects
the required standard deviation. The parameter values for each individual
may be generated by means of random numbers. A random number is a
number chosen as if by chance from some specified distribution such that
the selection of a large set of these numbers reproduces that distribution.
The subject of random number generation and testing is reviewed
extensively by Knuth (1997) and Hellekalek (2004). Many non-uniform
random number generators are available, even in such accessible
programmes as Excel and Minitab, whilst the source code for generating
these numbers in C++ is available free on the Internet. These
programmes generate random data from a normal distribution, given the
mean and standard deviation of the variate. It is also possible
simultaneously to generate values of correlated variates.
How Many Birds are Needed to Obtain a Realistic Result?
The length of time taken to simulate the performance of the population, in

order to calculate the population mean, is dependent on the number of
population samples simulated, the complexity of the programme and the
speed of the computer. Clearly, the more samples that are simulated, the
more representative the sample will be of the population, and therefore the
more accurate the population mean estimate will be. Also, the larger the
number of stochastic variables that will be varied in the simulation, the larger
the sample needs to be for an accurate estimate of the population mean. So
the number of individuals may need to be large, and the computational time
long, if up to six parameters are made stochastic. For this reason, a weighted
sampling method may be more practical. In this method, instead of choosing
individuals randomly selected from the multivariate normal distribution of
population parameters, a sampling ‘design’ is created, in which a fixed
number of individuals with fixed parameter settings are chosen to be
simulated. The number of individuals included for each parameter setting is
chosen so that they approximate the frequencies of these parameter settings
in a normal distribution. The advantage of such a method is that even with
relatively small numbers of individuals the resultant population will be
approximately normally distributed.
Here is an example of such a design. Suppose m is the mean and d the
standard deviation (SD) of one genetic parameter in the population. Three
‘points’ in the population are selected to simulate: those with parameters m,
(m1.5d), and (m+1.5d). That is, the mean, 1.5 SDs below the mean, and 1.5
SDs above the mean. In the normal distribution, individuals at 1.5 times the
SD from the mean occur with a frequency approximately 2/5ths of those that
occur at the mean. So we can approximate a normal distribution by
simulating nine individuals: five at m, and two each at (m1.5d), and
(m+1.5d). Of course, the simulation need only be run three times, not nine,
since the results for two individuals at the same point in the population
distribution are always the same.
To extend this idea to two parameters, say x and y, nine population

points may be used instead of three: one for each combination of m,
(m1.5d), and (m+1.5d) for each of the two parameters. The number of
individuals to be simulated at each point is just the product of the numbers
used in the single parameter case as shown in Table 5.1.
There are a total of 81 individuals in this population, but only nine
simulations are needed: one for each of the parameter combinations in the
table.
The two methods were compared in an exercise in which the means of
increasing numbers of individuals, generated with the use of random
numbers, were compared with the mean individual in the population from
which the samples were drawn, and with the mean of the sample
population generated with the use of weighted sampling. Because the
weighted sampling method produces a fixed population, the CV of
response variables does not differ unless the weightings are altered. Also,
although the number of simulations is low for a relatively large population
(nine simulations for 81 individuals, as described above), the number of
simulations required rapidly increases as the number of genetic
parameters is increased. For example, if six genetic parameters are varied,
and three weightings are used, e.g. m, (m1.5d) and (m+1.5d), the number
of simulations is increased to 36 = 729; and to 15625 where five weightings
are used. Clearly, this would not be a time-saving method, and the random
sampling method would be favoured under such circumstances.
Analysing the Sensitivity of the Genetic Parameters
As a first step in determining whether it should be necessary to simulate

the response of a population rather than that of the average individual, the
effect of variation in each of the genetic parameters describing the
individuals in a population should be simulated. If, by varying the
Table 5.1. An example of the use of the weighted sampling

method in which a population is made up of 81 individuals
using fixed parameter settings (the mean (m), and + or 1.5
standard deviations (d) from m) for two independent
genotypic parameters. The number of individuals included
for each parameter setting is chosen so that they
approximate the frequencies of these settings in a normal
distribution.
Parameter x
Parameter y (m1.5d) m (m+1.5d)
(m1.5d) 4 10 4
m 10 25 10
(m +1.5d) 4 10 4
magnitude of each of the parameters about the mean, a linear response is

obtained in the production variables of concern, then it would be
unnecessary to simulate the response of more than the average individual
in the population, as a sufficiently accurate assessment of the population
response could be obtained by applying some variation on either side of
the resultant mean. It is tempting to use such a technique to generate a
population response because of the saving in computation time, and for
this reason such a technique has been used in some models. However, if
the response is not linear, or if interactions occur between parameters,
then this method is invalid.
The sensitivity analysis technique proposed by Morris (1991), of
varying one factor at a time, was used (Berhe, 2003) to determine to what
extent variation in each of the six parameters that describe the genotype
would influence the performance of the broiler. After simulating the
response of the mean individual in a population, further simulations were
conducted in which each of the six genetic parameters was reduced, in
turn, by 0.05, 0.10, 0.15 and 0.20, and then increased by the same
proportions. The exercise was conducted separately on male and female
broilers, over two periods of growth (starter, 8–21 days, and finisher,
22–35 days) using two feeds limiting in lysine (9 or 16 g lysine/kg in the
starter period, and 7 or 11 g/kg in the finisher period, the lysine:protein
ratio being the same in both feeds within each period). The objective in
using two sexes, two periods of growth and two lysine contents was to
determine whether the responses to systematic changes in each genetic
parameter remained constant under all these conditions.
The genetic parameters that were varied were Pm, B and LPRm
(defined above), W0 (initial weight), MLG (maximum lipid in the gain) and
FR (a feathering rate multiplier). Some of the results of this exercise are
shown in Fig. 5.2, where the effect of variation in the six genetic
parameters on food intake in males and females of a feather-sexable strain,
in the starter and finisher periods, are illustrated. In both periods food
intake increased almost linearly with B, W0, MLG and Pm, the latter only at
the high lysine content in the starter, but at both lysine levels in the finisher
period. This would indicate that, if only these four parameters were being
used to describe a population of broilers, there would be no great
advantage in increasing computation time to simulate the population.
However, the effect of FR was non-linear in all cases, and more
pronounced in females than in males, especially in the starter period. Food
intake dropped more substantially with FR among males fed a low lysine
feed in the starter period than those fed the high lysine feed. The effects
were more pronounced and more uniform in the finisher than in the
starter period. The theory of food intake regulation of Emmans (1987)
predicts that food intake would decline with a higher rate of feather
growth, given that the processes of food intake and growth generate heat
that must be lost to the environment if the bird is to remain in thermal
neutrality and hence grow at its potential; if the bird cannot lose this heat
to the environment, food intake will be constrained. This would occur at
36
Food intake – Female (%)
Food intake – Female (%)

A B 40 E F
24
20
12
0
0
–20
–12
–40
–24
–60
–36
36 40
Food intake – Male (%)
Food intake – Male (%)

24 C D G H
20
12
0 0
–12 –20
–24
–40
–36
–20–10 0 10 20–20 –10 0 10 20 –20–10 0 10 20–20 –10 0 10 20
Parameter value (deviation from mean %) Parameter value (deviation from mean %)
Fig. 5.2. The relative effect on food intake in male and female broilers fed a lysine-limiting
feed containing 9 (A, C) or 16 g lysine/kg (B, D) from 7 to 21 days (left) or 7 (E, G) or 11 g
lysine/kg (F, H) from 22 to 35 days (right) when the means of each of six genetic parameters
were increased or decreased by 0.05, 0.1, 0.15 or 0.20 whilst holding the five remaining
genetic parameters constant (B = .-䉲.-; Fr = ..-䊊..-; LPRm = ..䊏..; MLG = --䊉--; Pm = —ⵧ—;
and Wo = —䉮—).
high environmental temperatures and particularly in a bird with an

extensive feather cover. The amount of variation introduced by varying FR
was often as great as that resulting from the same degree of variation in B,
except that food intake was affected only when FR was above the mean;
but the important distinction is between the linear effects of B and the
non-linear effects of FR.
As the CV of FR increases, an increasing proportion of the
population (those with the greatest feather cover) will have their food
intake constrained by their inability to lose heat to the environment. This
is illustrated in Fig. 5.3 by means of frequency distributions of the final
body weights of broilers given feeds sufficient to allow them to grow at
their potential; as FR is increased, the range of final body weights is
increased, but the distribution is negatively skewed resulting in a lower
mean body weight for the population. As the genotype, the feed and the
environment influence food intake and growth rate, it would be expected
that the effect of variation in FR would differ depending on the food
composition and the prevailing environmental conditions. These will be
referred to below.
40 40
CV = 0.0 CV = 0.05
30 30
Frequency
Frequency
20 20
10 10
0 0
1400 1600 1800 2000 2200 2400 2600 1400 1600 1800 2000 2200 2400 2600
Body weight (g) Body weight (g)
40 40
CV = 0.10 CV = 0.15
30 30
Frequency
Frequency
20 20
10 10
0 0
1400 1600 1800 2000 2200 2400 2600 1400 1600 1800 2000 2200 2400 2600
Body weight (g) Body weight (g)
Fig. 5.3. Effect of increasing the coefficient of variation (CV) of feathering rate on the
frequency distribution of body weights at 35 d of age in a simulated group of male broilers
given feeds designed to enable them to achieve their potential growth rate.
Effect of variation in FR on the response to dietary lysine at three

environmental temperatures
The simulated effects of increased variation in FR on food intake and

consequent protein gain of female broilers in the period 22 to 35 d of age
are illustrated in Fig. 5.4. Six lysine-limiting feeds (from 4 to 14 g lysine/kg
feed) were offered at three environmental temperatures. Many points of
interest arise from these simulations. At 29°C, food intake is severely
depressed on all feed treatments, yet food intake increases as the lysine
content of the feed is reduced, the ability of the birds to compensate for
the deficiency being constrained by, among others, the greater feather
cover. As the prevailing temperature declines so the overall food intake is
increased, but the differential between the highest and lowest intakes,
especially on the lowest lysine feeds, widens, i.e. the effect of the heavier
feather coat is relatively more severe in constraining intake as the
environmental temperature is decreased. The characteristic decline in food
intake on feeds with the lowest lysine contents (Gous and Morris, 1985;
Burnham et al., 1992) is evident only at the lower environmental
temperatures. The efficiency of utilization of lysine for body protein
growth remains the same at all temperatures on the marginally deficient
feeds, so protein gain is the same for a given lysine intake; it is only on
feeds with the highest lysine contents, and at the higher environmental
temperatures, that protein growth rate is curtailed due to the rapid
feathering rate, resulting in a separation of the maximum protein growth
rates at high dietary lysine contents.
Effect of variation in FR on the optimum dietary amino acid contents
Because of the non-linear effect of variation in FR on the performance of

broilers, it might be expected that the optimum amino acid contents of
feeds would differ for the mean individual in the population and for the
population itself, and that this difference would increase with variation in
FR. However, in spite of a considerable reduction in population
performance resulting from an increase in the CV of feathering rate (Table
5.2), the optimum amino acid contents of the three feeds used in the
feeding programme remained relatively similar for the individual and for
the five populations of broiler females. In the initial series of optimizations
a fixed feeding programme was used, namely, 600 g/bird starter, 1200
100 A 6
A
80 4
60 B 2
Protein gain (g/d)
Food intake (g/d)
120 9
B
100 6
80 3
160 C
12
140 9 C
120 6
100 3
4 6 8 10 12 14 16 400 600 800 1000 1200 1400 1600 1800
Lysine content (g/kg) Lysine intake (mg/bird day)
Fig. 5.4. The effect of variation in feathering rate (FR) on food intake (g/day) and body
protein gain (g/day) of female broilers from 22 to 35 days of age, fed lysine-limiting feeds at
environmental temperatures of 29°C (A), 25°C (B) and 21°C (C). Coefficients of variation
used were 0.0 (䊊—䊊), 0.05 (䉭---䉭), 0.10 (䊐-..-䊐) and 0.15 (*-.-*).
Table 5.2. Optimum 35-day performance, and optimum lysine contents in feeds, of an
individual and a population of broiler females, where margin/m2 annum was maximized at five
coefficients of variation (CV) of feathering rate, whilst the CVs of W0, Pm, LPRm and MLG
remained the same, as given in the text. Two feeding programmes were used: fixed amounts
of each feed, or a fixed number of days on each feed. Populations of 100 birds were
generated afresh for each optimization.
CV of feathering rate
Indiv. 0.00 0.05 0.10 0.15 0.25
Liveweight, g/bird 1968 1969 1966 1885 1775 1630
Food intake, g/bird 2844 2852 2883 2728 2565 2344
Breast meat, g 344 345 343 329 307 279
Abdominal fat, g 47 46 49 41 37 31
Cost of feeding, relative 100 101 101 97 91 84
Margin over feeding cost 100 100 98 95 90 82
Optimum lysine content, g/kg feed

Fixed amount of feed:
Starter (600 g/bird) 12.95 13.11 13.11 13.11 12.95 12.47
Grower (1200 g/bird) 8.72 8.94 9.17 9.17 9.17 9.40
Finisher (remainder) 8.11 7.89 7.16 7.89 7.89 8.08
Fixed number of days:
Starter (14 days) 13.14 13.11 13.11 13.11 12.64 13.11
Grower (10 days) 9.16 9.39 9.61 9.61 9.66 9.77
Finisher (remainder) 8.12 8.11 8.11 8.11 7.81 7.99
g/bird grower and the remainder finisher. The CVs of five of the genotypic
parameters were held constant (W0 = 0.1, Pm = 0.1, B* = 0.06, LPRm =
0.06 and MLG = 0.1) whilst FR was varied. A population of 100
individuals was generated afresh for each optimization.
Only one obvious trend emerged from this exercise, namely that the
optimum lysine content in the grower feed increased with variation in FR.
Interestingly, this happened also when a fixed number of days on each
feed was used in place of the fixed amount of each food (Table 5.2). The
inconsistent variation in optimum lysine contents (less when using a fixed
number of days in the feeding programme) probably reflects the variation
in the responses to different populations that were simulated for each
optimization. Generally, though, optimum lysine contents were the same in
the starter feeds, higher in the grower feeds, and lower in the finisher
feeds for populations than for the mean individual in the population.
The relatively small differences in amino acid contents of the three
feeds required to optimize performance of the mean individual and that of
the population are of more than passing interest, considering that the
mean performance of the population is much reduced when variation in
FR is high. Two precedents can be found to substantiate this observation.
The first is in Wethli and Morris (1978), where it was demonstrated that
the daily tryptophan required by a flock of laying hens does not decrease
during the first laying year, despite a decrease in mean rate of egg output.
The second is that when pigs are housed at high stocking rates, causing
reduced feed intakes and growth rates, efficiency of lysine utilization is
unaffected by the stress, and optimum biological performance is obtained
on feeds with lysine contents the same as those which optimize the
performance of individually housed pigs growing close to their potential
(B.A. Theeruth and R.M. Gous, 2005, unpublished results). It is not
unusual, therefore, for the optimum feeds to be unaffected by the reduced
mean performance of the animals in the population.
Effect of variation in FR on the optimum feeding programme
The non-linear effect of FR on the performance of a population was again

demonstrated where the feeding programme of broiler males and females
was optimized, with margin over feeding cost being maximized, first
when only five of the genetic parameters were varied (using the same CVs
as in the previous exercise, and the CV of FR = 0), and secondly, with the
CV of FR = 0.2. The results of the optimization process for an individual
and for the two CVs of FR are given in Table 5.3. Whereas the
performance at the optimum was the same for individuals and for a
population in which the CV of FR was 0.0, the performance of the
population with a higher CV of FR was markedly lower, this being the
result of the difficulty that broilers with a high FR experience in losing
heat to the environment, which results in a constrained growth rate. The
optimum amounts of each of the three feeds in the feeding programme
also reflect these differences in performance: with no variation in
feathering rate the optimum feeding programme for the population is
almost the same as for an individual, whereas with a large variation in
feathering in the population, approximately twice the amount of starter
feed is required by the population and, in males, approximately twice the
amount of grower feed is also required in order to maximize margin over
feeding cost.
Conclusions
Two conclusions may be reached from these exercises, both of which relate
to the non-linear effect of variation in FR, which both increases the
variation in response within a population and decreases its mean
performance. The first is that, where a fixed feeding programme is being
used, the optimum amino acid content in the feeds used in the programme
differs only marginally for a population and for the mean individual in the
population; whereas, if proprietary feeds of fixed composition are being
used, almost twice as much of the starter and grower feeds is needed to
optimize the performance of a population of broilers compared with that
required for the mean individual in the population. The second conclusion
Table 5.3. Simulated 35-day comparative performance of an individual and of a population of

male and female broilers at, and amounts of each feed needed to achieve, maximum margin
over feeding cost, when the coefficient of variation (CV) of feathering rate is increased from
0.0 to 0.2 with the CV of all other genotype parameters remaining the same in both
population simulations.
Male broilers Female broilers
Individual Population Individual Population
CV = 0.0 CV = 0.2 CV = 0.0 CV = 0.2
Body weight, g 2359 2336 2080 1920 1911 1679
Food intake, g 3762 3735 3335 3230 3230 2924
Breast meat, g 404 400 351 336 334 289
Margina 100 99 62 100 99 61
Starter, g/bird 225 300 465 143 185 340
Grower, g/bird 524 585 1160 428 408 470
Finisher, g/bird 3013 2850 1710 2659 2637 2114
aMargin over feeding cost, relative to individual.
is that, because the effect of FR is non-linear, whereas that of the other

genetic parameters is linear, if a population model is to be used when
optimizing the feeding programme of broilers, the population mean
should be generated by simulating the responses of individual animals
whose genotypes reflect the diversity found in a given population. It is not
possible to generate a population of broilers successfully without using this
technique.
References
Aerts, J.M., Berckmans, D., Saevels, P., Decuypere, E. and Buyse, J. (2000)
Modelling the static and dynamic responses of total heat production of broiler
chickens to step changes in air temperature and light intensity. British Poultry
Science 41, 651–659.
Al Homidan, A., Robertson, J.F. and Petchey, A.M. (1997) Effect of temperature,
litter and light intensity on ammonia and dust production and broiler
performance. British Poultry Science 38, S5–S6.
Al Homidan, A., Robertson, J.F. and Petchey, A.M. (1998) Effect of environmental
factors on ammonia and dust production and broiler performance. British
Poultry Science 39, S9–S10.
Berhe, E.T. (2003) Introducing stochasticity into a model of food intake and growth
of broilers. MScAgric thesis. University of KwaZulu-Natal, Pietermaritzburg,
South Africa.
Brody, S. (1945) Bioenergetics and Growth. Reinhold, New York.
Boshouwers, F.M.G., Develaar, F.G., Landman, W.J.M., Nicaise, E and Van Den
Bos, J. (1996) Vertical temperature profiles at bird level in broiler houses.
British Poultry Science 37, 55–62.
Burnham, D., Emmans, G.C. and Gous, R.M. (1992) Isoleucine responses in broiler
chickens. Interactions with leucine and valine. British Poultry Science 33, 71–87.
Cahaner, A., Druyan, S. and Deeb, N. (2003) Improving broiler meat production,
especially in hot climates, by genes that reduce or eliminate feather coverage.
British Poultry Science 44, S22–S23.
Dale, N. (1996) Variation in feed ingredient quality: oilseed meals. Animal Feed
Science Technology 59, 129–135.
Deming, W.E. (1986) Out of the Crisis. MIT Center for Advanced Engineering Study,
Cambridge, Massachusetts.
Duncan, M.S. (1988) Problems of dealing with raw ingredient variability. In:
Haresign, W. and Cole, D.J.A. (eds) Recent Advances in Animal Nutrition.
Butterworths, Boston, Massachusetts, pp. 3–11.
EFG Software Natal (1995) Advanced Computer Software for the Animal Industry.
http://www.efgsoftware.com/ (accessed February 2005).
animals, particularly poultry. In: Hillyer, G.M., Whittemore, C.T. and Gunn,
R.G. (eds) Computers in Animal Production. Occasional Publication No.5. British
Society of Animal Production, Edinburgh, UK, pp. 103–110.
Emmans, G.C. (1987) Growth, body composition, and feed intake. World’s Poultry
Science Journal 43, 208–227.
Emmans, G.C. and Fisher, C. (1986) Problems of nutritional theory. In: Fisher, C.
and Boorman, K.N. (eds) Nutritional Requirements and Nutritional Theory.
Butterworths, London, pp. 9–39.
Emmans, G.C. and Oldham, J.D. (1988) Modelling of growth and nutrition in
different species. In: Karver, S. and Van Arendonk, J.A.M. (eds) Modelling of
Livestock Production Systems. Kluwer Academic, Dordrecht, Netherlands,
pp. 13–21.
Fawcett, R.H., Webster, M., Thornton, P.K., Roan, S.W. and Morgan, C.A. (1992)
Predicting the response to variation in diet composition. In: Recent Advances in
Animal Nutrition. Butterworths, Boston, Massachusetts, pp. 137–158.
Science 14, 469–484.
Gous, R.M. and Morris, T.R. (1985) Evaluation of a diet dilution technique for
measuring the response of broiler chickens to increasing concentrations of
lysine. British Poultry Science 26, 147–161.
Gous, R.M., Pym, R.A.E., Mannion, P. and Wu, J.X. (1996) An evaluation of the
parameters of the Gompertz growth equation that describe the growth of eight
strains of broiler. In: Balnave, D. (ed.) Australian Poultry Science Symposium, Vol.
8. University of Sydney, Sydney, New South Wales, pp. 174–177.
Gous, R.M., Moran, E.T. Jr, Stilborn, H.R., Bradford, G.D. and Emmans, G.C.
(1999) Evaluation of the parameters needed to describe the overall growth, the
chemical growth and the growth of feathers and breast muscles of broilers.
Poultry Science 78, 812–821.
Hancock, C.E., Bradford, G.D., Emmans, G.C. and Gous, R.M. (1995) The
evaluation of the growth parameters of six strains of commercial broiler
chickens. British Poultry Science 36, 247–264.
Hellekalek, P. (2004) PLab. Theory and Practice of Random Number Generation.
http://random.mat.sbg.ac.at/ (accessed February 2005).
Hughes, R.J. and Choct, M. (1999) Chemical and physical characteristics of grains
related to variability in energy and amino acid availability in poultry. Australian
Journal of Agricultural Research 50, 689–701.
Jensen, L.S., Merrill, L.H., Reddy, C.V. and McGinnis, J. (1962) Observations on
eating patterns and rate of food passage of birds fed pelleted and unpelleted
diets. Poultry Science 41, 1414–1419.
Knap, P.W. (1995) Aspects of stochasticity: variation between animals. In: Moughan,
P.J., Verstegen, M.W.A. and Visser-Reyneveld, M.I. (eds) Modelling Growth in the
Pig. EAAP publication no 78, Wageningen Agricultural University,
Wageningen, Netherlands, pp.165–172.
Knuth, D.E. (1997) The Art of Computer Programming: Seminumerical Algorithms, Vol.
2, 3rd edn. Addison-Wesley, Reading, Massachusetts, USA.
Mack, S., Hohler, D. and Pack, M. (2000) Evaluation of dose-response data and
implications for commercial formulation of broiler diets. In: Balnave, D. (ed.)
Australian Poultry Science Symposium, Vol. 12. University of Sydney, Sydney, New
South Wales, pp. 82–87.
McCoy, R.A., Behnke, K.C., Hancock, J.D. and McEllhiney, R.R. (1994) Effect of
mixing uniformity on broiler chick performance. Poultry Science 73, 443–451.
Metayer, J.P., Grosjean, F. and Casting, J. (1993) Study of variability in French
cereals. Animal Feed Science and Technology 43, 87–108.
Mitchell, M.A. (1985) Effects of air velocity on convective and radiant heat transfer
from domestic fowls at environmental temperatures 20°C and 30°C. British
Morris, M.D. (1991) Factorial sampling plans for preliminary computational
experiments. Technometrics 33, 161–174.
Parmar, R.S., Diehl, K.C., Collins, E.R. and Hulet, R.M. (1992) Simulation of a
turkey house environment. Agricultural Systems 38, 425–445.
Reece, F.N. and Lott, B.D. (1982) The effect of environmental temperature on
sensible and latent heat production of broiler chickens. Poultry Science 61,
1590–1593.
Roush, W.B., Cravener, T.L. and Zhang, F. (1996) Computer formulation
observations and caveats. Journal of Applied Poultry Research 5, 116–125.
Senkoylu, N. and Dale, N. (1999) Sunflower meal in poultry diets: a review. World
Poultry Science Journal 55, 153–174.
Singh, P. and Trehan, P.K. (2002) Inheritance of rate of feathering and feather
density score and its relationship with body weight in broiler chicken. SARAS
Journal of Livestock and Poultry Production 18, 41–47.
Taylor, St. C.S. (1980) Genetic size scaling rules in animal growth. Animal Production
30, 161–165.
Van Beek G. and Beeking, F.F.E. (1995) A simple steady state model of the
distribution of vertical temperature in broiler houses without internal air
circulation. British Poultry Science 36, 341–356.
Wathes, C.M. and Clark, J.A. (1981) Sensible heat transfer from the fowl: radiative
and convective heat losses from a flock of broiler chickens. British Poultry Science
22, 185–196.
Weaver, W.D., Jr and Meijerhof, R. (1991) The effect of different levels of relative
humidity and air movement on litter conditions, ammonia levels, growth and
carcass quality for broiler chickens. Poultry Science 70, 746–755.
Wethli, E. and Morris, T.R. (1978) Effects of age on the tryptophan requirement of
laying hens. British Poultry Science 19, 559–565.
Xin, H., Berry, I.L., Tabler, G.T. and Barton, T.L. (1994) Temperature and
humidity profiles of broiler houses with experimental, conventional and
tunnel ventilation systems. Applied Engineering in Agriculture 10, 535–542.
6 Advancements in Empirical
Models for Prediction and
Prescription
W.B. ROUSH
USDA-ARS Poultry Research Unit, Mississippi State, MS 39762, USA
BRoush@msa-msstate.ars.usda.gov
From the beginning of efforts to model systems – long before the computer
era – linearity has dominated computation. This is not because anyone
believed that the systems of interest were truly linear, but simply for reasons of
computational tractability.
(Simon, 1990)
Nothing in Nature is random … A thing appears random only through the
incompleteness of our knowledge.
(Spinoza; quoted by Peitgen et al., 1992a, p. 319)
Introduction
The objective of this chapter is to present some thoughts on developments

in empirical prediction and prescription modelling. Each topic has
interesting implications for biological modelling. It is beyond the scope of
this chapter to go into detail. It is hoped that this brief introduction to the
developments in empirical models will serve as a catalyst in promoting and
considering their inclusion in the development of animal models.
Some Modelling Concepts
Modelling animal responses for prediction and prescription applications is

biologically complex. The ideal animal model for biological and economic
decisions has been identified as being composed of mechanistic, stochastic
and dynamic elements (Fisher, 1989; Theodorou and France, 2000).

98 W.B. Roush
Mechanistic and empirical models
Brown and Rothery (1994) define mechanistic models as concerned with the
‘nuts and bolts’ of biological processes and the way in which the
component parts fit together. They attempt to describe the observations in
terms of fundamental postulates about the biological processes. Theodorou
and France (2000) have noted that mechanistic models are constructed by
looking at the structure of the whole system, dividing it into its key
components and analysing the behaviour of the whole system in terms of
its individual components and their interactions with one another.
In contrast, empirical models provide quantitative descriptions of
patterns in the observations without attempting to describe the underlying
processes or mechanisms involved. In a sense, all models are empirical,
differing only in the degree of resolution and level of complexity.
Nevertheless, the broad distinction between mechanistic and empirical
models is a useful one (Brown and Rothery, 1994).
Theodorou and France (2000) have noted that the accuracy of
prediction of animal response using mechanistic models currently may be
lower than that achieved by the empirical methods used in practical
application. However, these research models are very useful in evaluating
the adequacy of current knowledge and data, identifying those areas
where research should be focused.
Empirical models are commonly referred to as black boxes. That is,
their mathematical workings are not transparent. Mechanistic models, at
the other extreme, are white boxes (or at least off-white boxes). The goal is
to have mechanistic models in which the workings are transparent. Artificial
neural networks (to be discussed later) are the ultimate empirical black box.
Recent research has suggested a merger between empirical and
mechanistic models. The merging of the two methods of modelling results
in a grey-box model. The grey-box model is intended to blend the better of two
worlds: knowledge-based modelling and black-box modelling (Oussar and
Dreyfus, 2001).
Stochastic and fuzzy logic models
Two of the tenets of classical science are order and precision. However
nature is not always orderly and precise. There is much variability and
imprecision. Casti (1994) comments:
… one of the great challenges to both science and philosophy is to provide a
rational, coherent account of the perceived uncertainty surrounding the events
of daily life. Classical probability theory offers one such approach but is riddled
with many well-known epistemological flaws and paradoxes. The theories of
fuzzy sets, satisfying and possibilities represent recent attempts to rectify some
of the deficiencies in the classical methods. Each of these newer schemes has at
its heart the basic fact that randomness is only one face of the mask of
uncertainty.
Advancements in Empirical Models for Prediction and Prescription 99
The stochastic and fuzzy concepts change the picture of the accurate
and precise answer of a deterministic model to a model that produces
answers that are based on probability and possibility. This presents a
challenge to the researcher in classical modelling and decision making. As
Zimmerman (1992) points out:
… until the 1960s, uncertainty, vagueness, and inexactness were features with
rather negative meanings. Nobody wanted to be called a ‘vague decision
maker’; a scientist that could not make precise and definite statements was not
regarded as a ‘true’ scientist, and uncertainty was considered to be something
disturbing that should, if at all possible, be avoided in models, theories, and
statements. The only theory that dealt with uncertainty was probability theory,
and this – predominately in its frequentative interpretation – was restricted to
situations in which the law of large numbers was valid and uncertainty could
be attributed to randomness.
Recent discoveries in nonlinear dynamics (chaos theory) further
complicate the matter of uncertainty by calling into question the nature of
randomness (Peitgen et al., 1992a).
Probability versus precision
Precision and accuracy in meeting nutrient levels and animal requirements

have been important goals; however, the inherent biological variance of
nutrients and requirements cannot be overlooked. Deming (2000) in his
studies on quality control has shown the futility of trying to get rid of
variability.
Precision and accuracy relate to average values. The variability of
nature makes the decisions associated with biology into risk problems
(Roush, 2001a). That is, in the case of nutrients, with what probability can
an animal’s requirement be met? Consideration of chance-constrained
programming, as an alternative to linear programming, to accomplish this
probabilistic approach to feed formulation is discussed below under the
heading Prescription Models.
Fuzzy set logic: dealing with imprecision
Fuzzy logic was introduced by Dr Lotfi Zadeh (1965) as a means of

defining the uncertainty of natural language to a computer. For
example, how does one define the concepts of hot and cold in a
computer program? To a human these words have meaning, though not
a precise meaning. The fuzzy set concept is related to set theory. In
traditional set theory an object is a member of a set of like objects. For
example, in comparing a baseball with a book, it is obvious that the
baseball would belong to a set of round objects and a book would belong
to a set of square objects. In the case of the colour spectrum, the
question becomes where does red become yellow and yellow become
100 W.B. Roush
red? From the fuzzy logic point of view the nanometres of light
representing red would be full membership in a red set and the
nanometres of light representing yellow would be full membership in
the yellow set. Nanometres representing colours between red and yellow
would have partial memberships in the red and yellow sets. Half way
between the red and yellow are the nanometres representing orange
which would have 0.5 membership in the red set and 0.5 membership in
the yellow set. Other imprecise concepts such as hot and cold, light and
heavy, short and tall, etc. can be represented by fuzzy logic in a similar
manner. Roush et al. (1989) and Roush and Cravener (1990) used fuzzy
logic to describe the imprecise concept of stress in a caged laying
situation. Fuzzy sets have been applied to the imprecision of human
nutrition and nutritional requirements by Wirsam and Uthus (1996),
Wirsam et al. (1997) and Gedrich et al. (1999).
There have been interesting fuzzy logic extensions to modelling of
control systems (Kosko, 1992) and simulations of social interaction of fish,
using fuzzy cognitive maps (Dickerson and Kosko, 1997).
Nonlinear dynamics: developmental history
Aristotle (c. 330 BC) pointed out that the ‘whole is greater than the sum of
its parts’. This is very evident in agricultural systems in which the response
of the organism results from the interaction of numerous inputs. Biological
and environmental inputs and the resulting outputs are not necessarily
additive or linear.
Sir Isaac Newton, early in his education, was a student of Aristotle’s
philosophy. However, Newton’s views of the world changed as he was later
influenced by the works of René Descartes and other mechanical
philosophers. The mechanical philosophers, in contrast to Aristotle, viewed
the world as composed entirely of particles of matter in motion and held
that all phenomena of nature result from their mechanical interaction
(Encyclopædia Britannica, 15th edn).
As a result, Newton and Descartes were advocates of a universe that
operates like clockwork, where everything is very orderly and mechanical.
The assumption was that if enough is known about a system (universe),
there is nothing that cannot be predicted for that system (universe). Under
this philosophy, the universe was a gigantic complicated clockwork
mechanism. The logic makes perfect sense that, if we understand each and
every part of the machine, then we can predict how the machine will act
and react. By taking apart the clock and studying each gear, an
understanding of the workings of the clock can be developed. Hence the
development of a reductionist approach to science.
Modern biology has inherited the reductionist approach. The living
organism is viewed as a complex biochemical machine. Examination is
made of the organs, tissues, cells and even the molecules in an effort to
define the mechanisms of life. Although the mechanistic view is desirable, it
has been recognized that the ideal model involves, in addition to defining
mechanisms, the inclusion of stochastic and dynamic concepts. Models
need to contain all three elements. The current focus has been mainly on
mechanistic models. There is a caveat that the exclusion of the dynamic
and stochastic elements in modelling promotes an illusion of precision in
attaining answers.
Mathematics of chaos theory
The firm concept of mathematical predictability, derived from experience

with linear equations, changed in 1963 when a weather researcher, Edward
Lorenz, discovered that small changes in the initial conditions (e.g.
1.00000 to 1.00001) of his mathematical model would result in
unpredictable changes in response over time. He developed a very simple
computer model of motion in which the air is heated from below and
cooled on top. Hot air rises and cold air falls. The air moves in rotating
cylinders that bring the hot air up and the cold air down on the other side.
The motion of the air mixes the hot and cold air, reducing the temperature
difference which is driving the motion of the air. Meanwhile the air is still
being heated from below and cooled at the top. After the cylinders of air
slow to a complete stop, they begin to rotate again and sometimes they
rotate in the opposite direction. That is, if the rotation was originally
clockwise, a switch is made to a counter clockwise direction. The rotation of
the cylinder speeds, slows and the rotation changes.
The Lorenz system consisted of three coupled differential equations
(Gleick, 1987):
dx/dt = 10 (y – x)
dy/dt = – xz +28x – y
dz/dt = xy – (8/3 z).
If a simple three equation model like that of Lorenz can exhibit surprising
dynamics, what does this infer about the dynamics involved in the
numerous differential equations involved in making a mathematical model
of a broiler, pig or cow?
The dynamics of the new science of chaos (Gleick, 1987) can be
illustrated with the following difference equation:
Xt+1= a Xt(1–Xt).
This (logistic) equation looks very predictable. However when iterated over
time as a difference equation the dynamics can be quite dramatic with
changes in the coefficient ‘a’.
May (1976) comments:
The fact that a simple deterministic equation can possess dynamical trajectories
which look like some sort of random noise has disturbing practical implications.
It means, for example that apparently erratic fluctuations … need not
necessarily betoken either the vagaries of an unpredictable environment or
102 W.B. Roush
sampling errors: they may simply derive from a rigidly deterministic …

equation such as (the logistic difference equation).
His sobering point is that the dramatic responses are not an influence
of the environment or of measurement error. The dramatic responses are
a phenomenon of nonlinear mathematics.
The reader is encouraged to examine the phenomenon with the
following BASIC program:
10 INPUT a
20 x = 0.54321
30 FOR n = 1 to 150
40 x = a*x*(1-x)
50 PRINT x
60 NEXT n
70 STOP
It is suggested to start ‘a’ at 2.0. Then increase the value of ‘a’ until it
equals 4.0. There will be a change in the output from periodic responses to
aperiodic (chaos) responses.
It seems as though there are different levels of the nonlinear dynamics.
Difference equations capture the moment-to-moment oscillation, while
differential equations capture the overall effect of interacting variables. It
seems that a three-dimensional system of first-order ordinary differential
equations is required for the manifestation of chaotic behaviour (May,
1976).
It is interesting that
Xt+1= a Xt (1–Xt)
is the difference equation form of the differential equation
dX
= aX t (1 − X t )
dt
which is the rate (velocity) form of the logistic equation of Robertson
(1908) (see Parks, 1982, p. 14). The equation is commonly used to model
growth in population studies. Derivation of the difference equation to
describe logistic growth from the Verhulst differential equation is outlined
in Solé and Goodwin (2000).
Because animal growth is often described as a logistic equation, the
next question is whether, in real life, growth shows the same moment-to-
moment dynamics as the difference equation form of the logistic equation.
Several studies have shown the day-to-day growth rate (velocity) of
broilers to be oscillatory (Roush et al., 1994; Roush and Wideman, 2000). The
oscillation has exhibited evidence of the new mathematics of chaos. Biological
systems, including heart rate dynamics and other physiological systems, have
also shown evidence of chaos (Degn et al., 1986; Glass and Mackey, 1988).
How does one deal with the nonlinear dynamics of chaos? A similar
problem in dealing with uncertainty and lack of precision occurs in
quantum theory. The suggested solution is to take a probabilistic point of
view. Mathematical chaos, along with randomness and imprecision, then

becomes a component of the variability inherent in the system.
For further information on nonlinear dynamics (chaos) and its related
topic, fractals, the following are suggested: Peitgen et al. (1992a,b); Moon
(1992); and Williams (1997).
Operations Research: the Science of Decision Making
Operations (Operational) Research (OR) is the formal discipline that

encompasses the development of the concepts, methods and tools for
decision modelling. It has a successful history for making efficient and
effective management decisions (Hillier and Lieberman, 2005). Roush
(2001b) has discussed OR from a poultry science point of view.
Much of the fundamental research for empirical prediction and
prescription decisions involves the discipline of OR. The development of
OR decision tools encompasses the disciplines of mathematics, statistics,
decision sciences, computer science and artificial intelligence. The well
known linear program was developed within the OR discipline.
Prediction models
Predictive models are designed to generate knowledge and come to the

truth without making any value judgments (Casti, 1989). Predictive models
are made for growth, feed intake, etc.
Traditionally, the tools for prediction have been algebraic and
regression equations, differential and difference equations. These are valid
and useful. However, research in artificial intelligence has added some
additional approaches including artificial neural networks, fuzzy logic and
genetic algorithms for modelling and prediction. In addition, the Kalman
Filter, a self-adjusting regression associated algorithm, has been suggested
for short term predictions (Roush et al., 1992).
Statistical analysis for biological research is sometimes taught as an
embellishment for a research project and it is often applied to the data as
an afterthought. In reality, the statistical design is an important tool for
obtaining a perspective of what is going on with the data and for efficient
and effective analysis. Traditionally, the student was taught to hold all
variables constant except for the variable of interest. This principle is
widely followed but, with the aid of modern statistical techniques it is now
possible to test many variables at the same time. It is interesting to note
that Beveridge (1957) gave this advice in 1957.
Regression analysis: response surface methodology
Factorial statistical models are known by most students who have taken a
graduate course in statistics. Usually the factorial model is analysed
104 W.B. Roush
qualitatively as an analysis of variance. However when factorial models

involve quantitative values such as levels of nutrients, an effective design is
to use response surface methods (RSM).
Response surface methods allow the simultaneous variation of two or
more variables to find the quantitative level that will give the most
desirable response. Box and Wilson (1951) were the first to report on this
topic. Recommended texts on response surface methodology are Cochran
and Cox (1957), Box and Draper (1987), Khuri and Cornell (1996) and
Myers and Montgomery (2002).
Yoshida et al. (1962, 1968, 1969) have used RSM to optimize the
growth and feed efficiency of chicks by varying protein and energy levels of
the ration. Mraz (1961a,b) used RSM to study the influence of calcium,
phosphorus and vitamin D3 on the uptake of several minerals. Waddel and
Sell (1964) used RSM designs to study the effects of calcium and
phosphorus on the utilization of iron by the chick. The studies by Mraz
and Waddel and Sell appear to have used RSM for their efficiency rather
than for their optimization capabilities.
Roush et al. (1979) used RSM to investigate the protein and energy
requirements of Japanese quail. The study showed an advantage of RSM
to identify optimal conditions outside the exploratory region covered in an
initial trial. A second experiment was run using the first trial predicted
optimum to pinpoint the optimal protein and energy levels for body
weight gain and feed conversion. Roush (1983) used RSM to examine the
protein levels in broiler starter and finisher diets and the optimal time of
ration change. Roush et al. (1986) investigated optimal calcium and
available phosphorus requirements for laying hens using RSM.
The response surface model allows the researcher to examine optimal
conditions such as the optimal levels of protein and energy to produce a
response. This is much more powerful and informative than just
examining differences between treatments.
The following is an example of a nonlinear quadratic equation to be
fitted by regression analysis:
ŷ = b0 + b1x1 + b2x 2 + b11x12 + b22x 22 + b12x1x 2
This model would probably give a useful approximation to the true

response surface. Three dimensional and two dimensional contour plots
can be drawn to define visually the optimum combination of variables and
the value at the optimum. These optimum values can also be found by
taking the first derivative of the equation for each variable and setting the
equations equal to zero.
Mixture models: making a cake
Mixture designs are a type of response surface design that have application
to problems in the animal sciences. In the general mixture problem, the
measured response is assumed to depend only on the proportions of the

ingredients present in the mixture and not on the amount of the mixture
(Snee, 1971; Cornell, 2002). However there are mixture designs where the
amount is also considered (Piepel and Cornell, 1985). Types of mixture
problems include cake formulations, content of construction concrete,
making flares, fruit punches, photographic film and gas blends. The model
adds to unity and does not have an intercept.
ŷ = b1x1 + b2x 2 + b3x 3 + b12x1x 2 + b13x1x 3 + b23x 2x 3
Gous and Swatson (2000) have used mixture experiments to study the
ability of the broiler to choose from three protein sources the combination
of ingredients that would maximize biological performance. Roush et al.
(2004) suggested that the optimal time to feed broilers the starter, grower
and finisher feeds could be viewed as a mixture problem with the objective
of finding the optimal proportion of time to feed each diet.
Artificial neural networks – the ultimate black box
Neural networks are an alternative to regression analysis. The neural

network was inspired by the structure and function of biological neurons.
Neural networks are trained through iteration of example patterns. The
neuron receives one or more inputs and transforms the sums of those
inputs to an output value which in turn is transferred to other neurons.
The artificial neural network is a set of processing units that simulate
biological neurons and are interconnected by a set of weights that allows
both serial and parallel processing through the network. The artificial
neuron works like a switch; when there is sufficient neurotransmitter
accumulated in the cell body, an action potential is activated. In the
artificial neuron, a weighted sum is made of the signals coming into a node
from other nodes. A comparison is then made to a threshold value. If the
threshold is exceeded, the node fires a signal that becomes the input for
another node or an output value. The key attribute of a neural network is
not the complexity of the neurons: power comes from the density and
complexity of the interconnections (Cross et al., 1995).
One of the challenges for neural networks is the over-training of the
model to the point that the model is not useful beyond the data on which it
was trained or developed. This is also true of regression polynomials. The
regression polynomial can be over developed by adding more variables to
the model. The neural network overcomes this problem in two ways. The
first method is to include a test set which represents a randomly chosen set
of data from the training data set (for example 20% of the training data
set) that is set aside. During the training procedure the model is
continually evaluated against the test data set and the error between the
input and output is determined. As the training proceeds, it is expected
that the error between predictions and actual values of the test set will
106 W.B. Roush
decrease. There will be a point at which the errors will start to increase
which is the point at which over-training has started to occur. That is the
point at which the neural network is saved. At this point, validation is made
of the new neural network on data that are independent of the training
and testing data sets. The second procedure for reducing over-training is
to use statistical Jack-Knife and bootstrap procedures, where the
experimental data are re-sampled in the process of development of the
neural network. In this way there is not a need for the test set.
The development of neural networks can incorporate training, testing
(to avoid over-training) and validation to make robust models. For
example Roush et al. (1996b) developed an artificial neural network to
predict the presence or absence of ascites in broilers. The neural network
was a three layer back propagation neural network with an input of 15
physiological variables. After developing the neural network with training
and test sets, the neural network predictive ability was validated with two
data sets that were not involved in the training. The neural network
accurately identified two false positives and one false positive in the first
and second evaluation data sets, respectively. The birds identified as false
positives were actually determined to be in the developmental stages of
ascites.
There are many different types of neural networks. These different
types can be generally classified as supervised and unsupervised networks.
In supervised learning, the neural network learns from an example. With
unsupervised learning, the neural network examines the data to define
clusters of information. The neural network is used to associate data,
classify data, transform data into a different representation and to model
data (Zupan and Gasteiger, 1993). The commercial neural network
package NeuralShell 2 (Ward Systems Group, 1996) contains 16 different
types of neural network, which include the following:
1. Backpropagation neural networks. This neural network is the standard.
Usually three layers are sufficient. The layers are the input, hidden and
output layers. Each layer is linked only to the previous layer.
2. Jump connection neural network. This type of backpropagation
network has every layer connected to every previous layer.
3. Recurrent network. This is a type of backpropagation neural network.
There is feedback to previous layers. These networks are often used for
time series data. Regular feed forward neural networks respond to a given
input with the same output each time. A recurrent network may respond
to the same input pattern differently from time to time, depending upon
the input patterns previously presented to it. The recurrent network
builds a long term memory based on the patterns presented.
4. Kohonen architecture. This is an unsupervised neural network. It is
able to learn without being shown correct output patterns. The use of this
type of network is for clustering problems. The network is able to separate
data into a specified number of groups or categories.
5. Probabalistic neural network (PNN). This is a powerful neural network
for classification problems. It is able to train on a sparse data set. PNN

separates data into a specified number of categories.
6. General regression neural networks (GRNN). GRNN are powerful
neural networks that have been shown to outperform backpropagation
methods. They train quickly on sparse data sets and are particularly useful
for continuous function approximation as in the case of examining the
relation between body weight and time. The GRNN is a three-layer neural
network with the number of hidden neurons equal to the number of
training patterns.
7. Group method of data handling (GMDH). This network derives a
mathematical formula which is a nonlinear polynomial expression relating
the values of the most important inputs to predict outputs. The network
works very much like the genetic algorithm in that the mathematical
expression is based on variables that survive.
For more information on neural networks (and fuzzy logic) see Tsoukalas
and Uhrig (1997).
Genetic algorithms
Genetic algorithms are search procedures that use the principles of natural
selection and genetics. The genetic algorithm was first developed by John
H. Holland in the 1960s. The search procedure is usually looking for an
optimum condition. The model to be optimized can be a formula or even a
neural network in which the maximum, minimum or a particular value is
required. The genetic algorithm works particularly well with problems that
are ‘not well behaved’. That is, situations where it may be difficult to find
the global optimum.
Commercial neural networks and genetic algorithms are available that
can be incorporated into a spreadsheet. The setup for the genetic
algorithm is based on an objective equation and constraints similar to the
setup of a linear program.
Kalman filter: tracking targets
The Kalman filter is a recursive algorithm for making short term

predictions. Biological monitoring is complicated by variation (noise) in
responses that may mask abrupt changes in responses. The monitoring of
changes in responses containing variation is a common problem in many
disciplines. An algorithm was developed by Kalman (1960) for application
to such problems. The algorithm has been used for navigation, missile
guidance, and satellite tracking. This type of problem requires short-term
prediction and adjustments. The Kalman filter has been applied to
biological problems such as monitoring renal transplants (Smith and Cook,
1980; Smith and West, 1983; Trimble et al, 1983), heart rates (Heath, 1984)
108 W.B. Roush
and poultry production responses (Garnaoui, 1987; Roush et al., 1992).

The Kalman Filter has predicted animal breeding values (Hudson, 1984),
estimated lactation curves (Goodall and Sprevak, 1985) and feed intake
and growth of beef cattle (Oltjen and Owens, 1987).
Prescription Models
Prescription models are designed to define values in making decisions.

Linear programming for feed formulation is one of the most commonly
used prescriptive models.
Linear programming
Since its inception in 1947, linear programming has been the workhorse of
decision-making algorithms. Numerous texts and applications have been
written about its use. Linear programs have been used for blending (e.g.
petroleum products), mixes (e.g. investments and budgeting), scheduling
(e.g. production to satisfy customer demand, production capacity, and
storage limitations), assignment (e.g. workers to tasks) and
transportation/dispatching (e.g. routing of pick up and deliveries). In the
animal sciences, the term ‘linear programming’ is considered by many as
synonymous with the mixing problem of feed formulation. The Sadia
company, the largest broiler producer in Brazil, used linear programming
and other operations research methods effectively to improve decision
making about production and product distribution in their business
(Taube-Netto, 1996).
The linear program consists of an objective equation and constraint
equations. For example in a feed formulation problem the objective is to
minimize the cost of ingredients subject to meeting the nutritional
constraints. The following is an example:
Objective equation:
Minimize cost: 0.08 Maize + 0.20 Soybean
Constraint equations:
87 Maize + 488 Soybean ≥ 230 (protein constraint)
Maize + Soybean = 1 (amount constraint)
where 0.08 and 0.20 represent the price ($/kg) of maize and soybean and
87 and 488 represent the protein contents (g/kg) of these ingredients. The
objective is to minimize the cost of the diet, subject to the constraints that
the protein supplied by the maize and soybean together must be ≥ 230
g/kg diet (the requirement of the animal) and the fractional amounts of
maize and soybean must total to 1.
Mathematically there are certain assumptions made about linear
programming (Render and Stair, 1982; Roush et al., 1996a):
1. Conditions of certainty exist; that is, the numerical values in the

objective and constraint equations are known with certainty and do not
change during the period being studied. It is assumed there is no
variability in the numerical values.
2. Proportionality exists in the objective and constraints. This means that
the units are consistent for each of the equations.
3. Additivity is assumed: that is, the total of the activities equals the sum of
each individual activity.
4. Divisibility is assumed: that is, the solutions need not be whole numbers.
Instead, they are divisible and may take a fractional value.
5. The answers or variables are non-negative.
Early in the advent of computer formulation with linear programming, it
was recognized that biological variability, particularly nutrient variability,
was a problem in meeting the nutrient requirements of animals. A linear
program solution based on an ingredient matrix of average nutrient values
has a 50% probability of not meeting the nutrient requirements of a group
of animals. In order to avoid this risk, some nutritionists incorporate a
margin of safety in the ingredient matrix. Nott and Combs (1967)
suggested an adjustment of the nutrient means by subtracting (or adding)
a fraction (they suggested 0.5) of the standard deviation from (or to) the
nutrient mean which would provide a probability of 69% or greater in
meeting the nutrient requirement. The adding or subtracting depends on
whether the constraint is for a maximum or minimum value.
Generally there is not a problem in meeting assumptions (2)–(5) for
feed formulation. However, the basic assumption of certainty (1) is violated
by the inherent nutrient variability of feed ingredients. A consequence of
this violation brings unexpected results. Usually there is an over-
formulation of the requested probability and requirements in the feed
formulation.
Chance constrained programming
A more appropriate approach is to use chance-constrained programming

(sometimes referred to as stochastic programming). Using the example
above the protein constraint becomes:
87 Maize + 488 Soybean – 0.5 (8 Maize)2 +(4 Soybean)2 ≥ 230
This method more accurately calculates the nonlinear nutrient variation.

An intuitive analogy is to compare the following two, similar looking, but
unequal equations:
9 + 16 =7 (6.1)
and
9+16 = 5 (6.2)
110 W.B. Roush
The linear program with a margin of safety corrects the nutrients by

adjusting the individual square root values as in Eqn 6.1 and the chance
constrained program adjustment is accomplished by a square root of the
summation (Eqn 6.2). The consequence is that Eqn 1 has a larger number
than Eqn 6.2. From a feed formulation point of view there would be an
over-correction using the Eqn 6.1 approach. The overcorrection results in
a higher cost ration and overshoots the requested probability for nutrient
level. Roush et al. (1996a) discuss examples of the difference between linear
programming with a margin of safety formulation and a chance-
constrained program formulation.
Goal programming: more than one objective
The goal program is a multiple objective approach to solving mathematical

programming problems. Ignizio and Cavalier (1994) point out that the goal
program may more accurately define real world problems than a rigid single
objective linear program. A single objective linear program sometimes results
in solutions that are infeasible. In contrast, the philosophy of a goal program
is of satisficing and not that of optimization. The concept of satisficing is an
attempt to seek an acceptable solution, that is, one that satisfies desired goals
(Ignizio and Cavalier, 1994). The approach is to make the goals into
constraints. Examples and illustrations of the methodology can be found in
Hillier and Lieberman (2005), Oberstone (1990) and other operations
research and management science texts.
Several animal science papers have been based on goal programming
including Rehman and Romero (1984, 1987), Lara and Romero (1992,
1994) and Zhang and Roush (2002).
It should be noted that a fuzzy linear program is a special case of a
goal program. Examples of such a program are given in Zimmermann
(1996).
Quadratic programming
Miller et al. (1986) and Pesti et al. (1986) combined broiler growth
equations obtained using response surface methodology with quadratic
programming. The result was that they were able to demonstrate that a
quadratic programming model would provide a method of ration
formulation that would take into account the productivity of the broiler.
They did this by defining a quadratic objective as the growth response to
intake of protein and energy. Live weight, transformed to feed input space,
was maximized subject to a given cost per bird and other common
constraints in linear programming of a feed mix. Linear programming
(LP) and quadratic programming (QP) results were compared. Pesti et al.
(1986) reported the energy concentrations of the diets were similar by both
methods of formulation (LP = 13.62 and QP = 13.232 MJ/kg). However,
the protein content was much higher when the QP method was used (LP
= 217 and QP = 244 g/kg). Miller et al. (1986) reported that analysis using
QP indicated that a leading broiler firm could have improved economic
efficiency by increasing protein density and (slightly) reducing energy
density of broiler finisher feed. Further if this was applied industry wide,
the savings would be US$120 million per year.
Decision analysis
Decision analysis is a technique for providing solutions to problems by

determining proper courses of action. The decision method is
accomplished by listing all available courses of action, expressing subjective
variables quantitatively, and determining possible returns based on each
action. Decision analysis is a framework upon which mathematical models
can be evaluated under different scenarios.
Roush (1986) showed how conventional decision analysis using a profit
potential equation based on different price situations for eggs and feed
could be used to examine the number of hens to place in laying hen cages.
In two subsequent papers, Roush et al. (1989) and Roush and Cravener
(1990) used fuzzy decision analysis to evaluate crowding of caged laying
hens based on cage space (in the first paper) and cage space and colony
size (in the second paper). Multicriteria decision analysis was applied by
Roush and Cravener (1992) to demonstrate how the choice of a
commercial laying hens strain could be made when the information used
in the comparison has incommensurate units.
Conclusion
This chapter has been an attempt to present some of the developments in

empirical models that may help in defining and making decisions with
animal models.
Casti (1989) lamented:
Should you have the misfortune to pick up a typical current textbook
purporting to address the arcane arts of mathematical modelling, the chances
are overwhelmingly high that the author will transport you back into the 1950s
with an account of how to model an oscillating pendulum, freeway traffic, or
dog food using the static, equilibrium-centered, linear techniques of
mathematical programming, regression analysis or, perhaps, elementary
functional analysis. My feeling is that the time is long overdue to bring the
mathematics of the 1980s into contact with the students of the 1980s and offer
courses on modelling that stress dynamics rather than statics, nonlinearity
rather than linearity and possibility rather than optimality.
Though there is still room for improvement, several modelling books
have modernized their mathematical approaches (e.g. Griffiths and
Oldknow, 1993; Brown and Rothery, 1994).
112 W.B. Roush
In the experience of the author, some of the mathematical results

appear to some researchers and practitioners to be ‘tricks with smoke and
mirrors’. One can be assured that the results of a concept like chance-
constrained programming are real. There are measurable differences in
the nutrients formulated with a margin of safety and with a chance-
constrained approach. There are mathematical reasons for this.
The nonlinear dynamics of chaos are also real. May (1976) suggested
… that people [should] be introduced to X(t+1)=aXt(1Xt) [i.e., the Logistic
Equation] early in their mathematical education. This equation can be studied
phenomenologically by iterating it on a calculator, or even by hand. Its study
does not involve as much conceptual sophistication as does elementary
calculus. Such study would greatly enrich the student’s intuition about
nonlinear systems.
The study of chaos, its implications and how it occurs, is a hot topic in
mathematics and physics. The dynamic results are not tricks.
Developments in empirical modelling are constantly expanding.
Nonlinear mathematics, Artificial Intelligence and a relatively new field,
Artificial Life (Levy, 1992), are areas where, in the opinion of the author,
there will be important melding of empirical and mechanistic modelling.
References
Aristotle (c.330 BC) Metaphysica 1045a, 10 f.
Beveridge, W.I.B. (1957) The Art of Scientific Investigation. Random House, New York.
Box, G.E.P. and Draper, N.R. (1987) Empirical Model-Building and Response Surfaces.
John Wiley, New York.
Box, G.E.P. and Wilson, K.B. (1951) On the experimental attainment of optimum
conditions (with discussion). Journal of the Royal Statistical Society B 13, 1–45.
Brown, D. and Rothery P. (1994) Models in Biology: Mathematics, Statistics and
Computing. John Wiley, New York.
Casti, J.L. (1989) Alternate Realities: Mathematical Models of Nature and Man. John
Wiley, New York.
Casti, J.L. (1994) Complexification: Explaining a Paradoxical World through the Science of
Surprise. Harper Collins, New York.
Cochran, W.G. and Cox, G.M. (1957) Experimental Designs, 2nd edn. John Wiley,
New York.
Cornell, J. (2002) Experiments with Mixtures: Designs, Models, and the Analysis of Mixture
Data, 3rd edn. John Wiley, New York.
Cross, S., Harrison, R.F. and Kennedy, R.L. (1995) Introduction to neural
networks. Lancet 346, 1075–1079.
Degn, H., Holden, A.V. and Olsen, L. F. (eds) (1986) Chaos in Biological Systems.
NATO ASI Series, Series A: Life Sciences Vol. 138. Plenum Press, New York.
Deming, W.E. (2000) Out of the Crisis. MIT Press, Cambridge, Massachusetts.
Dickerson, J.A. and Kosko, B. (1997) Virtual worlds in fuzzy cognitive maps. In:
Kosko, B. (ed.) Fuzzy Engineering. Prentice-Hall, Upper Saddle River, New
Jersey.
Encyclopædia Britannica, 15th edn, s.v. “Newton, Sir Isaac”.
Fisher, C. (1989) Use of models to describe biological function and to estimate

nutrient requirements for poultry. In: Report of Proceedings of First European
Symposium on EDP Applications in Poultry Management. WPSA Danish Branch,
DSR Forlag Copenhagen, pp. 52–63.
Garnaoui, K.H. (1987) Kalman filter detection of abnormal biological responses in
poultry management systems. MS thesis, Pennsylvania Sate University, USA.
Gedrich, K., Hensel, A., Binder, I. and Karg, G. (1999) How optimal are computer-
calculated diets? European Journal of Clinical Nutrition 53, 309–318.
Glass, L. and Mackey, M.C. (1988) From Clocks to Chaos: The Rhythms of Life.
Princeton University Press, Princeton, New Jersey.
Gleick, J. (1987) Chaos: Making a New Science. Penguin, New York.
Goodall, E.A. and Sprevak, D. (1985) A Bayesian estimation of the lactation curve
of a dairy cow. Animal Production 40, 189–193.
Gous, R.M. and Swatson, H.K. (2000) Mixture experiments: a severe test of the
ability of a broiler chicken to make the right choice. British Poultry Science 41,
136–140.
Griffiths, H.B. and Oldknow, A. (1993) Mathematics of Models: Continuous and Discrete
Dynamical Systems. Ellis Horwood, New York.
Heath, R.A. (1984) Detection of change in physiological measures using an
adaptive Kalman filter algorithm. Psychological Bulletin 96, 581–588.
Hillier, F.S. and Lieberman, G.J. (2005) Introduction to Operations Research, 8th edn.
McGraw-Hill, New York.
Hudson, G.F.S. (1984) Extension of a reduced animal model to recursive prediction
of breeding values. Journal of Animal Science 59, 1164–1175.
Ignizio, J.P. and Cavalier, T.M. (1994) Linear Programming. Prentice-Hall,
Engelwood Cliffs, New Jersey.
Kalman, R.E. (1960) A new approach to linear filtering and prediction problems.
Journal of Basic Engineering, Series D 82, 35–46.
Khuri, A.I. and Cornell, J.A. (1996) Response Surfaces, Designs and Analysis, 2nd edn.
Marcell Dekker, New York.
Kosko, B. (1992) Neural Networks and Fuzzy Systems: A Dynamical Systems Approach to
Machine Intelligence. Prentice-Hall, Englewood Cliffs, New Jersey.
Lara, P. and Romero, C. (1992) An interactive multigoal programming model for
determining livestock rations: An application to dairy cows in Andalusia Spain.
Journal of the Operational Research Society 43, 945–953.
Lara, P. and Romero, C. (1994) Relaxation of nutrient requirements on livestock
rations through interactive multigoal programming. Agricultural Systems 45,
443–453.
Levy, S. (1992) Artificial Life: The Quest for a New Creation. Pantheon, New York.
May, R.M. (1976) Simple mathematical models with very complicated dynamics.
Nature 261, 459–467.
Miller, B.R., Arraes, R.A. and Pesti, G.M. (1986) Formulation of broiler finishing
rations by quadratic programming. Southern Journal of Agricultural Economics
18, 141–150.
Moon, F.C. (1992) Chaotic and Fractal Dynamics: An Introduction for Applied Scientists
and Engineers. John Wiley, New York.
Mraz, F.R. (1961a) Influence of dietary calcium, phosphorus, and vitamin D3 on
Ca45, P32, and Sr85 uptake of chicks. Journal of Nutrition 73, 409–414.
Mraz, F.R. (1961b) Influence of phosphorus and some alkaline earth metals on the
uptake of Ca45, Sr85, Ba133, and P32 by tibiae of chicks. Poultry Science 40,
958–964.
114 W.B. Roush
Myers, R.H. and Montgomery, D.C. (2002) Response Surface Methodology, 2nd edn.
Nott, H. and Combs, G.F. (1967) Data processing of ingredient composition data.
Feedstuffs 14, 21.
Oberstone, J. (1990) Management Science: Concepts, Insights and Applications. West
Publishing, New York.
Oltjen, J.W. and Owens, F.N. (1987) Beef cattle feed intake and growth: Empirical
Bayes derivation of the Kalman filter applied to a nonlinear dynamic model.
Journal of Animal Science 65, 1362–1370.
Oussar, Y. and Dreyfus, G. (2001) How to be a gray box: dynamic semi-physical
modelling. Neural Networks 14, 1161–1172.
Peitgen, H-O., Jürgens, H. and Saupe, D. (1992a) Fractals for the Classroom. Part One:
Introduction to Fractals and Chaos. Springer, New York.
Peitgen, H-O., Jürgens, H. and Saupe, D. (1992b) Fractals for the Classroom. Part Two:
Complex Systems and Mandelbrot Set. Springer, New York.
Pesti, G.M., Arraes, R.A. and Miller, B.R. (1986) Use of quadratic growth response
to dietary energy and protein concentrations in least-cost feed formulation.
Piepel, G.F. and Cornell, J.A. (1985) Models for mixture experiments when the
response depends on the total amount. Technometrics 27, 219–227.
Rehman, T. and Romero, C. (1984) Multiple-criteria decision making techniques
and their role in livestock ration formulation. Agricultural Systems 15, 23–49.
Rehman, T. and Romero, C. (1987) Goal programming with penalty functions and
livestock ration formulation. Agricultural Systems 23, 117–132.
Render, B. and Stair, R.M. (1982) Quantitative Analysis for Management. Allyn and
Bacon, Boston, Massachusetts.
Roush, W.B. (1983) An investigation of the protein levels for broiler starter and
finisher rations and the time of ration change by response surface
methodology. Poultry Science 62, 110–116.
Roush, W.B. (1986) A decision analysis approach to the determination of
population density in laying cages. World’s Poultry Science Journal 42, 26–31.
Roush, W.B. (2001a) Stochastic programming: Implications for precision feeding.
Feedstuffs July 2, 14–15.
Roush, W.B. (2001b) An operations research approach to leveraging new
technology opportunities. Journal of Applied Poultry Research 10, 293–302.
Roush, W.B. and Cravener, T.L. (1990) Evaluation of colony size and cage space for
laying hens (Gallus domesticus) using fuzzy decision analysis. Poultry Science 69,
1480–1484.
Roush, W.B. and Cravener, T.L. (1992) Choosing a commercial laying hen strain
based on multicriteria decision analysis. Journal of Applied Poultry Research 1,
403–409.
Roush, W.B. and Wideman, R.F., Jr (2000) Evaluation of broiler growth velocity
and acceleration in relation to pulmonary hypertension syndrome. Poultry
Science 79, 180–191.
Roush, W.B., Petersen, R.G. and Arscott, G.H. (1979) An application of response
surface methodology to research in poultry nutrition. Poultry Science 58,
1504–1513.
Roush, W.B., Mylet, M., Rosenberger, J.L. and Derr, J. (1986) Investigation of
calcium and available phosphorus requirements for laying hens by response
surface methodology. Poultry Science 65, 964–970.
Roush, W.B., Bock, R.G. and Marszalek, M.A. (1989) Evaluation of crowding of
laying hens (Gallus domesticus) using fuzzy set decision analysis. Applied Animal
Behaviour Science 23, 155–163.
Roush, W.B., Tomiyama, K., Garnaoui, K.H., D’Alfonso, T.H. and Cravener, T.L.
(1992) Kalman filter and an example of its use to detect changes in poultry
production responses. Computers and Electronics in Agriculture 6, 347–356.
Roush, W.B., Barbato, G.F. and Cravener, T.L. (1994) A nonlinear dynamical
(chaos) approach to the analysis of broiler growth. Poultry Science 73,
1183–1195.
Roush, W.B., Cravener, T.L. and Zhang, F. (1996a) Computer formulation
observations and caveats. Journal of Applied Poultry Research 5, 116–125.
Roush, W.B., Kochera Kerby, Y., Cravener, T.L. and Wideman, R.F., Jr (1996b)
Artificial neural network prediction of ascites in broilers. Poultry Science 75,
1479–1487.
Roush, W.B., Boykin, D. and Branton, S.L. (2004) Optimization of phase feeding of
starter, grower and finisher diets for male broilers by mixture experimental
design: Forty-eight day production period. Poultry Science 83, 1264–1275.
Simon, H.A. (1990) Prediction and prescription in systems modelling. Operations
Research 38, 7–14.
Smith, A.F.M. and Cook, D.G. (1980) Straight lines with a change-point: a Bayesian
analysis of some renal transplant data. Applied Statistics 29, 180–189.
Smith, A.F.M. and West, M. (1983) Monitoring renal transplants: An application of
the multiprocess Kalman filter. Biometrics 39, 867–878.
Snee, R.D. (1971) Design and analysis of mixture experiments. Journal of Quality
Control 3, 159–169.
Solé, R. and Goodwin, B. (2000) Signs of Life: How Complexity Pervades Biology. Basic
Books, New York.
Taube-Netto, M. (1996) Integrated planning for poultry production at Sadia.
Interfaces 26, 38–53.
Theodorou, M.K. and France, J. (eds) (2000) Feeding Systems and Feed Evaluation
Models. CABI Publishing, New York.
Trimble, I.M., West, M., Knapp, M.S., Pownall, R. and Smith, A.F.M. (1983)
Detection of renal allograft rejection by computer. British Medical Journal 286,
1695–1699.
Tsoukalas, L.H. and Uhrig, R.E. (1997) Fuzzy and Neural Approaches in Engineering.
Waddel, D.G. and Sell, J.L. (1964) Effects of dietary calcium and phosphorus on
the utilization of dietary iron by the chick. Poultry Science 43, 1249–1257.
Ward Systems Group (1996) NeuroShell 2. Ward Systems Group, Frederick, MD,
USA.
Williams, G.P. (1997) Chaos Theory Tamed. Joseph Henry, Washington, DC.
Wirsam, B. and Uthus, E.O. (1996) The use of fuzzy logic in nutrition. Journal of
Nutrition 126 (9 Suppl.), 2337S–2341S.
Wirsam, B., Hahn, A., Uthus, E.O. and Leitzmann, C. (1997) Fuzzy sets and fuzzy
decision making in nutrition. European Journal of Clinical Nutrition 51, 286–296.
Yoshida, M., Hizikuro, S., Hoshii, H. and Morimoto, H. (1962) Effect of dietary
protein and energy levels on the growth rate, feed efficiency and carcass
composition of chicks. Agricultural and Biological Chemistry 26, 640–647.
Yoshida, M., Hoshii, H. and Morimoto, H. (1968) Breed difference in the pattern
of growth response to the change in the dietary energy and protein levels.
Japanese Poultry Science 5, 211–225.
116 W.B. Roush
Yoshida, M., Hoshii, H. and Morimoto, H. (1969) Age difference in the pattern of
growth response to the change in the dietary energy and protein levels.
Japanese Poultry Science 6, 7–21.
Zadeh, L.A. (1965) Fuzzy sets. Information and Control 8, 338–353.
Zhang, F. and Roush, W.B. (2002) Multiple-objective (goal) programming model
for feed formulation: an example for reducing nutrient variation. Poultry
Science 81, 182–192.
Zimmerman, H.J. (1992) In: Zadeh, L.A. and Kacprzyk, J. (eds) Fuzzy Logic for the
Management of Uncertainty. Wiley, New York.
Zimmerman, H.J. (1996) Fuzzy Set Theory and its Applications, 3rd edn. Kluwer
Academic, Boston, Massachusetts.
Zupan, J. and Gasteiger, J. (1993) Neural Networks for Chemists: An Introduction.
VCH, New York.
7 The Problem of Predicting the
Partitioning of Scarce Resources
during Sickness and Health
in Pigs
I. KYRIAZAKIS AND F.B. SANDBERG
Animal Nutrition and Health Department, Scottish Agricultural College
West Mains Road, Edinburgh, EH9 3JG, UK
Ilias.Kyriazakis@sac.ac.uk
Introduction
Pigs are often faced with the problem of partitioning one or more scarce
food resources. This situation may arise when they are offered
intentionally restricted amounts of food, but also when they are offered ad
libitum access to a food. In the latter case intake of scarce resources may
arise when pigs are given access to a poor quality food (e.g. bulky food,
Whittemore et al., 2001) or when environmental stressors unintentionally
constrain food intake (e.g. high environmental temperature, Wellock et al.,
2003a). The voluntary reduction of food intake that accompanies
subclinical infection (anorexia, Kyriazakis et al., 1998) is a special case that
may lead to intake of scarce resources.
We are interested in the problem of predicting the partitioning of
absorbed scarce protein (and energy) to protein (PR) and lipid (LR)
retention in healthy and ‘diseased’ growing pigs. Quantitative solutions to
this problem in healthy pigs have been evolving for over 30 years. For this
reason, we will start with a historical approach to resolving the issue. We
will then discuss the current solutions to it offered in the literature and aim
to reach a solution that appears able to predict satisfactorily protein and
lipid retention in healthy growing pigs. The preferred solution will form
the basis of a framework that will be developed to account for the
partitioning of absorbed scarce resources in pigs challenged by pathogens.
This part of the framework will mainly have a heuristic value, as it will be
pointing towards issues that need to be resolved in order to be able to
predict adequately protein and lipid retention in pigs challenged by
pathogens.

118 I. Kyriazakis and F.B. Sandberg
Inherent in the above is the definition of scarce food resources. We take

these to be resources that limit the pig from achieving its capacities, for
example for maximum protein retention (PRmax), as these are defined by its
genotype. For a description of these capacities and the resultant nutrient
requirements, the reader is referred to Emmans and Kyriazakis (2001).
A Historical Perspective to the Partitioning of Scarce Protein and

Energy in Healthy Pigs
The current solutions to the problem of partitioning scarce protein and
energy in pigs are summarized in Table 7.1. The solutions have been
grouped according to the ‘school of thought’ they have evolved from; this
is the key characteristic they invoke to resolve the problem.
Some of these solutions have now only historical value and their
proposers have replaced them with more recent ones, e.g. Whittemore and
Fawcett (1974, 1976) have been replaced by Green and Whittemore (2003).
Others represent a purely statistical, best-fit approach to a particular data
set (van Milgen and Noblet, 1999) and hence lack general applicability.
These solutions will not be considered any further here.
Three of the remaining proposed solutions may be rejected on
qualitative grounds alone. Whittemore (1995) and Green and Whittemore
(2003) propose the ratio of lipid to protein in the body as a possible
constraint to PR. Operating within a certain range, this solution does not
allow lipid to be lost whilst there is a gain in protein, despite the strong
evidence that this can occur (Stamataris et al., 1991; Kyriazakis and
Emmans, 1992a,b). The rule of Fuller and Crofts (1977) recognizes that
the efficiency of using protein above maintenance might be a function of
Table 7.1. The current solutions to the problem of predicting the partitioning of scarce
resources. The solutions have been grouped together according to the key characteristic they
invoke to resolve the problem.
Solutions Key characteristic
Whittemore and Fawcett (1974, 1976) ⎫

Whittemore (1995) ⎩ A minimum ratio of lipid to protein in gain
de Lange (1995) ⎧
Green and Whittemore (2003) ⎭
Fuller and Crofts (1977) Efficiency of using protein
Black et al. (1986) ⎫
de Greef and Verstegen (1995) ⎪
NRC (1998) ⎬ Marginal responses in protein retention to
van Milgen and Noblet (1999) ⎪ energy intake
van Milgen et al. (2000) ⎭
Kyriazakis and Emmans (1992a,b) Marginal responses in protein retention to
Sandberg et al. (2005a,b) } ideal protein intake
Partitioning of Scarce Resources during Sickness and Health 119
the energy to protein ratio of the food. It calls for the values of four
parameters in order to solve the problem; each of these values are stated to
be affected by genotype, state, liveweight and nutritional history. The
practical consequence of the solution is that each experiment needs to be
carried out across all of these factors, possibly in all combinations, in order
for the rule to apply in any given case. An enormous amount of
information is called for, and for this reason this rule will not be discussed
any further. The same criticism applies to the rule proposed by van Milgen
et al. (2000), as they estimate that 21 parameters are required in order to
predict protein and lipid retention. The information required by this rule
for any particular genotype, existing in the future, is unlikely ever to be
available. Sandberg et al. (2005a,b) have provided recently a more detailed
criticism on the deficiencies of the three rules and why they cannot have a
general application to the prediction of PR and LR.
The three solutions that survive qualitative testing against
experimental evidence (i.e. Black et al., 1986 (and its derivative by NRC,
1998); Kyriazakis and Emmans, 1992a,b and de Greef and Verstegen,
1995) identify the marginal response in protein retention to protein and
energy intakes as the key variable to solving the problem. Black et al. (1986)
proposed that:
PR = b . (MEI – (c . MEm)) g/day (7.1)
where MEI is metabolizable energy intake, c is a constant and MEm is the
metabolizable energy requirement for maintenance. On the other hand,
Kyriazakis and Emmans (1992a,b) propose that:
PR = ep. (IP – IPm) g/day (7.2)
where IP is the ideal protein intake and IPm the ideal protein requirements
for maintenance. The solution offered by de Greef and Verstegen (1995)
has been shown by Emmans and Kyriazakis (1997) to be algebraically
equivalent to that offered by Kyriazakis and Emmans (1992b) and has a
relatively high information requirement. For these reasons, their solution
will not be considered further here. The question then is whether either of
the above two key parameters, the marginal response in PR to energy
supply (b) and the marginal response in PR to protein supply (ep) are
affected by the pig, the environment in which it is kept and the composition
of the food it is offered. These factors will be considered in turn below.
The Marginal Response in Protein Retention to Energy Supply
The solution of Black et al. (1986) was developed for energy limiting foods,
i.e. when MEI is less than required for PRmax. An implicit assumption was
that for such foods the value of b would be independent of food composition,
i.e. it would attain its maximum value. Below we review the effects of
liveweight, genotype including sex, and environmental temperature on the
marginal response in protein retention to energy supply on such foods.
The effect of liveweight
The experiment of Quiniou et al. (1995) where pigs of different liveweight

(range 45 to 94 kg) were given access to four levels of feeding, all of which
provided a constant high supply of protein, is shown in Fig. 7.1.
A model that fitted a common slope of PR against MEI (i.e. b) was not
statistically different from a model that allowed for different slopes at
different liveweights. In addition, the different slopes did not support a
systematic change in the value of b with increasing liveweight. This finding
is consistent with subsequent experiments performed by Quiniou et al.
(1996) and Mohn et al. (2000), who concluded that stage of growth had no
significant effect on b.
The contradictory evidence comes from the experiment of Dunkin and
Black (1985) who estimated values of b, for pigs of a range of liveweight (30
to 90 kg) fed eight levels of an energy limiting food. The values of b were
8.25, 6.44, 5.75 and 6.75, respectively and therefore they too do not
support a systematic effect of liveweight on the value of b. The latter would
be the necessary quality of a rule that aims to have a general applicability.
The evidence, taken as a whole therefore, is more consistent with the view
that the marginal response in protein retention on energy limiting foods
does not vary with pig liveweight.
210
190
Protein retention, PR (g/day)
170
150
130
110
90
70
5 7 9 11 13 15 17 19 21
ME intake (MJ/day)
Fig. 7.1. The response in protein retention (PR) to metabolizable energy intake (MEI) above
maintenance of pigs of four different live weights (Quiniou et al., 1995): these were 45 kg
(–䊉–), 65 kg (- -䊊- -), 80 kg (–䊏–) and 94 kg (- -ⵧ- -). The four levels of MEI were achieved by
four levels of feeding, all of which provided a constant high supply of protein. A common
slope has been fitted for the relationship between PR and MEI.
The effect of genotype, including sex
Two experiments have addressed the effect of genotype on the marginal

response in PR to energy supply. Kyriazakis et al. (1995) used two very
different breeds of pig, Chinese Meishan and F1 Large White Landrace,
and gave them access to a high protein basal food that was diluted with
starch to different extents. The authors concluded that the values of b were
very similar between the two breeds (9.65 (SE 0.16) and 9.93 (SE 0.55)),
respectively. The experiment of Quiniou et al. (1996) is more
comprehensive as three genotypes, boars and castrates of a Large White
Pietrain breed and castrates of a Large White breed were used. The
responses in PR to four levels of energy intake at a constant high protein
intake were considered at four different liveweights. The authors
concluded that the response in PR to increasing supplies of ME was
independent of liveweight, but differed between genotypes, when the
intercept of the response was fixed. The highest marginal response was
observed in the Large White Pietrain boars. Recently, Sandberg et al.
(2005b) reanalysed the data of this experiment by assuming that both the
intercept and slope of the response were allowed to vary. A model with a
common slope was not statistically different from the model where the
slopes were different between genotypes. Based on the above, the evidence
on the effect of genotype on b is inconclusive.
The effect of environmental temperature
In their proposals for the nutrient requirements for swine, NRC (1998)
suggested that the marginal response to energy intake, on protein-
adequate foods, falls as the temperature increases. The argument follows
from the experiment of Close et al. (1978), whose data are plotted in Fig.
7.2.
Although it is difficult to be certain that the foods used in the
experiment were limiting in energy, the data are far from persuasive that
the response varied with temperature. This is consistent with the view of
Black et al. (1986) and Wellock et al. (2003a), who propose that
environmental temperature has no effect on the marginal response in PR
to energy intake on protein adequate foods.
The Marginal Response in Protein Retention to Protein Supply
Unlike the previous solution, the rule proposed by Kyriazakis and Emmans
(1992a,b) is intended to apply across protein- and energy-limiting foods.
According to this rule, such foods are defined according to their ratio of
the metabolizable energy content (MEC, MJ/kg) to the digestible crude
300
250
200
PR (kJ/kg0.75)
150
100
50
0
0 500 1000 1500 2000 2500
MEI (kJ/kg0.75.day)
Fig. 7.2. The response in protein retention (PR kJ/kg0.75.day) to metabolizable energy intake
(MEI kJ/kg0.75.day) for pigs given different allowances of the same food at five different
temperatures from Close et al. (1978); 10°C (䊉), 15°C (䊊), 20°C (䉲), 25°C (䉮), and 30°C (䊏).
The regression line for all the data is PR = 0.147 (0.0062).MEI – 44.07 (7.58).
protein content (DCPC, kg/kg) of the food, R (MJ ME/kg DCP). Foods with
a value of R > 72.55 are defined as being protein limiting and the
marginal response in PR to protein supply on such foods, ep, is assumed to
attain its maximum value (ep)max. Foods that are defined as energy limiting
have an ep that varies according to R:
ep = µ.R (7.3)
The above rule makes the distinction between the protein and energy
dependent phases of the marginal response in PR to protein supply. The
marginal response in PR when both the energy and the protein allowances
offered to pigs are varied is shown in Fig. 7.3.
Thus, liveweight, genotype and environmental temperature have the
potential to affect both (ep)max in protein limiting foods and ep in energy
limiting foods.
The effect of liveweight
Experiments that address the effect of liveweight on the marginal response

in PR to protein supply have usually employed protein-limiting foods.
Most of these experiments (e.g. Black and Griffiths, 1975; Campbell et al.,
PRmax
Protein retention (g/day)

E3
E2
E1
ep
Protein intake (g/day)
Fig. 7.3. The predicted rates of protein retention of a pig given access to feeds of different
protein contents at different levels of feed (and hence energy intake). Feeding levels E1 and
E2 constrain protein retention by being energy limiting, whereas level E3 is protein limiting
and hence allows the animal to reach its maximum rate of protein retention, PRmax.
1985a; Mohn et al., 2000; and de Lange et al., 2001) conclude that the
maximum ep, around 0.75, is not affected by liveweight. The only
experiment where a liveweight effect has been reported is that of Campbell
and Dunkin (1983) who found a high response in N retention to N intake
for pigs between 1.8 and 6.5 kg liveweight. Their value of (ep)max is closer to
0.90 and it is perhaps a reflection of the difficulty in measuring N balance
in very small pigs. Emmans and Kyriazakis (1996) have measured the ep for
small (12 kg) and large (72 kg) pigs given access to both protein- and
energy-limiting foods. The mean values of ep are not different between the
two different liveweights on any of the foods used.
The effect of genotype, including sex
Kyriazakis et al. (1994) used entire male Large White Landrace and pure
bred Chinese Meishan pigs to investigate the effect of genotype on the
relationship between food composition and ep in foods of varying
energy:protein ratios, R. The value of ep was found to be directly proportional
up to a maximum value of R. The overall constant of proportionality, µ, was
0.0108 and did not differ significantly between the two breeds. The
maximum value of ep was also similar between the two breeds.
Other experiments by de Greef et al. (1992) and Fuller et al. (1995) also
conclude that the maximum ep does not differ between different genotypes.
There is also evidence that different sexes of pigs use a limiting protein
supply with similar efficiency, ep. Campbell et al. (1984, 1985b) found no
difference in maximum ep between entire males and females. The data of
Batterham et al. (1990) shown in Fig. 7.4 lend strong support to this
conclusion.
It would, therefore, appear safe to conclude from the above
experiments that the marginal response in PR to protein supply does not
differ between different genotypes, including sexes.
10
Lysine retention (g/day)

6
0
0 2 4 6 8 10 12 14 16 18
Ileal digestible lysine intake (g/day)
Fig. 7.4. Response in lysine retention (Rlys g/day) to ileal digestible lysine intake (Ilys g/day):
the regression line is Rlys = 0.763.(Ilys – 1.245) for the combined data of male (䊉) and female
(䊊) Large White (20–45 kg) pigs used by Batterham et al. (1990). Separate plateaux for
males and females are shown.
The effect of environmental temperature
The data of the experiment by Ferguson and Gous (1997), who grew pigs
fed ad libitum from 13 to 30 kg on food with 93–230 g crude protein/kg at
18, 22, 26 and 30°C, are reproduced in Fig. 7.5.
The data clearly show that the marginal response in protein retention
was not affected by environmental temperature. They are consistent with
the findings of Berschauer et al. (1983) and Campbell and Taverner (1988)
who conclude that temperature does not affect ep.
Predicting Protein and Lipid Retention of Healthy Pigs
The implication of the above is that, irrespective of whether PR is made

either a function of energy or a function of protein intake, the information
required to predict the rates of protein and lipid retention are low. The
marginal response in protein retention would be unaffected by pig
liveweight and genotype, and even by the environmental temperature.
Given this, the application of either framework would predict that pigs
selected, for example, for different levels of fatness when they have been
given access to a non-limiting food, would perform identically when they
are given access to the same amount of food above maintenance, which is
140
120

100
80
60
40
20
50 100 150 200 250 300 350 400
Crude protein intake (g/day)
Fig. 7.5. The response in protein retention to crude protein intake for pigs fed ad libitum
foods that were limiting in protein at four different temperatures: 18°C (䊉), 22°C (䊊),
26°C (䉲), and 30°C (䉮) as found by Ferguson and Gous (1997). The solid line is described
by PR = 0.525 (CPI – 4.92) until the plateau of 117.4 g/day is reached.
limiting for both genotypes. The same would apply for male and female
pigs, despite the fact that females have the propensity to be fatter than
males of the same protein weight when fed ad libitum. Here, it would be
important to emphasize that we need to account for potential differences
between the genotypes (or sexes) in their maintenance requirements. This
has not always been taken into account in the interpretation of
experiments that have measured PR responses to different levels of
protein and/or energy intake (e.g. Quiniou et al., 1996, see above).
A second implication is that both frameworks are capable of making
predictions across a wide range of conditions. These include conditions
where the pig is depositing protein at the expense of lipid retention. Such
conditions may arise in weaned pigs consuming small amounts of a
relatively high protein content food (Kyriazakis and Emmans, 1992a,b).
Many current solutions in the literature are still unable to predict this (e.g.
de Lange, 1995; Whittemore, 1995; Green and Whittemore, 2003).
The framework that predicts PR as a function of protein intake does so
on the basis of ideal protein. It is, therefore, an implicit assumption in this
framework that the marginal response in PR will be the same for all amino
acids, when they are first limiting. This assumption has recently been
challenged by Heger et al. (2002, 2003), but there is considerable
uncertainty over their estimates of maximum efficiency of amino acid
retention (with values ranging from 1.17 to 0.66 for different amino acids).
For the time being it may be safe to assume a single overall efficiency for all
amino acids, whichever is the first limiting (Sandberg et al., 2005b).
Whilst the position that the marginal response in protein retention
may only be affected by food composition may be attractive from a
nutritional modelling point of view, it can be viewed as having unattractive
perspectives (Luiting and Knap, 2005). Animal breeders, interested in
genetic variation between individuals, have criticized it on the basis that it
does not ‘accommodate a genotype-specific drive towards body fatness as
well as a drive towards protein deposition’. The task is always to keep the
framework variables to the necessary minimum in order to lead to
tractable solutions and make predictions for populations of pigs. This
becomes more important when the framework is broadened to account, for
example, for the partitioning of nutrients during disease.
The Partitioning of Scarce Protein in Pigs Challenged by

Pathogens
When a pig that has not been previously exposed to a pathogen, i.e. an
immunologically naive pig, encounters the pathogen for the first time, it
may require nutrients for functions that will enable it to cope with the
challenge. Such functions may include the innate immune response, which
is one of the first lines of defence to pathogens, and the repair and
replenishment of damaged or lost tissue, such as blood plasma or cells.
Eventually, the pig will be expected to develop an acquired immune
response towards the pathogen, and nutrient resources will need to be
directed towards the maintenance of this function. When the pig is re-
exposed to the same kind of pathogen, the main additional resource
requirement would be due to the function of acquired immunity, as
investment towards this function would minimize the potential pathogen-
associated damage to the host.
When nutrient resources are scarce, the challenged pig can be seen as
having the problem of allocating these resources between its various
functions. These arise from the exposure to the pathogen, but they also
include the ‘normal’ functions of a healthy pig, such as maintenance and
growth. In this chapter, we concentrate upon the problem of allocating
scarce protein during exposure to pathogens. This is because: (i) current
evidence suggests that protein is often the first limiting resource in
challenged pigs; (ii) many components of the immune response are highly
proteinaceous (Houdijk et al., 2001); and (iii) as a consequence, most
evidence in the literature is in relation to the partitioning of protein or
amino acids during challenge. There is some limited evidence that energy
may become a limiting resource during pathogen challenge, as a
consequence of the increased energy requirements due to fever and the
requirements of the immune response. It is generally accepted, however,
that these requirements are relatively small for growing animals (chickens,
Klasing et al., 1987; mice, Demas et al., 1997; pigs, van Heugten et al., 1996).
The partitioning of scarce protein during pathogen exposure in naive pigs
Exposure to pathogens is usually accompanied by a voluntary reduction in

food intake in naive hosts, Fig. 7.6. The extent of anorexia is dependent
on pathogen dose, i.e. the number of pathogens that enter the host. There
is, however, a wide range of doses over which animals show a reduction in
the order of 15–20% of their normal voluntary food intake: this range of
dose is often that associated with sub-clinical disease. Higher pathogen
doses may lead to clinical disease and catastrophic reductions in voluntary
food intake. Currently there are at least two proposed mechanisms that
may lead to this pathogen-induced anorexia: (i) food intake is reduced
because the potential for growth (protein retention) of the challenged
animal is reduced (Wellock et al., 2003b); and (ii) the reduction in food
intake is a direct consequence of the exposure to the pathogen (Kyriazakis
et al., 1998). Mechanism (ii) implies that the animal will always be in a state
of nutrient (protein) scarcity during exposure to pathogens.
As stated previously, a naive pig exposed to a pathogen would be
expected to divert resources towards the functions of innate immunity and
repair; some resources will also have to be diverted towards the acquisition
of immunity. These increased requirements appear only as increases in
maintenance requirements, in experiments where naive pigs have been
exposed to a variety of antigenic challenges, including pathogens (Webel et
al., 1998a,b). In addition, the marginal responses in protein retention to
either protein or energy supply do not seem to be affected by exposure to
pathogens in naive pigs (Van Dam et al., 1998; Webel et al., 1998a,b). This
implies that the function of innate immunity, at least, is prioritized over the
function of growth in terms of nutrient allocation. The above further imply
that the framework developed to predict protein retention for healthy pigs
needs to be modified only slightly for naive pigs challenged by pathogens
Food
intake
(g/day)
Sub- Clinical
clinical
0
0 Pathogen dose or level
Fig. 7.6. A proposed schematic description of the effect of dose or level of parasites on the
daily rate of food intake by the host, reproduced from Kyriazakis et al. (1998). Food intake is
reduced once a threshold pathogen dose (or level) is reached which leads to sub-clinical
disease. Reductions in food intake become very severe at higher pathogen doses or levels
that lead to clinical disease.
(Eqn 7.2). The modification would be due to an increase in IPm in order to

account for the requirements of the innate immune response. This
suggestion assumes that the composition of the ideal protein required by
the innate immune response is similar to that required for maintenance.
Given current evidence that the requirements for the innate immune
response may be relatively small (Reeds et al., 1994), this position is a
tenable one.
However, both requirements for the functions of innate immunity and
repair are expected to be a function of pathogen challenge. This includes
the pathogen kind and level, which may be described as either pathogen
dose or load. The latter is the number of pathogens that establish, replicate
within the host and affect its function and metabolism. Although most
investigations usually refer to pathogen dose, there are now good models
that translate this into pathogen load for both pathogens that replicate
within the host (i.e. bacteria (Wilson and McElwain, 2004) and viruses
(Bocharov, 1998; Nowak et al., 1996)) or not (i.e. macroparasites, Louie et
al., 2005). Although relevant experiments that have investigated the
relationship between pathogen load and requirements for innate immunity
and repair have not been reported in the literature for pigs, evidence from
other species (Taylor-Robinson, 2000) suggests that they can be
represented as follows:
PIIm = ai. (PL – PL0) g/day (7.4)
where PIIm is the requirement for the innate immune response, PL (n) is
pathogen load and PL0 (n) is the minimal PL required to activate the
innate immune system. The PIIm is the requirement to reduce PL to PL0.
In cases where PL ≤ PL0 the animal would not be expected to require any
resources for the innate immune response, perhaps because the small
pathogen load does not warrant the effort. It is also expected that a
genetically determined maximum (PIIm)max exists, which denotes the
maximum capacity for response (rather than the maximum requirement)
and, therefore, may be genotype and size dependent. The above simple
linear relationship, which represents the response of a single pig, will lead
to a sigmoidal relationship for the response of a population of pigs (Fisher
et al., 1973; Pomar et al., 2003). The values of ai and PLo are expected to be
both pathogen and genotype specific.
A widely accepted relationship between PL and damage caused is that
proposed by Behnke et al. (1992):
Ploss = Prep = ek.PL – 1 g/day (7.5)
where Ploss is the amount of protein lost or damaged due to pathogen
exposure, and is equivalent to the amount of protein required for repair
(Prep). This protein loss may be actual tissue damage, such as
gastrointestinal tract (Yu et al., 2000) or blood constituent loss, including
plasma (Yakoob et al., 1983; Le Jambre, 1995). The expectation is that
given the available resources the pig will attempt to repair the ‘damage’
caused by the pathogen as fast as possible and return to a ‘healthy’ state.
Parameter k would be pathogen specific and may also reflect the virulence
of the pathogen. The exponential form of the relationship is consistent
with the view that small pathogen loads have little effect on hosts and that
an already weakened system would suffer more damage from increases in
its challenge (i.e. further exposure to pathogens). It should be emphasized
that there would be a (Ploss)max value equivalent to the amount of damage
that may lead to death.
The prediction of partitioning of scarce protein in naive pigs during
exposure to pathogens appears to be straightforward. Scarce protein
would be expected to be prioritized towards the functions of maintenance
and innate immune response; any protein above this should be used for
growth, after the protein needs for repair have been accounted for. The
parameterization of Eqns (7.4) and (7.5) should be relatively
straightforward to estimate for specific pathogens, providing that during
the course of the measurement the animal does not develop an acquired
immune response to the pathogen. As the development of the acquired
immune response may be very rapid, current information presented in the
literature where PR is measured over a number of days usually includes
both the phase of innate and acquired immune responses. This makes
interpretation of current experiments difficult (see below).
The partitioning of scarce protein during pathogen exposure in immune pigs
Information on the effect of pathogen (re-)challenge on the food intake of

already immune animals is significantly scarcer than for naive animals. The
very few experiments that have investigated the phenomenon (Takhar and
Farrell, 1979; Greer et al., 2005) suggest that re-exposure to pathogens
does not seem to be accompanied by anorexia. This raises the possibility,
yet untested, that re-exposed, immune animals may increase food intake in
order to meet their increased requirements due to exposure to a
pathogen. However, as maintenance of immunity is a dynamic
phenomenon (Anderson, 1994) and may be lost if hosts do not continue to
be exposed to a pathogen, it is possible that anorexia may reappear in
hosts that have lost their acquired immunity.
In parallel to the argument presented above for the requirements of
the innate immune response, the requirements for adaptive immunity can
be expressed as:
PAIm = aa. (PL – PL0) g/day (7.6)
where PAIm is the requirement for the acquired immune response, which
is also expected to attain a genetically determined maximum (PAIm)max at
a certain value of PL, when PL PL0. PAIm is the requirement to reduce
PL to PL0. However, an increase in the requirements for acquired
immunity does not manifest as an increase in the maintenance
requirements of immune pigs when they are re-exposed to a pathogen and
given access to scarce protein intake. There are now several experiments in
the literature to suggest that acquired immunity responds to increases in

protein intake, whilst at the same time animals increase protein retention
(pigs, Zijlstra et al., 1997; sheep, van Houtert et al., 1995; chickens,
Bhargava et al., 1971). This implies that a degree of competition for scarce
protein above maintenance exists between acquired immunity and protein
retention (growth). From an evolutionary point of view this is an attractive
proposition, since it would be consistent with the position that the animal
tries to grow as fast as it can, whilst it maintains some degree of acquired
immunity, which allows it to exert a degree of control over its pathogens
and their consequences (Coop and Kyriazakis, 1999). This is equivalent to
optimizing nutrient partitioning between long term (attainment of
reproductively mature size) and short term objectives (survival).
The above also implies that the simplicity of the framework developed
to account for the partitioning of protein in naive pigs, cannot be
maintained for immune pigs. Instead, one should be considering how to
represent the partitioning of scarce protein above maintenance between
the two competing functions of protein growth and acquired immunity.
Here, it is proposed that this partitioning may be represented by a
partitioning ratio, p, which is the amount directed towards the function of
acquired immunity and is expected to be a function of pathogen load. It
may attain a value of 0 < p < 0.5. This is because growth seems to have a
higher relative priority than the immune response, and hence is penalized
to a lesser extent during protein scarcity. The value of p is expected to
increase linearly as pathogen dose and/or virulence increase:
p = d. (PL – PLo) (7.7)
Any amount of protein invested in acquired immunity is expected to lead
to a reduction in pathogen load in the manner:
PL1 = PL – (PL. (PAAIm/PAIm)) n (7.8)
where PL1 and PL are pathogen loads post- and pre-immunity effects.
PAAIm is protein actually invested in immunity and may be less than
or equal to PAIm (as defined in Eqn (7.6)). In the latter case, when
PAAIm = PAIm, then PL1 = 0. If, on the other hand, PAAIm = 0 then
pathogen load would be unaffected, i.e. PL1 = PL. The new pathogen load,
PL1, can be used to calculate the actual damage achieved by a pathogen in a
pig that expresses acquired immunity (Eqn 7.5). Equation (7.8) can be
substituted into Eqn (7.5) to represent the relationship between damage
caused by a given pathogen load and the relative investment in acquired
immunity (PAAIm/PAIm), as shown in Fig. 7.7, given that:
Prep = ek.PL1 – 1 = ek. (PL – (PL. (PAAIm/PAIm))) – 1 g/day (7.9)
The resulting negative exponential relationship between damage and host
response is in agreement with the relationships proposed to hold for
common microbial pathogens (Casadevall and Pirofski, 1999).
If it is assumed that the efficiency of ideal protein use for acquired
immunity is the same as that for PR (i.e. ep), then the above framework can
be brought together to predict the PR of an animal exposed to a certain

pathogen dose (and hence PL) whilst given different amounts of protein,
Fig. 7.8.
The predictions are made over one particular time step (day). As ideal
protein intake increases, the amount of protein actually invested in
acquired immunity (PAAIm) is described by:
PAAIm = ep.p. (IPI – IPm) g/day (7.10)
Maximum
damage
Actual
damage
Level of investment in immunity
Fig. 7.7. The proposed relationship between the actual damage caused by a pathogen load
to a host in relation to the level of investment in immunity (see Eqn 7.9 in text). The maximum
damage caused by the pathogen load is achieved when there is no investment in immunity.
200
180 PRmax
160
140
120
100
80
60
40
20
0
0 50 100 150 200 250
Ideal protein intake (g/day)
Fig. 7.8. Predictions of protein retention in relation to different intakes of ideal protein for an
uninfected (solid line) and an immune pig challenged by a pathogen (dotted line). The increase
in the intercept for the challenged pig is a reflection of the costs associated with the pathogen,
when there is no investment of protein intake in immunity. Both healthy and challenged pigs
are expected to attain their genetically defined maximum protein retention, PRmax.
The increase in investment up to a maximum PAIm, predicts that the rate

of PR is brought closer to that of a healthy pig, due to the associated gains
(i.e. reduction) in the amount needed to be invested for repair. Further
investment in immunity is not associated with significant gains, and this is
reflected in the decline in the rate of PR. Finally, it is assumed that the
maximum capacity for protein retention (PRmax) is the same between
healthy and immune pigs challenged by a pathogen. As discussed
previously, the above predictions cannot be compared directly to actual
experiments, mainly because what is reported in the literature usually
aggregates data on both the naive and immune status in the same animal,
as well as the transition between the two phases.
Towards the Prediction of Protein and Lipid Retention in Pigs

Challenged by Pathogens
The use of the concept of ideal protein
The above framework was developed on the basis of ideal protein intake.
This assumes that the composition of the protein retained as either the
immune response or growth is similar. In Table 7.2 the amino acid
composition of some proteinaceous components of the immune response is
compared to the composition of pig body protein, which forms the basis of
the ideal protein system (ARC, 1981). The difference in amino acid
composition between these two body components is striking. Such
differences will be expected to have a significant effect on the prediction of
protein retention, if the contribution of the immune response to the
overall protein retention is relatively high (Wang and Fuller, 1989). In this
case, the framework will need to be modified to make predictions on the
basis of individual amino acid responses. Protein retention would then be
reconstituted on the basis of individual amino acid retention. This solution
will have exceedingly high information requirements and, therefore,
parameterization of the framework will be exceedingly difficult. An added
complication would also arise from the fact that stimulation by different
pathogens invokes different effector mechanisms of the immune response
(Dong and Flavell, 2001). These different effector mechanisms may have
different amino acid compositions (Table 7.2), which in turn may be
utilized by different efficiencies.
An alternative, but indirect solution to the problem that arises from
the above would be to retain the ideal protein system, but assume that
the efficiency with which protein is utilized for the purposes of the
immune response is modified. This would be in order to account for the
different ‘ideal’ protein composition of the two body components. This
seems to be a less onerous task than the above and experiments that can
be designed to contribute towards the parameterization of this solution
can be envisaged.
Table 7.2. A summary of the amino acid composition of different effector proteins associated
with the immune response, in relation to the reference pig body protein that is normally used
for calculating the biological value of a food (ideal protein).
Amino acid composition (g/kg protein)a
Average of
Pig protein 7 APPs IgA IgE MCP Mucin
Phe 36.7 71.8 26 30 29 15

Tyr 25.6 57.3 24 43 29 13
Trp 9.1 32.5 26 20 8 –
Leu 69.5 64.8 120 87 90 31
Ile 31.9 43.2 13 39 69 15
Val 44.2 55.3 96 80 73 222
Lys 64.2 73.3 39 57 53 18
Hist 28.0 28.3 11 18 29 –
Met 19.6 20.8 6 7 33 –
Thr 36.9 57.0 75 103 45 224
a Amino acid compositions for pig protein from average value calculated by Sandberg et al.
(2005), acute phase proteins from Reeds et al. (1994) and for IgA, IgE, sheep mast cell
proteases (MCP) and mucin were taken from Houdijk and Athanasiadou (2003).
The genetic ability to cope with pathogens
In the previous sections, several of the parameters identified in the

framework were proposed to be genotype dependent. These were the
minimum pathogen load that activates the immune system (Lo), the rates of
increase in the requirements for the innate (ai ) and acquired (aa) immune
responses, and the maximum capacities for these responses (PIIm and
PAIm, respectively). Pig genotypes that may be defined as resistant or
susceptible in terms of how they cope with a pathogen challenge (Knap
and Bishop, 2000), would be expected to differ in the values of these
parameters. It is likely that a degree of correlation exists between such
parameters. For example, a resistant pig genotype may have a lower
threshold for its immune system activation (PLo), whereas its maximum
capacity to cope with a pathogen may be reached at a higher pathogen
load. Thus, the additional parameter requirements for describing the host
genotype would be reduced if the correlation between parameter values as
affected by host resistance is known.
The idea of the lower PLo threshold for resistant genotypes may be
extended to the relationship between the partitioning ratio, p, and PL (Fig.
7.9). The consequence then would be that, with other things being equal,
resistant genotypes would be directing more nutrients towards the
acquired immune response than susceptible genotypes at a given pathogen
load (Kreukniet and van der Zijpp, 1989).
If the rate of increase in the value of p is also higher for the resistant
genotypes, then even more nutrients would be expected to be directed
PL0
Pathogen load
Fig. 7.9. The proposed relationship between the level of pathogen challenge and the
partitioning ratio of scarce resources (denoted by p) between growth and immune functions,
once the level of challenge has exceeded a threshold, PL0, in a susceptible pig (dotted line)
and resistant pigs (solid line and dashed lines). The two resistant pigs differ in the rate of
increase in the value of p for a given change in pathogen load.
towards the immune response at a given pathogen load. This could, in

part, account for the observed interactions between nutrition and host
genotype in their ability to cope with pathogens (Stewart et al., 1969; Rao et
al., 2003; Haile et al., 2004).
When nutrient (protein) intake is scarce, the pig genotype may be
described only according to the above parameters. As stated previously,
when protein intake is scarce there is no variation between individuals in
their drive towards body fatness and protein deposition. In these
circumstances, any differences in performance would be attributed to
differences in the above five parameters. Animal breeders should be
encouraged to describe pig genotypes accordingly, if progress is to be made
towards the prediction of the performance of pigs exposed to pathogens.
The continuum between naive and immune states
The above framework predicts the partitioning of scarce resources for

either naive or immune pigs exposed to pathogens at one point in time. As
such, there is an artificial distinction between these two states; this was
considered as the necessary first step for the purposes of the framework.
However, in most cases the transition between the two states is not abrupt
but continuous (Anderson, 1994), as they are linked by the phase of
acquisition of immunity. Therefore, in order to progress towards dynamic
predictions of protein and lipid retention one would need to account for
this transition over time.
There are currently a number of simulation models that predict
adequately the acquisition of immunity and make its onset a function of
pathogen load (for example, Barnes and Dobson (1993) for gastrointestinal
parasites, and Bocharov (1998), for viruses). In other words, how quickly
the animal starts to acquire immunity depends, within certain limits, upon
its pathogen load. This is assumed to be a reflection of the PL required to
stimulate the acquired immune response (see above). However, the
duration of the phase of acquisition of immunity is assumed to be
independent of pathogen load (Steel et al., 1980; Houdijk et al., 2005), and
only dependent on pathogen type. Duration of this phase can be very short
in certain pathogens, such as bacterial (Turner et al., 2002a,b) and viral
(Zijlstra et al., 1997) challenges.
The above models also assume that the duration of the acquisition of
immunity is independent of host nutrition. Sandberg et al. (2006) reviewed
the literature on this issue and concluded that whilst this variable may be
affected by food composition (food protein content in particular), no
general relationship could be proposed between the two. As such, it was
suggested that until experiments that address this issue are performed, the
above simple assumption should be retained. A simulation of how
pathogen load changes within the host as a function of time based on the
above assumptions is shown on Fig. 7.10. The change in pathogen load,
assuming normal acquisition and expression of immunity, is consistent with
experimental findings for bacterial (e.g. Kelly et al., 1996) and viral (e.g.
Bocharov, 1998; McDermott et al., 2004) pathogen challenges.
The effect of nutrition on pathogen load as shown in Fig. 7.10 is the
outcome of the proposed framework.
Future Directions
The problem of predicting partitioning of absorbed scarce resources to PR

and LR in healthy pigs has occupied animal scientists for over 30 years. We
Pathogen load
Time
Fig. 7.10. A prediction of the change in pathogen load (arbitrary units) over time of animals
given access to different levels of protein supply low (dashed line), medium (solid line) and
high (dotted line). No difference is observed in the early stage of infection whilst the animal
acquires immunity; however, once expression of acquired immunity commences, its level of
expression is proposed to be affected by the level of resource supply.
believe that the current solution offered in this chapter is capable of

predicting these responses in a variety of conditions and for pigs of
different genotypes. Experiments are warranted to fine-tune the solution
and these have been defined above. These, however, would not diminish
the strength, general applicability and heuristic value of the position
offered.
The solution offered for healthy pigs formed the basis of a framework
that was developed to predict the partitioning of absorbed scarce
resources to PR and LR in pigs challenged by pathogens. Unlike the
framework developed for healthy pigs, the completion of the framework
for ‘diseased’ pigs presents us with significant challenges. The information
requirements that would enable us to progress towards quantitative,
dynamic predictions from this framework have been described above. In
this chapter we purposely did not specify (describe in quantitative terms)
the pathogen challenge under consideration, as our aim was to develop a
generic framework. Parameterization of this framework may then be
possible by focusing on the characterization of specific pathogens and
their consequences, including disease. We considered this to be a more
fruitful approach than the creation of a model to account for the
consequences of exposure to a specific pathogen (e.g. the approach taken
by Black et al. (1999) to predict the consequences of pleuropneumonia in
pigs).
As reliance on chemoprophylaxis to control pathogens is decreasing,
due for example to consumer concerns or legislation (Waller, 1997; Olesen
et al., 2000), interest in the understanding of the performance of animals in
the presence of pathogens will increase. A framework that predicts the
performance of pigs during exposure to pathogens may then have a value
as a management tool to develop strategies, including breeding and
nutritional strategies, to deal with this challenge.
Acknowledgements
This work was in part funded by the Biotechnology and Biological Science
Research Council of the UK and PIC/Sygen. SAC receives support from the
Scottish Executive, Environment and Rural Affairs Department. We are
grateful to our colleagues Dimitris Vagenas, Andrea Doeschl-Wilson and
Will Brindle for comments on earlier versions of the manuscript and to
everyone in the Animal Nutrition and Health Department for their support
in this activity.
References
Anderson, R.M. (1994) Mathematical studies of parasitic infection and immunity.
Science 264, 1884–1886.
ARC (1981) The Nutrient Requirements of Pigs. Agricultural Research Council,
Commonwealth Agricultural Bureaux, Slough, UK.
Barnes, E.H. and Dobson, R.J. (1993) Persistence of acquired immunity to

Trichostrongylus colubriformis in sheep after termination of infection. International
Journal of Parasitology 23, 1019–1026.
Batterham, E.S., Andersen, L.M., Braigent, D.R. and White, E. (1990) Utilisation of
ileal digestible amino acids by growing pigs: effect of dietary lysine
concentration on efficiency of lysine retention. British Journal of Nutrition 64,
81–94.
Behnke, J.M., Barnard, C.J. and Wakelin, D. (1992) Understanding chronic
nematode infections: evolutionary considerations, current hypotheses and the
way forward. International Journal of Parasitology 22, 861–907.
Berschauer, F., Close, W.H. and Stephens, D.B. (1983) The influence of protein:
energy value of the ration and level of feed intake on the energy and nitrogen
metabolism of the growing pig. 2. N metabolism at two environmental
temperatures. British Journal of Nutrition 49, 271–283.
Bhargava, K.K., Hanson, R.P. and Sunde, M.L. (1971) Effects of threonine on
growth and antibody production in chicks infected with Newcastle virus.
Black, J.L. and Griffiths, D.A. (1975) Effects of live weight and energy intake on
nitrogen balance and total N requirement of lambs. British Journal of Nutrition
33, 399–413.
Black, J.L., Davies, G.T., Bray, H.J., Giles, L.R. and Chapple, R.P. (1986)
Black, J.L., Bray, H.J. and Giles, L.R. (1999) The thermal and infectious
environment. In: Kyriazakis, I. (ed.) A Quantitative Biology of the Pig. CAB
Bocharov, G.A. (1998) Modelling the dynamics of LCMV infection in mice:
Conventional and exhaustive CTL responses. Journal of Theoretical Biology 192,
283–308.
Campbell, R.G. and Dunkin, A.C. (1983) The effects of energy intake and dietary
protein on nitrogen retention, growth performance, body composition and
some aspects of energy metabolism of baby pigs. British Journal of Nutrition 49,
221–230.
Campbell, R.G. and Taverner, M.R. (1988) Relationships between energy intake
and protein and energy metabolism, growth and body composition of pigs
kept at 14 or 32 degree C from 9 to 20 kg. Livestock Production Science 18,
289–303.
Campbell, R.G., Taverner, M.R. and Curic, D.M. (1984) Effect of feeding level and
dietary protein content on the growth, body composition and rate of protein
deposition in pigs growing from 45 to 90 kg. Animal Production 38, 233–240.
Campbell, R.G., Taverner, M.R. and Curic, D.M. (1985a) The influence of feeding
level on the protein requirement of pigs between 20 and 45 kg live weight.
Campbell, R.G., Taverner, M.R. and Curic, D.M. (1985b) Effects of sex and energy
intake between 48 – 90 kg live weight on protein retention in growing pigs.
Casadevall, A. and Pirofski, L.A. (1999) Host-pathogen interactions: Redefining the
basic concepts of virulence and pathogenicity Infection and Immunity 67,
3703–3713.
Close, W.H., Mount, L.E. and Brown, D. (1978) The effects of plane of nutrition
and environmental temperature on the energy metabolism of the growing pig.
2. Growth rate, including protein and fat deposition. British Journal of Nutrition
40, 423–431.
Coop, R.L. and Kyriazakis, I. (1999) Nutrition-parasite interaction. Veterinary
Parasitology 84, 187–204.
de Greef, K.H. and Verstegen, M.W.A. (1995) Evaluation of a concept on energy
partitioning in growing pigs. In: Moughan, P.J., Verstegen, M.W.A., Visser-
Reyneveld, M.I. (eds) Modelling Growth in the Pig. EAAP Publication No. 78.
Wageningen Press, Wageningen, Netherlands, pp. 137–149.
de Greef, K.H., Kemp, B. and Verstegen, M.W.A. (1992) Performance and body
composition of fattening pigs of two strains during protein deficiency and
subsequent realimentation. Livestock Production Science 30, 141–153.
de Lange, C.F.M. (1995) Framework for a simplified model to demonstrate principles
of nutrient partitioning for growth in the pig. In: Moughan, P.J., Verstegen,
M.W.A., Visser-Reyneveld, M.I. (eds) Modelling Growth in the Pig. EAAP
Publication No. 78. Wageningen Press, Wageningen, Netherlands, pp. 71–85.
de Lange, C.F.M., Gillis, A.M. and Simpson, G.J. (2001) Influence of threonine
intake on whole-body protein deposition and threonine utilisation in growing
pigs fed purified diets. Journal of Animal Science 79, 3087–3095.
Demas, G.E., Chefer, V., Talan, M.I. and Nelson, R.J. (1997) Metabolic cost of
mounting an antigen-stimulated immune response in adult and aged
C57BL/6J mice. American Journal of Physiology 273, R1631–R1637.
Dong, C. and Flavell, R.A. (2001) Th1 and Th2 cells. Current Opinion in Hematology
8, 47–51.
Dunkin, A.C. and Black, J.L. (1985) The relationship between energy intake and
nitrogen balance in the growing pig. In: Moe, P.W., Tyrrell, H.F. and Reynolds,
P.J. (eds) Energy Metabolism of Farm Animals: Proceedings of the 10th Symposium.
EAAP Publication No 32. Rowman and Littlefield, Totowa, N, pp. 110–113.
Emmans, G.C. and Kyriazakis, I. (1996) The effect of live weight on the efficiency
of using ideal protein for protein retention. Proceedings of the British Society of
Animal Science, pp 13.
Emmans, G.C. and Kyriazakis, I. (1997) Models of pigs growth: problems and
proposed solutions. Livestock Production Science 51, 119–129.
Emmans, G.C. and Kyriazakis, I. (2001) Consequences of genetic change in farm
animals on food intake and feeding behavior. Proceedings of the Nutrition Society
60, 115–125.
Ferguson, N.S. and Gous, R.M. (1997) The influence of heat production on
voluntary food intake in growing pigs given protein-deficient diets. Animal
Science 64, 365–378.
Science 14, 469–484.
Fuller, M.F. and Crofts, R.M.J. (1977) The protein-sparing effect of carbohydrate.
1. Nitrogen retention of growing pigs in relation to diet. British Journal of
Nutrition 38, 479–488.
Fuller, M.F., Franklin, M.F., McWilliam, R. and Pennie, K. (1995) The responses of
growing pigs, of different sex and genotype, to dietary energy and protein.
Green, D.M. and Whittemore, C.T. (2003) Architecture of a harmonized model of
requirements and of excretions into the environment (IMS Pig). Animal Science
77, 113–130.
Greer, A.W., Stankiewicz, M., Jay, N.P., McNulty, R.W. and Sykes, A.R. (2005) The
effect of concurrent corticosteriod induced immuno-suppression and infection
with the intestinal parasite Trichostrongylus colubriformis on food intake and
utilisation in both immunologically naïve and competent sheep. Animal Science
80, 89–99.
Haile, A., Anindo, D.O., Tembley, S., Mukasa-Mugerwa, E., Tibbo, M., Yami, A.,
Baker, R.L. and Rege, J.E.O. (2004) Effects of dietary protein supplementation
and infection with gastrointestinal nematode parasites on some nutritional and
metabolic parameters in Ethiopian Menz and Horro sheep Livestock Production
Science 91,183–195.
Heger, J., van Phung, T. and Krizova, L. (2002) Efficiency of amino acid utilisation
in the growing pig at sub-optimal levels of intake: lysine, threonine, sulphur
amino acids and tryptophan. Journal of Animal Physiology and Animal Nutrition
86, 153–165.
Heger, J., van Phung, T., Krizova, L., Sustala, M. and Simecek, K. (2003) Efficiency
of amino acid utilisation in the growing pig at sub-optimal levels of intake:
branched chain-chain amino acids, histidine and phenylalanine + tyrosine.
Journal of Animal Physiology and Animal Nutrition 87, 52–65.
Houdijk, J.G.M. and Athanasiadou, S.L. (2003) Direct and indirect effects of host
nutrition on ruminant gastrointestinal nematodes. In: t’Mannetje, L., Ramirez-
Aviles, L., Sandoval-Castro, C.A. and Ku-vera, J.C. (eds) Proceedings of the VI
International Symposium on the Nutrition of Herbivores: Matching Herbivore Nutrition
to Ecosystems Biodiversity. Merida, Mexico.
Houdijk, J.G.M., Jessop, N.S. and Kyriazakis, I. (2001) Nutrient partitioning
between reproductive and immune functions in animals. Proceedings of the
Nutrition Society 60, 515–525.
Houdijk, J.G.M., Anderson, D.H. and Kyriazakis, I. (2005) Induction of
experimental sub-clinical post-weaning colibacillosis in pigs. Proceedings of the
British Society of Animal Science 47, 84.
Kelly, B.P., Furney, S.K., Jessen, M.T. and Orme, I.M. (1996) Low-dose aerosol
infection model for testing drugs for efficacy against Mycobacterium tuberculosis.
Antimicrobial Agents and Chemotherapy 40, 2809–2812.
Klasing, K.C., Laurin, D.E., Peng, R.K. and Fry, M. (1987) Immunologically
mediated growth depression in chicks: Influence of feed intake, corticosterone
and interleukin-1. Journal of Nutrition 117, 1629–1637.
Knap, P.W. and Bishop, S.C. (2000) Relationships between genetic change and
infectious disease in domestic livestock. Proceedings of the British Society of Animal
Science: Occasional Publication 27, 65–80.
Kreukniet, M.B. and van der Zijpp, A.J. (1989) Effects of different doses of sheep
erythrocytes on the humoral immune response of chicken lines selected for
high or low antibody response. Poultry Science 69, 608–614.
Kyriazakis, I. and Emmans, G.C. (1992a) The effects of varying protein and energy
intakes on the growth and body-composition of pigs. 1. The effects of energy-
intake at constant, high protein-intake. British Journal of Nutrition 68, 603–613.
Kyriazakis, I. and Emmans, G.C. (1992b) The effects of varying protein and energy
intakes on the growth and body-composition of pigs. 2. The effects of varying
both energy and protein-intake. British Journal of Nutrition 68, 615–625.
Kyriazakis, I., Dotas, D. and Emmans, G.C. (1994) The effect of breed on the
relationship between feed composition and the efficiency of protein utilisation
in pigs. British Journal of Nutrition 71, 849–859.
Kyriazakis, I., Emmans, G.C. and with the technical assistance of Anderson, D.H.
(1995) Do breeds of pig differ in the efficiency with which they use a limiting
protein supply? British Journal of Nutrition 74, 183–195.
Kyriazakis, I., Tolkamp, B.J. and Hutchings, M.R. (1998) Towards a functional
explanation for the occurrence of anorexia during parasitic infections. Animal
Behaviour 56, 265–274.
Le Jambre, L.F. (1995) Relationship of blood loss to worm numbers, biomass and
egg production in Haemonchus infected sheep. International Journal of
Parasitology 25, 269–273.
Louie, K., Vlassoff, A. and Mackay, A. (2005) Nematode parasites of sheep:
extension of a simple model to include host variability. Parasitology 130,
437–446.
Luiting, P. and Knap, P.W. (2005) Comparison of pig growth models – the genetic
point of view. Recent Advances in Pig and Poultry Modelling. Kwazulu-Natal,
South Africa.
McDermott, A.B., Mitchen, J., Piaskowski, S., Souza, I., Yant, L.J., Stephan, J.,
Furlott, J. and Watkins, D.I. (2004) Repeated low-dose mucosal simian
immunodeficiency virus SIVmac239 challenge results in the same viral and
immunological kinetics as high-dose challenge: a model for the evaluation of
vaccine efficacy in nonhuman primates. Journal of Virology 78, 3140–3144.
Mohn, S., Gillis, A.M., Moughan, P.J. and de Lange, C.F.M. (2000) Influence of
dietary lysine and energy intakes on body protein deposition and lysine
utilisation in the growing pig. Journal of Animal Science 78, 1510–1519.
Nowak, M.A., Bonhoeffer, S., Hill, A.M., Boehme, R., Thomas, H.C. and McDade,
H. (1996) Viral dynamics in hepatitis B virus infection. Proceedings of the
National Academy of Sciences 93, 4398–4402.
NRC (1998) Nutrient Requirements of Swine, 10th edn. National Academy Press,
Washington, DC.
Olesen, I., Groen, A.F. and Gjerde, B. (2000) Definition of animal breeding goals
for sustainable production systems. Journal of Animal Science 78, 570–582.
Pomar, C., Kyriazakis, I., Emmans, G.C. and Knap, P.W. (2003) Modeling
stochasticity: Dealing with population rather than individual pigs. Journal of
Animal Science 81, E178–E186.
Quiniou, N., Noblet, J., van Milgen, J. and Dourmad, J-Y. (1995) Effect of energy
intake on performance, nutrient and tissue gain and protein and energy
utilisation in growing boars. Animal Science 61, 133–143.
Quiniou, N., Dourmad, J-Y. and Noblet, J. (1996) Effect of energy intake on the
performance of different types of pig from 45 – 100 kg body weight. 1. Protein
and lipid deposition. Animal Science 63, 277–288.
Rao, S.V.R., Praharaj, N.K., Reddy, M.R. and Panda, A.K. (2003) Interaction
between genotype and dietary concentrations of methionine for immune
function in commercial broilers. British Poultry Science 44, 104–112.
Reeds, P.J., Fjeld, C.R. and Jahoor, F. (1994) Do the differences between the amino
acid compositions of acute-phase and muscle proteins have a bearing on
nitrogen loss in traumatic states? Journal of Nutrition 124, 906–910.
Sandberg, F.B., Emmans, G.C. and Kyriazakis, I. (2005a) Partitioning of limiting
protein and energy in the growing pig: description of the problem, possible
rules and their qualitative evaluation. British Journal of Nutrition 93, 205–212.
Sandberg, F.B., Emmans, G.C. and Kyriazakis, I. (2005b) Partitioning of limiting
protein and energy in the growing pig: testing quantitative rules against
experimental data. British Journal of Nutrition 93, 213–224.
Sandberg, F.B., Emmans, G.C. and Kyriazakis, I. (2006) A model for predicting
food intake of growing animals during exposure to pathogens. Journal of
Animal Science (in press).
Stamataris, C., Kyriazakis, I. and Emmans, G.C. (1991) The performance and
body-composition of young pigs following a period of growth retardation by
food restriction. Animal Production 53, 373–381.
Steel, J.W., Symons, L.E.A. and Jones, W.O. (1980) Effects of level of larval intake
on the productivity and physiological and metabolic responses of lambs
infected with Trichostrongylus colubriformis. Australian Journal of Agricultural
Research 31, 821–838.
Stewart, T.B., Hale, O.M. and Johnson J.C., Jr (1969) Failure of parasitised gilt and
barrow pigs on different planes of nutrition to respond alike to a superimposed
infection with Strongyloides ransomi. The Journal of Parasitology 55, 1055–1062.
Takhar, B.S. and Farrell, D.J. (1979) Energy and nitrogen metabolism of chickens
infected with either Eimeria acervulina or Eimeria tenella. British Poultry Science
20, 197–211.
Taylor-Robinson, A.W. (2000) Increased production of acute-phase proteins
corresponds to the peak parasitemia of primary malaria infection. Parasitology
International 48, 297–301.
Turner, J.L., Dritz, S.S., Higgins, J.J., Herkelman, K.L. and Minton, J.E. (2002a)
Effect of a Quillaja saponaria extract on growth performance and immune
function of weanling pigs challenged with Salmonella typhimurium. Journal of
Turner, J.L., Dritz, S.S., Higgins, J.J. and Minton, J.E. (2002b) Effects of
Ascophyllum nodosum extract on growth performance and immune function on
young pigs challenged with Salmonella typhimurium. Journal of Animal Science 80,
1947–1953.
Van Dam, J.T.P., Schrama, J.W., Vreden, A., Verstegen, M.W.A., Wensing, T., van
der Heide, D. and Zwart, A. (1998) The effect of previous growth retardation
on energy and nitrogen metabolism of goats infected with Trypanosoma vivax.
British Journal of Nutrition 77, 427–441.
van Heugten, E., Coffey, M.T. and Spears, J.W. (1996) Effects of immune challenge,
dietary energy density, and source of energy on performance and immunity in
weanling pigs. Journal of Animal Science 74, 2431–2440.
van Houtert, M.F.J., Barger, I.A., Steel, J.W., Windon, R.G. and Emery, D.L. (1995)
Effects of dietary protein intake on responses of young sheep to infection with
Trichostrongylus colubriformis. Veterinary Parasitology 56, 163–180.
van Milgen, J. and Noblet, J. (1999) Energy partitioning in growing pigs: the use of
a multivariate model as an alternative for the factorial analysis. Journal of
van Milgen, J., Quiniou, N. and Noblet, J. (2000) Modelling the relation between
energy intake and protein and lipid deposition in growing pigs. Animal Science
71, 119–1130.
Waller, P.J. (1997) Nematode parasite control of livestock in the Tropics/Subtropics:
the need for novel approaches. International Journal of Parasitology 27,
1193–1201.
Wang, T.C. and Fuller, M.F. (1989) The optimum dietary amino acid pattern for
growing pigs. 1. Experiments by amino acid deletion. British Journal of Nutrition
62, 77–89.
Webel, D.M., Johnson, R.W. and Baker, D.H. (1998a) Lipopolysaccharide-induced
reductions in body weight gain and feed intake do not reduce the efficiency of
arginine utilization for whole-body protein accretion in the chick. Poultry

Science 77, 1893–1898.
Webel, D.M., Johnson, R.W. and Baker, D.H. (1998b) Lipopolysaccharide-induced
reductions in food intake do not decrease the efficiency of lysine and threonine
utilization for protein accretion in chickens. Journal of Nutrition 128,
1760–1766.
Wellock, I.J., Emmans, G.C. and Kyriazakis, I. (2003b) Predicting the consequences
81, 2995–3007.
Wilson, D.P. and McElwain, D.L. (2004) A model of neutralization of Chlamydia
trachomatis based on antibody and host cell aggregation on the elementary
body surface. Journal of Theoretical Biology 226, 321–330.
Whittemore, C.T. (1995) Modelling the requirement of the young growing pig for
dietary protein. Agricultural Systems 47, 415–425.
Whittemore, E.C., Kyriazakis, I., Emmans, G.C. and Tolkamp, B.J. (2001) Tests of
two theories of food intake using growing pigs 1. The effect of ambient
temperature on the intake of foods of differing bulk content. Animal Science 72,
351–360.
Yakoob, A., Holmes, P.H. and Armour, J. (1983) Patho-physiology of
gastrointestinal Trichostrongyles in sheep – plasma losses and changes in plasma
pepsinogen levels associated with parasite challenge of immune animals.
Research in Veterinary Science 34, 305–309.
Yu, F., Bruce, L.A., Calder, A.G., Milne, E., Coop, R.L., Jackson, F., Horgan, G.W.
and MacRae, J.C. (2000) Subclinical infection with the nematode
Trichostrongylus colubriformis increases gastrointestinal tract leucine metabolism
and reduces availability of leucine for other tissues. Journal of Animal Science 78,
380–390.
Zijlstra, R.T., Donovan, S.M., Odle, J., Gelberg, H.B., Petschow, B.W. and Gaskins,
H.R. (1997) Protein-energy malnutrition delays small-intestinal recovery in
neonatal pigs infected with rotavirus. Journal of Nutrition 127, 1118–1127.
8 Nutrient Flow Models, Energy
Transactions and Energy Feed
Systems
J. VAN MILGEN
INRA – UMR SENAH, Domaine de la Prise, 35590 Saint-Gilles, France
jaap.vanmilgen@rennes.inra.fr
Introduction
Systems of nutrient and energy utilization are widely used in the animal feed
industry. Existing energy systems such as digestible energy (DE),
metabolizable energy (ME) and net energy (NE) attribute a single energy
value to a feed but differ in the way that energy losses are accounted for.
Although these systems are simple to use and relatively robust, they have
been criticized because they cannot account for interactions between the feed
and the animal. This has led to the development of more mechanistic models
that can be used for feed evaluation or that predict the response of an
animal to a changing nutrient supply. The objective of this chapter is to
describe different aspects of energy loss and energy utilization in animals and
to challenge this knowledge with empirical and mechanistic approaches to
feed evaluation. Mechanistic models of animal nutrition undoubtedly
provide a more solid theoretical basis for feed evaluation. However, they
only explain part of the efficiency with which nutrients are used for different
productive purposes. Moreover, although ‘scientific truth’ may be a reason to
push for a more mechanistic approach to feed evaluation systems, there are
also reasons to be somewhat conservative. Past history, established
knowledge and the use of an established reference base can be expected to
play major roles in the acceptance or failure of future energy systems.
Energy Values and Energy Losses
Gross energy
There are different ways to express the energy value of a feed. In one of its
more basic forms, it can be represented by its gross energy (GE) value,

144 J. van Milgen
which can be estimated by combustion in a calorimeter. Alternatively, it can

be estimated from the chemical composition of the feed by multiple
regression. Using data from Noblet et al. (1994), the following equation
estimates the GE content of a feed (kJ/g):
GE = 22.6•CP + 38.8•EE + 17.5•starch + 16.7•sugars + 18.6•residue
where CP is the proportion of crude protein (g/g), EE the ether extract,
and residue is the difference between organic matter and the other
identified fractions in the equation (essentially fibre). Equations of this
form have the advantage that they attribute an energy value to each
category of nutrient, which bears close resemblance to their theoretical
energy values. For example, the energy value of glucose is 2820 kJ/mol (1
mol = 180 g). The theoretical energy value of starch, a polymer of glucose,
is then 2820 / (180–18) = 17.4 kJ/g, which corresponds closely to the
coefficient in the equation given above.
Depending on the amino acid composition of protein and the fatty acid
composition of lipid, their GE content may vary (Boisen and Verstegen,
2000; van Milgen, 2002). Moreover, there are other nutrients (e.g. volatile
fatty acids or lactic acid in fermented products) that also contribute to the
GE content of a feed, although their contribution is not specifically
accounted for in the equation above. Tran and Sauvant (2002) and Noblet
et al. (2003) included correction factors for each group of feed ingredients
to account for these deviations.
Faecal energy losses
In contrast to current practice for amino acids, digestibility of energy is not

considered at the ileal but at the faecal level. The faecal energy digestibility
in typical pig diets ranges from 70 to 90%. The energy digestibility of feed
ingredients can be much more variable. Energy digestibility declines
linearly with neutral detergent fibre (NDF) content (Noblet and Perez,
1993; Le Goff and Noblet, 2001; Lindberg and Pedersen, 2003). The
extent of this decline is such that NDF contributes little to the energy
supply to the animal and mainly acts by diluting the energy content. For
growing pigs, a 1% increase in NDF content reduces the energy
digestibility by 0.9% (Le Goff and Noblet, 2001). This does not necessarily
mean that fibre or NDF is not digested by growing pigs. It is possible that
the actual digestibility of fibre is offset by increased endogenous secretions
or by a reduction in digestibility of other nutrients, so that the overall
contribution of fibre to the energy supply is close to zero. The fact that
NDF appears to be a good indicator for energy digestibility should not be
interpreted as meaning that all NDF is the same. Le Goff et al. (2002)
observed that the NDF digestibility in growing pigs ranged from 38% for
maize bran to 71% for sugar beet pulp. The low digestibility of fibre in
growing pigs seems mainly due to the high rate of passage when feeding
fibre-rich diets (Le Goff et al., 2002). Adult sows are much more capable of
Flow Models, Energy Transactions and Feed Systems 145
digesting fibre than growing pigs, due to the four-times greater retention
time of digesta in the gastro-intestinal tract. Similarly, finishing pigs
typically digested fibre better than did growing pigs (Le Goff et al., 2002).
This means that, in contrast to GE, the DE value of a diet is not a
characteristic of the diet itself but is influenced by the animal.
Urinary and gaseous energy losses
The metabolizable energy (ME) value of a diet accounts for energy losses
that occur from fermentation gases and in the urine. Energy losses as
methane and hydrogen are relatively small for diets fed to growing pigs
(typically less than 0.5% of DE). Noblet et al. (2002) assumed that energy
losses in fermentation gases correspond to approximately 0.67 kJ/g of
fermented dietary fibre. Energy losses in the urine are of the order of 3.5%
of DE. Most of this energy is lost as urea (or uric acid in birds), which
originates from amino acid catabolism in the liver. Amino acids given in
excess of requirements will be deaminated and result in additional urinary
energy loss. There is a relatively straightforward relation between the
nitrogen and energy contents in the urine. In the INRA tables (Noblet et
al., 2002), the following equation is used for pigs:
(energy in urine; kJ/kg DM intake) = 192 + 31 × (nitrogen in urine; g/kg
DM intake)
Under the assumption that approximately half of the digestible nitrogen is
excreted in the urine, the equation above can easily be converted into a
relation between urinary energy losses and dietary protein content.
Heat production
All metabolizable energy that is not retained by the animal is lost as heat.
Heat production can be measured through calorimetry, whereas energy
retention can be measured using the serial slaughter technique. Indirect
calorimetry is based on the measurement of gas exchanges between the
animal and its environment. When nutrients are oxidized, animals
consume oxygen and produce carbon dioxide and methane. These gas
exchanges and the nitrogen excretion originating from protein catabolism,
combined with the stoichiometry of carbohydrate, protein, and lipid
oxidation allow calculation of heat production (Brouwer, 1965).
Calorimetry has the advantage over the serial slaughter technique that it
can be used to measure energy balance over successive short periods of
time. We have further refined this technique in order to obtain estimates of
different components of heat production related to fasting, physical activity
and the thermic effect of feeding (van Milgen et al., 1997). The energy and
nitrogen balance techniques typically give higher retention values than the
comparative slaughter technique (Quiniou et al., 1995) and are thought to
overestimate actual lipid and protein retention.
146 J. van Milgen
Total heat production is not to be confused with the heat increment (or
thermic effect of feeding). As indicated above, total heat production is the
difference between ME intake and retained energy. The heat increment is
the change in heat production associated with a change in ME intake.
Energy Systems
The purpose of an energy system is to attribute an energy value to a feed

so that it can be compared with the energy requirement for a specific
function, which should be expressed on the same scale. Expressing the
feed and requirement value as a single entity has the advantage of being
simple to use.
Several systems of energy utilization have been proposed. The DE
system accounts for differences in digestive utilization. The ME system also
accounts for energy losses in the urine and as combustible gasses. The NE
system is calculated as the sum of fasting heat production (FHP) and the
retained energy, or as ME minus the heat increment of feeding. Because it
is difficult and costly to measure directly the energy value of a diet, several
research groups have proposed equations to estimate the energy (DE, ME
or NE) values from the chemical composition of the diet. Because there
can be important differences in approaches and methodologies (e.g. within
NE systems), energy values are not necessarily interchangeable between
laboratories. The equations below estimate the DE, ME and NE values (in
kJ/g) from digestible nutrients (as a proportion of the diet). These
equations originate from statistical relationships between the measured
energy value and the (digestible) nutrient composition for 61 different
diets (see Noblet et al., 1994 for a description of methods):
DE = 23.25•dCP + 38.73•dEE + 17.45•starch + 16.77•sugars +
16.68•dresidue
ME = 20.40•dCP + 39.28•dEE + 17.45•starch + 16.47•sugars +
15.45•dresidue
NE = 12.08•dCP + 35.01•dEE + 14.32•starch + 11.94•sugars +
8.64•dresidue
where dCP is the faecal digestible crude protein, dEE is the digestible ether
extract, and dresidue is the digestible residue (i.e. the difference between
digestible organic matter and the digestible CP, EE, starch and sugar
contents). Starch and sugars are both assumed to be completely digestible at
the faecal level. The coefficients estimated for the nutrients have some
interesting features. First, because the DE value is based on the digestible
nutrient content, the coefficients correspond closely to the GE values for
each nutrient. When the coefficients for the ME equation are compared
with those of DE, it is clear that the difference between the DE and ME
values is mainly due to the protein and residue contents. The energy lost in
the urine (from protein) and as combustible gases (from fibre fermentation)
is the cause of this. Comparing the coefficients of NE with those of ME
illustrates the differences in energetic efficiencies between nutrients (see also

the section ‘Confronting stoichiometry with experimental data’, p.150).
Dietary fat is used with the highest efficiency (89.1%), followed by starch
(82.0%) and sugar (72.5%), and finally fibre (55.9%) and protein (59.2%).
Excess dietary protein is thus a relatively inefficient source of energy.
The consequence of this is that the contribution of protein to the energy
value of the diet diminishes in the order DE, ME and NE. For example, in
a typical European cereal-based diet, soybean meal may contribute 25% to
the GE or DE value (16.2 and 13.5 kJ/g, respectively), 23% to the ME value
(13.0 kJ/g) and only 19% to the NE value (9.6 kJ/g). Fat, on the other hand,
may contribute 4.8% to GE, 4.9% to DE, 5.2% to ME and 6.2% to the NE
value of the diet. In other words, the ranking of different feed ingredients
is affected by the system of expressing the energy value. Feed ingredients
rich in protein (or fibre) have relatively lower values in an NE system,
whereas those rich in fat will be attributed a higher value.
Apart from establishing energy values for a diet, requirement values
have to be defined in an energy system. The following two equations
indicate the ME and NE requirements (kJ/d) for growing animals:
ME = MEm + PD/kp + LD/kf
NE = FHP + PD + LD
where MEm is ME for maintenance, PD is protein deposition (kJ/d), LD is
lipid deposition (kJ/d) and kp and kf are the efficiencies of protein and lipid
deposition, respectively. On average, these efficiencies are close to 60 and
80%, respectively (Noblet et al., 1999). The ME system does not account for
the fact that these efficiencies may be affected by the diet. The difference
between both equations also illustrates that in the ME system the heat
increment is accounted for in the requirements (albeit in a very crude
way), whereas in the NE system this is accounted for in the feed value. The
effective energy system (Emmans, 1994) uses an intermediate approach by
attributing part of the heat increment to the effective energy feed value
and another part to the effective energy requirement. Heat increments for
urinary nitrogen excretion, faecal organic matter excretion, and methane
production are deducted from the ME supply (corrected for zero nitrogen
retention), whereas heat increments for protein and lipid deposition are
included in the effective energy requirement. In addition, on the supply
side a distinction is made between the heat increment of using dietary
lipid, as opposed to non-lipid sources, for lipid deposition
Shortcomings of Classical Feed Evaluation Systems
In classical energy systems, energy values and requirements are reduced to

a single number. This has the advantage that the system is simple to use
and relatively robust. Tabular values of feed values and animal
requirements can be established and the user only has to ensure that these
match. The disadvantage is that the system is not necessarily consistent
148 J. van Milgen
with biological reality and the following examples will illustrate this. As
indicated earlier, dietary protein affects the ME value of a diet because
amino acids that are not deposited are deaminated, and urea is excreted in
the urine. In other words, it is not the dietary protein itself that will
determine its ME value, as it will also depend on what fraction of that
protein the animal retains. A similar example can be given for the NE
system. The NE requirement is based on a requirement for maintenance
(FHP) and energy retention. The maintenance requirement is essentially
an ATP requirement. If glucose and lipid are compared, it can be shown
that their potential to produce ATP is similar (~ 74 kJ/ATP produced; van
Milgen, 2002). When these nutrients are used for maintenance, glucose
and lipid should therefore be attributed similar energy efficiencies.
However, when these nutrients are used for lipid deposition, large
differences occur. The biochemical efficiency of using glucose for lipid
deposition is close to 81%, whereas the efficiency of using dietary lipid for
lipid deposition is around 98%. This indicates that the final utilization will
eventually determine the energy value.
The fact that there are interactions between the animal and its diet
implies that classical energy systems are theoretically incorrect. This
problem can only be overcome if we acknowledge that there is no such
thing as ‘an energy value of a diet’ expressed in such a way that it can be
related to a requirement. The only energy value that would be an attribute
of the diet itself would be its GE value. However, this is of no use when
expressing a requirement.
Mathematical models are ideally suited to account for interactions
between the diet and the animal. In the past and also recently, several
models have been published that address this issue. It is beyond the scope
of this paper to compare these models, but Luiting and Knap (Chapter 13,
this volume) deal with this in some detail. This chapter will focus only on
how recent models deal with nutrient and energy transactions and to what
extent this mechanistic approach is more useful than classical systems of
nutrient evaluation.
Stoichiometry of Energy Transactions

When developing a more mechanistic approach to energy evaluation,
several issues have to be taken into account. These include the
stoichiometry of energy transactions, the pathways and organs involved in
the transaction and the energy cost of physiological functions. The
stoichiometry of the quantitatively most important transactions is well
established. For the interested reader, I would recommend the work of
Salway (1994), which helps to get an overview of major pathways of
metabolism including some aspects of regulation. In the past, different
publications addressed the issue of quantifying the stoichiometry of
nutrient transactions (Armstrong, 1969; Krebs, 1972; Schulz, 1978;
Livesey, 1984). We expanded on this by providing a generic framework
that can be used to construct and quantify the energetics of different

metabolic pathways (van Milgen, 2002). The model, a simple spreadsheet,
is used in this presentation to quantify the energetics of nutrient
transactions; the model is available from the author on request.
Although stoichiometry may seem merely a matter of book keeping,
some differences between approaches can occur depending on the
pathways involved (especially for catabolism of essential amino acids). For
example, the different pathways of tryptophan catabolism may lead to
differences in stoichiometric results. Also catabolism of glycine, methionine
(methyl-groups) and cysteine has been quantified in different ways (van
Milgen, 2002). In addition, a decision has to be made about which
nutrients will be represented specifically in the model. For example, the
complete oxidation of glucose in the TCA cycle is a process involving (at
least) 18 intermediate steps involving the carbon chain (glucose, glucose-
6-phosphate, fructose-6-phosphate, etc.). It goes without saying that it
would be of little use to include all these steps in whole animal models. In
our framework (van Milgen, 2002), we used six carbon chain ‘pivots’ and
eight co-factors (e.g. ATP, NADH) to quantify nutrient transactions. Most
models use considerably fewer energy pivots to express these transactions
and acetylCoA and/or ATP are most frequently used (Boisen, 2000; Chudy,
2000; Birkett and de Lange, 2001a; Lovatto and Sauvant, 2003; Green and
Whittemore, 2003; Halas et al., 2004). AcetylCoA may seem an obvious
choice as a carbon-chain pivot as it plays a central role in the use of
nutrients for oxidation in the Krebs cycle (ATP synthesis) or for fatty acid
synthesis required for lipid deposition. A necessary condition in the choice
of pivots is of course that the energy released (or required) from
transforming a nutrient to the pivot is accounted for. For example, the
transformation of 1 mol of glucose to acetylCoA requires 2 ATP but also
releases 4 ATP and 4 NADH. If the NADH is oxidized in the mitochondria,
it can be expressed as an ATP equivalent.
By reducing complete pathways to pivot equivalents, some information
will be lost. For example, NADH can be produced both in the cytosol and
the mitochondrion. As ATP synthesis from NADH occurs in the
mitochondrion, and the transfer of NADH from the cytosol to the
mitochondrion implies a loss of energy (Salway, 1994), one may opt to
include this energy loss directly in the stoichiometric balance. However,
when glucose is used for fatty acid synthesis, the NADH released during
glucolysis can be re-used (in the cytosol) in the pyruvate/malate cycle,
thereby avoiding the energy loss. Consequently, cytosolic NADH can be
used in different metabolic pathways, having different energetic
efficiencies. One option to solve this problem is the inclusion of zero-pools
in the model. These pools are present only for accounting purposes and
are programmed so as to maintain a zero size. For example, a zero-pool of
cytosolic NADH may be programmed so to transfer its surplus to a zero-
pool of mitochondrial NADH (while accounting for the corresponding
energy loss) which, in turn, transfers its surplus to a zero-pool of ATP.
Zero-pools should be programmed so that when positive, they transfer
150 J. van Milgen
their content to another nutrient pool, whereas if negative, they pull

energy from a preceding nutrient pool. For example, if insufficient
NADPH is available for fatty acid synthesis, the system may pull from a
glucose pool to generate NADPH in the pentose-phosphate shunt. The use
of zero-pools is a convenient way for bookkeeping without making a priori
decisions on pathways involved in metabolism.
Short-term Energetics of Nutrient Transactions
Metabolism includes several pathways that are designed to cope with the
short-term dynamics of nutrient supply and requirement, or that deal with
nutrient transfers between different organs. Apart from the physiological
cost of using these pathways, there is also a biochemical cost related to the
nutrient transformations. Few, if any, of the published mechanistic whole
animal models include the examples that will be developed here. These
examples are not given to criticize these models, but to illustrate that
energy metabolism in animals is a very complex process, which goes far
beyond the approach employed in most models.
Although it may seem evident not to include glycogen in models that
are based on empirical estimates of efficiency (as the energy loss is
included, empirically, in the efficiency estimate and glycogen stores vary
little from day to day), this is less so for models that are based on
biochemistry. Depending on the site of glycogen storage and utilization
(muscle or liver), ATP synthesis from glucose is stoichiometrically 3 to 6%
less efficient if this glucose is stored as glycogen before being oxidized.
Type IIb muscle fibres have a high glycogen storage capacity. As their
mitochondrial oxidative capacity is limited these fibres, when solicited,
produce ATP mainly through glycolysis and yield lactate as an end-
product. Although lactate can be used directly by some tissues (e.g. heart
muscle but also skeletal muscle itself), it can also be used by the liver to
regenerate glucose through gluconeogenesis. In the transformation of
glucose to lactate (in muscle) and back to glucose (in the liver), a total of
four ATP will be lost, hence an energy loss of more than 10%.
Another example is the temporary storage of energy as lipid (see also
Baldwin, 1995 for a discussion of this topic). Type I (red) muscle fibres are
used to provide sustained muscular work (e.g. for maintaining posture).
Although these oxidative fibres can use both lipid and glucose as fuel, most
of the energy storage occurs as lipid. The temporary energy storage of
glucose as lipid is rather inefficient and requires 30% more energy than the
direct utilization of glucose.
Confronting Stoichiometry with Experimental Data
Stoichiometry explains only part of the observed heat increment. Table 8.1
shows some results of an experiment in which either starch, maize gluten
meal (an unbalanced protein source), casein or lipid was added to a basal
diet limiting in lysine supply (van Milgen et al., 2001). Also the coefficients
of the ME and NE equations described above can be used to evaluate the
contribution of stoichiometry to observed energy efficiencies.
The stoichiometric efficiency of using glucose for lipid varies between
81 and 84%, depending on whether the ATP synthesis during lipogenesis
from glucose is accounted for or not (Baldwin, 1995; van Milgen, 2002).
For starch, the NE/ME ratio from the equations used above is 0.82,
whereas van Milgen et al. (2001) observed a value of 0.84 (Table 8.1). This
suggests that the heat increment of glucose metabolism not related to
stoichiometry (e.g. intake, hydrolysis of starch, absorption, maintaining
blood glucose) is relatively small or that this cost is accounted for
elsewhere, for example in the maintenance energy expenditure.
Larger differences occur when comparing the stoichiometric and
experimental efficiencies for lipid metabolism. The theoretical efficiency of
lipid deposition from dietary lipid is very high. The only energy loss is due
to the re-activation of fatty acids to acyl-CoA during re-esterification.
Assuming that dietary lipids are hydrolysed and re-esterified twice, the
energy loss does not exceed 3%. However, the NE/ME ratio of the
equations above suggests that more than 10% of the energy in dietary lipid
is lost when it is used for lipid deposition. The results in Table 8.1 also
indicate that the observed efficiency is much lower than the stoichiometric
efficiency. The relatively low efficiency of using dietary lipid may be due to
the oxidation of dietary lipids (e.g. for ATP synthesis) combined with de
Table 8.1. Utilization of energy by growing pigs (van Milgen et al., 2001).
Maize gluten
Starch meal Casein Lipid
Chemical composition (%)

Nitrogen content 0.13 10.89 15.28 n.a.
Lysine content n.a. 0.56 8.13 n.a.
Fat n.a. 2.5 n.a. n.a.
Starch 97.7 17.7 n.a. n.a.
Energy utilization
Ileal digestibility 0.988 0.848 0.971 0.900
Faecal digestibility 1.001 0.889 0.965 0.859
Metabolizability 1.002 0.843 0.884 0.985
Fraction of ME used for PD 0.044 0.050 0.420 –0.025
Efficiency of lipid deposition 0.842 0.520 0.520 0.883
Energy cost of protein deposition
(kJ NE/kJ PD) 0.484 0.484 0.484 0.484
Energy values (kJ/g)
GE 17.27 24.33 22.86 39.47
DE (faecal) 17.29 21.63 22.06 33.90
ME 17.32 18.23 19.50 33.40
NE – excluding the energy cost of PD 14.71 9.92 14.07 29.39
NE – including the energy cost of PD 14.34 9.48 10.11 29.80
152 J. van Milgen
novo lipid synthesis from other nutrients. The scenario is energetically

much more expensive than the use of dietary glucose for ATP synthesis
and the use of dietary lipids for lipid deposition. Lizardo et al. (2002)
assumed that only 85% of digested lipids were deposited in growing pigs
but recent data from our laboratory suggest an even lower value of 70%
(Kloareg et al., 2005).
An even greater difference between stoichiometry and experimental
data is observed for protein deposition. Synthesis of a peptide bond from
amino acids requires at least 5 ATP and, based on the efficiency of ATP
synthesis, the maximum efficiency of protein deposition ranges between 85
and 90%. However, experimental values of the energetic efficiency of
protein deposition are much lower. We estimated that 0.484 kJ of NE (as
ATP) was required to support the deposition of 1 kJ of protein (van Milgen
et al., 2001). Because the efficiency of using ME for NE (ATP) varies with
the nutrient source, the estimated experimental efficiency of protein
deposition ranges from 0.52 (using amino acids for the support costs) to
0.63 (using glucose for the support costs). It is thought that protein
turnover (i.e. the repeated hydrolysis and synthesis of peptide bonds)
contributes to a large extent to the low efficiency of protein deposition.
Based on the hypothesis that 5 ATP are required to synthesize a peptide
bond, 18–22 ATP would be required to explain the observed efficiency of
protein deposition. This value would correspond to the synthesis of four
peptide bonds for each peptide deposited. Lobley (2002) estimated that
whole animal protein synthesis in pigs could reach a value of 600 g/day,
which is indeed about four times the protein deposition rate in pigs.
The reasoning above applies to the case of using dietary protein for
protein deposition. However, typically not more 50% of dietary protein will
be deposited, which means that the remainder will be deaminated and
used for other purposes. Although the synthesis of urea requires ATP, the
energy cost involved is not sufficient to account for the heat increment of
excess protein. In our experiment, excess protein was used with an
efficiency of 0.52 for lipid deposition and the NE/ME ratio was 0.59 (Table
8.1). This may be due to the fact that providing unbalanced dietary protein
stimulates protein turnover. This protein will first induce an ATP cost due
to the synthesis and hydrolysis of peptide bonds and finally induce an ATP
cost for urea synthesis. It is interesting to note that the experimental
efficiencies of using dietary protein for protein deposition or for lipid
deposition are almost identical (van Milgen et al., 2001).
The difference between experimental values of energy efficiency and
stoichiometry has, to some extent, been accounted for in models of nutrient
metabolism. Green and Whittemore (2003) included a ‘residual efficiency’
in their model in order to account for differences between stoichiometric
and experimental efficiencies. Nutrient transport was seen as a major
contributor to the residual efficiency. Birkett and de Lange (2001a,b) also
calibrated experimental efficiency values to stoichiometric ones. However,
in contrast to Green and Whittemore (2003), they chose to relate the
unaccounted efficiency to physiological processes such as intake of digestible
nutrients, excretion of faeces and urine, and costs of protein and lipid
deposition, all expressed on an ATP basis. As such, the model resembles the
concepts used in the effective energy system (Emmans, 1994), although the
units of expression are different. The model of Halas et al. (2004) also relies
on the stoichiometry of nutrient transactions. However, rather than
estimating efficiency values for different processes through calibration,
energy expenditures were attributed a priori to different physiological
processes. Energy costs for absorption and transport were attributed to
different nutrients. Protein synthesis was assumed to cost 4 ATP per
peptide, whereas hydrolysis of a peptide bond also required 1 ATP. Because
the model includes four different protein pools (muscle, skin–backfat,
organs and bone), each with their own turnover characteristics, differences
in body protein composition will result in differences in ATP requirement
and, hence, energy expenditure. Other costs specifically accounted for
include urea synthesis and bone mineralization.
Support Costs (Maintenance)
The most difficult part of establishing a system (or a model) of nutrient

evaluation is the quantification of the energy cost of physiological
functions. Processes such as muscle contraction, ion transport, peptide
synthesis all require energy, most of which will come directly or indirectly
from ATP. The largest contribution to this energy expenditure comes from
maintenance functions. Especially nervous functions, maintaining
membrane potential, and protein resynthesis contribute largely to the basal
metabolic rate (Table 8.2). Many physiological functions are driven by a
Na+ gradient. For example, during active transport of glucose Na+ enters
the cell, which will be pumped out of the cell by a Na/K-ATPase at the cost
of ATP. Consequently, the Na+ gradient by itself represents an energy
reserve, which has to be maintained.
It is virtually impossible to specifically include the energy costs of all
physiological functions in a model. Although certain aspects may be
Table 8.2. Energy expenditure of several maintenance functions

(Baldwin, 1995).
% of BMR
Service functions
Kidney (Na+ transport) 6–7
Heart 9–11
Nervous tissues 15–20
Respiration 6–7
Repair functions
Protein re-synthesis 10–15
Lipid re-synthesis 1–2
Na+ transport (membrane potential) 20–25
154 J. van Milgen
included (e.g. the cost of protein turnover discussed above), we will have to
rely to a large extent on empirical estimates of maintenance. From a
biological and physiological point of view, it is very difficult to define and
measure maintenance unambiguously. Theoretically, maintenance corre-
sponds to a situation in which energy retention equals zero (i.e. energy
intake equals heat production) but, for a growing animal, this corresponds
to a non-physiological situation. Moreover, zero energy retention may
theoretically be achieved while depositing protein and catabolizing body
lipid and therefore does not correspond to maintaining a constant body
weight.
The FHP is closely related to the maintenance energy requirement and
serves as the maintenance energy requirement in NE systems (Noblet et al.,
1994). During fasting, animals mobilize body reserves in order to supply
energy for vital body functions. Measured values for FHP after a 1-day
fasting period range from 700 to 800 kJ/((kg BW)0.60•d) in growing pigs
offered feed close to ad libitum prior to fasting. Genotype (or leanness)
appears to have an important impact on FHP with lower estimates for
obese Meishan barrows and higher estimates for lean Piétrain boars (van
Milgen et al., 1998). Part of the difference between genotypes seems to be
due to differences in body composition. In particular, the viscera make an
important contribution to FHP. This is consistent with the observation that
the previous plane of nutrition, and thus the viscera mass, affects FHP
(Koong et al., 1982) and justifies the use of different protein pools to
explain energy expenditure (Halas et al., 2004).
Part of the differences observed in FHP may be due to differences in
the body reserves that are mobilized during fasting. It is likely that during
fasting, the animal first mobilizes glycogen reserves, followed by the
mobilization of labile protein stores (visceral proteins) and lipids.
Consequently, the length of the fasting period may affect FHP. During
prolonged fasting, the animal is likely to adapt its metabolism so as to
minimize energy expenditure. It is for this reason that in our laboratory
we measure FHP after a 1-day fasting period so that metabolism still bears
some resemblance to a producing animal.
The energy cost of physical activity can be an important source of
variation between animals. Energy expenditure per hour of standing
appears at least fourfold greater in pigs than in most other domestic
species (Noblet et al., 1993). Different techniques exist to measure physical
activity including measurement of standing duration, motion detection,
and force detection. Heat production due to physical activity is estimated
from a statistical relation between variation in heat production and
variation in recorded physical activity. Our current estimate (based on
force detection) is approximately 200 kJ/((kg BW)0.60 •d), which
corresponds to 3 h of standing per day (van Milgen et al., 2001); this value
corresponds to approximately 20% of the maintenance energy
requirement. Physical activity appears rather variable between animals and
may be affected by feeding level, type of diet, and genotype (Susenbeth
and Menke, 1991; Schrama et al., 1996). Because of its contribution to heat
production and thus its effect on energy retention, it is important to obtain

reasonable indicators of physical activity.
Energy Status and Energy Scales
In the literature, there are different approaches to evaluating the energy

status of an animal. How does the animal know where it is and how does it
know where it has to go? Although these issues are probably more of
philosophical than of practical interest, it is nevertheless interesting to see
how different modellers view this issue.
It has been known for a long time that energy supply affects the
partitioning of energy between protein and lipid deposition. In one of the
first models of pig growth, Whittemore and Fawcett (1976) assumed the
existence of a minimum ratio between lipid and protein retention. In their
view, the animal ‘monitors’ the potential composition of the gain, which
may result in the catabolism of additional dietary protein to favour lipid
deposition. Moughan et al. (1987) and de Lange (1995) used a similar
logic, but assumed the existence of a minimum lipid to protein mass ratio.
Wellock et al. (2003) and Green and Whittemore (2003) assumed that
growing animals have a preferred lipid to protein mass ratio that they will
try to achieve or maintain. These models therefore include a control on
the ‘receiving’ side of the equation. The animal (in the modeller’s view)
somehow wants to control its body composition and therefore changes the
partitioning of nutrients between protein and lipid deposition.
However, also on the ‘supply side’, the energy status of the animal has
to be quantified. The response curve of protein to energy will be different
when the latter is expressed on a DE, ME, NE or ATP scale. In addition, the
interpretation of a given energy scale may change during growth. Does a
30-kg pig interpret an additional kJ of DE in the same way as a 100-kg pig?
Black et al. (1986) scaled protein deposition as a function of energy intake
and body weight on a linear MJ/day scale whereas van Milgen et al. (2005)
expressed the same relation as multiples of maintenance (i.e. related body
weight raised to the power 0.60). Halas et al. (2004) scaled many equations
to metabolic empty body weight (kg0.75) thereby implying a scaling to
maintenance. The use of metabolic body weight (kg0.75) originates from the
comparison of maintenance between different species of mature animals.
When MEm or FHP is compared for animals of different BW within a
species, the value of ‘b’ is typically lower than 0.75 and, for growing pigs, a
value close to 0.60 is often found. Although this may seem a minor issue, it
has important consequences for the dynamics of heat production during
growth. For example, suppose that one has obtained a reliable estimate of
Mem at 60 kg of BW. If maintenance is constant per kg0.60, it would result
in an 18% higher maintenance requirement at 20 kg BW compared to
assuming a constant maintenance requirement per kg0.75. However, at
greater BW, the ranking is reversed so that at 120 kg of BW, Mem is 10%
lower when using the power 0.60 compared to using 0.75.
156 J. van Milgen
The choice of an appropriate energy scale is also important when

modelling or describing voluntary feed intake. Most applied models of
animal production consider voluntary feed intake as a user input, and it is
often described as an empirical function of body weight. When it is
assumed that feed intake is regulated by quantity, feeding an energy-rich
diet (e.g. by adding vegetable oil) will thus provide more energy resulting
in a higher growth rate. On the other hand, if feed intake is assumed to be
regulated on an NE basis, feeding an energy-rich diet will result in similar
NE intakes (for different quantities of feed) and similar performance. The
energy scale for expressing voluntary feed intake (DE, ME, NE) thus has
important consequences for model predictions. Using NE as the energy
scale to describe voluntary feed intake implies that animals eat for energy
retention. The model of Wellock et al. (2003) expands on this by
considering that animals eat to attain a desired body composition.
A novel approach to representing the energy status of the animal was
proposed by Lovatto and Sauvant (2003). They considered that growth of
an animal is regulated by homeostatic and homeorhetic controls.
Homeorhetic regulation ensures the long-term control of growth and is
driven by the current state of the animal. Protein and lipid deposition are
represented as the difference between catabolism and anabolism:
dS/dt = (A – C) • S
where S is the current state of the animal (i.e. protein or lipid mass) and A
and C are the fractional anabolic and catabolic rates, respectively, which
were represented by:
A = k1 + k2 • exp (–k3 • time)
C = k1 + k4 • exp (–k5 • time).
At maturity, both fractional rates are equal to k1, whereas during growth A
is greater than C due to the fact that k2 > k4. Thus, at maturity both
catabolism and anabolism continue to operate but at an equal homeorhetic
rate. The equations used by Lovatto and Sauvant (2003) have two
interesting features. First, the equations for A and C resemble the
Gompertz function, which is often used to describe growth. However, in
the Gompertz function, the asymptote (k1) equals zero and the function is
not used to describe anabolism and catabolism separately. Secondly, ‘time’
is specifically represented in this model (in addition to ‘state’). This means
that animals cease to grow due to the fact that they age (and lose the
potential to grow), rather than the fact that they approach a mature body
state. Although there is some debate on whether growth is determined by
age or by state (van Milgen et al., 2000; Wellock et al., 2003), Lister and
McCance (1967) concluded that pigs severely undernourished for a year
(i.e. 5.5 kg of body weight after 1 year) stopped growing at the same
physiological age as normal pigs and did not attain the same adult size.
Nevertheless, growth during the re-feeding phase was virtually identical to
that of normal pigs. This seems to suggest that both age and state play a
role in determining the status of the animal.
The homeostatic control for carcass protein proposed by Lovatto and

Sauvant (2003) is depicted in Fig. 8.1. At the homeostatic equilibrium (the
solid dots), protein anabolism will be greater than catabolism, resulting in a
net positive protein deposition. Increasing the plasma amino acid
concentration above the equilibrium value increases both protein
anabolism and catabolism, although the former increases to a greater
extent. This mechanism allows the animal to respond quickly to a
changing nutrient supply. The integration step in the model of Lovatto
and Sauvant (2003) was 0.001 days (1.44 min), which is ten times smaller
than that used in the model of Halas et al. (2004). In the latter model, the
supply of acetylCoA may push certain anabolic processes, but it relies on a
constant supply of nutrients. A good challenge for testing metabolite-
driven models is to see how they operate under different conditions of
nutrient input (e.g. frequent meal patterns vs continuous nutrient input).
Perhaps these different situations will result in completely different model
predictions. Consequently, when assuming a constant nutrient input, the
numerical values of model parameters (e.g. Vmax in the model of Halas et
al., 2004) have a meaning only when applied to long-term (daily)
phenomena and most likely have no meaning at the tissue or cellular level.
Where Should We Go from Here with Energy Evaluation?
The shortcomings of current energy evaluation systems are currently

leading to proposals for new systems or models that are based on the
biochemical utilization of energy sources. It is clear that these systems
provide a more solid theoretical basis for feed evaluation. However,
scientific soundness is not sufficient for adopting a new system. Past
history, established knowledge and the use of a reference basis play a major
role in the acceptance or failure of future energy systems. For example, the
concept of maintenance used in many energy systems has been heavily
nutrient flow
equilibrium
nutrient concentration
Fig. 8.1. Homeostatic regulation of amino acid catabolism (solid line) and anabolism (dashed
line) of carcass proteins. At equilibrium nutrient concentration (the solid dots), anabolism and
catabolism are regulated by a time-dependent homeorhetic process (Lovatto and Sauvant,
2003).
158 J. van Milgen
criticized because it does not correspond to a physiological reality for

producing animals. On the other hand, it is sufficiently robust and well-
established that a statement like ‘I offered pigs feed at a level of 10 times
MEm’ will be readily qualified as nonsense. On the other hand, how many
nutritionists are able to evaluate a statement such as ‘a 60-kg pig requires
300 moles of ATP per day’? Expressing requirements and feed values as
‘ATP equivalents’ will probably require considerable time before one is
accustomed to this mode of expression. However, one may say that, by
doing so, we separate some of the ‘known’ (stoichiometry) from the
‘unknown’ (turnover, ion pumping, etc.) and are thus making progress.
However, a quantitatively important part of what we thought was ‘known’
is now considered to be highly questionable. The synthesis of ATP in the
mitochondria from NADH and FADH2 has been considered for a long
time as fixed (i.e. 3 ATP/NADH and 2 FADH2/ATP) and many (older)
biochemistry text books still use these values. It is now more common to
assume lower ATP yields from NADH and FADH2 and, due to uncoupling
of mitochondrial membrane potential from ATP production, it even may
be variable. This means that the ATP yield from most nutrients will be
much lower (e.g. 31 rather than 38 ATP from glucose). Stated otherwise,
we would require more energy from nutrients to provide the same amount
of ATP. This means that the ATP by itself is not a stable currency to express
an energy value. For models that use ATP and calibrate the model for
unaccounted energy expenditure, this finding would lower the ATP
requirement while increasing the heat increment of ATP synthesis.
Apart from changing the reference base of expressing energy values,
there is another, probably more important reason to be conservative in
adopting new systems of energy evaluation. As mentioned earlier, the
classical system reduces energy to a single value. Although too simplistic,
the user can easily manipulate and work with feed values in his
calculations. The system is additive and transparent to the end-user. If a
system is adopted based on biological reality, both the diet and the animal
will determine the energy value of the diet. The only way to deal with the
interaction between the diet and the animal is by using modelling
techniques. It will essentially change the type of question from ‘what does
the animal need?’ to ‘how does the animal respond to a changing nutrient
supply?’. Although there is considerable progress to make by changing to
this type of questioning, it will make the system less transparent to the end-
user. Model developers have to be aware that development of the model
structure is only a very small part of proposing a new system of evaluating
animal feeds. In order to gain confidence, the model should be as
transparent as possible so that end-users can follow the modeller’s logic.
This requires a considerable investment in terms of interface development
and ensuring appropriate user training. Although these new systems
definitely have something to offer, it goes without saying that there is still a
long way to go.
I do not want to leave the reader with the final impression that I feel
there is no use in developing mechanistic models. I do think that there is a
tremendous future for these models, especially when studying nutrient

transfer and interactions between organs. At the cellular level, and
especially in combination with the enormous amount of information
becoming available from genomics and proteomics, mechanistic models
based on stoichiometry will have a very important role to play in
structuring and understanding the data. At the whole animal level,
mechanistic models can be very useful for research and education
purposes. However, in our systems approach, we have to be careful not to
become reductionists only for the purpose of including the cause. The
purpose of many operational nutrition models of animal production and
feed evaluation systems is to control: control growth, control fatness,
control the weight at slaughter, etc. The following example that Pattee
(1997) gives concerning the light in his room may also apply to establishing
applied models of animal nutrition.
The electrical power that provides the light in my room is ultimately caused by
nuclear fission in the sun that drives the water cycle and photosynthesis, or by
nuclear fusion on earth. Many complex machines and complex power
distribution systems are also necessary in the causal chain of events lighting my
room. So why do I think that the cause of the light in my room is my turning
the switch on at the wall? Because that is where I have proximal, focal control,
and also because switching is a simple act that is easy to model, as contrasted
with the complexities of nuclear reactions and power distribution networks.
References
Armstrong, D.G. (1969) Cell bioenergetics and energy metabolism. In: Lenkeit, W.,
Breirem, K. and Crasemann, E. (eds) Handbuch der Tierernährung. Parey,
Hamburg, Germany, pp. 385–414.
Baldwin, R.L. (1995) Modeling Ruminant Digestion and Metabolism. Chapman and
Hall, London.
Birkett, S. and de Lange, K. (2001a) A computational framework for a nutrient
flow representation of energy utilisation by growing monogastric animals.
Birkett, S. and de Lange, K. (2001b) Calibration of a nutrient flow model of energy
utilisation by growing pigs. British Journal of Nutrition 86, 675–689.
Boisen, S. (2000) A simple nutrient-based production model for the growing pig.
In: McNamara, J.P., France, J. and Beever, D.E. (eds) Modelling Nutrient
Utilisation in Farm Animals. CAB International, Wallingford, UK, pp. 183–196.
Boisen, S. and Verstegen, M.W.A. (2000) Developments in the measurement of the
energy content of feed and energy utilisation in animals. In: Moughan, P.J.,
Verstegen, M.W.A. and Visser-Reyneveld, M.I. (eds) Feed Evaluation. Principles
and Practice. Wageningen Pers, Wageningen, Netherlands, pp. 57–76.
Brouwer, E. (1965) Report of sub-committee on constants and factors. In: Blaxter,
K.L. (ed.) Energy Metabolism. Proceedings of the 3rd Symposium held at Troon,
160 J. van Milgen
Scotland, May, 1964. European Association for Animal Production Publication

No. 11. Academic Press, London, pp. 441–443.
Chudy, A. (2000) Model for the interpretation of energy metabolism in farm
animals. In: McNamara, J.P., France, J. and Beever, D.E. (eds) Modelling Nutrient
Utilisation in Farm Animals. CAB International, Wallingford, UK, pp. 329–346.
de Lange, C.F.M. (1995) Framework for a simplified model to demonstrate
principles of nutrient partitioning for growth in the pig. In: Moughan, P.J.,
Wageningen Pers, Wageningen, Netherlands, pp. 71–85.
Emmans, G.C. (1994) Effective energy: a concept of energy utilisation applied
77, 113–130.
Halas, V., Dijkstra, J., Babinszky, L., Verstegen, M.W.A. and Gerrits, W.J.J. (2004)
Modelling of nutrient partitioning in growing pigs to predict their anatomical
body composition. 1. Model description. British Journal of Nutrition 92,
725–734.
Kloareg, M., Le Bellego, L., Mourot, J., Noblet, J. and van Milgen, J. (2005)
Deposition of dietary fatty acids and of de novo synthesised fatty acids in
growing pigs – effects of high ambient temperature and feeding restriction.
Koong, L.J., Nienaber, J.A., Pekas, J.C. and Yen, J.T. (1982) Effects of plane of
nutrition on organ size and fasting heat production in pigs. Journal of Nutrition
112, 1638–1642.
Krebs, H.A. (1972) Some aspects of the regulation of fuel supply in omnivorous
animals. Advances in Enzyme Regulation 10, 397–420.
Le Goff, G. and Noblet, J. (2001) Comparative digestibility of dietary energy and
nutrients in growing pigs and adult sows. Journal of Animal Science 79,
2418–2427.
Le Goff, G., van Milgen, J. and Noblet, J. (2002) Influence of dietary fibre on
digestive utilisation and rate of passage in growing pigs, finishing pigs and
adult sows. Animal Science 74, 503–515.
Lindberg, J.E. and Pedersen, C. (2003) Predictions of energy digestibility in pig
feeds. In: Souffrant, W.B. and Metges, C.C. (eds) Progress in Research on Energy
and Protein Metabolism held at Rostock-Warnemünde, September, 2003. EAAP
publication N° 109. Wageningen Academic, Wageningen, Netherlands,
pp. 201–204.
Lister, D. and McCance, R.A. (1967) Severe undernutrition in growing and adult
animals. 17. The ultimate results of rehabilitation: Pigs. British Journal of
Livesey, G. (1984) The energy equivalents of ATP and the energy values of food
proteins and fats. British Journal of Nutrition 51, 15–28.
Lizardo, R., van Milgen, J., Mourot, J., Noblet, J. and Bonneau, M. (2002) A
nutritional model of fatty acid composition in the growing-finishing pig.
Lobley, G.E. (2002) Protein turnover – what does it mean for animal production?
In: Lapierre, H. and Ouellet, D.R. (eds) Amino Acids: Milk Meat and More! Held
at Québec, Canada, July, 2002. Canadian Society of Animal Science, Québec,
Canada, pp. 1–15.
Lovatto, P.A. and Sauvant, D. (2003) Modeling homeorhetic and homeostatic

controls of pig growth. Journal of Animal Science 81, 683–696.
Moughan, P.J., Smith, W.C. and Pearson, G. (1987) Description and validation of a
model simulating growth in the pig (20–90 kg liveweight). New Zealand Journal
of Agricultural Research 30, 481–489.
Noblet, J. and Perez, J.M. (1993) Prediction of digestibility of nutrients and energy
values of pig diets from chemical analysis. Journal of Animal Science 71,
3389–3398.
Noblet, J., Shi, X.S. and Dubois, S. (1993) Energy cost of standing activity in sows.
Noblet, J., Fortune, H., Shi, X.S. and Dubois, S. (1994) Prediction of net energy
value of feeds for growing pigs. Journal of Animal Science 72, 344–354.
Noblet, J., Karege, C., Dubois, S. and van Milgen, J. (1999) Metabolic utilisation of
energy and maintenance requirements in growing pigs: effects of sex and
genotype. Journal of Animal Science 77, 1208–1216.
Noblet, J., Sève, B. and Jondreville, C. (2002) Valeurs nutritives pour les porcs. In:
Sauvant, D., Perez, J.M. and Tran, G. (eds) Tables de Composition et de Valeur
Nutritive des Matières Premières Destinées aux Animaux d’Élevage. INRA Editions,
Paris, pp. 25–35.
Noblet, J., Bontems, V. and Tran, G. (2003) Estimation de la valeur énergétique des
aliments pour le porc. INRA Production Animales 16, 197–210.
Pattee, H.H. (1997) Causation, control, and the evolution of complexity. In:
Andersen, P.B., Christiansen, P.V., Emmeche, C. and Finnemann, N.O. (eds)
Downward Causation. Minds, Bodies and Matter. Aarhus University Press, Aarhus,
Netherlands, pp. 63–77.
Quiniou, N., Dubois, S. and Noblet, J. (1995) Effect of dietary crude protein level
on protein and energy balances in growing pigs: comparison of two
measurement methods. Livestock Production Science 41, 51–61.
Salway, J.G. (1994) Metabolism at a Glance. Blackwell Science, Oxford, UK.
Schrama, J.W., Verstegen, M.W.A., Verboeket, P.H.J., Schutte, J.B. and Haaksma, J.
(1996) Energy metabolism in relation to physical activity in growing pigs as
affected by type of dietary carbohydrate. Journal of Animal Science 74,
2220–2225.
Schulz, A.R. (1978) Simulation of energy metabolism in the simple-stomached
animal. British Journal of Nutrition 39, 235–254.
Susenbeth, A. and Menke, K.H. (1991) Energy requirements for physical activity in
pigs. In: Wenk, C. and Boessinger, M. (eds) Energy Metabolism of Farm Animals.
Proceedings of the 12th Symposium held at Kartause Ittingen, Switzerland, 1991.
European Association for Animal Production Publication N° 58. Juris Druck
and Verlag, Zürich, pp. 416–419.
Tran, G. and Sauvant, D. (2002). Données chimiques et de valeur nutritive. In:
Sauvant, D., Perez, J.M. and Tran, G. (eds) Tables de Composition et de Valeur
Nutritive des Matières Premières Destinées aux Animaux d’Élevage. INRA Editions,
Paris, pp. 17–24.
van Milgen, J. (2002) Modeling biochemical aspects of energy metabolism in
mammals. Journal of Nutrition 132, 3195–3202.
van Milgen, J., Noblet, J., Dubois, S. and Bernier, J.F. (1997) Dynamic aspects of
oxygen consumption and carbon dioxide production in swine. British Journal of
van Milgen, J., Bernier, J.F., Lecozler, Y., Dubois, S. and Noblet, J. (1998) Major
determinants of fasting heat production and energetic cost of activity in
162 J. van Milgen
growing pigs of different body weight and breed/castration combination. British

Journal of Nutrition 79, 509–517.
van Milgen, J., Quiniou, N. and Noblet, J. (2000) Modelling the relation between
71, 119–130.
van Milgen, J., Noblet, J. and Dubois, S. (2001) Energetic efficiency of starch,
protein, and lipid utilisation in growing pigs. Journal of Nutrition 131,
1309–1318.
van Milgen, J., Noblet, J., Valancogne, A., Dubois, S. and Dourmad, J.Y. (2005)
InraPorc : un modèle pour analyser les performances et évaluer les stratégies
alimentaires chez le porc en croissance. Journées de la Recherche Porcine en France
37, 291–298.
Wellock, I.J., Emmans, G.C. and Kyriazakis, I. (2003) Modelling the effects of
9 Evaluating Animal Genotypes
through Model Inversion
A.B. DOESCHL-WILSON, P.W. KNAP AND B.P. KINGHORN
Sygen International, Scottish Agricultural College, Bush Estates, Penicuik,
Edinburgh, EH26 0PH, UK.
andrea.wilson@sac.ac.uk
Introduction
Specification of the animal’s genetic potential for growth, efficiency and
robustness to environmental stressors is the key to successful livestock
production and breeding. The current practice in breeding is to record
performance traits on individuals and use regression to disentangle the
genetic from the environmental influence (Fig. 9.1, top). Traits considered
are typically related to growth, feed efficiency and, more recently,
robustness to various stressors. This widely used technique has several
shortcomings.
First, as noted by Bourdon (1998), the obtained estimated breeding
values (EBVs) do not represent the ‘true genetic potential’ (the maximum
performance given optimal conditions), because the data were measured
under sub-optimal conditions. Second, regression models are designed to
fit a specific data set rather than to represent the underlying biological
processes. This usually results in useful statistics (EBVs) appropriate for
the prevailing conditions. However, it provides a narrow scope of use; in
particular, empirical models assume simple linear relationships between
(combinations of) individual model components, which can cause problems
when extrapolating to conditions not covered in the data. In order to
avoid unexpected poor performance in environments that differ from the
data conditions, the covariances required for the EBVs need to be re-
estimated whenever the production conditions change.
Illustrating the genotype-phenotype relationship is Fig. 9.1, with
phenotypic performance at one end of the spectrum and the genotype at
the opposite end; mechanistic animal growth models represent this
relationship in the reverse direction to the mathematical models currently

164 A.B. Doeschl-Wilson et al.
Genes Genotype
specification Mathematical methods Measurable traits
Traditional method of pig breeders

Recorded phenotypic
performance
Estimated genetic
components of Multi-variate regression, Recorded environmental
phenotypic traits covariance estimates, BLUP conditions
(EBVs)
Recorded pedigree
information
Mechanistic growth model
Simulation model
Genetic potential
environmental, Phenotypic
for growth, Potential × dietary and performance
energy efficiency performance physiological
and robustness conditions
Model Model
input output
Fig. 9.1. Illustration of the different approaches for specifying genotypes.
used in breeding. Mechanistic models use as inputs descriptions of the

genetic potential for growth and efficiency in optimal conditions, together
with descriptions of the environmental and dietary conditions. Genotypic
and environmental specifications are assumed mutually independent. The
growth model then simulates the (non-linear) interaction between
environmental conditions and genetic potential. Model outputs are
phenotypic predictions of the performance traits. By contrast, phenotypic
data serve as inputs for the statistical models normally used in animal
breeding.
The description of the genotype in mechanistic growth models is free
from environmental influences, which makes it theoretically more closely
associated with the true genotype than the phenotypic traits. As a
consequence of the explicit description of genotype by environment
interactions in mechanistic models, these can provide for proper
extrapolation outside the data. As Black (1995) states, ‘mechanistic pig
growth models have proved great value to research and industry as they
combine the present knowledge about the influencing factors of pig
growth and predict performance where knowledge is limited’.
All this raises the question of why current methods in breeding are still
focusing on phenotypic performance traits instead of the more adequate
Evaluating Animal Genotypes through Model Inversion 165
‘universal description of the genotype’ (Bourdon, 1998) of mechanistic

growth models. The main reason is that the underlying physiological traits
required for growth model parameterization are difficult to measure.
However, accurate estimates for the underlying trait levels are not only
essential for accurate description of the genotype, but also for accurate
model predictions, and thus for the appropriate application of growth
models.
The genotypic model parameters intend to describe the pig’s intrinsic
ability to grow and cope with stressors. Luiting and Knap (Chapter 14, this
volume) list the growth descriptors of the main pig growth models in
current use. Most of these models simulate growth performance in a non-
limiting environment with two to five genotypic parameters. The
prediction of performance under the influence of physical, social, climatic
or pathological stressors requires additional parameters describing the
ability to cope with these (see for example Wellock et al., 2003b; Kyriazakis
et al., Chapter 7, this volume).
The number of genotypic model parameters depends also on the
purpose of the model. Models that predict the performance of an
individual animal or of a population average describe the genotype by one
parameter for each of the underlying traits. Models that simulate a
population of distinct individuals require parameters representing the
means of and covariances among the underlying traits (Knap, 2000b;
Pomar et al., 2003; Wellock et al., 2004). If various generations are
involved, such as in problems in pig breeding, heritabilities of the
underlying traits are of additional interest. Thus, every underlying trait
requires several model parameters (mean, covariances and heritability),
which can add up to a large number of parameters to be determined.
Methods for Estimating Genotypic Specifications
Knap et al. (2003) proposed two methods for estimating genotypic model
parameters: (i) direct measurements of the underlying traits represented
by the model parameters, combined with controlled environmental
conditions; and (ii) serial measurement of performance traits and fitting
the data to the mechanistic growth model itself (Model inversion).
Using direct measurements of traits represented by the genotypic model

parameters
Ferguson and Gous (1993a) described techniques to obtain estimates of the

parameters of the Gompertz equation, which they use to model the
potential rate of body protein growth and, from that, potential rates for
other components using allometry. These involve serial slaughter trials
with measurement of chemical body composition. The crux of these trials
is to minimize environmental load, maximizing the chance that the growth
potential can be achieved. Dietary and environmental conditions must be

carefully balanced, so that protein retention can reach its maximum
without causing lipid retention to exceed the animal’s intrinsic desire.
Knap et al. (2003) proposed a controlled experiment that would allow the
genetic potential for protein and lipid retention, and for intrinsic
maintenance energy requirements, to be fully expressed either at different
growth stages or in different individuals from the genotype. Such an
experiment would enable direct measurement of traits whose extremes are
unlikely to be reached in normal conditions.
Such techniques have provided estimates for population means of model
parameters. Direct estimates for population means of Gompertz curve
parameters for protein and lipid deposition and of maintenance energy
requirements in pigs were also obtained by Knap (2000a), who analysed
serial slaughter data from the literature for a wide range of genotypes, and
by Landgraf et al. (2002), who used deuterium dilution techniques and
computer tomography in vivo combined with chemical analysis.
The estimation of the within-population variation of these parameters
requires serial measurement of the related traits in several animals over a
significantly wide body weight range, as demonstrated by results from
computer tomography and deuterium dilution measurements (Knap et al.,
2003). These data indicate a substantial variation between animals in the
underlying traits of the growth model used by these authors (coefficients of
variation ranged between 8 and 27%), which was interpreted to be of
genetic nature.
The estimation of heritabilities of underlying traits requires serial
measurement of the related traits in several animals with a known family
structure. Roehe et al. (2002) exemplified this method by estimating
heritabilities of Gompertz parameters for body weight growth.
The combined results of these studies suggest that it is possible to
specify genotypic model parameters from direct experimental measure-
ment, but that it requires complex experimental design and expensive,
time consuming measurement techniques.
Fitting phenotypic performance trait measurements to the growth model

(model inversion)
The second method proposed by Knap et al. (2003) involves, instead of

complex experimental design for direct measurement of underlying traits,
readily obtained performance data and the growth model itself. Model
inversion builds upon the idea that phenotypic performance data can be
used to estimate those particular values of the genotypic model parameters
that produce model predictions for the performance traits that match the
observations. The model is ‘inverted’ in the sense that the conventional
model input traits are treated as model outputs that need to be determined
in the inversion process, and the conventional model output traits are
treated as inputs through substitution of performance data.
Knap et al. (2003) mention two alternative methods of performing the

inversion process. The first is an iterative process, first described by
Baldwin (1976), which is essentially a trial and error approach, in which
the model parameters are systematically adjusted so that the
corresponding model predictions approach and eventually give a best fit to
the performance data. We refer to this process as model inversion by
optimization, and will deal with it below.
The second method involves rearrangement of the model equations so
that, upon substitution of observations into the model output variables, the
resulting equation system can be solved for the genotypic model
parameters. This we call algebraic model inversion.
Algebraic Model Inversion
Assuming that the growth model is able to reproduce performance data,

the possibility arises of substituting these observations for the model
predictions and solving the system for the genotypic parameter values.
Most mechanistic growth models use algebraic equations to relate genotype
and environmental conditions to observable production traits. These
equations must then be rearranged so that the system of equations can be
solved for the unknown parameters.
Compared to numerical optimization routines, methods for calculating
unique, exact solutions are more robust as they do not require subjective
criteria that determine when an approximate solution is acceptable and
which solution is classified as optimal. Animal growth models that use
relatively simple algebraic expressions to describe the growth mechanism
may lend themselves to this method for determining appropriate
parameter values.
However, unique exact solutions of a system of algebraic equations
only exist if certain assumptions are met. This section presents the
conditions for the algebraic invertibility of mechanistic growth models, and
demonstrates how the theory applies to a specific pig growth model. The
model of de Lange (1995) was originally prepared for educational
purposes, and was later used as the basis for an attempt to algebraically
invert a model to relate pig genotype to nutritional requirements
(unpublished results; see www.defra.gov.uk/science/LINK/agriculture for
background information). The case study below represents further
development of that work.
Mathematical Theory for Algebraic Model Inversion
A prerequisite for determining genotypic parameters by algebraic

inversion is a one-to-one relationship between model input and model
output as expressed by the model equations. If various combinations of
input parameters produce the same predictions, or if several predictions
are possible for a given set of input parameters, the inversion process
cannot produce unique values for the input parameters. All mechanistic
pig growth models fall into this category. Stochastic models, for example,
which integrate a random component, cannot be inverted by this method.
Most pig growth models are a composite of several algebraic equations
with an intrinsic hierarchical structure. This prevents a straightforward
reversion of the multiple modelling steps. It is therefore desirable to
represent the entire model by one system of equations with the model
input parameters as unknowns and the observable model outputs assumed
as known. The inversion process then corresponds to solving the system of
equations for the genotypic parameters for given values of predictable
performance traits. Exact solutions of a system of algebraic equations
require that the system is exactly determined, i.e. that the number of
variables to be solved for equals the number of equations.
The Implicit Function Theorem of multivariate calculus
Mechanistic growth models are generally non-linear. The conditions for

the existence of a unique solution of a non-linear system of equations are
given by the Implicit Function Theorem of multivariate calculus (see for
example Krantz, 1951). A local unique solution of the equations system
exists if the local linear approximation of the system has a unique solution.
The theorem provides further information about the solution space by
stating that solutions corresponding to similar performance have the same
form, i.e. can be expressed through the same multivariate function. The
term ‘local’ means that the existence, uniqueness and unique expression of
the solutions hold only within a limited region in the parameter space, it
does not guarantee that the specific solution is the only solution in the
entire parameter space. The restriction from global to local properties is a
generic characteristic of non-linear problems, which prevails for every
method of calibrating model parameters.
Representing Pig Growth Models by a System of Algebraic

Equations
Most mechanistic growth models share properties that determine whether
they can be represented by a system of equations that can be solved for the
genotypic parameters, and which simplifications are necessary to derive
such a system.
Reduction to few equations
Most pig growth models represent the genotype of the pig by two to five
parameters. This implies that for model inversion, the model should be
reformulated to a system with two to five equations and the genotypic

parameters as the only unknowns. Other model input parameters, for
which no measurements exist, must be either eliminated from the system
or described empirically.
Recurrence relations and long time spans between observations
Dynamic growth models describe growth through time, but genotypic

parameters are generally assumed to be time invariant. Although dynamic
systems are most accurately represented by differential equations (for
example, see Parks, 1982), most growth models approximate the growth
changes by first order difference equations, which describe the change of
model variables between two specific points in time (usually 1 day apart) or
two specific body weight values. The estimation of the time invariant
genotypic parameters through algebraic model inversion then requires
information about the animal’s performance at two instances of time (or
weight).
Many growth models use a 1-day time step, but daily observations of
performance traits are rare. Using observations over a longer period than
the model’s therefore requires either reformulation of the model equations
(e.g. cumulative instead of daily feed intake), or an interpolation of the
production traits for the model’s time interval from longer-term
observations. In either approach, short-term environmental fluctuations
cannot be accounted for, which reduces the accuracies of the calculated
parameters.
Often, more than two measurements of performance traits are
available. The use of repeated measurements of performance traits for
deriving genotypic parameters would result in different equation systems
for different periods, and thus most likely to different values of the
genotypic parameters for different growth phases. These would need to be
combined to an individual value to match the model assumptions of time-
invariant genotypic parameters.
Hierarchical model structure
Animal growth is the result of various mechanical and biochemical

processes, many of which are modelled sequentially. For example, most
growth models give priority to the satisfaction of body maintenance
requirements over processes related to growth. Reformulating sequential
processes to one system of equations is not straightforward as the processes
generally depend on conditions specified at intermediate steps in the
model. For example, most models include a process that reduces the
potential rates of protein and lipid retention if the environmental or
dietary conditions are found limiting. Other examples of events triggered
under certain conditions are thermoregulatory processes, which come only
into play when the ambient temperature is outside the pig’s comfort zone
(Bruce and Clark, 1979; Black et al., 1986; Knap, 2000b; Wellock et al.,
2003a), and abrupt changes in the rates of lipid retention under
malnutrition (Green and Whittemore, 2003; Wellock et al., 2003a).
Since many of these criteria depend on intermediate model results and
are therefore not known a priori, the representation of the model by a
system of equations suited for algebraic model inversion can become
problematic. Model inversion may thus require prior assumptions about
the conditions for which the modelled responses divide. These
assumptions should be based on the conditions from which the data are
obtained, and need to be re-validated for the calculated genotypic values
once the inversion has been carried out.
Implemented thresholds
Implemented thresholds cause problems for algebraic model inversion as

different parameter values can produce the same function value and as the
abrupt slope change at the location where the threshold is reached violates
the assumption of differentiability in the Implicit Function Theorem.
Implemented thresholds could represent, for example, a maximum value
for protein deposition, a minimum ratio of lipid to protein retention or a
critical value of net efficiency for using ideal protein for protein retention
(see Luiting and Knap, Chapter 14, this volume).
Two simple alternatives are proposed to overcome these problems.
First, to restrict the domain to one side of the threshold values, based on
the nature of the conditions from which the data are obtained. For
example, for an experiment involving restricted feed intake, the
assumption that the maximum protein retention rate was not achieved by
the animal would be appropriate. Second, more elegantly, to replace the
thresholds by mathematical functions that gradually approach an
asymptote, as exemplified through the substitution of a linear plateau
function for protein retention (Whittemore and Fawcett, 1974) by a
Gompertz function (Emmans, 1988).
The above described generic properties of pig growth models
demonstrate the need for model simplification before algebraic model
inversion can be tackled. Functions with inappropriate mathematical
properties could be replaced by more appropriate ones, or diverse a priori
assumptions must be made, whose validity can only be justified after the
model inversion has been carried out.
Solving the System of Equations
The Implicit Function Theorem states the conditions for the existence of a
unique solution, but does not provide an explicit expression of the
solution. In fact, only few non-linear systems of equations can be solved by
analytical means. Most systems require a numerical solver.

Finding the solutions of the corresponding system of equations is a
mathematical problem. There is no guarantee that the solutions are
biologically realistic. But if the estimates for genotypic parameters from
algebraic inversion differ greatly from those expected from empirical
studies, fundamental flaws in the model equations or in the prior
assumptions must be suspected.
Case Study 1: Algebraic Inversion of a Pig Growth Model
The model of de Lange (1995), which was previously used to estimate

genotypic parameters from given diet and food intake data (www.defra.
gov.uk/science/LINK/agriculture), is a simple mechanistic, deterministic pig
growth model that contains all the features that should be included to
predict with reasonable accuracy the performance of pigs under defined
conditions. It has two genotypic input parameters LPmin, representing a
minimal lipid to protein ratio, and Prmax, the animal’s upper limit to
protein retention. Both parameters are assumed time-invariant. In
addition to estimates of these parameters, the model requires as input the
body lipid and protein masses at a time t0 and information about food
intake and dietary composition. The model then recursively predicts the
pig’s protein and lipid contents together with the body weight at a future
time.
Derivation of the system of equations
Under the assumption that the body protein and lipid mass of the pig at
time t0 are known and the required dietary information is available, the
hierarchical model was reformulated into systems of two equations, from
which unique values for LPmin and Prmax were obtained. The derivation
was built upon the assumptions of the original research project, i.e. that
information about feed intake, diet composition, body weight and backfat
depth were available and that initial body composition could be estimated
from initial body weight.
A system of equations with unique solutions for LPmin and Prmax could
only be obtained if the two constraints C1 and C2 below were satisfied. The
final Eqns 9.1a and 9.1b contain LPmin and Prmax as the only unknowns. All
other components are expressed in terms of initial and final body weight
(W0 and W1, respectively) or final backfat depth (BF1).
0.95 (W1 – W0) / dt = (0.001 × ((1.189 – E) × PRmax + B(W0))
+ 4.889 × (0.95 W0 / (A(W0) + LPmin) + 0.001 PRmax) 0.855
– 4.889 × (0.95 W0 / (A(W0) + LPmin) ) 0.855 (9.1a)

0.001dt × (B(W0) – E × PRmax) = 8.244 + 2.053 BF1 – 0.95 LPmin ×

W0 / (A(W0) + LPmin) (9.1b)
with constraints
BPg (W0) PRmax (C1)
B(W0) (LPmin + E) × PRmax (C2)
where dt denotes the time span t1 – t0 between the measurements W0 and

W1 and the expressions E, BPg (W), A(W) and B(W) are defined as follows:
E = (Epd – Epe) / Eld,
A(W) = 1.189 + 4.889 × (1.4 + 0.15 × W) –0.145 ,
BPg (W) = 0.85 × [ minimum ( API , ALI / LysBalP ) – 0.9375 W 0.75],
B(W) = 1/ Eld × [EPFI + (Ep – Epe) × API – (0.9375 × (Ep – Epe) + 458) × W0.75].
where API = FI DProt AAa is the available protein intake (g/day),
ALI = FI x DLys AAa the available lysine intake (g/day) and EPFI = FI
DDE Ep API is the protein-free digestible energy intake (kJ/day).
Food intake (FI, kg/day) values and the description and values of the
remaining constants are specified in Tables 9.2. and 9.1. respectively.
Solving the equations
Weekly least square means of one of three genotypes from a previous

growth trial (Green et al., 2003; Whittemore et al., 2003) were used for this
case study, providing information on nutrient intake and repeated
measurements of body weight and backfat depth for a weight range
between 25 and 115 kg (Table 9.2). Dietary specifications are in Table 9.1.
For each of the 12 weeks, the weekly least square means of food intake,
body weight and backfat depth were substituted into system (Eqns 9.1a, 9.1b),
yielding different equation systems for every week. The conditions of the
Table 9.1. Constants of the model. Except for the dietary constants, DProt, DLys and DDE,
which were specified in Whittemore et al. (2003), all other constants are taken from de Lange
(1995).
Constants Explanation Estimated value
Aaa Post absorptive efficiency for utilizing amino acids and
dietary protein for growth 0.85
Dprot Dietary protein content (g/kg) 194
Dlys Dietary lysine content (g/kg) 11.4
Ep Gross energy content of protein (kJ/g) 23.6
DDE Dietary energy content (kJ/g) 14.5
LysBalP Lysine content of balanced protein (%) 7
Epe Energy cost of available protein excretion (kJ/g) 12.1
Epd Energy cost for protein deposition (kJ/g) 43.9
Eld Energy cost for lipid deposition (kJ/g) 52.8
Table 9.2. Weekly least square means for daily food intake, body weight and
backfat depth used for the inversion of the model. The values are the results of the
statistical models for ‘Landrace type’ pigs, published in Green et al. (2003).
Week Daily food intake (kg) Body weight (kg) Backfat depth (mm)
1 1.29 30.89 5.30

2 1.48 35.41 5.77
3 1.68 40.14 6.39
4 1.87 45.10 7.04
5 2.05 50.28 7.72
6 2.22 55.70 8.43
7 2.38 61.36 9.17
8 2.53 67.26 9.94
9 2.66 73.40 10.75
10 2.77 79.80 11.58
11 2.86 86.46 12.46
12 2.93 93.37 13.36
13 100.56 14.30
Table 9.3. Weekly estimates of the genotypic parameters PRmax and LPmin from algebraic
model inversion (columns 2 and 3) and inversion through optimization (columns 4 and 5).
RMSD is the root of the mean of the squared deviations of predicted body weight and backfat
depths from the observations (Table 9.2) using the optimized parameter values.
Week PRmax LPmin (alg. PRmax LPmin RMSD
(alg. inversion) inversion) (optimization) (optimization) (optimization)
1 113.5 0.531 101.5 0.529 < 1018

2 114.7 0.642 102.6 0.641 < 10–18
3 116.0 0.732 104.5 0.732 8.9 × 1016
4 117.5 0.810 106.8 0.811 < 1018
5 120.3 0.880 110.5 0.881 –1.4 × 1014
6 123.8 0.941 115.0 0.942 < 1018
7 128.0 0.996 120.3 0.997 < 1018
8 132.9 1.046 126.4 1.049 < 1018
9 140.0 1.094 134.9 1.095 –1.2 × 1014
10 148.6 1.139 144.7 1.141 < 1018
11 158.3 1.181 155.8 1.184 < 1018
12 170.7 1.222 169.7 1.225 –3.6 × 1015
1–12 21.4 –1.895 111.8 1.129 5.70
Implicit Function Theorem for the existence of a unique solution with the
weekly least square means were satisfied. The numerical solver fzero of Matlab
(Matlab 6.5, 2002) was used to calculate values for PRmax and LPmin as
solutions of the corresponding systems of equations. The weekly solutions are
in Table 9.3 (columns 1 to 3). All solutions satisfied constraints C1 and C2.
The value for PRmax obtained through inversion of the de Lange
model ranged between 114 and 171 g/day (Table 9.3), which agrees with
estimates provided in the literature, i.e. 99 g/day (SE 4 g/day) to 212 g/day
(SE 29 g/day) (Knap, 2000a). Literature estimates for the lipid to protein
ratio at maturity vary between 0.97 (SE 0.28) and 5.16 (SE 0.50). In
comparison with these values, the range between 0.53 and 1.22 for the
estimates for the minimal lipid to protein ratio obtained from the algebraic
model inversion also appears realistic. In the last row of Table 9.3,
solutions corresponding to a time span of 12 weeks are shown. The
calculated values are unrealistic, demonstrating that algebraic model
inversion fails if observations are far apart.
Although the solutions of PRmax and LPmin appear realistic, they are
not, as assumed, constant over time. Over the 12 weeks, PRmax and LPmin
both increase systematically. This apparent time trend contradicts the
assumption that genotypic parameters are time invariant. Agreement of
the model predictions for body weight and backfat depths with the
observed least square means is only possible if the genotypic parameters
are allowed to vary over time, which indicates that the growth trends for
body weight and backfat depth predicted by the model differ from the
observed ones. This suspicion is confirmed in Fig. 9.2, which plots those
growth curves corresponding to fixed values of the genotypic parameters
together with the least square means from the data. For the parameter
combinations considered, the predicted backfat depth curves differ
strongly from the data. Sensitivity analysis indicates that no combination of
genotypic parameters can produce a trend similar to the data.
Influence of different types of errors on the solutions
Although the genotypic parameter estimates are accurate solutions of the

derived equation system, they may nevertheless not be an accurate
description of the genotype. The algebraic inversion process integrates
various sources of error, which must be taken into account when interpreting
the results. The most significant types of error for algebraic model inversion
are: (i) modelization errors; and (ii) errors from simplifications of the model
equations and from inverting regression equations.
Realistic time invariant values for the genotypic parameters can only be
obtained if the model can reproduce observed trends. This model could not
generate backfat depth predictions that matched the data over a time span
of 12 weeks. The predicted trends associated with various starting values for
genotypic parameters should be validated before calculating exact estimates
of genotypic parameters through model inversion. In the inversion process,
differences between predicted and observed growth trends show up in a
progressive change of the estimated parameters with time.
Most mechanistic growth models incorporate the results from
empirical studies through regression equations. The algebraic inversion
process often requires rearrangement of these, and thus a change of the
role of the dependent and independent regression variables. However,
swapping these roles in a least squares approach results in different
equations from those obtained by solving the initial regression equation for
the independent variable, and this introduces more errors.
(a)
110
observed LSM
100 algebraic inversion, week 12
algebraic inversion, week 6
inversion through optimization, weeks 1–12
90 inversion through optimization, week 12
Body weight (kg)

80
70
60
50
40
30
0 1 2 3 4 5 6 7 8 9 10 11 12 13
Week
(b)
16
observed LSM
algebraic inversion, week 1
14 algebraic inversion, week 12
inversion through optimization, weeks 1–12
inversion through optimization, week 12
Backfat depth (mm)
12
10
4
0 1 2 3 4 5 6 7 8 9 10 11 12 13
Week
Fig. 9.2. (a) Predicted body weight for different solutions for the genotypic parameters of the
model. The estimates of the genotypic parameters used to generate these plots are
presented in Table 9.3. (b) Predicted backfat depth for different solutions for the genotypic
parameters of the model. The estimates of the genotypic parameters used to generate these
plots are presented in Table 9.3.
Model Inversion through Optimization
Algebraic model inversion relies upon the assumption that a unique set of
model input parameters exists for which the model predictions equal the
data. Uncertainties and errors in the model itself or in the data are
ignored. Also, if the number of parameters to be determined through
inversion exceeds the number of parameters for which a unique solution
of the inverse model exists, the algebraic approach forces the modeller to
restrict the number of unknowns by guessing the values of some
parameters. This restriction may lead to poor estimates for the remaining
parameters, since they are dependent on these guesses.
By treating the estimation of genotypic parameters by model inversion
as an optimization problem, various sources of uncertainty can be taken
into account. Instead of determining the set of genotypic parameters for
which the model predictions exactly match the data, inversion through
optimization determines the set of parameters for which the discrepancy
between predictions and data is minimal.
Description of the process
Inversion by optimization is an iterative process, and as such very different

from algebraic model inversion. The process starts with an initial guess for
the unknown model parameters. The model then calculates the
predictions associated with these parameter values. In contrast to algebraic
inversion, which inverts the model equations, no change in the growth
model is required. The optimization criterion involves a quantitative measure
(such as sum of squares of deviations between predictions and data) that
describes the discrepancy between predictions and data. The best set of
model parameters has been specified when this measure is minimal.
Otherwise, the optimization algorithm determines the next set of model
parameters, for which the process is repeated until a minimum is reached.
The process is similar to a trial and error approach, but the optimization
algorithm searches through the parameter space non-exhaustively but
methodologically, so that the discrepancy gradually decreases until a
minimum is reached.
The minimal discrepancy between predictions and data is not known a
priori. The minimum is assumed to be achieved when the ‘minimal
discrepancy’ set of parameters does not change over many iterations, and
the algorithm is then said to have converged.
Common problems
Optimization algorithms differ in their methods for searching through

parameter space, and the most appropriate method differs from one
problem to another. Many optimization algorithms have been proposed in
the mathematical, computational and applied literature. A review of

existing algorithms and their application to agricultural models is provided
by Mayer et al. (2005) and by Green and Parsons (Chapter 15, this
volume).
The challenge for model inversion by optimization is not only in
determining the most appropriate optimization algorithm for the specific
model, but also in establishing an appropriate optimization criterion. Various
combinations of genotypic parameters may correspond to the same or
similar discrepancies between predictions and data. In contrast to algebraic
model inversion, the conditions for a unique ‘minimum discrepancy’
parameter set are not mathematically specified. It is generally true that a
higher number of independent observations contributing to the optimization
criterion increase the chance of obtaining unique parameter estimates, but
certainty can only be achieved by response surface analysis and by repeating
the process for various starting values for the input parameters.
Case Study 2: Estimating the Genotypic Parameters of the de

Lange Model by Inversion through Optimization
We applied the differential evolution algorithm (Price and Storn, 1997;
Mayer et al., 2005) to obtain estimates for the genotypic parameters LPmin
and PDmax of the de Lange model.
Weekly estimates for LPmin and PDmax were calculated through the
optimization procedure, using the least square means of body weight as
initial values. Daily food intake was estimated from weekly food intake
(Table 9.2). The optimization algorithm calculated the weekly estimates for
LPmin and PDmax that correspond to weekly model predictions for body
weight and backfat with minimal deviation from the data in Table 9.2. The
deviation between observed and predicted body weight and backfat depth
was represented by the root mean square deviation (RMSD).
The weekly estimates for LPmin and PDmax, and the RMSD are in Table
9.3, columns 4 to 6. The LPmin estimates from inversion by optimization are
very similar to the ones from algebraic inversion, whereas some of the weekly
estimates for PRmax are up to 10% lower. The differences in the parameter
estimates are probably due to the fact that, in the optimization process, the
growth model operates on a daily basis, whereas in the algebraic approach,
model equations were rearranged to represent 1 week. Taking the differences
in the approaches into account, their genotypic parameter estimates are
similar, which increases confidence in the validity of either method.
Specifying time invariant values for the genotypic parameters
In contrast to the algebraic model inversion, which could only take

observations corresponding to two different instances in time into account,
observations of body weight and backfat depth for all 12 weeks can be
included in the optimization criterion. The optimization method then

produces one estimate for each genotypic parameter that corresponds to
the minimal discrepancy between predictions and data during the entire
12-week growth period. The LPmin and PRmax estimates are 1.129 and
111.8, with a RMSD significantly higher than the RMSDs associated with
individual weeks (Table 9.3). Fig. 9.2a shows that the corresponding body
weight growth curve is a good approximation of the curve produced by the
statistical model. The fit of the corresponding backfat depth curve to the
observed least squares means (Fig. 9.2b) is however very poor, pointing
again to inappropriate model equations for calculating backfat depth.
Case Study 3: a More Complex Problem for Model Inversion by

Optimization
The Pig Genetic Growth Model (PGGM), developed by Knap (1999), is a
stochastic mechanistic growth model for a population of pigs. It was
designed to predict the effects of different environmental conditions
(performance testing regimes) on the reliability of estimating the genetic
potential for growth rate, feed efficiency and body composition from
current performance traits. Model parameters describing genetic growth
potential are assumed to represent physiological traits with additive genetic
and permanent environmental components.
The four physiological traits specifying the genotype
The model uses four traits to describe the pig’s intrinsic capacity for
growth and energy utilization. The genetic potentials for protein and lipid
growth are modelled by Gompertz functions and full allometry between
body protein and lipid is assumed. The description of the genotype’s
growth potential then requires three independent parameters: mature
protein mass (Pmat), the ratio between mature lipid and protein mass
(LPmat) and the standardized rate parameter B* = BGomp × Pmat0.27.
The resource-demanding processes other than protein and lipid
growth are categorized as ‘maintenance processes’. The current version of
the model explicitly deals with the energy requirements for protein
turnover and thermoregulation, which depend on body composition. The
requirements for all other maintenance processes, including physiological
service functions, are combined into a single parameter (MEmaint)
calculated as a simple function of metabolic body weight: (MEm0 + b.BW)
× BW0.75, where MEm0 is the fourth genotype-specific model parameter.
Introducing variation between individual animals
The phenotypic value for each of the physiological traits can be de-
composed according to
Pij = µ + Ai + PEi + eij (9.2)
where Pij is the phenotype of animal i on day j, µ is the population mean for
the trait, and Ai, PEi and eij are its additive genetic and permanent
environmental deviations, and its random environmental deviation on day j.
The genetic deviation is determined according to
A i = rannorA i × h2 × σ P (9.3)
where rannorAi is a random deviation drawn from the standard Normal

distribution, h2 is the heritability of the trait and σP its phenotypic standard
deviation.
The permanent environmental deviation is determined according to
PEi = rannorPE i × 1 − h2 × σ P (9.4)
with rannorPEi also being a random deviation drawn from the standard
Normal distribution. The random environmental variation is modelled
according to
e ij = Ri j × σ e (9.5)
where R represents either a random drawing from the standard normal

distribution or, more realistically, an autoregressive function involving
random drawings.
The generated variation in the driving model parameters between
individual animals of the population leads to variation in (and covariation
between) model output traits such as growth rate, feed intake and body
composition.
This formulation implies that the same model form applies to both
genetic and environmental effects. In truth, these are likely to differ.
However, power to discover such individual underlying growth models is
likely to be very poor, so that the assumption of a single model is probably
inevitable.
Model inversion: methods
The specification of four physiological traits (Pmat, LPmat, B* and MEm0),

which describe the genotype of the pig according to Eqn 9.1, each with its
population mean, heritability and phenotypic standard deviation, leads to
4 × 3 = 12 model parameters to be determined. The parameters are
assumed as mutually uncorrelated, so that the covariances between them
can be set to zero. Empirical estimates exist for population means and
standard deviations of the Gompertz growth traits (Knap et al., 2003), but
not for their heritabilities nor for the trait describing maintenance
requirements. Model inversion is thus the only possibility for obtaining

these estimates.
Determining the 12 parameters using inversion through optimization
presents a 12-dimensional optimization problem. The differential
evolution algorithm above was used due to its efficiency with high-
dimensional complex optimization problems (Mayer et al., 2005).
The data used as criteria for estimating the population means, standard
deviations and heritabilities for the four genotypic traits were phenotypic
means, genetic correlations and heritabilities for the three performance
traits: days to reach 110 kg body weight (DAYS), average daily feed intake
(DFI) and backfat depth (BF). These were estimated from data of a PIC pig
sireline measured at 110 kg body weight. The statistical estimates for these
multiple traits constitute multiple objectives for the parameter calibration,
which need to be simultaneously achieved. Various methods exist to
generate solutions that simultaneously satisfy multiple goals, some producing
a continuous solution space, others individual solutions (e.g. Steuer, 1986).
We used a combined objective of minimizing the sum of squared relative
deviations of the phenotypic means, genetic covariances and heritabilities
derived from model predictions and from data. Alternative objectives, such
as minimizing the maximum of the relative deviations or applying
punishment functions to the deviations produced similar results.
The stochastic nature of the model implies that its predictions are
influenced by the specific random drawings used in the simulation. Two
simulations with identical values for all input parameters will produce
different predictions and possibly different covariance estimates. The
influence of individual random drawings can be reduced by increasing the
number of replicates in the simulated population, but this also increases
model runtime and the time of statistical calculations. In the optimization
process the model is called many thousands of times and statistical results
must be produced for every run, so computing time is a real issue. Fig. 9.3
gives the runtime of the optimization process, relative to population size,
together with the average coefficient of variation (CV) of the covariances
and heritabilities of the predictions used in the optimization criterion. The
results given below are based on 7000 replicates (20 days runtime on a
standard PC, CV < 10%).
The model inversion aimed to specify the genotypic parameters
associated with a PIC pig sireline. Table 9.4a shows the characterization of
this genotype in terms of empirical estimates of the heritabilities,
correlations and phenotypic means of the three performance traits used in
the optimization procedure.
Model inversion: results
Estimates for population means, CVs and heritabilities for the four
underlying traits, produced by four optimization runs with different
random number sequences and different starting points for the
100000 0.30
0.25
10000
Runtime (days)
0.20
CV of solutions
1000
runtime
0.15
100
0.10
10
0.05
CV of soutions
0 0.00
0 10000 20000 30000 40000 50000
Population size
Fig. 9.3. Runtime for the optimization process and average coefficient of variation (CV) for
the calculated covariances and heritabilities for the predicted performance traits as affected
by population size.
optimization algorithm are shown in Table 9.5. The solutions of runs C

and D are very similar, whereas runs A and B yield substantially different
results for some of the parameters. Estimates for population averages in
Pmat vary from 37 to 56.5 kg, in LPmat from 0.65 to 0.89, and in B* from
0.029 to 0.035 kg/day per kg0.73. Except for run B, which gives an average
MEm0 estimate of 506 kJ/kg0.75 per day, three runs produce similar
estimates between 638 and 652 kJ/kg0.75 per day.
Knap (2000a) and Knap et al. (2003) estimated average values for Pmat,
LPmat and B* from earlier published data on five widely different pig
genotypes. Their Pmat estimates range from 20.0 to 40.7 kg, lower than our
present estimates. In contrast, their LPmat estimates range from 0.97 to
5.16, higher than ours. Their B* estimates range from 0.022 to 0.044
kg/day per kg0.73, similar to ours. The genotype simulated here is
genetically more advanced than the ones in the earlier studies, so we
would expect the above Pmat and LPmat differences; in fact, the B* values
would be expected to be higher, too.
The coefficients of variation produced by model inversion are all
between 3 and 12%, which agrees with the estimates obtained by Knap
(2000b), but are below the empirical estimates reported in Knap et al.
(2003), which range between 14 and 27%.
Estimated heritabilities for the four physiological traits vary among the
optimization runs, but exhibit a consistent pattern. In all four runs, LPmat
has the highest heritability. The heritability for MEm0 is much lower than
the heritability of the other three traits. These estimates are slightly lower
Table 9.4. Genetic correlations (upper triangle of the white area), heritabilities (on
the diagonal) and phenotypic correlations (lower triangle of the white area),
together with phenotypic means (grey area) for a PIC Sire Line as predicted from
data analysis and from the growth model combined with optimization routines from
four different simulations.
(a) Genotype specifications in run A.

Days to 110 kg DFI (kg/day) BF (mm)
Days to 110 kg Data 0.396 –0.725 –0.042
Model 0.364 –0.697 –0.044
DFI Data –0.435 0.273 0.405
Model –0.853 0.270 0.396
BF Data –0.044 0.237 0.550
Model –0.016 0.246 0.554
Phenotyp. means Data 177 2.39 10.3
Model 175 2.36 10.25
(b) Genotype specifications in run B.

Days to 110 kg Data 0.396 –0.725 –0.042
Model 0.393 –0.720 –0.043
DFI Data –0.435 0.273 0.405
Model –0.694 0.275 0.409
BF Data –0.044 0.237 0.550
Model 0.033 0.266 0.563
Model 175 2.36 10.27
(c) Genotype specifications in run C.

Days to 110 kg Data 0.396 –0.725 –0.042
Model 0.384 –0.711 –0.042
DFI Data –0.435 0.273 0.405
Model –0.896 0.280 0.404
BF Data –0.044 0.237 0.550
Model 0.007 0.274 0.536
Model 176 2.32 10.52
(d) Genotype specifications in run D.

Days to 110 kg Data 0.396 –0.725 –0.042
Model 0.399 –0.724 –0.042
DFI Data –0.435 0.273 0.405
Model –0.668 0.272 0.405
BF Data –0.044 0.237 0.550
Model –0.028 0.275 0.553
Model 177 2.02 10.33
Table 9.5. Estimates of the genetic parameters of the growth model corresponding to a PIC
Sire Line obtained from four simulations.
Simulation results Parameter Pmat LPmat B* MEm0
Run A Phen. mean 56.53 0.89 0.029 637.9

h2 0.46 0.47 0.21 0.03
Phen. CV 0.03 0.09 0.09 0.03
Run B Phen. mean 50.55 0.78 0.029 505.7
h2 0.42 0.51 0.31 0.12
Phen. CV 0.07 0.11 0.09 0.08
Run C Phen. mean 38.29 0.68 0.034 650.0
h2 0.37 0.58 0.26 0.07
Phen. CV 0.03 0.06 0.07 0.03
Run D Phen. mean 37.04 0.65 0.035 652.6
h2 0.37 0.52 0.39 0.11
Phen. CV 0.05 0.12 0.10 0.07
than those reported by Knap (2000a), who reported literature estimates

from 0.1 to 0.7, with the majority lying between 0.1 and 0.4.
Tables 9.4a to 9.4d show that all four optimization runs produced
similar phenotypic means, genetic correlations and heritabilities for all
three performance traits used in the optimization criterion. In addition to
those criteria used in the optimization process, the model with the
specified parameter estimates also produces realistic phenotypic
correlations (white fields below the diagonal). Only the phenotypic
correlation between growth rate and daily feed intake is stronger (in the
negative direction) according to the model, which might indicate
inconsistencies in the feed intake data.
The good match between predictions and data implies that, in contrast
to the de Lange model, PGGM realistically predicts various traits up to the
level of their covariation, provided that the genotype is appropriately
specified. This increases confidence in its predictions. Nevertheless, despite
this good fit, the inversion process used here does not produce a unique
set of model parameters that fully characterize the genotype.
Choosing the most appropriate specification of the genotype: stricter criteria
Our optimization criteria exclusively refer to data measured at 110 kg

body weight. The identification of discrepancies in the predictions from
different parameter combinations, and thus of the most appropriate
combination, would require a more extensive description of the
phenotype.
The benefit of data covering multiple growth stages is illustrated in
Fig. 9.4. It shows the predicted trends in average daily gain from the four
parameter combinations A, B, C and D, together with the growth curve
derived from data of a genotype, bred from the same sireline as above but
with a different damline. Differences between the parameter sets A and B
versus the sets C and D (higher values for Pmat and LPmat) are clearly
reflected in the associated growth curves (upper versus lower curves in Fig.
9.4). If genetic correlations together with heritabilities, and repeated
measurements from various growth stages, were simultaneously available
for a genotype (which is not the case here), the most appropriate
parameter combination could be chosen according to the model fit to
empirically established growth curves. Alternatively, phenotypic growth
trends could be directly included in the optimization criteria using
dynamic control of the objective functions (as pioneered by Kinghorn et al.,
2002).
Conclusion
Until now, mechanistic animal growth models have been primarily

recognized as a valid method for predicting animal performance under
conditions that are not covered by available data. The present study
suggests a novel use of such models for quantifying genotypes in a way that
could be advantageous to animal breeding and management.
Appropriate quantification of the genotype-specific model parameters
1.2
1.0
ADG (kg/day)
0.8
Empirical estimates
Model, parameter set A
0.6
Model, parameter set B
Model, parameter set C
Model, parameter set D
0.4
60 80 100 120
Body weight (kg)
Fig. 9.4. Average daily gain relative to body weight as predicted from empirical studies for a
crossbred type and as predicted for the pure bred type by the growth model according to the
parameter sets A to D specified in Table 9.5. The crossbred was produced from the same
sireline as that used for model inversion, but from a different damline.
is crucial for accurate model predictions and thus for the appropriate use
of growth models. Growth models also give a window on to underlying
physiological traits that are intrinsic drivers of observed phenotype,
whatever environment it is expressed in. Their approach can be used for
making genetic evaluation of animals for these underlying traits. There
are prospects that this will give better gains – through an integration of
different observed traits in a more intelligent manner than the linear
statistical approach that conventional selection index theory in animal
breeding involves, but also through a better choice of observed traits to
measure (including choices of ages and diets), leading to a more accurate
evaluation of both underlying physiological traits, and observed traits as
expressed in different environments, and thus of the genetic potential.
Combined with phenotypic performance data, which are relatively easy
to measure, model inversion has been identified as a promising tool to
derive the desired specification of the genotype intrinsic physiological
traits. The case studies presented here demonstrate that conclusive
estimates for the genotypic specifications are only possible if: (i) the growth
model simulates the physiological mechanisms of pig growth sufficiently
accurately that observed growth trends can be reproduced; and (ii) if
sufficient data expressing the genetic potential for growth and energy
utilization are available.
Acknowledgements
The authors would like to thank Professor Colin T. Whittemore for

initiating the work on algebraic model inversion and for his useful remarks
throughout the development of this work. We are also thankful to Anthea
Springbett for sharing the concepts and results from the original research
project on algebraic model inversion and for various inspiring discussions.
Many thanks are also due to Dr Darren Green for his comments on an
earlier version of this chapter and to Dr David Parsons for sharing his
experience in constructing a framework for model optimization and
control.
References
Baldwin, R.L. (1976) Principles of modelling animal systems. Proceedings of the
NewZealand Society of Animal Production 36, 128–139.
Black, J.L. (1995) The testing and evaluation of models. In: Moughan, P.J.,
Bourdon, R.M. (1998) Shortcomings of current genetic evaluation systems. Journal
of Animal Science 76, 2308–2323.
de Lange, C.F.M. (1995) Framework for a simplified model to demonstrate
Emmans, G.C. (1988) Genetic components of potential and actual growth. In:
Land, R.B., Bulfield, G. and Hill, W.G. (eds) Animal Breeding Opportunities.
British Society of Animal Production, Edinburgh, UK, pp. 153–181.
Ferguson, N.S. and Gous, R.M. (1993) Evaluation of pig genotypes. 1. Theoretical
aspects of measuring genetic parameters. Animal Production 56, 233–243.
Green, D., Brotherstone, S., Schofield, C.P. and Whittemore, C.T. (2003) Food
intake and live growth performance of pigs measured automatically and
continuously from 25 to 115 kg live weight. Journal of the Science of Food and
Agriculture 83, 1150–1155.
requirements and of excretions into the environment. (IMS Pig). Animal Science
77, 113–130.
Kinghorn, B.P., Meszaros, S.A. and Vagg, R.D. (2002) Dynamic tactical decision
systems for animal breeding. In: Proceedings of the 7th World Congress on Genetics
Applied to Livestock Production 33, 179–186.
Knap, P.W. (1999) Simulation of growth in pigs: evaluation of a model to relate
thermoregulation to body protein and lipid content and deposition. Animal
Science 68, 655–679.
Knap, P.W. (2000a) Time trends of Gompertz growth parameters in “meat type”
pigs. Animal Science 70, 39–49.
Knap, P.W. (2000b) Stochastic simulation of growth in pigs: relations between body
composition and maintenance requirements as mediated through protein
turnover and thermoregulation. Animal Science 71, 11–30.
Knap, P.W., Roehe, R., Kolstad, K., Pomar, C. and Luiting, P. (2003)
Characterization of pig genotypes for growth modelling. Journal of Animal
Science 81 (E-suppl. 2), E187-E195. www.asas.org/symposia/03esupp2/jas2593.pdf.
Krantz, S.G. (1951) The Implicit Function Theorem: History, Theory, and Applications.
Birkhäuser, Boston, Massachusetts.
Landgraf, S., Roehe, R., Susenbeth, A., Baulain, U., Knap, P.W., Looft, H., Plastow,
G.S. and Kalm, E. (2002) Biological growth model as a new selection strategy
for improvement of feed efficiency in swine. Veterinarija ir Zootechnica 18,
84–89.
Matlab, 6.5. (2002) The Language of Technical Computing 1984–2002. The
MathWorks, Inc.
Mayer, D.G., Kinghorn, B.P. and Archer, A.A. (2005) Differential evolution – an
easy and efficient evolutionary algorithm for model optimisation. Agricultural
Systems 83, 315–328.
Pomar, C., Kyriazakis, I., Emmans, G.C. and Knap, P.W. (2003) Modeling
stochasticity: dealing with populations rather than individual pigs. Journal of
Animal Science 81 (E-suppl. 2), E178–E186.
Price, K. and Storn R. (1997) Differential evolution. Dr. Dobb’s Journal, April, 18, 24
and 78.
Roehe, R., Lorenzo Bermejo, J., Landgraf, S., Knap, P.W. and Kalm, E. (2002)
Estimation of growth curves using linear random regression and nonlinear

two-stage models for experimental data of protein deposition in swine. In:
Proceedings of the 53rd Annual Meeting of the European Association of Animal
Production, Cairo. Paper G3.0.
Steuer, R.E. (1986) Multiple Criteria Optimisation: Theory, Computation and Application.
Wiley, New York.
Wellock, I. J., Emmans, G.C. and Kyriazakis, I. (2003a) Modelling the effects of
thermal environmental and dietary composition on pig performance: model
81, 2995–3007.
Wellock, I.J., Emmans, G.C. and Kyriazakis, I. (2004) Modeling the effects of
stressors on the performance of populations of pigs. Journal of Animal Science
82, 2442–2450.
Whittemore, C.T., Green, D.M., Wood, J.D., Fisher, A.V. and Schofield, C.P. (2003)
Physical and chemical composition of the carcass of three different types of
pigs grown from 25 to 115 kg live weight. Animal Science 77, 235–245.
10 Considerations for Representing
Micro-environmental Conditions
in Simulation Models for Broiler
Chickens
O.A. BLANCO AND R.M. GOUS
Animal and Poultry Science, School of Agricultural Sciences and
Agribusiness, University of KwaZulu-Natal, Private Bag X01, Scottsville
3209, South Africa.
201508550@ukzn.ac.za
Introduction
The thermal relationship between a homeothermic animal and its

surrounding environment presents a challenging level of complexity for
anyone attempting to analyse or model it. A conflicting issue, most of the
time overlooked, is the way in which the environment surrounding the
animal is described. The common approach is to account only for the
values of air temperature (Tair) and relative humidity (RH), which
constitutes an extremely naive procedure (Charles, 1994; Mitchell, 2005).
Considering that the values of the variables characterizing the
environment surrounding the animal become meaningful only when
expressed in the context of the animal (Monteith, 1974), we discourage the
adoption of simplistic approaches. Furthermore, if the spatial and temporal
variations of the environmental conditions, as well as the variability of their
biological implications, are considered it is evident that the way in which the
environmental conditions are usually expressed should be revised.
There are three objectives to this chapter. First, we wish to review and
clarify concepts related to the physiology of heat exchange in broiler
chickens; secondly, we aim to give some guidelines for selecting ways in
which the specification of the environment should be tackled; and finally,
and most importantly, we wish to create an awareness of the negative
consequences of using superficial and improper descriptions of the
environment on predictions made by models dealing with the energy
balance of chickens. In order to accomplish these objectives, this work
analyses the expression of environmental conditions considering the
peculiarities of heat dissipation in broiler chickens.

Micro-environmental Conditions in Simulation Models 189
The Thermal Relationship between the Animal and its

Surroundings
Homeothermic animals are able to keep their body temperature fairly

constant over a wide range of environmental conditions (Mount, 1979;
Mitchell, 2005). Their success depends, on the one hand, on the demand
of heat and water vapour that the surrounding environment inflicts on the
animal and, on the other, on the physical characteristics of the animal and
its capacity to respond to such an exigency through thermoregulatory
mechanisms (Mount, 1979; Mitchell, 2005).
The heat and water vapour dissipated during thermoregulation modify
the climatological conditions in the animal’s immediate environment. This
becomes particularly important in animals reared in confined conditions,
such as poultry (Charles et al., 1994). Blanco et al. (2004a,b) proposed that
such changes in environmental conditions would induce new thermo-
regulatory responses in the animal, producing further environmental
modifications. Charles (1974) proposed a partition of the environment
centred on the chicken, based on differentiation of the environmental
conditions and on the way in which each of these divisions affects the
animal.
The properties of the mass of air immediately surrounding an animal
play a decisive role in determining the magnitude of the net heat
dissipation from that animal. Charles (1974) defined this fraction of the
environment as the micro-environment. Due to the spatial variability of the
environmental conditions inside a broiler house (Czaric and Tyson, 1990;
Xin et al., 1994), micro-environmental conditions, i.e. the conditions at
chicken height, need to be specified at the moment of analysing the
thermal balance of the animal.
Charles (1974) referred to the outdoor part of the environment as the
macro-environment. The properties of the macro-environmental variables
would influence those of the micro-environment depending on the
insulating properties of the building and its design (Xin et al., 1994).
Finally, the area of transition between the micro-environment and the
macro-environment was termed the meso-environment (Charles, 1974).
The environmental conditions should be specified in the
environmental context that best suit the purposes of the model.
Note: As a convention, the term micro-environment, which comprises
not only micro-climatological conditions but also aspects such as air
pollutants and presence of pathogens, has been used in this work as a
synonym of micro-climate.
The Heat Balance of a Farm Animal
Homeothermic animals dissipate heat in the form of both sensible (H) and
latent (λE) heat (Mount, 1979). The heat flow density from the animal to
the environment (expressed normally in W m2) depends on the structure
190 O.A. Blanco and R.M. Gous
and physiological state of the animal (body size, feather cover, respiratory
rate, etc.) and on the thermal gradient (in the case of H) or the vapour
pressure gradient (∆eV) (in the case of λE) between the surface of the
animal and the surrounding environment (Mount, 1979). The level of
micro-environmental radiation affects H, whilst air velocity (u) affects both
λE and H (Monteith and Unsworth, 1990) (see Fig. 10.1).
According to the peculiarities of heat exchange with its surroundings,
a farm animal can be modelled as a system with two interfaces, i.e. the
‘body surface’ (Ib) and the ‘surface of the anterior respiratory tract’ (Ir)
Fig. 10.1. Partition of heat dissipated by a chicken. The circular detail shows the
components of the heat exchange between breathed air and mucous surface of the upper
respiratory tract, where the double-headed arrow indicates the movement of the circulating
air. λEf: dissipation of latent heat from feathered areas; Hf: dissipation of sensible heat from
feathered areas; rHRf: thermal resistance of the boundary layer of feathered areas to
convective and radiant heat transfer; rvf: resistance of the boundary layer of feathered areas
to water vapour transfer; λEa: dissipation of latent heat from bare appendages; Ha:
dissipation of sensible heat from bare appendages; rHRa: thermal resistance of the boundary
layer of bare appendages to convective and radiant heat transfer; rva: resistance of the
boundary layer of bare appendages to water vapour transfer; λEr: dissipation of latent heat
from the surface of the anterior respiratory tract (interface Ir); Hr: dissipation of sensible heat
from interface Ir; rHRr: thermal resistance of the boundary layer of interface Ir to latent and
radiant heat transfer; rvr: resistance of the boundary layer of the interface Ir to water vapour
transfer; Rni: micro-environmental isothermal net radiation; u: micro-environmental air
velocity; C: thermogenic core; Sk: skin; F: feather coat; BL: Boundary-layer (note: due to
space restrictions, it was impossible to represent rva, rHRa, eVa and Ta on the wattles) (based
on Campbell and Norman, 1998 and Blanco et al., 2004a,b).
(Blanco, 2004a,b). A thin layer of air over the surface of each of these
interfaces, called the boundary-layer (BL), imposes resistance to heat and
water vapour loss to the environment (Monteith and Unsworth, 1990) (Fig.
10.1). The micro-environmental variables may impose dissimilar demands
on each of the interfaces due to the different values of boundary-layer
resistances (Bakken, 1981; Blanco et al., 2004a,b), the various processes
controlling the heat exchange at each interface, and the anatomical and
physiological differences between respiratory tract and body surface.
Sensible heat loss
A homeothermic animal uses three mechanisms for dissipating sensible

heat from its surface, i.e. conduction, convection and radiation.
Conductive heat exchange

Conduction depends on the thermal gradient between the animal and the
surface in contact with it, for example bedding material, as well as on the
coefficient of thermal conductivity of the mass in contact with the animal
(Mount, 1979; Monteith and Unsworth, 1990). The amount of heat
conducted to the surrounding air is minimal, but the magnitude of the
heat transferred to some poorly insulated surfaces, such as floors and
walls, can be quite substantial (Monteith and Unsworth, 1990). The total
heat lost by conduction depends also on the size of the contact area
between the animal and the contacting surface (Mount, 1968), which is
directly related to the posture of the animal (Mount, 1968). Since that
posture is unpredictable (Turnpenny, 2000b), Blanco et al. (2004a,b)
considered, in the first steps of the construction of their model for
analysing thermoregulatory responses of a broiler, that the animal stands
on its feet on a surface with very low thermal conductivity, thereby
minimizing the heat loss by conduction and ignoring this means of heat
transfer.
Convective heat exchange

Convective heat loss depends on the thermal difference between the
animal surface and the surrounding air (or circulating air in the particular
case of the respiratory tract) (Mount, 1979; Wathes and Clark, 1981a;
Monteith and Unsworth, 1990; Campbell and Norman, 1998). It is the
result of the combined effect of natural convection, which depends on the
buoyant forces produced by the warming of the air in contact with the
animal surface, and forced convection, which occurs when an airstream
affects any of the interfaces (Mount, 1979). The BL imposes a resistance to
heat transfer by convection (rH) directly proportional to the diameter of
the animal and inversely proportional to air speed (u) (McArthur, 1981;
Monteith and Unsworth, 1990). Therefore, smaller animals, and animals
exposed to higher u, will dissipate more heat by convection per unit of

surface area.
Radiant heat exchange

The dissipation of radiant heat from a confined animal depends on the
emissivities of the surface of the animal and the internal surface of the
building, the surface areas of the animal and of the enclosure where the
animal is kept, and the thermal gradient between the surface of the animal
and the surrounding environment (Cena, 1974; Wathes and Clark, 1981a;
Mitchell, 1985; Campbell and Norman, 1998).
McArthur (1987), in a model of the thermal interaction between
animal and microclimate, utilized the parameter isothermal net radiation
(Rni) to quantify the net radiant heat exchange between the animal and the
environment. The parameter represents the net radiation that would be
exchanged if the temperature of the animal’s surface were equal to the air
temperature. The mathematical expression for animals housed in
buildings where the solar radiation is negligible is presented in Eqn 10.1.
Rni = ρcp (τe – Ts) (rR)1 (10.1)
where ρcp is the volumetric heat capacity of the air (kJ m3 K1), τe is the
mean radiative temperature of the environment (°C), Ts is the temperature
of the animal surface (°C), and rR is the thermal resistance of BL to radiant
heat transfer (s m1).
Monteith and Unsworth (1990) performed a more comprehensive
analysis of this parameter, and the reader is referred to that work for
further information.
Latent heat loss
Homeothermic animals dissipate latent heat from Ib and Ir (Mount, 1979;

Monteith and Unsworth, 1990; Campbell and Norman, 1998; Willmer et
al., 2000) (Fig. 10.1). The driving force is the gradient of vapour pressure
(eV) between the exposed surface (eVb in the case of the body surface, and
eVr in the case of the anterior respiratory tract) and the micro-
environmental air (eVair) (Mount, 1979; Monteith and Unsworth, 1990;
Campbell and Norman, 1998). (Note that, as will be explained later, the
surface of the animal has been divided into feathered areas and bare
appendages, which is the reason why, in Fig. 10.1, eVb is represented by
eVf and eVa, which accounts for vapour pressure of the surface of
feathered areas and bare appendages, respectively.) The boundary layers
corresponding to the body surface and the surface of the anterior
respiratory tract impose resistances to water evaporation, which are
symbolized as rvs and rvr, respectively (McArthur, 1981). (The expression
rvs includes the resistance to water vapour transfer from the surface of the
bare appendages, or rva, and from the surface of the feathered areas, or
rvf.) The value of rvs (i.e. rva and rvf) is inversely related to u, whilst rvr is
inversely proportional to the respiratory rate (RR) (McArthur, 1981;
McArthur, 1987; Turnpenny et al., 2000a,b).
Peculiarities of Heat Dissipation in Poultry
Chickens have distinctive anatomical and physiological characteristics

related to additional thermoregulatory properties that transform the
analysis of the physiology of their heat exchange into an exciting
challenge.
In the first place, although latent heat loss from both Ir and Ib is a
substantial component of the heat balance of chickens in environments
where the sensible heat loss is not restricted (Bernstein, 1971; Marder and
Ben-Asher, 1983; Mitchell, 2005), chickens are not able to sweat when
exposed to conditions above least thermoregulatory effort (Mount, 1979;
Willmer et al., 2000). In such environments, birds rely on active latent heat
loss from Ir in order to control their body temperature (Mount, 1979;
McArthur, 1981; Barnas and Rautenberg, 1987; Willmer et al., 2000).
When the capacity to dissipate sensible heat is reduced, the respiratory rate
(RR) increases secondary to an elevation of body temperature (Tb) (Zhou
and Yamamoto, 1997). As a consequence, the properties of BL of Ir change
and, as a result, its resistance to water vapour transfer (rvr) is reduced
(McArthur, 1981). Finally, the amount of latent heat dissipated from the
interface Ir increases.
A second distinctive feature of commercial chickens is the substantial
feather coat with high insulation properties that covers approximately
0.80–0.85 of their body surface area (McArthur, 1981). This coat is
responsible for controlling the sensible heat dissipation from the feathered
areas (Richards, 1970, 1974; McArthur, 1981; Wathes and Clark, 1981b).
Since the high thermal resistance afforded by this coat cannot be widely
varied, the bare appendages of the animal, i.e. legs, combs and wattles,
play an important role in thermoregulation (Richards, 1974; Bakken,
1981; McArthur, 1981; Willmer et al., 2000; Turnpenny et al., 2000a). The
differential redistribution of blood to these bare areas, secondary to the
autonomic regulation of vasodilation and vasoconstriction in the
subcutaneous arterioles, helps the animal to manage its skin resistance (rs)
in order to modify its total sensible heat loss (Mount, 1979; Bakken, 1981).
The shape and size of the bare appendages (comparatively smaller than
the body) result in a high surface:mass ratio, which facilitates the heat
dissipation, as well as in a reduced resistance of their BL to convective and
evaporative heat transfer (Bakken, 1981; Monteith and Unsworth, 1990).
The combination of a lower rH, variable rs, high surface:mass ratio, and
direct exposure of the skin surface of the bare appendages plays an
important role in the determination of the total sensible heat dissipation
from the animal.
Summary of the Facts to Consider when Modelling the Micro-

environmental Conditions Surrounding a Chicken
In view of the concepts previously discussed, several considerations should

be made before deciding the way in which to specify the micro-
environmental conditions in a simulation model for poultry.
First, in a model dealing with the heat balance of a chicken, the
environmental variables air temperature (Tair), air speed (u), vapour
pressure of the air (eVair) and isothermal net radiation (Rni) need to be
specified in the micro-environmental context.
Secondly, since each micro-environmental variable plays an important
role in the dynamics of the dissipation of sensible and latent heat from the
animal, neglecting one or more of these variables may negatively affect the
final estimation of heat production or loss when modelling the heat
balance of the animal.
The important role played by the thermal properties of the chicken, in
particular the thermal and vapour resistances of the BL, should also be
considered. Monteith (1974) acknowledged that a satisfactory specification
of the environment would account for the close relationship between the
micro-environment and the animal interface, whilst Mount (1979)
recognized that the ratio of thermal gradient to thermal insulation
determines the magnitude of sensible heat flow (H) from the body surface.
The role of respiratory thermoregulation in micro-environmental
situations above least thermoregulatory effort should definitely be
considered. The importance of evaporative heat loss on heat balance
under non-stressing conditions should also be contemplated.
In chickens, it is also necessary to take into account the proportion of
the total body surface represented by bare appendages. These structures
have a decisive influence on the value of the total thermal resistance,
through modification of the thermal skin resistance and the
characteristically reduced thermal resistance of their BL.
Finally, the differences between interfaces, mainly with respect to the
properties of their boundary layers, has been shown to be as important as
the environmental variables themselves (Monteith, 1974), and they should
be included in the specification of the environment.
Therefore, it is evident that it is not sufficient to account only for the
micro-environmental variables. The following sections deal with issues
such as what needs to be modelled when specifying the micro-
environment, and the format that we consider appropriate in the
construction of such models.
Modelling the Action of the Micro-environment
As stated above, thermal and vapour resistances of BL regulate heat and

vapour dissipation from the surface of the interfaces Ir and Ib. Therefore,
it is only by considering the value of the micro-environmental variables in
the context of BL that a proper description of the micro-environment can

be obtained (Monteith, 1974). In doing so, the specification of the
environment becomes the description of the action of the micro-
environment on the animal.
Furthermore, we believe it necessary to differentiate between the action
and the effect of the micro-environment on the animal. Whilst the action
refers to the value of the micro-environmental variables considered in the
context of the BL, the effect of the environmental conditions on heat
balance in the animal refers to the activation of thermoregulatory
mechanisms and the changes in thermal and physiological parameters of
the animal, secondary to the micro-environmental action. This is more
than semantic vagary; they are two dissimilar concepts, and they have to be
properly differentiated in order to clarify the aim of the modelling
procedure. Once the magnitude of the action is known, further modelling
steps can be formulated in order to estimate the effect on, for example, heat
balance and the energy requirements of the animal.
A diagram showing the factors considered at the moment of accounting
for the action of the micro-environmental conditions is presented in
Fig. 10.2.
Regarding the form in which the action of the micro-environment
should be represented, the adoption of a parametric approach is highly
recommended, as described below.
Selection of the Proper Parameter for Describing the Action of

the Micro-environmental Variables on the Animal
There are a great variety of environmental parameters in the literature
describing the way in which the micro-environmental variables combine
and affect the animal, each of which is suitable for a specific purpose.
Therefore, the selection should be performed by means of a careful
analysis. Even then, it is not guaranteed that the parameter that suits the
requirements of the modeller will be found, and pertinent modifications
should be performed.
Bearing in mind that animal characteristics are as important as
environmental variables in determining the heat balance of a homeotherm
(Monteith, 1974), a reasonable description of the environment should
neglect neither of these two aspects.
Monteith (1974) and Mount (1979) proposed the following three main
characteristics that a suitable parameter for specifying micro-
environmental conditions should fulfil:
1. Applicability to all the species in any physiological state.
2. Validity in both indoor and outdoor conditions.
3. Independence from the characteristics of the exchanging interface (i.e.
surface of feather cover in feathered regions, skin surface in the bare regions,
and the surface of mucous membrane in the upper respiratory tract).
⎫
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎬
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎭
Fig. 10.2. Diagram showing the components of the action of the micro-environment
and the consequent micro-environmental effect or animal response. rHR: Thermal resistance
of the boundary layer to radiant and convective heat transfer; rv: resistance of the boundary
layer to water vapour transfer. Note that BL is clearly included in the environment.
to which we have added three further requisites:

4. That it should account for all the micro-environmental variables.
5. That it should consider that the interaction between animal and
surrounding environment is produced through the boundary layer. Thus a
variable accounting for the properties of such boundary layer should be
included in the parameter.
6. If possible, the parameter should account for the interaction of both
interfaces into which the animal has been divided, i.e. Ir and Ib, and their
respective micro-environments.
According to these criteria, an initial selection was performed among the
parameters reported in the literature, considering those that have been
developed following proper scientific procedures. Such parameters were
analysed and classified in two groups.
The first group recognized is the family of discomfort indices and is of
an empirical nature. These indices usually combine terms accounting for Tair
and for the water vapour content of the air. Their empirical nature is the
result of the inclusion of multipliers experimentally derived. Considering the
importance of the physical and thermoregulatory characteristics of the
broiler on the thermal relationship between bird and environment, as well as
the variation of such properties according to factors such as maturity
(Poczopko, 1981), breed (Özkan et al., 2002), and adaptation at genetic,
phenotypic and epigenetic levels (Nichelman and Tzschentke, 2002), the
usefulness of these parameters is very limited, unless new experimental
multipliers are calculated for each new case. Therefore, these discomfort
indices do not fulfil the first criterion of being a suitable parameter.
In addition, the inductive approach used in the construction of the
parameters included in this first group neither allows for understanding the
animal–micro-environment interaction, nor for analysing the role of each
variable as part of such a relationship. This is basically because they have
been conceived for purposes other than the creation of models for the
analysis of the action and effect of micro-environmental conditions on the
heat balance of a broiler. They were originally used for classifying animals as
heat stressed and non-heat stressed, according to experimentally determined
comfort ranges, constituting a safer way of controlling the environment and
a moderately accurate way of predicting production results. We have not
found them useful for our model, but do not discard the possibility that such
parameters could be applied in models approaching the relationship
between animal and micro-environment from another viewpoint. Examples
of such parameters are the Discomfort Index (Tselepidaki et al., 1995), the
Thermal Hygrometric Index (Esmay, 1978; Tao and Xin, 2003) and the
Wind Chill Factor (Siple and Passel, cited by Starr, 1981).
A second group of parameters, which we term a family of equivalent
temperatures, has been identified. These parameters, derived following
the criteria proposed by Monteith (1974) and Mount (1979), include
environmental apparent equivalent temperature (θ*e) (Mount, 1974),
environmental temperature (Te) (McArthur, 1981), effective temperature
(Tef) (Monteith and Unsworth, 1990), equivalent temperature (Monteith
and Unsworth, 1990) and operative temperature (Campbell and Norman,
1998). Such parameters are characterized by the deductive approach used
in their construction.
The equations for the parameters of the second group combine the
value of two or more micro-environmental variables with the factors
accounting for the physical properties of the animal. These parameters
have two advantages over those from the former groups:
1. They have not been developed inductively; hence they do not have an
empirical character. Therefore, the three first characteristics of an ideal
parameter are, at least, partially fulfilled.
2. The inclusion of the thermal resistances of the boundary layer in some
of these parameters results in a thermal equivalent of the strain imposed
by the immediate surroundings on the bird, and accounts therefore for the
action of the environment on the animal.
According to Campbell and Norman (1998), these equivalent temperatures,

on the one hand, allow for a better interpretation of the interaction between
animal and micro-environment due to their dimension of temperature and,
on the other, they facilitate the application of experimental results obtained
in controlled conditions to real situations.
Effective Temperature (Tef) (Monteith and Unsworth, 1990)

Blanco et al. (2004a,b) found effective temperature (Tef) (Monteith and
Unsworth, 1990) to be the most suitable parameter for specifying the
action of the micro-environment in meeting their objectives. This 1990
version of environmental apparent equivalent temperature (θ*e)
(Monteith, 1974) does not include water vapour pressure as a variable, and
expresses the parameter in the animal context by replacing the coefficients
of heat transfer with thermal resistances of BL, which, as expressed above,
depends directly on the dimensions of the animals and on the value of the
variable u.
Effective temperature (Tef) follows the mathematical structure of the
parameters of the second group, preserved since Monteith (1974), i.e. the
combination of Tair and an increment (expressed in units of temperature),
which in this case accounts for the joint effect of u and Rni. The result is the
temperature value that should be reached in a non-radiant environment in
order for an animal to exchange the same amount of sensible heat (H) as it
would if it were situated in a radiant environment (both with identical u)
(Monteith and Unsworth, 1990). This parameter includes the thermal
resistance of BL to radiant and convective heat (rHR) as a factor. Equation
10.2 is the mathematical expression for Tef according to Monteith and
Unsworth (1990).
Tef = Tair + ζ = Tair + rHR (ρcp)–1 Rni (10.2)
where Tair is air temperature (°C), ζ is the radiation increment (°C), rHR is
the thermal resistance to radiant and convective heat transfer in the
boundary layer (s m1), ρcp is the volumetric heat capacity of the air (kJ m3
K1) and Rni is the isothermal net radiation affecting the animal (W m2).
The second term of the linear Equation 10.2, called radiation
increment (ζ) (Monteith and Unsworth, 1990), expresses, in units of
temperature, the combined action of the micro-environmental radiation
air velocity in the context of BL. The factor rHR is the combined resistance
to radiant and convective heat flux imposed by BL. The resistance to
radiant heat flux (rR) has a constant value of 2.1 s m1 for animals with the
characteristic dimension of a broiler (McArthur, 1981; Monteith and
Unsworth, 1990). The resistance to convective heat transfer (rH), which
differs from rR, depends on both the dimension and shape of the animal’s
body and on the value of u (McArthur, 1987; Monteith and Unsworth,
1990), as well as on the smoothness of the interface (Wathes and Clark,
1981a; McArthur, 1987). According to Monteith and Unsworth (1990),
when u is such that the Reynolds number is higher than 103, rH is

represented by Eqn 10.3.
rH = d (κ Nu)1 ∝ d0.4 u–0.6 (10.3)
where rH is the thermal resistance of the boundary-layer to convective heat
transfer (s m1), d is the characteristic dimension of the animal (diameter
in the broiler, for being considered as a sphere, in m), κ is the thermal
diffusivity of still air (m2 s1) and Nu is the Nusselt number (non-
dimensional value).
For a mature broiler chicken, with a body diameter of approximately
0.25 m, rH and rR are of comparable magnitude at the low values of u
usually found in a proper tunnel-ventilated broiler house (from 0.25 to
3 m/s at chicken level) (Monteith and Unsworth, 1990). Considering that,
in the boundary layer, these two resistances are working in parallel, the
total thermal resistance to radiant and convective heat exchange, rHR, is
calculated, in analogy with Ohm’s law (Wathes and Clark, 1981a;
McArthur, 1987; Monteith and Unsworth, 1990; Campbell and Norman,
1998), according to Eqn 10.4.
rHR = (rR–1 + rH–1)–1 (10.4)
Equation 10.2 is represented in the diagram of temperature vs heat flux
density (Monteith and Unsworth, 1990) in Fig. 10.3 for given values of u
and Rni. The slope is directly proportional to rHR. Tef is graphically
determined by the intersection between the line with intercept Tair and
slope rHR (ρcp)–1, with the Rni flux density.
The impact of the variation in air velocity on the final value of Tef is
represented in Fig. 10.4. The effect is produced through a modification of
rHR, which changes the slope of equation 10.2.
The variations of Tef secondary to the modifications of Rni for the same
Tair and u are shown in Fig. 10.5. Higher values of Rni result in higher Tef,
which has important consequences on the control of the macro-
Temperature (°C)
Tef
⎧

⎨
T⎩ air
Rni
Heat flux density (W m–2)
Fig. 10.3. Temperature vs heat flux density diagram for the graphical determination of Tef.
Tair: air temperature, Tef: effective temperature, Rni: isothermal net radiation, ζ: Radiation
increment (after Monteith and Unsworth, 1990).
u1>u2>u3
Tef3
Temperature (°C)
rHR1<rHR2<rHR3
3
Tef2
Tef1<Tef2<Tef3
2
Tef1
1
Tair
Rni
Fig. 10.4. Temperature vs heat flux density diagram for the graphical determination of Tef
considering three different u, with the same Tair and Rni. Tair: air temperature, Tef: effective
temperature, Rni: Isothermal net radiation, u: air velocity, rHR: thermal resistance to radiant
and convective heat transfer through the boundary layer (after Monteith and Unsworth,
1990).
Tef(n-ins)
Temperature (°C)
Tef(ins)
Tair
Rni(ins) Rni(n-ins)
Fig. 10.5. Comparison between Tef values reached in two comparable hypothetical houses
with and without insulated roofs in the same season, same air temperature (Tair) and same
air velocity. Rni(ins): isothermal net radiation loaded on chickens under insulated roof; Rni(n-ins):
isothermal net radiation loaded on chickens under non-insulated roof; Tef(ins): effective
temperature for chickens under insulated roof; Tef(n-ins): effective temperature for chickens
under non-insulated roof (after Monteith and Unsworth, 1990 and Blanco et al., 2003).
environmental conditions in a broiler house: for the same value of Tair and
u, a broiler exposed to higher Rni, would experience higher Tef.
Limitations of Tef
1. Inability to account for the action of the micro-environmental

vapour pressure
The most evident limitation of Tef (Monteith and Unsworth, 1990) is its
inability to account for the action of the micro-environmental vapour
pressure. In order to include such a variable, Monteith and Unsworth
(1990) returned to the concept of apparent equivalent temperature (Teq)
introduced by Monteith (1974). As in the case of the equation developed in
1974, this apparent equivalent temperature (here symbolized with Te*) is
obtained after the addition of a humidity increment (ξ) to Tair, as shown in
Eqn 10.5.
Te* = Tair + ξ= Tair + e (Tair) (γ*)–1 (10.5)
where e(Tair) is the vapour pressure of the micro-environmental air (kPa) at
a given Tair (°C), and γ* is the psychrometric constant in its apparent form
(kPa K1). The mathematical expression for the apparent psychometric
constant γ* is γ* = γ [ rv (rHR)1], where γ is the psychrometric constant
(=0.066 kPa K1). The form γ* has the advantage of accounting for the
thermal resistance for convective and radiant heat exchange through BL
(rHR) and the resistance to water vapour transfer through BL (rv) in a same
term.
Monteith and Unsworth (1990) obtained an environmental apparent
equivalent temperature (Ter*), after combining Te* and a radiant
increment (ζ). The mathematical expression of Ter*, comparable to that for
environmental apparent equivalent temperature (θ*e) reported by
Monteith (1974), is presented here as Eqn 10.6.
Ter* = Te* + rHR (ρcp)–1 Rni = Tair + eVair rHR (rv γ*)1 + rHR (ρcp)–1 Rni (10.6)
where Ter* is the environmental apparent equivalent temperature (°C).
As in the case of θ*e (Monteith, 1974), in steady state conditions, Ter*
can be applied to both interfaces of the system, i.e. Ib and Ir. In the
particular case of Ir, it should be remembered that the value of the
resistance to water vapour transfer through its boundary, rvr, depends on
the respiratory rate (RR). McArthur (1981) published an empirical
equation based on experimental data by Hutchinson (1954), reproduced
here as Eqn 10.7.
rvr = [(9.1 × 105 RR) + 9.0 × 103]–1 (10.7)
Problems appear when attempting to create a unique value of Ter* to
represent the conditions of both interfaces. Since these two interfaces act in
parallel, and the conditions affecting each surface are substantially
different, it is almost impossible to find a valid common mathematical

expression.
The task of finding a common Ter* becomes almost infeasible in
dynamic models such as that developed by Blanco et al. (2004a,b).
According to equation 10.7, rvr depends directly on RR. The value of RR,
as well as other physiological parameters, varies according to the body
temperature (Tb) of the animal, which at the same time depends on the
micro-environmental conditions as well as on the time the animal has been
exposed to that environment (Hutchinson, 1954; McArthur, 1987; Zhou
and Yamamoto, 1996). Paradoxically, the micro-environmental conditions
affecting RR are those to be included in Ter*. The derivation of a
parameter to account for the action of the micro-environment on both
interfaces proves to be a situation similar to the old riddle about the
ontogeny of hen and egg. It is impossible to define what the exact value of
RR, to be used to estimate hB, should be, mainly because it depends on the
still unknown action of the micro-environmental conditions and their
variations in time, as well as on the consequent changes in Tb.
In models developed to simulate the time-variation of respiratory
parameters, equation 10.7 may be used in the process of modelling the
heat dissipation from the interface Ir. But, again, because of the cause and
effect dilemma, the action of the micro-environment on the interface Ib
and the variable rvr cannot (and should not) be combined in the same
general parameter.
The use of Ter* should be limited to situations in which water
evaporation and sensible heat dissipation occur on the same surface. In
fact, Monteith and Unsworth (1990) stated that the latent heat dissipated
from the respiratory interface should be measured or estimated separately.
This is expressed in Eqn 10.8 where a value x, representing the latent heat
dissipation from the anterior respiratory tract, is subtracted from the total
heat produced by the animal (M) before estimating the total heat
dissipation from the interface Ib.
(1 – x) M = ρcp (Teo* – Ter*) (rHRb)–1 (10.8)
where x is the proportion of metabolic heat (M) that is dissipated as latent
heat (both in W m2), Teo* is the apparent equivalent temperature of the
interface Ib (°C), Ter* is the environmental apparent equivalent
temperature (°C), and rHRb is the thermal resistance to convective and
radiant heat transfer through BL of interface Ib.
Campbell and Norman (1998), in the construction of the parameter
operative temperature, managed the latent heat loss in the same manner.
In this case, the proportion of the metabolic heat dissipated as latent heat
is assumed to be constant, formulating the equation for the animal’s heat
balance based on that proportion of the metabolic heat assumed to be
dissipated by sensible means. This parameter has exactly the same form as
Tef (Monteith and Unsworth, 1990), but in this case the authors have
worked with thermal conductances instead of thermal resistances.
Neither of these approaches gives a solution to the main limitation of
Tef and, consequently, the creation of a parameter combining the action of

the micro-environmental conditions on both interfaces remains infeasible.
2. Modifications of the total thermal resistance of the boundary layer

introduced by the bare appendages
According to the concepts already discussed in this chapter, the BL of the

bare appendages, which has a surface area smaller than the feathered
body, would have a lower thermal resistance to convective and radiant heat
transfer than the boundary layer of the feathered body (Bakken, 1981;
Monteith and Unsworth, 1990).
Considering that the radiation increment of Tef depends on the
thermal resistance of the boundary layer, Eqn 10.2 should be revised.
Blanco et al. (2004a,b) expressed rHR as a serial sum of the individual
thermal resistances of the feathered body and bare appendages. Each
resistance was multiplied by a factor proportional to the area of total body
surface area represented by feathers and appendages. Equation 10.9 shows
the modified form of Eqn 10.2.
Tef = Tair + ζ = Tair + (0.85 rHRf–1 + 0.15 rHRa–1)1 (ρcp)–1 Rni (10.9)
where rHRf and rHRa are the thermal resistances to convective and radiant
heat transfer of the boundary-layer of the feathered body and bare
appendages, respectively. The numbers 0.85 and 0.15 represent the
proportion of the total surface area corresponding to feathered body and
bare appendages, respectively. It should be noted that since rHRb and rHRa
are working in parallel, they have been added according to Eqn 10.4.
This approach seems to be sensible. The estimations performed by
Blanco (2004b) using equation 10.8 have been shown to be reasonably
realistic.
3. The problem of accounting for the micro-environmental effect without a

joint parameter expressing the micro-environmental action on interfaces Ir
and Ib
In the development of their simulation model for estimating the effect of

micro-environmental conditions above least thermoregulatory effort on
broiler chickens, Blanco et al. (2004b) have used the parameter Tef to
account for the action of the micro-environment.
As previously mentioned, Tef cannot be used for both interfaces of the
system, i.e. Ir and Ib, in the same equation. Blanco et al. (2004b) have
overcome this limitation by using Tef to account for the action of micro-
environment on Ib, considering Ir as a mere ‘regulating device’.
Depending on the effect of the micro-environmental conditions on the
physiological parameters (mainly Tb and RR) through the action on Ib, Ir
dissipates the excessive heat that cannot be lost by sensible means from Ib.
This dynamic model estimates the action of the micro-environmental

variables on Ib and their effect on the heat balance of a single broiler in 1-s
cycles. As a first step, the model estimates the experimental Tef
(represented as Tef*). After calculating the amount of sensible heat
dissipated from Ib for that Tef* (H(Tef*)), the heat excess (HE) is calculated
by subtracting H(Tef*) from the sensible heat hypothetically lost from Ib at
least thermoregulatory effort (H(TNT)). If HE is higher than the capacity of
the respiratory tract for dissipating latent heat (λEr), which depends on the
RR, the model calculates the heat excess and re-estimates Tb, RR and rvr.
These outputs are used in the following cycle of simulations. Fig. 10.6
synthesizes in a flux diagram the main steps of this model.
Taking this approach, a parameter for estimating the action of the
micro-environmental conditions on interface Ir is not necessary. Instead,
when considering the evaporation from interface Ir as a process secondary
to the variation of Tb, the modeller would only need to know the values of
eVair and RR.
Boundary Model of a thermoregulating broiler

layer
Tef Estimation
of H from
interface
Micro- Estimation ‘body Estimation
environmental of TNT surface’ of HE
variables
Estimation of
latent heat
Re dissipation from
estimation respiratory
of rvr interface
(λEr)
1 s cycle
Tb RR Estimation
of S
(S=H-λEr)
Information
used in the OUTPUT Recalculation Recalculation
next cycle of Tb of RR of Tb
estimations RR
Fig. 10.6. Diagram of a simulation model for estimating the effect of micro-environmental
conditions on the heat balance of a broiler, including an alternative approach for estimating
the latent heat dissipation from the respiratory tract. The estimations are performed once
every second, as indicated by the circular arrow; Tef: Effective temperature estimated for a
particular broiler under given environmental conditions; TNT: particular value of Tef at which
the animal is under least thermoregulatory effort; H: sensible heat loss; HE: Heat excess; λEr:
latent heat loss from the respiratory tract; S: heat storage; Tb: body temperature; RR:
Respiratory rate; rvr: resistance of the boundary-layer of interface Ir to water vapour transfer.
Conclusions
The authors of this chapter consider that the way in which micro-
environmental conditions are usually accounted for in simulation models
needs to be revised.
In the first place, and bearing in mind the differentiation of the
environment directly surrounding the animals, micro-environmental
conditions should always be considered in models strongly influenced by
the heat exchange between animal and environment as, for example, those
dealing with heat balance and energy utilization.
Secondly, considering the importance of each of the micro-
environmental variables on the heat exchange between animal and
environment, models should at least consider Tair, eVair, u and Rni in order
to produce balanced estimations.
A third point to take into account: since the BL of the animal plays a
fundamental role in the relationship with its surroundings, its
characteristics should be considered when accounting for the
environmental conditions. The description of the micro-environmental
conditions in the context of the BL allows for the estimation of what we
have termed micro-environmental action. This can be a starting point for
further modelling procedures in order to estimate the consequent effect of
those micro-environmental conditions on the heat balance of the chicken.
Regarding mathematical expressions that account for the micro-
environmental action, two groups of parameters have been identified in
the literature. Among the parameters belonging to the group of equivalent
temperatures, effective temperature (Tef) (Monteith and Unsworth, 1990)
best suited the needs of the authors of this chapter. Although most of the
parameters of this group adequately account for the micro-environmental
action by considering the properties of the boundary layer, they need to be
adapted to the following peculiarities of the thermoregulatory processes of
non-sweating animals such as poultry:
1. Water vapour dissipation is important from both the body surface and
the upper respiratory surface.
2. Both body surface and mucous membranes of the upper respiratory tract
are separate interfaces of a system tract (Ib and Ir, respectively).
3. Because different factors control heat and vapour exchange, the action of
the environment on those two interfaces cannot be included in the same
parameter.
4. As RR is directly related to the value of Tb, which is a product of the
action of the environmental conditions to be specified by the parameter,
the formulation of a parameter accounting for the action of the micro-
environment in both interfaces is infeasible.
An alternative approach in dynamic models would be to estimate the latent
heat loss from the respiratory tract as a consequence of the effect of the
micro-environmental conditions on the sensible heat dissipation. By doing
so, the introduction of a respiratory component on a parameter accounting
for the action of the micro-environment would not be necessary, since the
latent heat loss from the respiratory tract could be estimated by
considering the value of the resulting respiratory rate and the gradient of
vapour pressure between the circulating air and the surface of the anterior
respiratory tract.
Finally, we strongly believe that encouraging the interaction between
animal scientists and environmental physiologists would be a fruitful
means of improving the way in which the environmental conditions are
specified in models dealing with the energy balance of poultry.
References
Bakken, G.S. (1981) A two-dimensional operative temperature model for thermal
energy management by animals. Journal of Thermal Biology 6, 23–30.
Barnas, G.M. and Rautenburg, W. (1987) Temperature control. In: Seller, T.J. (ed.)
Bird Respiration, Vol. I. CRC Press, Boca Raton, Florida, pp. 131–153.
Bernstein, M.H. (1971) Cutaneous and respiratory evaporation in the painted
quail, Excalfactoria chinensis, during ontogeny of thermoregulation. Comparative
Biochemistry and Physiology (Part A) 38, 611–617.
Blanco, O.A., Gous, R.M. and Savage, M.J. (2003) Biophysical basis for controlling
the effect of microenvironment on the heat balance of broilers. In: Proceedings
of the 22nd Scientific Day WPSA Southern African Branch, Pretoria, pp. 123–138.
Blanco, O.A., Gous, R.M. and Savage, M.J. (2004a) A simulation model for estimating
the effect of warm micro-environmental conditions on the heat balance of a
broiler. In: Proceedings of the XII World Poultry Congress, Istanbul, Turkey.
Blanco, O.A., Gous, R.M. and Savage, M.J. (2004b) Modelling the physiological
responses of a broiler exposed to micro-environmental conditions above
thermoneutrality. In: Proceedings of the 23rd Scientific Day, World Poultry Science
Association Southern African Branch, Pretoria, pp. 10–23.
Campbell, G.S. and Norman, J.M. (1998) Environmental Biophysics. Springer, New
York.
Cena, K. (1974) Radiative heat loss from animals and man. In: Monteith, J.L. and
Mount, L.E. (eds) Heat Loss from Animals and Man: Assessment and Control.
Charles, D.R. (1974) The definition and measurement of the climatic environment
in poultry houses. In: Morris, T.R. and Freeman, B.M. (eds) Energy
Requirements for Poultry. British Poultry Science, Edinburgh, UK, pp. 73–77.
Charles, D.R. (1994) Comparative climatic requirements. In: Wathes, C.M. and
Charles, D.R. (eds) Livestock Housing. CAB International, Wallingford, UK,
pp. 3–24.
Charles, D.R., Elson, H.A. and Haywood, M.P.S. (1994) Poultry housing. In:
Wathes, C.M. and Charles, D.R. (eds) Livestock Housing. CAB International,
Czaric, M. and Tyson, B.L. (1990) Temperature Variations in a High Density, Curtain-
sided Laying House. ASAE Paper No. 90–4059. ASAE, Michigan.
Esmay, M.L. (1978) Principles of Animal Environment. AVI Publishing, Westport,
Virginia.
Hutchinson, J.C.D. (1954) Evaporative cooling in fowls. The Journal of Agricultural
Sciences (Cambridge), 45, 48–59.
Marder, J. and Ben-Asher, J. (1983) Cutaneous water evaporation – Its significance in

heat-stressed birds. Comparative Biochemistry and Physiology (Part A) 75, 425–431.
McArthur, A.J. (1981) Thermal insulation and heat loss from animals. In: Clark,
J.A. (ed.) Environmental Aspects of Housing for Animal Production. Butterworths,
London, pp. 37–60.
McArthur, A. J. (1987) Thermal interaction between animal and microclimate: a
comprehensive model. Journal of Theoretical Biology 126, 203–238.
Mitchell, M.A. (1985) Effects of air velocity on convective and radiant heat transfer
from domestic fowls at environmental temperatures of 20° and 30° C. British
Mitchell, M.A. (2005) The thermoneutral zone: Where should we be? Integrating
temperature effects into the productive cycle. In: Proceedings of Poultry beyond
2010, Brisbane, Australia, pp.213–236.
Monteith, J.L. (1974) Specification of the environment for thermal physiology. In:
Monteith, J.L. and Mount, L.E. (eds) Heat Loss from Animals and Man: Assessment
and Control. Butterworths, London, pp. 1–17.
Monteith, J.L. and Unsworth, M.H. (1990) Principles of Environmental Physics.
Edward Arnold, London.
Mount, L.E. (1968) The Climatic Physiology of the Pig. Edward Arnold, London.
Mount, L.E. (1979) Heat production. In: Adaptation to Thermal Environment: Man
and his Productive Animals. Edward Arnold, London, pp. 14–39.
Nichelman, M. and Tzschentke, B. (2002) Ontogeny of thermoregulation in
precocial birds. Comparative Biochemistry and Physiology (Part A) 131, 751–763.
Özkan, S., Smith, W.K and Bath, H.M. (2002) The development of thermal
resistance of the feather coat in broilers with different feathering genotypes
and feeding regimes. British Poultry Science 43, 472–481.
Poczopko, P. (1981) The environmental physiology of juvenal animals. In: Clark,
J.A. (ed.) Environmental Aspects of Housing for Animal Production. Butterworths,
London, pp. 109–130.
Richards, S.A. (1970) The role of hypothalamic temperature in the control of
panting in the chicken exposed to heat. Journal of Physiology 211, 341–358.
Richards, S.A. (1974) Aspects of thermoregulation in the fowl. In: Monteith, J.L.
and Mount, L.E. (eds) Heat Loss from Animals and Man: Assessment and Control.
Starr, J.R. (1981) Climate and need for housing. In: Clark, J. A. (ed.) Environmental
Aspects of Housing for Animal Production. Butterworths, London, pp. 19–35.
Tao, X. and Xin, H. (2003) Acute synergistic effects of air temperature, humidity,
and velocity on homoeostasis of market-size broilers. Transactions of the ASAE
46, 491–497.
Tselepidaki, J.G., Asimakopoulus, D.N., Katsouyani, K., Moustris, C., Touloumi, G.
and Pantazopoulou, A. (1995) The use of a complex thermohygrometric index
in predicting adverse effects in Athens. International Journal of Biometeorology 38,
194–198.
Turnpenny, J.R., McArthur, A.J., Clark, J.A. and Wathes, C.M. (2000a) Thermal
balance of livestock: 1. A parsimonious model. Agricultural and Forest Meteorology
101, 15–27.
Turnpenny, J.R., McArthur, A.J., Clark, J.A. and Wathes, C.M. (2000b) Thermal
balance of livestock: 2. Applications of a parsimonious model. Agricultural and
Forest Meteorology 101, 29–52.
Willmer, P., Stone, G. and Johnston, I. (2000) Environmental Physiology of Animals.
Blackwell Science, Oxford, UK.
Wathes, C.M. and Clark, J.A. (1981a) Sensible heat transfer from the fowl:
boundary-layer resistance of a model fowl. British Poultry Science 22, 161–173.
Wathes, C.M. and Clark, J.A. (1981b) Sensible heat transfer from the fowl: thermal
resistance of the pelt. British Poultry Science 22, 175–185.
Xin, H., Berry, I.L., Tabler, G.T. and Barton, T.L. (1994) Temperature and
humidity profiles of broiler houses with experimental conventional and tunnel
ventilation systems. Transactions of the ASAE 10, 535–542.
Zhou, W.T. and Yamamoto, S. (1997) Effects of environmental temperature and
heat production due to food intake on abdominal temperature, shank skin
temperature and respiration rate of broilers. British Poultry Science 38, 107–114.
11 Using Physiological Models to
Define Environmental Control
Strategies
M.A. MITCHELL
Roslin Institute, Roslin, Midlothian, EH25 9PS, UK
malcolm.mitchell5@btinternet.com
Introduction
Biologists develop and use mathematical models to understand observed

phenomena more fully, to characterize mechanisms and to predict the
behaviour of systems under a range of conditions or in the face of changing
inputs. Models may be derived from theory or empirical data or both, but
ultimately should be applicable to ‘real world’ situations and systems.
Models may be generally classified according to their theoretical origins, the
nature of the input-output variables and probability components and their
descriptive or predictive functions. Thus models describing a process or
response may be regarded as analytical, numerical or observational.
Analytical models are based upon fundamental equations, with some
approximations, to provide an explicit closed form solution. Numerical
models are derived from first principles and may be detailed and complex
whilst observational models are inferred from measured data (Gershenfeld,
1999). Such approaches may require the application of established
mathematical techniques including linear and quadratic programming and
ordinary and partial differential equations or more novel methodologies
such as neural networks. Other model descriptions, frequently encountered
in biology and agriculture, include the terms static and dynamic. Static
models have the mathematical form that relates a dependent to an
independent variable each taking different numerical values. Dynamic
models incorporate an element of time and frequently describe a time
course of events or responses. Dynamic models are commonly represented
in their differential form, i.e. by one or more first order differential
equations (France and Thornley, 1984). Finally models may be stochastic
(probabilistic), being based upon random trials, or deterministic, where
random phenomena are excluded. The selection of appropriate models is
based upon a range of criteria including goodness of fit, freedom from

210 M.A. Mitchell
systematic errors and simplicity. A given model is usually fitted and the
optimal parameters determined by minimization of an objective function,
usually the sum of squared errors (Garfinkel and Fegley, 1984).
Models of all types are commonly applied in animal biology and
agriculture and are employed to describe a wide range of processes
including growth and the influence thereon of genetics, nutrition and
husbandry (Wellock et al., 2003a, 2004a); reproductive performance; the
effects of environmental and other stressors upon animal production and
well-being (Wellock et al., 2003b, 2004b); behavioural and physiological
needs; factors influencing animal health; toxicity and environmental
hazard and management and risk extrapolation. The outputs of such
models may provide the basis for the setting of standards, optimum
practices and nutritional strategies, operational control systems and
practical guidelines and legislation. In the current review attention is
focused particularly upon the thermal environments to which livestock are
exposed and the modelling of animal responses in the context of
environmental specification and control.
Animals and the Thermal Environment
A major factor influencing both the efficiency of production of animals and

their welfare is the thermal environment. Problems may be encountered in
both extensive outdoor systems and in more intensive artificial or
‘controlled’ environments. An important objective of modern sustainable
production systems should be the matching of environmental conditions to
the biological needs or requirements of the animals (Wathes, 1994;
Mitchell and Kettlewell, 2004a). Such an approach should minimize the
stress to which livestock are exposed and reduce the physiological
homeostatic demands that are imposed. It is clear that modelling of
animals’ responses to the thermal environment, and the concomitant
development of models describing the thermal environment and its
components, will facilitate the design of effective control systems based
upon definition of acceptable ranges and limits for thermal loads and
optimum thermal micro-environments.
Livestock housing offers an opportunity to control the internal
environment and to maximize productivity if optimum conditions can be
achieved. Intensive systems, however, have attendant problems associated
with the precise specification of those conditions, animal welfare standards
and the impact upon the external environment. Even in ostensibly
‘environmentally controlled buildings’ animals may be subject to adverse
thermal loads during demanding meteorological conditions (Seedorf et al.,
1998). It is proposed that future design and management of livestock
environments should involve fully integrated systems and that these should
incorporate elements of the animals’ responses to rate limiting environmental
components in addition to other production and environmental concerns
(Frost et al., 1997, 2003; Wathes et al., 2001). Even in the specific case of the
Using Physiological Models to Define Environmental Control 211
thermal micro-environment the responses of housed livestock are complex

and dynamic. Animals housed under intensive conditions may have
limitations imposed upon the behavioural responses that individual free or
wild animals might exploit in the face of environmental challenge. Also,
genetic selection for desirable production traits may result in altered
homeostatic capabilities or in more extreme cases pathologies that limit
adaptive responses. Specification of practical environmental conditions and
control criteria should be cognizant of these issues in addition to utilizing
models of the animals’ physiological and behavioural reactions and
requirements. The development of truly integrated control systems
incorporating comprehensive simulation models of all the inputs in a
production system is dependent upon characterization and mechanistic
modelling of each of the sub-processes that must be combined to represent
the larger more complex system. The thermal micro-environment is one such
sub-process. It in turn consists of interactive components each of which may
exert either a direct influence upon the animal or a summative effect in
conjunction with other variables. Each sub-process analysis should also be
regarded in a more holistic context since biological responses to thermal
inputs may be markedly affected by non-thermal environmental variables.
The Thermal Micro-environment and Ventilation
In intensive system livestock buildings ventilation is the primary controller

of thermal conditions, although the major source of heat and moisture may
be the animals. The ventilation system must regulate the air temperature
and moisture content, provide air mixing and appropriate velocities of air
movement at animal level and remove air-borne contaminants. The control
strategy may be based upon simple measures such as air temperature with
correctly sited sensors or may involve simultaneous measurement of several
environmental inputs and processing through predetermined set points
and models based upon environmental, production or biological
requirements. Future control system algorithms may incorporate
information relating to physiological characteristics of the animals as well as
control variables such as heat, moisture and carbon dioxide balance (Allison
et al., 1991; Frost et al., 1997; Pedersen et al., 1998; Aerts et al., 2003). The
application of fuzzy logic based control systems (Gates et al., 2001) and
Computational Fluid Dynamics (CFD) should facilitate ever greater
sophistication and accuracy of ventilation of livestock housing provided the
appropriate optimum set points and acceptable ranges of the
environmental variables have been clearly defined.
Defining the Optimal Thermal Conditions
Several approaches have been adopted to determine the required thermal

conditions in livestock houses that are consistent with optimal productivity,
212 M.A. Mitchell
health and welfare. Clearly the heat exchange between an animal and its
immediate environment and the thermoregulatory responses to the
imposed conditions are the major consideration when determining
optimum temperature and humidity ranges in addition to predictions of
effects upon production parameters. Thus, in cattle, dynamic models of
responses to thermal loads have been developed based upon fractal
analysis of tympanic temperature responses that allow identification of the
environmental temperature at which heat stress is apparent (Hahn, 1999).
The fractal dimensions for body temperature may be employed to
categorize thermal stress (environmental temperature) from slight through
moderate to severe. Measurements of respiration rate changes were used
to further refine this ‘stress’ model. Dynamic responses provide the basis
for evaluating stress thresholds and differentiation of stress levels in cattle
and therefore can be used as criteria for environmental management
(Korthals et al., 1997; Hahn, 1999). In pigs, stochastic models have been
developed and evaluated that predict the effects of the thermal
environment upon food intake, daily weight gain, food conversion
efficiency and metabolic heat production (Wellock et al., 2003a). The model
indicated that within the apparent zone of thermal comfort there are no
temperature effects upon the production variables. A heat balance model
of the pig proposed by Fiahlo et al. (2004) and derived from first principles
of heat and mass transfer, simulates heat loss, heat balance and deep body
temperature responses in animals exposed to different thermal
environments. It is suggested that this model might be integrated into a
more comprehensive physiological model of the pig for precise definition
of optimum production environments and limits for imposed thermal
loads. Important advances in understanding the relationships between
animals and their thermal environments can be made through entirely
theoretical models of heat exchange which cannot always be evaluated by
comparison with experimental or empirical data. An idiosyncratic,
although useful, example has been presented by Phillips and Heath (2001)
in which the heat loss from ‘Dumbo’ (© Walt Disney Company) has been
modelled. Using fundamental principles, these authors have demonstrated
that Dumbo may often be at risk of losing more heat than he can produce
at rest and therefore will be at risk of hypothermia. The large pinnae area
associated with this character may have evolved to favour heat loss during
the hypermetabolism of flying!
Mathematical models describing both evaporative and non-evaporative
heat losses from livestock in relation to the thermal environment have been
developed (e.g. Ehrlemark, 1993). Such models provide valuable predictive
data based on the thermal properties of the animals but without reference
to the internal mechanisms of the thermoregulatory systems. The model
outputs can underpin determination of the thermoneutral zone and upper
critical temperature and ultimately may be employed as the basis for control
systems or algorithms. Thermal balance models for livestock, including
cattle and sheep (outdoors) and pigs and chickens (indoors), have also been
described by Turnpenny et al. (2000a,b). These mathematical models,
examining the effects of temperature, vapour pressure, wind speed and

solar radiation, were also derived from heat transfer principles and
incorporated physiological and thermoregulatory responses to thermal
challenge. The outputs of the models were successfully validated by
simulating the experimental thermal conditions, for each species, used by
previous workers. Whilst the models may be improved by incorporation of
data from more modern commercial strains of livestock, valuable outputs
are the prediction of thermal comfort zones for each species and the
sensitivity of heat exchange to individual environmental components such
as humidity that may be important in commercial production conditions.
From these models the prediction of upper and lower critical temperatures,
evaporation thresholds and thermoneutral zones facilitates the definition of
optimum ‘in house’ thermal micro-environments.
Stress and the Thermal Environment
It is apparent that the imposition of thermal challenges upon an animal

constitutes a stress. It may be argued that minimizing stress by controlling
the thermal micro-environment to within the thermoneutral zone the
physiological needs of the animal will be met and both productivity and
welfare will in turn be optimized. Thermal stress is often defined by some
index incorporating two or more of the primary components of the thermal
environment, e.g. temperature, water vapour pressure, radiant temperature
and air velocity (Budd, 2001). It may be suggested that thermal stress, or
more correctly strain, should be quantified by means of animal responses,
particularly those associated with physiological stress mediators and the
homeostatic systems directly influenced by the stress stimulus (Von Borell,
2001). Integration of physiological stress indices with concurrent behavioural
observations will allow identification of thermal conditions that impose
minimum threat to homeostasis and least risk of undue stress and thus, by
definition, the optimum thermal environment. Stress assessment will
therefore identify combinations of temperature, humidity, air movement and
radiant exchange that produce discomfort or distress and, as a corollary,
those combinations constituting thermal ‘comfort’.
Thermal Comfort
The term ‘thermal comfort’ has evolved in the discipline of human thermal
physiology. The ‘Zone of Thermal Comfort’ in man is defined as
the range of ambient temperatures, associated with specified mean radiant
temperature, humidity and air movement within which a human in specified
clothing expresses indifference to the thermal environment for an indefinite
period.
(Commission for Thermal Physiology of the International Union of
Physiological Sciences, 2003)
214 M.A. Mitchell
Models of thermal comfort have been developed based upon human

heat and mass balance and physiological responses to the thermal
environment. These models range from simple one-dimensional steady
state simulations to complex, transient, finite element codes with
thousands of nodes (Jones, 2002; Tanabe et al., 2002; Zhang et al., 2004;
Kaynakli and Kilic, 2005). The limitations of these models lie in the
accuracy of the inputs and the difficulty in relating the comfort perceptions
to the physiological variables simulated in the thermal models. Calibration
of models in terms of the perceived comfort is often performed in humans
by means of the Predicted Mean Vote (PMV), as proposed by Fanger
(1970, 1973) and Kjerulfjensen et al. (1975). PMV is a simple numeric scale
(3 to +3) used to designate perceived thermal comfort from cold to cool,
slightly cool, neutral, slightly warm, warm to hot. PMV is often related to
the Predicted Percentage Dissatisfied (PPD). For a numerical deviation in
PMV of ±1.5 approximately 50% of individuals are dissatisfied. Attempts
have been made to develop sensors or physical models which can integrate
the effects of temperature, humidity and air movement to yield ‘standard
effective temperatures’ or ‘equivalent temperatures’ and through this to
predict PMV (Ye et al., 2003; Mendes and da Silva, 2004). Clearly such
systems can form the basis of environmental control. A recent study has
employed thermodynamic analysis of human heat and mass transfer using
a two-node thermal model to identify the combinations of thermal
variables at which energy consumption for physiological and metabolic
functions is minimal (Prek, 2005). It was demonstrated that expected
thermal sensation expressed as PMV exhibited a strong correlation with
energy consumption. In man it is thus possible to base environmental
specifications both upon physiological or metabolic response models and
upon human perceptions of thermal comfort. It has been shown, using a
static linear model, that thermal comfort in humans is heavily determined
by skin temperature whilst physiological adaptive responses are
determined by changes in core temperature (Bulcao et al., 2000). Such
studies and all other human models using subjective correlates of
environmental and physiological variables can provide a mechanistic
understanding of thermoregulatory responses and definition of ‘thermal
comfort zones’.
In livestock, however, it is not possible to directly ask the animals for a
subjective assessment of thermal comfort and thus to determine PMV or
PPD. Other approaches are possible that can indirectly determine comfort.
Behavioural methods, including passive avoidance and continuous choice,
have been applied to assess aversion to multiple concurrent stressors in
poultry (Abeyesinghe et al., 2001a,b; MacCaluim et al., 2003). The stressors
included thermal variables. This approach might form the basis of a
behavioural model yielding preferred temperature-humidity combinations
and those producing mild to severe aversion and thus constituting
unacceptable thermal loads in commercial practice.
At present, models utilizing energy balance, thermal exchange and
physiological and stress responses are more likely to be employed to assess
thermal comfort in livestock and to form the basis of environmental

control systems. The comprehensive models described by McArthur (1987,
1991) and Turnpenny et al. (2000a,b) may facilitate determination of
thermoneutral zones and upper and lower critical temperatures and thus
thermal conditions in which heat or cold stress may occur. Randall (1993)
used models of heat, moisture and carbon dioxide production of pigs,
cattle and sheep to recommend appropriate thermal envelopes to ensure
‘comfort’ during transportation of the animals. The study demonstrated
the profound effects of humidity upon thermal comfort, particularly at
higher dry bulb temperatures, and stressed the importance of control of
the thermal environment by appropriate ventilation strategies. Other
studies examining the thermal environment during transport have
employed physical models to assess thermal comfort (Webster et al., 1993;
Weeks et al., 1997). A model chicken was constructed to simulate heat
exchanges between poultry and the transport environment. Measurements
of sensible heat loss from the model at different temperatures were
compared with published estimates of thermoneutral heat production to
estimate the range of thermal stresses experienced by chickens in transit. It
was concluded that thermal conditions ensuring comfort rarely occurred
in commercial transport. This model was also used to compare thermal
comfort upon different vehicle types and the efficacy of ventilation
regimes.
Indices of Thermal Loads
It is abundantly clear that the heat and mass exchange between an animal
and its immediate environment is dependent not only upon the dry bulb
temperature but upon the water vapour pressure or density, air movement
and radiant temperature. Whilst it is possible to model the effects of each
variable upon animal performance, physiological and metabolic responses
and thermal comfort, it is often desirable to develop a single integrated
index of imposed thermal load that will allow prediction of the biological
response. Various indices have been proposed for man and other animals
with a particular focus upon the prediction of the risk of heat stress or cold
stress. The situation may be summarized thus:
Ta is only one of several physical factors determining heat exchange with the
environment which occurs via ‘dry’ (conduction, radiant, convection) and ‘wet’
(evaporation) mechanisms. In addition to depending upon Ta each mechanism
also depends upon one or more of the following physical factors, humidity, air
velocity, barometric pressure, contact with housing structures and material e.g.
bedding/litter and effective radiant field.
(Romanovsky et al., 2002)
The age, physiological status, acclimation state, diet and food intake and
behavioural freedom of the animals will all contribute to the integrated
heat exchange of the animals and thus to any assessment of desirable
thermoneutral conditions.
216 M.A. Mitchell
It may be concluded that a more integrated and biologically

meaningful index of the thermal environment is essential to define
optimum thermal conditions or thermal comfort zones. Simple measures
of Ta as dry bulb temperature must no longer be considered adequate.
There have been many attempts to incorporate the various factors into
more comprehensive indices of environmental thermal load and the
consequent effects upon thermal balance of the animal. Theoretical models
have been developed which allow prediction of thermal balance and body
temperature in livestock (including poultry) over a wide range of practical
thermal environments into which the influence of humidity and air
movement can be incorporated (Turnpenny et al., 2000a,b; Fialho et al.,
2004). Temperature-humidity indices have been proposed as the main
descriptor of factors influencing heat exchange (e.g. Hubbard et al., 1999).
The Wet Bulb Globe Thermometer (WBGT) has been employed for a
similar purpose (Schroter et al., 1996), although alternative measures to
the WBGT integrating humidity and temperature effects have been
proposed more recently, such as the environmental stress index or ESI
(Moran et al., 2001, 2003). An effective temperature index for poultry was
proposed by Tzschentke and Nichelmann (2000) that incorporated the
effects of only temperature and air speed. Effective temperature was
originally designed for use in man and was intended to identify
combinations of temperature, humidity and air movement that produced
the same ‘sensation of cold or warmth’. This strategy may be very useful in
animals if physiological measures and behavioural assessments are used to
determine the biological equivalence of different permutations of the
relevant thermal factors (vide infra). The importance of this more
integrated environmental description has been emphasized in recent
industry and practical publications (e.g. Garden, 2004; Hulzebosch, 2004).
Indeed, a practical effective temperature scale for broiler production
environments has been produced by Barnwell (1997). This study tabulates
temperature, humidity and air speed and on the basis of thermal
equivalence indicates the necessary air velocities required to produce the
optimum conditions (an effective temperature of 21°C). It is tempting to
suggest that similar studies could accurately define the biologically
optimum conditions in terms of these environmental variables and thus
provide insight into the effective temperature equivalents of the true
thermo-neutral zone (TNZ). Another approach has been the development
of ‘standard operative temperatures’ integrating several environmental
variables including temperature, humidity, air movement and solar
radiation. The values represent an index of potential heat flow between an
animal and its environment (Beaver et al., 1996) and require accurate
measurement of all the relevant parameters in the production
environment. Physical models have been employed to derive standard
operative temperatures (e.g. O’Connor, 2000; Bartelt and Peterson, 2005)
and the application and performance of such models has been reviewed
recently (Dzialowski, 2005). Devices have been developed for field
measurement of standard operative temperature for selected avian species
(Bakken et al., 2001) and the principles involved in the design of

environmental control systems. Tao and Xin (2003) have proposed a
temperature-humidity-velocity index as a basis for management decisions
and control strategies in commercial broiler chicken production.
Physiological Response Modelling: Transport Thermal

Environments
A specific animal production thermal environment that has presented
many challenges to the industry and researchers alike is that encountered
during transportation of livestock. It has been stated that in order to
define comfort zones for animals in transit all the relevant thermal factors
must be taken into account (Randall, 1993).
In order to address these issues, physiological response modelling
must first involve full characterization of the commercial transport
environment and identify the stressors most likely to compromise the
welfare of the transported animals. Laboratory based studies then produce
quantitative, predictive response models over ranges of magnitude of
individual stressors such as thermal loads. The selection of an appropriate
spectrum of physiological variables, which reflect disturbances in all the
major homeostatic systems, is crucial. Measurements must indicate not
only the adequacy of homeostatic responses but the homeostatic effort
required for maintenance of controlled variables. Homeostatic responses
may be categorized as adequate compensation, inadequate compensation
or decompensation approximately corresponding to mild, moderate and
severe physiological stress. Thus models may contain assessments of
cardiovascular, respiratory, thermoregulatory and metabolic function in
addition to indices of hydration, electrolyte and acid-base or blood gas
status and the degree of any tissue dysfunction or pathology. Both point
measurements and continuous monitoring of physiological variables
should be employed. Radio-telemetric systems have been developed for
the continuous remote monitoring of physiological variables for use in
both laboratory and commercial transport studies (Kettlewell et al., 1997;
Mitchell et al., 2001a). Point measures and continuous remote monitoring
have been applied to studies in poultry, pigs and calves and the findings
used as the basis of recommendations for ‘thermal comfort zones’ and
‘thermal limits’ for animals in transit.
In the case of poultry, birds are transported from hatchery to farm, for
relocation and to slaughter (Mitchell and Kettlewell, 2004a,b). During
these journeys the birds may be exposed to hostile thermal micro-
environments resulting in mortalities, pathology, production losses and
reduced welfare (Mitchell and Kettlewell, 2003). Research in this area
required the development of an index of thermal load upon broilers in
transit that integrated the two most essential variables of this unique
environment, namely temperature and humidity. This index could then be
employed to establish the biological equivalence of different temperature-
218 M.A. Mitchell
humidity combinations through physiological modelling (Mitchell and

Kettlewell, 1998). In these studies models incorporating the use of
Apparent Equivalent Temperature (AET) have been employed to define
the ‘thermal comfort zones’ for birds under commercial conditions. Clearly
thermal environments which cause deep body temperature to approach
the upper or lower lethal limits will increase mortality and must be
regarded as totally unacceptable. Other thermal loads, however, will result
in the bird exhibiting a range of thermoregulatory responses aimed at
minimizing the change in deep body temperature. The adequacy of these
responses may be judged by measurement of deep body temperature and
the thermoregulatory effort involved can be assessed from physiological
monitoring of panting rate in the heat, shivering rate in the cold, changes
in metabolic rate and disturbances in blood gases and acid-base balance.
The physiological stress response model may thus be based upon the
relationships between these responses and a single integrated index of the
thermal load imposed upon the birds. The hypothetical relationship
between a given physiological variable, thermal load and the severity of
stress is shown in Fig. 11.1.
Studies were undertaken to determine the stress imposed upon
broilers by different thermal loads and to define thermal comfort zones for
birds in transit. Broiler birds were held in commercial transport crates and
placed in controlled climate chambers for a period of 3 h (typical of
commercial journeys). Various combinations of temperature and humidity
were employed in the range 10–35°C and 30–95% RH. Thermoregulatory
success (deep body temperature) and thermoregulatory effort (blood pH
and gas disturbances) were correlated with the actual imposed thermal
load (temperature-humidity combination).
AET was used as an index of thermal load. This parameter is derived
from the temperature, water vapour pressure and the corrected
Physiological stress
Severe
Moderate
Mild
Integrated Index of thermal load
Fig. 11.1. Physiological stress response vs integrated thermal load.

psychrometric constant (Eqn 11.2) and describes the total heat exchange
between a wetted surface and the environment (Eqn 11.1).
θ*app. = T + ( e/γ* ) (11.1)
where θ*app = AET
T = absolute temperature (K)
e = water vapour pressure (mbar)
γ* = corrected psychrometric constant (mbar/K)
γ* = γ (rv/rh) (11.2)
where rv = the resistance to water vapour transfer (s/m)
rh = the resistance to heat transfer (s/m).
AET may be calculated from the dry bulb temperature and the relative
humidity (RH) alone (Eqn 11.3).
10
(30.5905 −8.2⋅Log 10 ( K )+ 0.0024804 K − 3142.31 / K ) ⋅Φ (11.3)
AET = T +
(
0.93 ⋅ 0.0006363601K + 0.472 )
where
T = Observed temperature (°C)
K = T corrected to Kelvin (°C + 273.16)
Φ = Observed relative humidity (RH/100).
The derivation and application of AET is perhaps best described
graphically (Fig. 11.2). If a sample of air at 18°C (t) and 10 mbar vapour
pressure (e) is represented by point X then the line YXZ with a slope of -γ
yields the wet bulb temperature Y (12°C) and the equivalent temperature Z
(33.3°C). The line QX gives the dew point temperature (7.1°C) and the line
XP gives the saturation vapour pressure (20.6 mbar). If water vapour was
condensed adiabatically from a sample of saturated air (point Y) then the
temperature and water vapour pressure changes are described by the line
YX. When all the water vapour has condensed (e = 0) then t = Z
(Equivalent Temperature). In the calculation of AET the corrected
psychrometric constant replaces γ as described above.
Using the AET approach, the combinations of temperatures and
humidities that produce equivalent biological effects were determined. The
relationship between change in deep body temperature and AET is
presented in Fig. 11.3. It is clear that the response to thermal load is
similar to the hypothetical curve presented in Fig. 11.1 and as such allows
definition of temperature-humidity combinations imposing mild, moderate
and severe physiological stress. Similar response patterns were observed
for changes in blood pH and pCO2 which are used to assess homeostatic or
thermoregulatory effort.
Physiological stress response modelling, exploiting the concept of AET,
has thus allowed identification of ‘safe’, ‘alert’ and ‘danger’ combinations of
temperature and humidity which equate to mild, moderate and severe
220 M.A. Mitchell
30
Vapour pressure (mbar)

P
20
Q
10 X
Slope
–γ
M Z
0
0 10 20 30
Temperature (°C)
Fig. 11.2. The curve PQ gives saturation vapour pressure as a function of ambient
temperature. Point X represents a sample of air at 18°C and 10 mbar vapour pressure. The
slope of the line YZ, passing through X, corresponds to the psychometric constant. Other
symbols are explained in the text.
Change in body temperature (°C)
0
0 20 40 60 80 100
AET (°C)
Fig. 11.3. Relationship between change in deep body temperature and AET.
physiological stress. The model thus permits the definition of thermal

comfort zones for broilers in transit as presented in Fig. 11.4. At
temperature-humidity combinations yielding AET values of 40°C or less,
thermal stress will be minimal in transit. At temperature-humidity
combinations giving AETs between 40 and 45°C moderate thermal stress
will occur with some degree of hyperthermia and acid-base disturbances.
At AETs of 65°C or greater physiological stress may be deemed severe and
100
80
Relative humidity (%)

DANGER
60
ALERT
40
SAFE
20
0
10 15 20 25 30 35 40
Dry bulb temperature (°C)
Fig. 11.4. ‘Thermal Comfort Zones’ for broiler transport. Safe limit AET = 40°C; danger limit
AET = 65°C or greater.
mortalities will increase. Such thermal loads must be considered

unacceptable.
In a parallel series of experiments the effects of low environmental
temperatures accompanied by wetting and air movement were examined.
The experimental protocol was as for heat stress but the temperature range
employed was 4°C to +12°C with a constant air speed (supplied from a
laminar flow cabinet) of 0.7 m/s. For each temperature employed, birds
were either exposed to dry conditions or were intermittently wetted by
spraying for 1 min every 30 min. The experimental conditions were based
upon characterization of the environments on typical commercial broiler
transporters in the UK. Measurements of deep body temperature were
made before and after exposure to each temperature-air speed-wetting
combination. In addition some birds were surgically implanted with deep
body temperature data loggers to allow continuous recording of this
variable. The degree of physiological stress imposed was assessed by the
extent of the reduction in body temperature in each set of conditions.
In the cold stress studies, surface wetting (at constant air speed of 0.7
m/s) had a profound effect upon thermoregulatory success with an
increasing degree of hypothermia across the whole range of chamber
temperatures employed. At 12°C the fall in rectal temperature in wetted
birds was 3.03 ± 1.75°C, an additional decrease of 2.1°C compared to the
corresponding ‘dry’ group. In wetted birds the relationship between the
change in rectal temperature (Tr) and Te could be described by:
y = (3)(1013)(e–0.1065K) (11.4)
where
y = change in Tr, °C
K = Te corrected to Kelvin.
222 M.A. Mitchell
Marked hypothermia was therefore induced at all other chamber

temperatures and ranged from a reduction of 4.4 ± 3.2°C at Te = 8°C to a
maximum and life threatening fall of 14.2 ± 5.4°C when Te = 4°C. As
lethal deep body temperature in the fowl is 24°C then from the derived
relationship between environmental temperature and the change in core
temperature (Fig. 11.5) it is apparent that a lethal hypothermia will occur
in all wetted birds when Te = 9.5°C and marked elevations of mortalities
will be observed if Te falls to only +1.0°C. The findings from this model
define the lower limits for temperature exposure in transported broilers in
wet and in dry conditions (Hunter et al., 1997; Mitchell et al., 1997, 2000,
2001a). Improved ventilation strategies based upon the outputs of this
model can reduce the risk of cold stress and all the sequelae in terms of
animal welfare, product quality and animal mortality.
Conclusions
In summary, physiological modelling evaluates both homeostatic success

and homeostatic effort and through the determination of biological
equivalence of thermal loads consisting of different combinations of
temperature and humidity or other environmental parameters can define
thermal comfort zones. AET has been used already in a technical manual
(Ross Tech 99/34, 1999) as a guide to the temperature profiles for broiler
rearing taking into account the important interactions of temperature and
humidity in determining heat exchange. Incorporation of air movement
into a similar model of broiler house environments (see Tao and Xin, 2003;
Yahav et al., 2004) should provide the sound scientific basis for setting in-
house thermal conditions to optimize not only productivity but also bird
Change in body temperature (°C)
15
10
0
265 270 275 280 285 290
Environmental temperature (K)
Fig. 11.5. Relationship between change in deep body temperature in wetted birds and
environmental temperature (K).
comfort and welfare. Physiological models may be complemented by

behavioural studies in which birds’ preferences for temperature-humidity
combinations may be assessed by established choice techniques
(Abeyesinghe et al., 2001a,b; MacCaluim et al., 2003). Bird monitoring and
accurate measures of bird responses to well defined thermal loads will play
an increasing role in the development of advanced automated
environmental control and management systems (Frost et al., 1997; Naas,
2002; Aerts et al., 2003; Taylor et al., 2004).
The novel experimental modelling approaches described herein have
provided data which are specific to transportation conditions and directly
applicable to commercial practice. Using physiological response modelling
(Mitchell et al., 1996, 2001b; Mitchell and Kettlewell 1998, 2004a,b) the
optimum thermal envelope for broiler carriage has been defined in terms
of the ‘physiological thermal comfort zones’ and factors precipitating or
contributing to the incidence of thermal stress have been identified. The
outputs of the models have been employed as the basis for environmental
control strategies on animal transport vehicles using mechanical ventilation
systems with the controlled capacity to remove the imposed heat and
moisture loads thus matching the ‘on-board’ thermal environments to the
physiological requirements of the ‘passengers’ (Kettlewell and Mitchell,
2001a,b; Kettlewell et al., 2001a,b; Mitchell and Kettlewell, 2004a,b,c). It
may thus be proposed that this very successful physiological modelling
approach may be applied to other areas of animal production in order to
define the optimum thermal conditions in terms of biological response and
requirements and to underpin the development of integrated monitoring
and control systems for livestock environments.
Acknowledgements
The work described in this review was supported by DEFRA. The author is
greatly indebted to Peter Kettlewell and Richard Hunter for their efforts
and inputs in all of the research projects in this area and to all the staff
involved at the Roslin and Silsoe Research Institutes for their technical
assistance, which ensured the successful completion of the modelling
research programme.
References
Abeyesinghe, S.M., Nicol, C.J., Wathes, C.M. and Randall, J.M. (2001a)
Development of a raceway method to assess aversion of domestic fowl to
concurrent stressors. Behavioural Processes 56, 175–194.
Abeyesinghe, S.M., Wathes, C.M., Nicol, C.J. and Randall, J.M. (2001b) The
aversion of broiler chickens to concurrent vibrational and thermal stressors.
Applied Animal Behaviour Science 73, 199–215.
Aerts, J.M., Wathes, C.M. and Berckmans, D. (2003) Dynamic data based modeling
224 M.A. Mitchell
of heat production and growth of broiler chickens: development of an

integrated management system. Biosystems Engineering 84, 257–266.
Allison, J.M., White, J.M., Worley, J.W. and Kay, F.W. (1991) Algorithms for
microcomputer control of the environment of a production broiler house.
Transactions of the American Society of Agricultural Engineers 34, 313–320.
Bakken, G.S., Boysen, A.F., Korschgen, C.E., Kenow, K.P. and Lima, S.L. (2001)
Design and performance of a rugged standard operative temperature
thermometer for avian studies. Journal of Thermal Biology 26, 595–604.
Barnwell, R. (1997) Evaporative cooling systems: understanding and maximizing
summer ventilation. Technical News: Quarterly publication of Cobb-Vantress,
Incorporated 5, 1–3.
Bartelt, P.E. and Peterson, C.R. (2005) Physically modeling operative temperatures
and evaporation rates in amphibians. Journal of Thermal Biology 30, 93–102.
Beaver, J.M., Olson, B.E. and Wraith, J.M. (1996) A simple index of standard
operative temperature for mule deer and cattle in winter. Journal of Thermal
Biology 21, 345–352.
Budd, G.M. (2001) Assessment of thermal stress – the essentials. Journal of Thermal
Biology 26, 371–374.
Bulcao, C.F., Frank, M.F., Srinivasa, N.R., Tran, K.M. and Goldstein, D.S. (2000)
Relative contribution of core and skin temperatures to thermal comfort in
humans. Journal of Thermal Biology 25, 147–150.
Commission for Thermal Physiology of the International Union of Physiological
Sciences (2003) Glossary of terms for thermal physiology. Journal of Thermal
Biology 28, 75–106.
Dzialowski, E.M. (2005) Use of operative temperature and standard operative
temperature models in thermal biology. Journal of Thermal Biology 30, 317–334.
Ehrlemark, A.G. (1993) Calculation of heat and moisture dissipation from animals
using basic thermal properties. In: Collins, E. and Boon, C. (eds) Proceedings of
the 4th International Livestock Environment Symposium, University of Warwick,
Coventry, 6–9 July 2003. American Society of Agricultural Engineers,
pp. 835–842.
Fanger, P.O. (1970) Thermal Comfort: Analysis and Applications in Environmental
Engineering. McGraw-Hill, London.
Fanger, P.O. (1973) Assessment of man’s thermal comfort in practice. British Journal
of Industrial Medicine 30, 313–324.
Fialho, F.B., Bucklin, R.A., Zazueta, F.S. and Myer, R.O. (2004) Theoretical model
of heat balance in pigs. Animal Science 79, 121–134.
France, J. and Thornley, J.H.M. (1984) Mathematical Models in Agriculture.
Frost, A.R., Schofield, C.P., Beaulah, S.A., Mottram, T.T., Lines, J.A. and Wathes,
C.M. (1997) A review of livestock monitoring and the need for integrated
systems. Computers and Electronics in Agriculture 17, 139–159.
Frost, A.R., Parsons, D.J., Stacey, K.F., Robertson, A.P., Welch, S.K., Filmer, D. and
Fothergill, A. (2003) Progress towards the development of an integrated
management system for broiler production. Computers and Electronics in
Garden, M. (2004) The optimum environment for the modern broiler.
http://www.asian-agribiz.com
Garfinkel, D. and Fegley, K.A. (1984) Fitting physiological models to data. American
Journal of Physiology (Regulatory, Integrative and Comparative Physiology) 246,
R641–R650.
Gates, R.S., Chao, K. and Sigrimis, N. (2001) Identifying design parameters for
fuzzy control of staged ventilation control systems. Computers and Electronics in
Gershenfeld, N. (1999) The Nature of Mathematical Modeling. Cambridge University
Press, Cambridge, UK, pp. 1–22.
Hahn, G.L. (1999) Dynamic responses of cattle to thermal heat loads. Journal of
Animal Science 77 (supplement 2), 10–20.
Hubbard, K.G., Stooksbury, D.E., Hahn, G.L. and Mader, T.L. (1999) A
climatological perspective on feedlot cattle performance and mortality related
to the temperature-humidity index. Journal of Production Agriculture 12,
650–653.
Hulzebosch, J. (2004) What affects the climate in poultry houses. World Poultry 20,
36–38.
Hunter, R.R., Mitchell, M.A. and Matheu, C. (1997) Distribution of ‘dead on
arrivals’ within the bio-load on commercial broiler transporters: correlation
with climate conditions and ventilation regimen. British Poultry Science 38,
Supplement, S7–S9.
Hunter, R.R., Mitchell, M.A. and Matheau, C. (2001) Mortality of broiler chickens
in transit – correlation with the thermal micro-environment. In: Stowell, R.R.
(ed.) Proceedings of the 6th International Livestock Environment Symposium,
Louisville, Kentucky, USA, 21–23 May 2001. American Society of Agricultural
Engineers, pp. 542–549.
Jones, B.W. (2002) Capabilities and limitations of thermal models for use in
thermal comfort standards. Energy and Buildings 34, 653–659.
Kaynakli, O. and Kilic, M. (2005) Investigation of indoor thermal comfort under
transient conditions. Building and Environment 40, 165–174.
Kettlewell, P.J. and Mitchell, M.A. (2001a) Comfortable ride: concept 2000 provides
climate control during poultry transport. Resource Engineering and Technology for
a Sustainable World 8, 13–14.
Kettlewell, P.J. and Mitchell, M.A. (2001b) Mechanical ventilation: improving the
welfare of broiler chickens in transit. Journal of the Royal Agricultural Society of
England 162, 175–184.
Kettlewell, P.J., Mitchell, M.A. and Meeks, I.M. (1997) An implantable radio-
telemetry system for remote monitoring of heart rate and deep body
temperature in poultry. Computers and Electronics in Agriculture 17, 161–175.
Kettlewell, P.J., Hoxey, R.P., Hampson, C.J., Green, N.R., Veale, B.M. and Mitchell,
M.A. (2001a) Design and operation of a prototype mechanical ventilation
system for livestock transport vehicles. Journal of Agricultural Engineering
Research 79, 429–439.
Kettlewell, P.J., Hoxey, R.P., Hampson, C.J., Green, N.R., Veale, B.M. and Mitchell,
M.A. (2001b) Heat and moisture generation of livestock during transportation.
In: Stowell, R.R. (ed.) Proceedings of the 6th International Livestock Environment
Symposium, Louisville, Kentucky, USA, 21–23 May, 2001. American Society of
Agricultural Engineers, pp. 519–526.
Kjerulfjensen, P., Nishi, Y., Fanger, P.O. and Gagge, A.P. (1975) Investigation on
man’s thermal comfort + physiological response. Journal of the American Society
of Heating Refrigerating and Air-Conditioning Engineers 17, 65–68.
Korthals, R.L., Chen, Y.R., Hahn, G.L. and Eigenberg, R.A. (1997) Calculation of
fractal dimension from cattle thermoregulatory responses. Journal of Thermal
Biology 22, 285–293.
MacCaluim, J.M., Abeyesinghe, S.M., White, R.P. and Wathes, C.M. (2003) A
226 M.A. Mitchell
continuous-choice assessment of the domestic fowl’s aversion to concurrent

transport stressors. Animal Welfare 12, 95–107.
McArthur, A.J. (1987) Thermal interaction between animal and microclimate: a
comprehensive model. Journal of Theoretical Biology 126, 203–238.
McArthur, A.J. (1991) Thermal interaction between animal and microclimate:
specification of a ‘Standard Environmental Temperature’ for animals outdoors.
Journal of Theoretical Biology 148, 331–343.
Mendes, J.C.A.F. and da Silva, M.C.G. (2004) Development of a new thermal
environment meter responding both to sensible and latent heat fluxes.
Measurement Science and Technology 15, 839–847.
Mitchell, M.A. and Kettlewell, P.J. (1998) Physiological stress and welfare of broiler
chickens in transit: Solutions not problems! Poultry Science 77, 1803–1814.
Mitchell, M.A. and Kettlewell, P.J. (2003) Sistemas de transporte e Bem-Estar de
Frangos de Corte. APINCO Conferencia de Ciencia e Technologia Avicolas.
Campinas, Brazil, 7–9 May 2003, pp. 199–215.
Mitchell, M.A. and Kettlewell, P.J. (2004a) The poultry transport thermal
environment – matching ‘on-board’ conditions to the birds’ physiological
requirements. Proceedings of the Australian Poultry Science Symposium 16, 175–178.
Mitchell, M.A. and Kettlewell, P.J. (2004b) Transport and Handling. In: Weeks,
C.A. and Butterworth, A. (eds) Measuring and Auditing Broiler Welfare. CABI
Publishing, Wallingford, UK, pp.145–160.
Mitchell, M.A. and Kettlewell, P.J. (2004c) Transport of chicks, pullets and spent
hens. In: Perry, G.C. (ed.) Welfare of the Laying Hen. CABI Publishing,
Mitchell, M.A., Kettlewell, P.J., Carlisle, A.J. and Matheu, C. (1996) The use of
apparent equivalent temperature (AET) to define the optimum thermal
environment for broilers in transit. Poultry Science 75 (Supplement 1), 18.
Mitchell, M.A., Carlisle, A.J., Hunter, R.R. and Kettlewell, P.J. (1997) Welfare of
broilers during transportation: cold stress in winter – causes and solutions. In:
Koene, P. and Blokhuis, H.J. (eds) Proceedings of the Fifth European Symposium on
Poultry Welfare. University of Wageningen and Institute of Animal Science and
Health, Netherlands, pp. 49–52.
Mitchell, M.A., Carlisle, A.J., Hunter, R.R. and Kettlewell, P.J. (2000) The responses
of birds to transportation. Proceedings of the World’s Poultry Congress, Montreal,
Canada, 22–24 August 2000, 1–14.
Mitchell, M.A., Kettlewell, P.J., Lowe, J., Hunter, R.R., King, T., Ritchie, M. and
Bracken, J. (2001a) Remote physiological monitoring of livestock – An
implantable radio-telemetry system. In: Stowell, R.R. (ed.) Proceedings of the 6th
International Livestock Environment Symposium, Louisville, Kentucky, USA, 21–23
May 2001. American Society of Agricultural Engineers, pp. 535–541.
Mitchell, M.A., Kettlewell, P.J., Hunter, R.R. and Carlisle, A.J. (2001b) Physiological
stress response modeling – application to the broiler transport thermal
environment. In: Stowell, R.R. (ed.) Proceedings of the 6th International Livestock
Environment Symposium, Louisville, Kentucky, USA, 21–23 May 2001. American
Society of Agricultural Engineers, pp. 550–555.
Moran, D.S., Pandolf, K.B., Shapiro, Y., Heled, Y., Shani, Y., Mathew, W.T. and
Gonzalez, R.R. (2001) An environmental stress index (ESI) as a substitute for the
wet bulb globe temperature (WBGT). Journal of Thermal Biology 26, 427–431.
Moran, D.S., Pandolf, K.B., Shapiro, Y., Laor, A., Heled, Y. and Gonzalez, R.R.
(2003) Evaluation of the environmental stress index for physiological variables.
Journal of Thermal Biology 28, 43–49.
Naas, I.D.A. (2002) Applications of mechatronics to animal production. Agricultural

Engineering Review; CIGR Journal of Scientific Research and Development. Vol. IV
2002.
O’Connor, M.P. (2000) Extracting operative temperatures from temperatures of
physical models with thermal inertia. Journal of Thermal Biology 25, 329–343.
Pedersen, S., Takai, H., Johnsen, J.O., Metz, J.H.M., Koerkamp, P.W.G.G., Uenk,
G.H., Phillips, V.R., Holden, M.R., Sneath, R.W., Short, J.L., White, R.P.,
Hartung, J., Seedorf, J., Schroder, M., Linkert, K.H. and Wathes, C.M. (1998)
A comparison of three balance methods for calculating ventilation rates in
livestock buildings. Journal of Agricultural Engineering Research 70, 25–37.
Phillips, P.K. and Heath, J.E. (2001) Heat loss in Dumbo: a theoretical approach.
Journal of Thermal Biology 26, 117–120.
Prek, M. (2005) Thermodynamic analysis of human heat and mass transfer and
their impact on thermal comfort. International Journal of Heat and Mass Transfer
48, 731–739.
Randall, J.M. (1993) Environmental parameters necessary to define comfort for
pigs, cattle and sheep in livestock transporters. Animal Production 57, 299–307.
Romanovsky, A.A., Ivanov, A.I. and Shimansky, Y.P. (2002) Molecular biology of
thermoregulation. Selected contribution: ambient temperature for
experiments in rats: a new method for determining the zone of thermal
neutrality. Journal of Applied Physiology 92, 2667–2679.
Ross Tech 99/34 (1999) www.aviagen.com Ross Techs.
Schroter, R.C., Marlin, D.J. and Jeffcott, L.B. (1996) Use of the wet bulb globe
temperature (WBGT) index to quantify environmental heat loads during
three-day-event competitions. Equine Veterinary Journal 22, 3–6.
Seedorf, J., Hartung, J., Schroder, M., Linkert, K.H., Pedersen, S., Takai, H.,
Johnsen, J.O., Metz, J.H.M., Koerkamp, P.W.G.G., Uenk, G.H., Phillips, V.R.,
Holden, M.R., Sneath, R.W., Short, J.L., White, R.P. and Wathes, C.M. (1998)
Temperature and moisture conditions in livestock buildings. Northern Europe
Journal of Agricultural Engineering Research 70, 49–57.
Tanabe, S., Kobayashi, K., Nakano, J., Ozeki, Y. and Konishi, M. (2002) Evaluation
of thermal comfort using combined multi-node thermoregulation (65MN) and
radiation models and computational fluid dynamics (CFD). Energy and
Buildings 34, 637–646.
Tao, X. and Xin, H. (2003) Acute synergistic effects of air temperature, humidity
and velocity on homeostasis of market-size broilers. Transactions of the American
Society of Agricultural Engineers 46, 491–497.
Taylor, C.J., Leigh, P., Price, L., Young, P.C., Vranken, E. and Berckmans, D. (2004)
Proportional-integral-plus (PIP) control of ventilation rate in agricultural
buildings. Control Engineering Practice 12, 225–233.
Turnpenny, J.R., Wathes, C.M., Clark, J.A. and McArthur, A.J. (2000a) Thermal
balance of livestock. 1: A parsimonious model. Agriculture and Forest Meteorology
101, 15–27.
Turnpenny, J.R., McArthur, A.J., Clark, J.A. and Wathes, C.M. (2000b) Thermal
balance of livestock. 2: Applications of a parsimonious model. Agriculture and
Forest Meteorology 101, 29–52.
Tzschentke, B. and Nichelmann, M. (2000) Influence of age and wind speed on
total effective ambient temperature in three poultry species (Gallus domesticus,
Cairina moschata, Meleagris gallopavo). Archiv fur Geflugelkunde 64, 1–8.
Von Borell, E.H. (2001) The biology of stress and its application to livestock housing
and transportation assessment. Journal of Animal Science 79 (E supplement),
E260–E267.
228 M.A. Mitchell
Wathes, C.M. (1994) Animals in man’s environment: a question of interest. Outlook

on Agriculture 23, 47–54.
Wathes, C.M., Abeyesinghe, S.M. and Frost, A.R. (2001) Environmental design and
management for livestock in the 21st century: resolving conflicts by integrated
solutions. In: Stowell, R.R. (ed.) Proceedings of the 6th International Livestock
Environment Symposium, Louisville, Kentucky, USA, 21–23 May 2001. American
Society of Agricultural Engineers, pp. 5–14.
Webster, A.J.F., Tuddenham, A., Saville, C.A. and Scott, G.B. (1993) Thermal-stress
on chickens in transit. British Poultry Science 34, 267–277.
Weeks, C.A., Webster, A.J.F. and Wyld, H.M. (1997) Vehicle design and thermal
comfort of poultry in transit. British Poultry Science 38, 464–474.
81, 2995–3007.
Wellock, I.J., Emmans, G.C. and Kyriazakis, I. (2004a) Describing and predicting
potential growth in the pig. Animal Science 78, 379–388.
Wellock, I.J., Emmans, G.C. and Kyriazakis, I. (2004b) Modeling the effects of
stressors on the performance of a population of pigs. Journal of Animal Science
82, 2442–2450.
Yahav, S., Straschnow, A., Luger, D., Shinder, D., Tanny, J. and Cohen, S. (2004)
Ventilation, sensible heat loss, broiler energy and water balance under harsh
environmental conditions. Poultry Science 83, 253–258.
Ye, G.D., Yang, C.Z., Chen, Y.M. and Li, Y.G. (2003) A new approach for
measuring predicted mean vote (PMV) and standard effective temperature
(SET*). Building and Environment 38, 33–44.
Zhang, H., Huizenga, C., Arens, E. and Wang, D. (2004) Thermal sensation and
comfort in transient non-uniform thermal environments. European Journal of
Applied Physiology 92, 728–733.
12 Modelling Egg Production in
Laying Hens
S.A. JOHNSTON AND R.M. GOUS
Animal and Poultry Science, School of Agricultural Sciences and Agribusiness,
University of KwaZulu-Natal, Private Bag X01, Scottsville 3209, South Africa.
sajohnston@nitrosoft.co.za
Introduction
The processes of ovulation and oviposition in domestic hens have been the
subjects of much interest and of numerous investigations by poultry
scientists over the past century. The associated temporal relationships
make an intriguing study, both for the modeller aiming to predict rate of
lay and for the physiologist attempting to integrate the hormonal and
neurological pathways. Although tremendous progress has been made in
recent years in understanding the hormonal control, mainly due to the
advent of radioimmunoassay techniques, some aspects of the hen’s
reproductive system continue to elude scientific explanation.
Domestic hens lay their eggs in sequences; a sequence being made up
of a number of consecutive daily ovipositions followed by one or more
pause days, when no egg is laid. The first egg of a sequence is usually laid
early in the morning, roughly 8 h after the onset of darkness. Each
subsequent oviposition occurs progressively later in the day, at intervals
slightly longer than 24 h, on successive days. The difference in times of day
at which eggs are laid on consecutive days is known as the lag, as is the
corresponding delay in the ovulation cycle (Fraps, 1955). Total lag is a
measure of the proportion of the day used for egg laying and can be
regarded as an indication of the length of the open period for luteinizing
hormone (LH) release and hence ovulation (Lillpers and Wilhelmson,
1993). The reason why hens lay eggs in sequences is that the ovulatory
cycle is due to the interaction of two asynchronous physiological systems
(Etches, 1984; Johnson, 1984). These are the maturation of the largest
ovarian follicle and a circadian rhythm that restricts the preovulatory surge
of LH to a limited portion of the day. Ovulation occurs when the
maturation of a follicle coincides with a certain phase of the circadian
rhythm.

230 S.A. Johnston and R.M. Gous
Much controversy surrounds the issue of whether circadian rhythms

are indeed involved in timing the preovulatory surge of LH, since
conclusive evidence is still lacking in poultry. Such evidence would lend
weight to the mathematical theory of the ovulatory cycle proposed by
Etches and Schoch (1984). In their model, the concentration of the
regulator substance is assumed to have a circadian rhythm, being
reinitiated at the same time daily. Birds are known to have a circadian
rhythm in photosensitivity that is involved in their reproductive responses
to light (Siopes and Wilson, 1980). This may be partially attributed to
melatonin, the pineal hormone thought to link the reproductive system to
the light:dark cycle, because it has a circadian rhythm of release that
entrains to the light:dark cycle (Menaker et al., 1981). However, efforts to
relate the pineal gland to photoperiodic control of LH have failed thus far
(Liou et al., 1987). In the meantime, it is convenient to assume that the
restriction of LH release to a limited period of the day is brought about by
a hormone or neurotransmitter that has a circadian rhythm of secretion or
activity.
A number of publications have shown that either a single or two
mathematical functions may be used to describe the egg production curve
(Johnston, 1993). The Adams-Bell model, for example, uses a logistic
growth curve to describe the initial rise to peak and a negative linear
function to mimic the decline in egg production post peak (Adams and
Bell, 1980). These empirical models show no awareness of the underlying
reasons for the shape of the curve, nor do they take into account the fact
that rate of egg production for the flock is derived from individual
performances.
Rate of lay is determined by the ovulation rate, which varies amongst
individuals and over time. The ovulation rate, in turn, is established by an
interaction between the rate of follicle maturation and the rhythmic release
of LH. It therefore makes sense that a model designed to predict the egg
production of a flock of hens should start by predicting the ovulation rate
for each hen. There are two main benefits to this approach. First, the large
number of variables that play a role in egg production are accounted for in
the model. In this manner the process of egg production is acknowledged
as being an extremely complex one. Secondly, the variation between
individuals within a population is brought to light. Most poultry producers
record performance indicators in terms of means (e.g. mean egg weight,
rate of lay, egg output and feed intake) without having the vaguest idea of
the extent of the variation within the flock. The peril is that unsatisfactory
performances from some individuals may be masked by the population
means, with the result that overall flock performance efficiency is reduced.
In a stochastic model, each variable has an associated standard deviation or
coefficient of variation. These are either measured during experimentation
or derived from published scientific papers, failing which educated guesses
need to be made by the modeller.
The development of a mechanistic model of this nature therefore
requires a thorough understanding of the system to be modelled, which in
Modelling Egg Production in Laying Hens 231
turn entails a comprehensive review of the available literature. During this

procedure, several unknown quantities usually become evident. These may
relate to the response of a selected strain to environmental stimuli (e.g.
how the age at first egg for the strain is affected by different lighting
programmes), or to the values of constants (e.g. the constants defining the
relationship between yolk weight and albumen weight) and of the variables
required by the model (such as the prevalence of double-yolked or soft-
shelled eggs during the course of the laying cycle). The modeller is thus
likely to challenge current knowledge and to direct further research
towards providing precise answers to relevant questions.
The Ovulatory Cycle
A mathematical model of the hen’s ovulatory cycle
Based on the work of Fraps (1955), a mathematical model of the ovulatory

cycle was presented by Etches and Schoch (1984), to provide support for
the theory of two independent biological systems interacting to produce
the observed pattern of sequential laying in commercial hens. The model
was able to predict times of ovulation close to actual times of oviposition
recorded under experimental conditions for different sequence lengths, on
the assumption that the ovulatory and ovipository cycles were displaced by
the oviducal term. Furthermore, the accepted features of lag were
demonstrated by the mathematical model. The calculated lag between
predicted ovulation times decreased initially between successive ovulations,
was minimal for mid-sequence ovulations and increased towards the end of
a sequence, so that the greatest lag was between the penultimate and last
ovulations. The predicted total lag given by their model for the different
sequence lengths fell within the prescribed 8- to 10-h open period. Total
lag was also positively related to sequence length, with a two-ovulation
sequence having a relatively short total lag of 4 h 32 min and a nine-
ovulation sequence, 8 h 7 min. Mean lag was negatively related to the
sequence length, so that longer sequences had comparatively shorter
intervals between successive ovulations (Fig. 12.1).
A three-compartmental model was used to describe how the
production of a hypothetical regulator substance could restrict the release
of LH to a limited period of the day. The regulator is assumed to be
produced in compartment one and instantaneously released into
compartment two at regular intervals determined by a circadian oscillator.
The regulator is then released into compartment three at a given rate and
cleared from this compartment at a different rate. The concentration of
the substance (Fraps’ threshold of response of the neural component) in
compartment three at time t depends on the rates at which it is entering
and leaving the compartment and is given by:
R3(t) = a1 •e 1 (t-S1) – a2 •e 2 (t-S1) (12.1)
Fig. 12.1. The observed (shaded columns) and predicted (blank columns) mean lag for eight
sequence lengths (data from Etches and Schoch, 1984).
where a1 is a constant and S1 is the time of day when the cycle is completed
and the regulator function is reinitiated. The events leading up to
ovulation take place some hours beforehand, but the inclusion of the
parameter S1 shifts the curve so that R3(t) exceeds a threshold value from
about 07:00 to early afternoon, not during the night. Thus the curve given
by Eqn 12.1 also represents the period of the day during which ovulation
may take place. The values of the parameters 1, 2 and S1 and of the
constant a1 used to predict ovulation sequences of varying lengths are
given in Table 12.1.
The value of a2 depends on the values allocated to parameters 1 and
2 and may be calculated from the function
a2 = (a1(a1 .e (λ1 . 24) )) / (1e (λ2 . 24)) (12.2)
when the light:dark cycle is of 24 h duration.
The second equation in the ovulation model of Etches and Schoch
(1984) represents the maturation of the largest ovarian follicle (Fraps’
excitation hormone concentration) and takes the form of a Gompertz
function:
G(t) = b1 .eb2 eb3 (tS2) (12.3)
where S2 represents the time taken after the previous ovulation to
reinitiate G(t). This interval is necessary because the follicle is not sensitive
to gonadotrophic stimulation for a number of hours after the previous
ovulation. The values for the parameters b1, b2, b3 and S2 to predict
sequences of different lengths are shown in Table 12.1.
Follicle maturation refers to the competence of the follicle to ovulate in
response to an LH surge, rather than to an increase in weight or diameter
of the follicle. This phase is brought about by a number of physiological
Table 12.1. The values of the parameters and constants used by Etches and Schoch (1984),
University of Guelph data, to predict the times of ovulation and oviposition.
Seq. length a1 1 2 S1 b1 b2 b3 S2
2 2.175 0.14 0.25 8.0 0.75 4.5 0.22 17.0
3 2.175 0.14 0.25 7.5 0.75 5.1 0.27 17.0
4 2.175 0.14 0.25 6.75 0.75 5.2 0.30 17.0
5 2.175 0.15 0.285 6.5 0.75 5.15 0.315 17.5
6 2.175 0.15 0.285 6.5 0.76 5.05 0.32 17.5
7 2.175 0.15 0.285 6.5 0.78 5.2 0.325 17.5
8 2.175 0.15 0.285 6.5 0.78 5.15 0.33 17.5
9 2.175 0.15 0.285 6.5 0.785 5.17 0.334 17.5
processes. For instance, the granulosa cells in hierarchical follicles

differentiate so that they are capable of secreting progesterone (Johnson,
1996), the hormone needed to initiate and maintain the ovulation-
inducing surge of LH. Prior to ovulating, the F1 follicle changes from an
FSH- to a LH-dominated phase (Tilly et al., 1991). It is of interest to note
that the rate of maturation of follicles is affected by advancing hen age,
since Johnson et al. (1986) found that the sensitivity of the follicle to LH
and hence its ability to ovulate, declines with age.
At the time of day when the two functions intersect, ovulation takes
place. This is based on the hypothesis that a sufficient concentration of the
regulator coincides with a sufficient concentration of receptor to induce
ovulation (Etches and Schoch, 1984). The intersection is brought about by
using a mirror image of the curve produced by Eqn 12.1, so that the
moment of ovulation occurs when G(t) (1 R3(t)) = 0. The greater the
concentration of regulator and the more advanced the stage of follicle
maturation, the greater is the likelihood of intersection of the two curves.
The earlier in the day each consecutive ovulation occurs, the shorter the
lag and hence the longer the sequence is likely to be (Johnston and Gous,
2003).
A graphical representation of an ovulation sequence may be achieved
by plotting the two functions given by Eqns 12.1 (in the form 1 R3(t))
and 12.2 on the same axes, with time of day along the x-axis. Where the
two curves intersect, ovulation takes place. Fig. 12.2 shows the series of
events over a 6-day period for a three-ovulation sequence.
On day 2 the predicted ovulation time is 08:31. On days 3 and 4
ovulations occur at 11:30 and 15:22, respectively, resulting in corre-
sponding lag values of 2 h 59 min and 3 h 52 min. The two functions do
not meet on day 5; hence there is no ovulation, resulting in a pause day.
Day 6 sees the start of a new sequence, with an ovulation taking place at
08:31. The total lag for the three-ovulation sequence is thus 6h 51 min and
the mean lag is 3 h 25 min.
Credit is therefore due to Etches and Schoch (1984) for successfully
achieving their objective, which was to show mathematically why hens lay
their eggs in sequences. Indirect support is given to the theory of a
Fig. 12.2. An illustration of the asynchronous cycle of ovulation for a three-egg sequence;
ovulation occurs when the regulator concentration 1R3(t) (solid line) intersects with follicle
maturation G(t) (dotted line).
circadian rhythm restricting the release of LH to a portion of the day,

although the hypothesis still needs to be confirmed.
Although this model satisfactorily accounts for the pattern of
sequential laying, it does have limitations, because the table of values used
for the parameters in the model restricts the simulated sequences to two to
nine ovulations. Most individuals in a population of hens produce
considerably longer egg sequences around peak rate of lay. If the ovulatory
model is to be used to predict flock egg production, it needs to be able to
reproduce these longer sequences. The way to do this is to replace the
table of values with a set of continuous functions; one function for each of
the seven parameters in the ovulatory model. In addition, a method of
introducing inter-sequence pauses longer than 1 day needs to be found,
because experimental observations on individually-caged birds indicate
that a number of hens do pause for several days at a time (Johnston, 2004).
An enhanced version of the ovulatory model
The first step in amending the model is to convert sequence length to a

continuous variable, namely ovulation rate. For example, a sequence of
two ovulations followed by a single pause day gives an ovulation rate of
0.667 (two divided by three) and a nine-ovulation sequence, 0.90 (nine
ovulations in 10 days). A long sequence of 119 eggs in 120 days is
equivalent to an ovulation rate of 0.992. A relationship may now be
established between the ovulatory cycle parameters and ovulation rate.
The ovulatory cycle parameters are able to change predicted ovulation

rate by changing the lag between successive ovulations. As the ovulation
rate approaches 1.0 (for very long sequences), increasingly smaller changes
to the values of the parameters are required to cause increasingly smaller
decreases in lag values. If, as is seen in Fig. 12.1, mean lag decreases at a
diminishing rate as the sequence length or ovulation rate increases, the
parameters themselves presumably need to show a diminishing rate of
change with increasing rate of ovulation. This trend is evident in the values
for b3 (the rate parameter in the Gompertz equation representing follicle
maturation) given in Table 12.1. Three of the other parameters, i.e. λ1, λ2
and S2, were each allocated two values by Etches and Schoch (1984). It
seems reasonable that these and the remaining three parameters (S1, b1
and b2) should also alter gradually to bring about steady changes in
sequence length or ovulation rate. A Gompertz function of the form
y = A + C . exp (exp (B . (xM))) (12.4)
fulfils the requirements. It has two asymptotes, so that as the ovulation rate
approaches zero the parameter values remain positive and as the ovulation
rate approaches 1.0, the parameter values continue to increase but at a
diminishing rate. The procedure used to solve simultaneous equations for
the seven ovulatory cycle parameters is discussed by Johnston and Gous
(2003). The values allocated to the parameters A, C, B and M from Eqn
12.4 are listed in Table 12.2.
Figure 12.3 illustrates the values for the rate parameter b3 given by
Etches and Schoch (1984) and the fitted Gompertz function that may be
used to predict the value of b3 required for any ovulation rate from 0.5 (a
one-ovulation sequence) to 1.0. Fig. 12.4 shows that, although a
continuous function is now used to predict 2 rather than the two values
listed in Table 12.1, the range in the values for two- to nine-ovulation
sequences remains the same as used by Etches and Schoch (1984).
The set of continuous functions, representing the changes required to
the values of the different parameters, makes the prediction of any
sequence length or ovulation rate possible.
Table 12.2. Values assigned to the Gompertz function parameters for each of the seven
ovulatory cycle parameters.
Parameter B M C A
λ1 8 0.66667 0.020456 0.132475
λ2 8 0.66667 0.071596 0.223662
S1 8 0.66667 3.068414 9.128776
b1 8 0.66667 0.071596 0.723662
b2 6.148166 0.708292 21.097792 0.760415
b3 8 0.66667 0.233199 0.134213
S2 8 0.66667 3.596213 17.080953
Fig. 12.3. The curvilinear relationship between b3 and ovulation rate, showing the fitted
Gompertz function (—) and the given parameter values ().
Fig. 12.4. The values for λ2 given by Etches and Schoch (1984) (䉬) and the Gompertz
function for predicting λ2 from ovulation rate (—).
A Mechanistic, Stochastic Population Model
Age at first egg
The starting point for any layer model aiming to predict flock egg
production is to estimate the mean age at first egg. This is because a hen’s
chronological age determines her ovulation rate and yolk weight, as well as
the incidence of internal ovulations, soft shells and double-yolked eggs.
Sequence length and rate of lay are therefore also influenced by the mean
age at sexual maturity. Although other factors do play a role, light is the
single most important factor determining the age at sexual maturity in

pullets (Lewis et al., 1997). The Bristol-Reading model (Lewis et al., 2002)
predicts the mean age at first egg of a flock of pullets, based on the genotype
and the lighting programme applied during rearing. The function
SD = 8.76 + 0.124 . mean AFE (12.5)
estimates the associated standard deviation and predicts that delayed
photostimulation causes an increase in the standard deviation about the
mean. The model also calculates the proportion of the flock capable of
responding to photostimulation and, depending on the age at photo-
stimulation, produces either a normal or a bimodal distribution of ages at
first egg.
The distributions of ages at first egg for three theoretical Hy-Line
Brown flocks, photostimulated at either 12, 15 or 18 weeks of age, are
illustrated in Fig. 12.5. The mean ages at first egg (± SD) as predicted by
the Bristol-Reading model are 117.0 (± 5.75), 129.4 (± 7.29) and 142.04
days (± 8.85), respectively.
Internal cycle length
A hen has an internal cycle length, which is expressed as the interval

between successive ovulations. For example, if an ovulation occurs on day
1 at 08:00 and on day 2 at 10:00, then the internal cycle length is 26 h.
Fig. 12.5. Theoretical distributions of ages at first egg for Hy-Line Brown pullets
photostimulated at 12 weeks (hatched columns), 15 weeks (blank columns) and 18 weeks
(solid columns).
This interval is determined by the synchrony between the follicular

hierarchy and the circadian rhythm of LH release. A hen that has a
properly maintained hierarchy containing sufficient follicles, capable of
maturing at the rate of one every 24 h, will have an internal cycle of 24 h
and will lay an egg every day. With advancing age the follicles take longer
to mature, with the result that the internal cycle length increases. There
may also be age-related changes to the circadian clock system (Turek et al.,
1995). Emmans and Fisher (1986) proposed that the change in a hen’s
internal cycle length (ICL) over time may be calculated from the following:
ICL = ICLo – Lag + 1/((1/Lag) – kt) (12.6)
where ICLo = initial internal cycle length, k = a decay factor and t = time
from first egg, in days. If the internal cycle length is less than or equal to
the 24 h daylength or external cycle length (EXCL), rate of ovulation or
rate of lay (R) is given by:
R = 24/EXCL (12.7)
If the internal cycle length is greater than the external cycle length:
R = Lag/((ICL – EXCL) (1+ (Lag /(ICL-EXCL)))) (12.8)
These functions allow a decrease in egg production with advancing hen
age, but they do not permit the simulation of the shorter egg sequences
produced by many hens at onset of lay. It is evident from trials that
involve recording time of lay that these short egg sequences are not all
due to internal ovulations (Lewis and Perry, 1991; Johnston, 2004). It
would appear that there may be insufficient FSH at sexual maturity to
maximize follicle growth and to promote recruitment into the hierarchy
(Zakaria, 1999). In order to reproduce these short sequences, the internal
cycle length initially needs to be longer than 24 h, before decreasing with
advancing time from first egg to close to or below the daylength and
subsequently increasing. Quadratic-by-linear equations of the form:
ICL = A + B/(1 + D. x) + C . x (12.9)
where x = time from first egg (in days), give the required curvilinear shape.
The theoretical relationship between internal cycle length and time
from first egg for two individuals is illustrated in Fig. 12.6. The predicted
rates of lay for the same two hens are illustrated in Fig. 12.7. Because the
initial internal cycle lengths are longer than 24 h, both of these birds will
lay a few short egg sequences before producing their prime sequences.
Quadratic-by-linear functions therefore have the advantage over the
function given by Emmans and Fisher (1986) (Eqn 12.6), in terms of being
able to produce short ovulation sequences at onset of lay. Table 12.3 lists
possible values of the parameters that may be substituted in Eqn 12.9, to
predict changes in internal cycle length over time for Hy-Line Silver and
Hy-Line Brown birds. These values have been found to produce
acceptable rates of lay and sequence lengths for the two strains (Johnston,
2004).
Fig. 12.6. Curvilinear relationships between ICL and time from first egg for two hens.
Fig. 12.7. The rates of lay for two hens, using the internal cycle lengths shown in Fig. 12.6.
Table 12.3. Parameter values for the quadratic-by-linear

functions to predict ICL for Hy-Line Silver and Hy-Line
Brown hens.
Variable Hy-Line Silver Hy-Line Brown
A 22.4416 22.5416
B 2.0469 2.0469
D 0.02279 0.02279
C 0.011 0.011418
In order to introduce variation in the rate of change in internal cycle

length within a population, a coefficient of variation of 1% may be used to
generate normal distributions for each of the four parameters. The rates of
lay of individual hens will then differ; a small proportion of the flock will
lay exceptionally long sequences and a proportion will produce short
sequences, even at peak egg production.
Linking ovulation time to sunset
The ovulatory cycle parameter S1 (from Eqn 12.1) determines the time
when the circadian rhythm for the regulator concentration function is
reinitiated. Because of the phase-setting effect of sunset, S1 needs to be
linked to the time that the lights are turned off in the poultry house, so
that predicted ovulation and oviposition times are relative to the onset of
darkness. The predicted mean time of lay on 15- to 16-h photoperiods is
about 13–14 h after sunset, as reported by Lillpers (1991) and Patterson
(1997), although as the duration of the scotoperiod increases, the interval
between dusk and the mean oviposition time increases (Etches, 1996).
The open period for LH release is thought to commence about 2-h
after sunset and to last for about 8–9 h (Williams and Sharp, 1978). The
preovulatory surge of LH can take place at any time during the open
period. Between 4 and 6 h after the LH peak, ovulation occurs. There is a
corresponding open period of about 8–9 h during which ovulation can
take place. Presumably the earliest ovulations at the start of a sequence
take place about 6–8 h after the onset of darkness. In a population model
each hen may be allocated a unique value for the sunset-ovulation interval
by creating a normal distribution, using a mean of 7 h and a coefficient of
variation of 5%. Suggested temporal relationships between the onset of
darkness, LH release and ovulation are shown in Fig. 12.8.
The method used to link S1 to the onset of darkness is explained in
detail by Johnston (2004). Briefly, the ovulatory cycle parameter S1 is itself
Open period for ovulation
Lights off
Open period for LH release
18:00 21:00 24:00 3:00 6:00 9:00 12:00
Time of day
Fig. 12.8. An illustration of the temporal relationships between lights off (21:00), the start of
the open period for LH release (23:00) and the start of the open period for ovulation (04:00).
described by four parameters A, C, B and M. Instead of using the value for

A given in Table 12.2 (i.e. 9.128776), the formula
A = lights off + sunset-ovulation interval – daylength + 3 (12.10)
where daylength = 24 h, is applied. If, for example, the lights are turned
off at 21:30 (i.e. 21.5 h) and the sunset-ovulation interval for a given hen is
7.25 h (7 h 15 min), then A = 21.5 + 7.25 – 24 + 3 = 7.75. This
modification still allows the value of S1 to change with rate of ovulation, so
that longer sequences are linked to an earlier initiation of the circadian
rhythm, but the circadian rhythm itself will now phase-shift in accordance
with the time the lights are turned off.
Predicting the ovulation rate
An ovulation needs to occur the day before the hen lays its first egg. At
sexual maturity the time from first egg is zero, and this determines the
bird’s initial internal cycle length (Eqn 12.9), which in turn determines the
ovulation rate (using Eqns 12.7 and 12.8). A normal distribution of lag
values with, for example, a mean of 8.5 h and a coefficient of variation of
2%, introduces further variation. The ovulation rate establishes the time of
each consecutive ovulation by using the appropriate values for the seven
ovulatory cycle parameters. In a flock of hens, the mean ovulation rate at
onset of lay will be determined by the distribution of ages at first egg and
also by the initial sequence lengths, or individual ovulation rates. The
predicted flock ovulation rate over a laying cycle for a hypothetical flock of
birds is illustrated in Fig. 12.9. In this example the ovulation rate reaches a
peak 9 weeks after the onset of lay.
Fig. 12.9. The predicted ovulation rate for a theoretical flock of 100 Hy-Line Silver hens.
In theory, the ovulation rate and the rate of lay are identical unless internal
ovulations occur, or the production of double-yolked or soft-shelled eggs
causes interruptions to egg sequences. Johnston (2004) found that 8% of
double yolks and 18% of soft shells were associated with pauses in egg
sequences, and these percentages increased with earlier photostimulation.
Predicting time of lay
The time that an egg is laid is determined by the time of the associated
ovulation (Gilbert and Wood-Gush, 1971), which means that, except for
the terminal egg of a sequence, the ovulation times may be used to
estimate oviposition times. Oviposition normally occurs about half an hour
before the next ovulation, with a range of 7–75 min (Warren and Scott,
1935; Fraps, 1955). A normal distribution of oviposition-ovulation intervals
may be produced using a mean of 30 min and a coefficient of variation of
30%, so that the range of intervals is from 3 to 57 min. In a population,
each hen may therefore be allocated a different value for the interval and
this value may change daily. This interval needs to be subtracted from the
predicted ovulation time to give the estimated time of the associated
oviposition.
Estimation of the oviposition time for the last egg of a sequence needs
to be handled in a different way, because there is no associated ovulation.
The relationship between the lag for the last two eggs of sequences and
sequence length is illustrated in Fig. 12.10.
Fig. 12.10. The relationship between sequence length and lag (the oviposition interval minus
24 h) for the last two eggs of a sequence. Data from Morris, unpublished (×) and Johnston,
2004 (). The solid line shows the fitted linear-by-linear function.
Shorter sequences tend to have longer intervals between the last two
ovipositions. The large variation in lag for the longer egg sequences is due
to the fact that the sample sizes are very small.
A fitted linear-by-linear equation of the form
y = 1.75 – 0.9 / ( 1 – 1.969 . x ) (12.11)
predicts the lag between the last two ovipositions of a sequence (y, h) from
ovulation rate (x) which, in the population model, determines sequence
length. Consequently for an ovulation rate of 0.75 (a three-ovulation
sequence), the lag for the last two eggs is estimated to be 03 h 33 min.
The predicted distribution of oviposition times for a theoretical flock of
100 Hy-Line Silver hens is illustrated in Fig. 12.11. The distribution is
bimodal, because the shorter sequences that come with advancing age lead
to an increasing number of late afternoon ovipositions (Foster, 1968). The
mean time of lay is 10:28, 12 h 58 min after the sunset signal at 21:30.
Predicting internal ovulations
Prior to the rupture of the follicle and the release of the ovum, the follicle is
usually engulfed by the infundibulum so that ovulation occurs directly into
the infundibulum. This process is under hormonal and neurological control
and its effectiveness may therefore be impaired at puberty and with ageing
(Gilbert, 1972). Internal laying occurs when, for some reason, an ovum is
not grasped by the infundibulum and therefore remains in the body cavity.
The yolk material is usually resorbed within 24 h (Sturkie, 1955).
Fig. 12.11. Frequency distribution of the predicted times of lay for a theoretical flock of 100
Hy-Line Silver hens for 500 days, with sunset at 21:30.
The frequency of internal ovulations within a population may be

estimated by studying times of lay from individually caged birds, although
a physiological measurement, such as a change in body temperature, is
required to confirm that ovulation has occurred.
The percentage of internal ovulations is expected to decrease with age
in the early stages of lay. Pullets subjected to early photostimulation and
hence coming into lay at a young age are more likely to exhibit asynchrony
between ovary and oviduct. However, older hens are also increasingly
prone to internal ovulations. A quadratic-by-linear function such as
y = 58.08 + 67.17 / (1 + 0.001585 . x) + 0.04426 . x (12.12)
where y = % internal ovulations and x = hen age in days, may be used to
determine the percentage of internal ovulations expected at a particular
age for Hy-Line Silver hens. For example, at 126 days of age 3.5% of total
ovulations are expected to occur internally. Similarly the per cent internal
ovulations for Hy-Line Brown birds may be predicted from the quadratic-
by-linear function
y = 29.78 + 38.32 / (1 + 0.002352 . x) + 0.02659 . x (12.13)
These functions are based on the findings of Johnston (2004) for the two
strains of laying hens at onset of lay. No useful information is available on
the precise frequency of internal ovulations for different breeds and at
specific ages in older hens, so the predicted increase in per cent internal
ovulations towards the end of the laying cycle, used to derive these
functions, is conjecture. The curves given by these functions are shown in
Fig. 12.12. In addition, 48% of Hy-Line Silver and 28% of Hy-Line Brown
birds were thought to be prone to internal ovulations. In a population
model, a proportion of the flock needs to be identified and marked as
5
% Internal ovulations
0
12 17 22 27 32 37 42 47 52 57 62 67 72
Age (weeks)
Fig. 12.12. The relationship between expected per cent internal ovulations and hen age for
Hy-Line Silver (solid line) and Brown (dotted line) birds.
potential internal ovulators. Equations 12.12 and 12.13 may be used in a

population model to produce random internal ovulations over time within
the designated proportion of the flock.
Double-yolked eggs
The occurrence of double-yolked eggs in young hens at onset of lay is a

common phenomenon and may be attributed to some irregularity in the
process of ovulation or in the regulation of the follicular hierarchy. The
most likely cause is simultaneous development and ovulation of two ova
(Christmas and Harms, 1982). The earlier the pullets are brought into lay,
the greater the incidence of double yolks. Johnston (2004) found that 36%
of the experimental hens laid one or more double yolks, the proportion
being the same for both strains of Hy-Line birds.
An exponential function, based on the experimental data, may be used
to predict the per cent double yolks for both strains of Hy-Line birds:
y = 0.000475 + 8786.66 . (0.9403199 x) (12.14)
where y = % double-yolked eggs and x = hen age, in days. The curve
given by this equation is shown in Fig. 12.13. Hy-Line pullets receiving
increasing daylength at 12 weeks and coming into lay at about 15 weeks
may be expected to produce 13.7% double yolks. At 18 weeks of age, the
predicted per cent double yolks drops to 3.8%.
The distribution of double yolks up to 70 weeks of age within a
theoretical flock of Hy-Line Silver hens photostimulated at 18 weeks is
illustrated in Fig. 12.14. Twelve hens laid one double-yolked egg and nine
Fig. 12.13. An illustration of the relationship between expected per cent double yolks and
hen age for Hy-Line birds.
Fig. 12.14. The distribution of double yolks in a theoretical flock of 100 Hy-Line Silver hens
photostimulated at 18 weeks of age.
hens laid two double-yolked eggs. Only 23% of the population produced
double yolks. This proportion is substantially lower than anticipated, the
model having identified 36 of the 100 hens as potential multiple ovulators,
but the figure is presumably influenced by the relatively short period
allocated to the production of double yolks. These 37 double-yolked eggs
accounted for 0.11% of the total eggs laid to 70 weeks of age.
Soft-shelled eggs
Soft-shelled eggs have been found to be associated with short ovulation

intervals (van Middelkoop, 1972). If two or more ovulations occur within a
24-h period, the resulting eggs are unlikely to have normal shells unless a
double-yolked egg is produced. Premature expulsion of eggs from the
uterus may also lead to the production of soft-shelled eggs.
Older hens have a higher incidence of soft shells or breakages, due to
thinner shells, than young hens. This is as a result of their changing
hormone profiles and decreased ability to transport calcium at the
duodenum (Hansen et al., 2003). The reduced populations of oestrogen
receptors mean that old hens are less efficient, both in their absorption of
dietary calcium and in their utilization of calcium in the uterine fluid for
shell formation. Premature photostimulation of pullets also causes an
increase in the number of soft-shelled eggs (Johnston, 2004).
Johnston (2004) reported that 38% of Hy-Line Silver and 26% of Hy-
Line Brown birds produced soft-shelled eggs at onset of lay. Although
these proportions will vary according to nutritional and environmental
factors, it is useful to have some value for the purpose of testing the model.
Many hens lay normal-shelled eggs throughout their laying year. The
population model therefore needs to set apart a proportion of the hens
that will randomly produce soft-shelled eggs.
Two line-plus-exponential functions, derived from experimental data
(Johnston, 2004), may be used to predict the percentage of soft shells for
the Silver and Brown birds, respectively:
y = 0.8006 + 1202.6 . (0.952622 x) + 0.0036895 . x (12.15)
y = 0.8134 + 714.6 . (0.953762 x) + 0.0037948 . x (12.16)
where y = % soft shells and x = hen age, in days. The curves given by
these functions are shown in Fig. 12.15. Hy-Line Silver birds that are
photostimulated at 12 weeks and come into lay at about 15 weeks of age
may be expected to produce 6.9% soft shells in the first week of lay. At 18
weeks the frequency will be reduced to 2.3%.
A hypothetical distribution of soft shells within a population of 70-
week-old Hy-Line Silver hens photostimulated at 18 weeks of age is shown
in Fig. 12.16. Three birds laid one soft-shelled egg each and one bird laid
eight. A total of 141 soft shells (0.43% of total eggs) was produced by 36%
of the hens; 64 of the 100 hens did not lay any soft-shelled eggs. Fig. 12.17
illustrates how the per cent soft shells changed during the course of the
laying year for the same flock. Because the pullets were photostimulated at
the recommended age, there were very few soft shells at onset of lay. An
increasing number were produced towards the end of the productive
period.
Fig. 12.15. An illustration of the relationship between expected per cent soft shells and hen
age for Hy-Line Silver (solid line) and Brown (dotted line) birds.
Fig. 12.16. The distribution of soft shells in a theoretical flock of 100 Hy-Line Silver hens
photostimulated at 18 weeks of age.
Fig. 12.17. The per cent soft shells produced over a full laying cycle, by a theoretical flock of
100 Hy-Line Silver hens photostimulated at 18 weeks.
Predicting egg weight
In a layer model, egg weight may be predicted as a unit or as the sum of

the weights of the three components – yolk, albumen and shell. In view of
the fact that the proportions of the components change with advancing age
and that the three components have very different chemical compositions,
a model that predicts the weights of each separately may be of more
interest to a nutritionist making a study of the hen’s changing nutritional
requirements.
At a given age, and over a range of egg weights, the weights of all three
components increase as egg weight increases, with the proportion of
albumen increasing at the expense of yolk and shell. In a sample of eggs
collected on a specific day from a single-age flock, it may therefore be
expected that the larger eggs will contain proportionately more albumen
than the smaller eggs. As the laying cycle progresses and egg weight
increases, the component weights all increase, but the percentage of yolk
increases to the detriment of albumen and shell. These proportional
changes need to be accounted for by the layer model, and this is
accomplished via allometry. A satisfactory method for estimating the
weights of the three egg components is given by Hussein et al. (1993).
Allometric functions of the form
y = a. x b (12.17)
are used by biologists working with variables that are scaled relative to
body size. Allometry refers to non-isometric scaling, i.e. changes in size of
irregular-shaped organisms. A useful property of allometric functions is
that when the two variables x and y are plotted on logarithmic coordinates,
the result is a straight line:
ln y = ln a + b. ln x (12.18)
The exponent b from Eqn 12.17 represents the slope of the straight line
obtained in a logarithmic plot. Emmans and Fisher (1986) proposed that
albumen and shell weights may be predicted from these allometric
functions:
AW = a1 . YW b1 (12.19)
and
SW = a2 . ECW b2 (12.20)
where AW = albumen weight, YW = yolk weight, SW = shell weight and
ECW = the weight of the egg contents, yolk plus albumen. Egg weight is
given by YW + AW+ SW.
The first step in predicting egg weight is to predict the weight of the
yolk. This may be achieved by using a Gompertz function (Eqn 12.4) or a
logistic function:
y = A + C / (1 + exp (–B . (x –M))) (12.21)
where y = yolk weight, in grams and x = hen age, in days. The parameters
of both functions need to be defined for each genotype, because strains
differ both in their initial yolk weight and in their rate of increase in yolk
weight with advancing hen age. In addition, the values allocated to each of
the parameters may need to be adjusted from time to time in line with
genetic advancements, since the component weights are likely to be
affected. The relationship between yolk weight and hen age, for three
commercially available genotypes, is illustrated in Fig. 12.18.
It is interesting to note that, although the eggs from Hy-Line Brown
hens are heavier, they have a smaller proportion of yolk than those from
Hy-Line Silver hens. Some suggested values of the parameters are given in
Table 12.4.
The next step is to predict albumen weight from yolk weight, using
Eqn 12.19 and substituting the relevant values for the two parameters for
the specific genotype. Subsequently, shell weight may be estimated from the
weight of the egg contents (yolk plus albumen) using Eqn 12.20. Finally, egg
weight is calculated as the sum of yolk, albumen and shell weights. The
suggested values for the allometric function parameters for the same three
strains of bird are summarized in Table 12.5. The method used to derive the
values for the parameters is explained by Johnston (2004).
Fig. 12.18. The curvilinear relationship between yolk weight and hen age, for Amber-Link
(bold line), Hy-Line Silver (solid line) and Hy-Line Brown (dotted line) hens.
Table 12.4. Values of the parameters derived for the logistic (Amber-Link) and Gompertz
(Hy-Line Silver and Brown) functions used to predict yolk weight from age.
Strain A C B M
Amber-Link 224.7 243.2 0.01268 116.4
Hy-Line Silver 51107 51123 0.01771 370.1
Hy-Line Brown 101.1 116.0 0.01972 15.36
Table 12.5. Estimates of the parameter values in the allometric functions used to predict
albumen and shell weight.
Parameter
a1 (albumen) b1 (albumen) a2 (shell) b2 (shell)
Hy-Line Silver 9.473515 0.5044 0.138207 0.9180
Hy-Line Brown 10.8906 0.5020 0.133187 0.9310
Amber-Link 10.9900 0.4491 0.33875 0.6896
The positive relationship between the percentage of yolk and time

from first egg, for a simulated flock of 100 Hy-Line Silver hens
photostimulated at 18 weeks of age is illustrated in Fig. 12.19. Figure 12.20
shows how the percentage of albumen decreases as the same flock
progresses in its laying cycle. The strong positive relationship between
albumen weight and egg weight, for eggs produced by hens at 163 days of
age (23.2 weeks), is illustrated in Fig. 12.21.
A population model that makes use of Eqns 12.19 to 12.21,
substituting the relevant parameter values for the genotype, is able to
correctly predict changes in the egg component weights and proportions
with advancing hen age. Moreover, the relationships between the weights
of the three components and egg weight, at a given hen age, are accurate.
Variation in yolk, albumen and shell weights within the population may be
introduced by allocating standard deviations to each of the parameters
listed in Tables 12.4 and 12.5.
Fig. 12.19. An illustration of how predicted mean yolk percentage increases with age, for a
hypothetical flock of 100 Hy-Line Silver hens.
Fig. 12.20. An illustration of how predicted mean albumen percentage decreases with age
for a hypothetical flock of 100 Hy-Line Silver hens.
Fig. 12.21. A scatter diagram illustrating the positive relationship between albumen weight
and egg weight (r2 = 0.99), for a hypothetical flock at 163 days of age.
Mean sequence length
The mean sequence lengths for two theoretical flocks, both photostimulated
at 18 weeks, one consisting of 100 Hy-Line Silver hens and the other of 100
Hy-Line Brown hens, are shown in Fig. 12.22. It may be seen that the
predicted mean sequence length for both strains is initially short at onset of
lay, owing to the quadratic-by-linear function used to predict changes in
internal cycle length. The maximum mean sequence lengths are 86.7 and
76.8 days for the Hy-Line Silver and Hy-Line Brown flocks, respectively.
The trends simulated by the model are similar to those observed in practice
and reported by Johnston (2004) and Lewis and Perry (1991). As expected,
the model predicts that the Silver strain has slightly longer mean sequence
lengths than the Brown strain for most of the laying cycle.
The predicted sequence lengths per bird over an extended laying
period may be imported into the Sequence Analyzer program (Zuidhof et
al., 1999). This software calculates (as well as other variables) the mean
sequence length, the number of sequences, the prime sequence length and
the mean pause length for each bird, and is useful for comparing the
performances of individuals or groups of birds subjected to different
treatments. The performance indicators for the same Hy-Line Silver and
Brown flocks discussed above are summarized in Table 12.6. The Silver
strain tends to have longer prime sequences and to lay fewer but longer
sequences than the Brown strain. These factors contribute to the higher
rate of lay seen in the Hy-Line Silver hens. The large standard deviations
indicate the huge amount of variation between individuals in terms of
sequence lengths.
Fig. 12.22. Predicted mean sequence lengths for 100 Hy-Line Silver (solid line) and 100
Hy-Line Brown (dotted line) hens.
Table 12.6. Summary of mean sequence characteristics (± SD) for the theoretical flocks.
Prime sequence Number of Sequence
Strain length sequences length Pause length
Hy-Line Silver 82.76 (±60.20) 27.41 (±9.15) 13.60 (±5.56) 1.02 (±0.03)
Hy-Line Brown 72.70 (±61.69) 28.47 (±9.82) 13.02 (±5.90) 1.02 (±0.03)
A mean pause length of 1.16 days was reported by Robinson et al.

(2001) for SCWL hens. Johnston (2004) found mean pause lengths of 1.59
and 1.51 days for Hy-Line Silver and Hy-Line Brown birds, respectively.
The model of Etches and Schoch (1984), both in its original and in its
revised form, does not create pauses longer than 1 day. In the population
model, however, a few longer pauses are introduced by the inclusion of
random internal ovulations.
Discussion
The improved mathematical model of the ovulatory cycle provides the

ideal starting point for a layer model. It is capable of producing ovulation
sequences of any length and of restricting the total lag between the first
and last ovulations of a sequence to between 8 and 10 h. The predicted
ovulation times are therefore constrained to occur within an 8–10 h open
period. The model provides a sound mathematical explanation for the
long-observed phenomenon of sequential laying in domestic hens. On its
own, however, the ovulatory model is not able to show changes in
ovulation rate with advancing hen age. This is brought about by utilizing
the theory of Emmans and Fisher (1986), namely that a hen has an
internal cycle length that increases with time from first egg, causing the
ovulation rate and rate of lay to decline. The assumption is made that all
hens commence laying by producing their prime sequence, which is not
always the case.
The quadratic-by-linear functions, used in the model to predict
changes in internal cycle length in ageing hens, are able to reproduce
shorter sequences at onset of lay. This is because the functions allow the
internal cycle length to decrease towards 24 h in the first few weeks of lay,
before increasing. However, the primary disadvantage of these quadratic-
by-linear functions is that it is relatively difficult to find suitable values for
the four parameters that enable the population model to produce an
acceptable mean sequence length and rate of lay for the genotype. It is felt
that, while these quadratic-by-linear functions give the desired result, an
improved or simplified method of creating shorter sequences needs to be
found.
It may be helpful to develop a model of the follicle hierarchy, the size
of the hierarchy being determined by the processes of recruitment and
atresia. The absence or presence of a mature follicle may then dictate
whether or not ovulation takes place. If young hens have fewer follicles in
the hierarchy at onset of lay, the result would be shorter sequences.
Another benefit to be derived from modelling the hierarchy is that yolk
weight may be predicted with a greater degree of accuracy. The follicles
may be allowed to continue to accumulate yolk up to the precise time of
ovulation. In this way the yolk of the first egg of a sequence is likely to be
heavier than in subsequent eggs.
It would be useful to know why some birds have inter-sequence pauses
longer than one day. These may well be due to inadequate maintenance of
the follicular hierarchy, but there may also be times during the laying cycle
when the preovulatory surge of LH does not proceed as expected. Stresses
that are imposed on the hens over a number of days, such as the
withdrawal of feed, cause apoptosis and subsequent cell proliferation in the
tissues of the anterior pituitary, with corresponding changes to the LH
secretions (Chowdhury and Yoshimura, 2002). Short-term stresses may
also have the ability to prevent the secretion of LH. The current model
presumes that the circadian rhythm of LH release continues without
interruption. A method of blocking the rhythm of the regulator substance
can be found such that, for one day or a number of days, ovulation will not
take place. This will allow the model to respond to a limiting environment
by altering rates of ovulation and lay. Accordingly, pauses longer than one
day can be introduced into sequences. Increasing the incidence of internal
ovulations or allowing soft-shelled and double-yolked eggs to disrupt egg
sequences will also result in longer inter-sequence pauses, as long as two or
more disturbances occur on consecutive days. A well-designed trial,
focusing on the poor producers, could provide valuable information on the
underlying reasons for their low rate of lay.
The proportional changes in yolk, albumen and shell with increasing
age have been observed and accepted by poultry scientists for more than
half a century (Romanoff and Romanoff, 1949; Harms and Hussein, 1993).
It is important to be able to model these changes successfully, so that
increases in egg weight with advancing hen age are brought about by
increases in the three components occurring at different rates. The
parameters in the allometric functions need to be carefully defined for
each genotype, since strains differ in the weights and proportions of yolk,
albumen and shell and this may entail different nutritional requirements.
Moreover, it is necessary to reassess the information every few years in
order to keep pace with genetic advancements.
A great many factors are taken into account by the layer model
presented in this chapter. The mean age at first egg is calculated from the
lighting programme applied during rearing. Ovulation times are predicted
for a theoretical flock of 100 birds over a laying year. The flock ovulation
rate is seen to increase rapidly to reach a peak about 8 weeks from onset of
lay, before declining in a linear fashion. Oviposition times are predicted
from the associated ovulation times. For all but the last egg of a sequence,
oviposition occurs roughly half an hour before the associated ovulation.
The lag between the penultimate and ultimate eggs of a sequence is
calculated from an equation relating the length of the interval to ovulation

rate; short sequences have a longer interval between the last two
ovipositions. The open period for LH release entrains to the onset of
darkness, so that mean time of lay occurs 13–14 h after sunset. The user is
able to observe the effect of different times of sunset on the mean time of
lay. The distribution of oviposition times is seen to be unimodal and
positively skewed in young flocks, when a greater proportion of the eggs
are laid in the morning, and bimodal in older flocks, when due to the
shorter sequences, comparatively more eggs are laid in the afternoon. Flock
rate of lay follows a similar pattern to ovulation rate, and also follows the
trend given in breed manuals. Yolk weight increases with hen age.
Allometric functions are used to predict albumen weight from yolk weight,
and shell weight from the combined weight of yolk plus albumen, using the
parameters appropriate for the genotype. Egg weight is calculated as the
sum of the three component weights and may be seen to be in line with
breed standards. The model correctly predicts the increasing proportion of
yolk and the decreasing proportions of albumen and shell with advancing
hen age. Random occurrences of internal ovulations, restricted to a
proportion of the flock, cause interruptions to egg sequences. The number
of internal ovulations in the population is higher at the beginning and at
the end of the laying cycle, when asynchrony between ovary and oviduct is
more likely. Double-yolked eggs are produced by a proportion of the flock
at the start of the laying period; the earlier the hens are brought into lay,
the greater the incidence of double yolks. A number of hens are prone to
laying soft-shelled eggs, with a greater frequency occurring at onset of lay
and in older hens. The mean sequence length is initially low, but it increases
to about 70–80 eggs at peak rate of lay and subsequently decreases.
To broaden the scope of the model, a number of parameters still need
to be defined for a selection of genotypes. For instance, the Lohmann
Silver and Brown strains possess a reasonably large share of the South
African market for commercial pullets and layers and yet little academic
research seems to have been conducted with these hens. They may
respond to environmental stimuli in a similar manner to the Hy-Line
birds, but this needs to be tested. No information appears to be available
on the weights of the egg components for the Lohmann birds. In a slightly
different vein, it is not known whether yolk, albumen and shell weights can
be manipulated by the environment, although yolk weight is strongly
influenced by hen age. Thus, while the model addresses many issues
associated with egg production, it is hoped that it may be continually
improved as new information comes to light.
Despite the fact that several improvements can still be made to the
layer model, it is hoped that the model in its current form will be of benefit
to sections of the poultry community. As a teaching aid to poultry science
students it may be invaluable, given its ability to illustrate the complexities
of egg production and to draw attention to the extent of variation within a
population. By allowing the user to change various inputs, the model
demonstrates the ability of hens to respond to their environment. The egg
producer may be interested in predicting the response of the birds, in

terms of laying performance, to changes in the lighting programme
applied during rearing. The importance of applying stimulatory
daylengths to pullets at the recommended age for the breed is underlined
by the effect of age at photostimulation on the incidence of internal
ovulations, soft shells and double-yolked eggs. He or she would also gain a
better understanding of the process of egg production and the effect that
poor layers may have on flock efficiency. The producer may benefit from
seeing how mean time of lay is influenced by hen age and the time the
lights are switched off at night; daily egg collections could then be timed so
that the majority of eggs are collected as soon as possible after being laid.
The model may also be used to predict voluntary food intake on a daily
basis and over the laying year; food intake being less on a day when no egg
is laid. The nutritionist may be interested in estimating the nutrient
requirements of laying hens according to the predicted weights of yolk,
albumen and shell. Ultimately, this model may assist nutritionists to
optimize the nutrient requirements for the various commercially available
strains. With a few modifications, the population model may be adapted
for use with broiler breeder hens that have notoriously erratic egg
production cycles, with long and frequent pauses between egg sequences.
References
Adams, C.J. and Bell, D.D. (1980) Predicting poultry egg production. Poultry Science
59, 937–938.
Chowdhury, V.S. and Yoshimura, Y. (2002) Cell proliferation and apoptosis in the
anterior pituitary of chicken during inhibition and resumption of laying.
General and Comparative Endocrinology 125, 132–141.
Christmas, R.B. and Harms, R.H. (1982) Incidence of double yolked eggs in the
initial stages of lay as affected by strain and season of the year. Poultry Science
61, 1290–1292.
and Boorman, K.N. (eds) Nutrient Requirements of Poultry and Nutritional
Etches, R.J. (1984) Maturation of ovarian follicles. In: Cunningham, F.J., Lake, P.E.
and Hewitt, D. (eds) Reproductive Biology of Poultry. Longman Group, Harlow,
UK, pp. 51–73.
Etches, R.J. (1996) Reproduction in Poultry. Cambridge University Press, Cambridge,
UK.
Etches, R.J. and Schoch, J.P. (1984) A mathematical representation of the ovulatory
cycle of the domestic hen. British Poultry Science 25, 65–76.
Foster, W.H. (1968) The effect of light-dark cycles of abnormal lengths upon egg
production. British Poultry Science 9, 273–284.
Fraps, R.M. (1955) Egg production and fertility in poultry. In: Hammond, J. (ed.)
Progress in the Physiology of Farm Animals. Butterworths, London, pp. 661–740.
Gilbert, A.B. (1972) The activity of the ovary in relation to egg production. In:
Freeman, B.M. and Lake, P.E. (eds) Egg Formation and Production. British
Poultry Science, Edinburgh, UK, pp. 3–17.
Gilbert, A.B. and Wood-Gush, D.G.M. (1971) Ovulatory and ovipository cycles. In:
Bell, D.J. and Freeman, B.M. (eds) Physiology and Biochemistry of the Domestic
Fowl, Vol. 3. Academic Press, London, pp. 1353–1378.
Hansen, K.K., Kittok, R.J., Sarath, G., Toombs, C.F., Caceres, N. and Beck, M.M.
(2003) Estrogen receptor-α populations change with age in commercial laying
hens. Poultry Science 82, 1624–1629.
Harms, R.H. and Hussein, S.M. (1993) Variations in yolk: albumen ratio in hen
eggs from commercial flocks. Journal of Applied Poultry Research 2, 166–170.
Hussein, S.M., Harms, R.H. and Janky, D.M. (1993) Research Note: Effect of age
on the yolk to albumen ratio in chicken eggs. Poultry Science 72, 594–597.
Johnson, A.L. (1984) Interactions of progesterone and luteinizing hormone leading
to ovulation in the domestic hen. In: Cunningham, F.J., Lake, P.E. and Hewitt,
D. (eds) Reproductive Biology of Poultry. Longman Group, Harlow, UK,
pp.133–143.
Johnson, A.L. (1996) The avian ovarian hierarchy: a balance between follicle
differentiation and atresia. Poultry and Avian Biology Reviews 7, 99–110.
Johnson, P.A., Dickerman, R.W. and Bahr, J.M. (1986) Decreased granulosa cell
luteinizing hormone sensitivity and altered thecal estradiol concentration in
the aged hen, Gallus domesticus. Biology of Reproduction 35, 641–646.
Johnston, S.A. (1993) Simulation modelling as a means of predicting the
performance of laying hens. MSc Agric thesis, University of Natal, South Africa.
Johnston, S.A. (2004) A stochastic model to predict annual egg production of a
flock of laying hens. PhD thesis, University of KwaZulu-Natal, South Africa.
Johnston, S.A. and Gous, R.M. (2003) An improved mathematical model of the
ovulatory cycle of the laying hen. British Poultry Science 44, 752–760.
Lewis, P.D., Morris, T.R. and Perry, G.C. (2002) A model for predicting the age at
sexual maturity for growing pullets of layer strains given a single change in
photoperiod. Journal of Agricultural Science 138, 441–458.
Lewis, P.D. and Perry, G.C. (1991) Oviposition time: correlations with age, egg
weight and shell weight. British Poultry Science 32, 1135–1136 (Abstr.).
Lewis, P.D., Perry, G.C. and Morris, T.R. (1997) Effect of size and timing of
photoperiod increase on age at first egg and subsequent performance of two
breeds of laying hen. British Poultry Science 38, 142–150.
Lillpers, K. (1991) Genetic variation in the time of oviposition in the laying hen.
British Poultry Science 32, 303–312.
Lillpers, K. and Wilhelmson, M. (1993) Genetic and phenotypic parameters for
oviposition pattern traits in three selection lines of laying hens. British Poultry
Science 34, 297–308.
Liou, S.S., Cogburn, L.A and Biellier, H.V. (1987) Photoperiodic regulation of
plasma melatonin levels in the laying chicken (Gallus domesticus). General and
Comparative Endocrinology 67, 221–226.
Menaker, M., Hudson, D.J. and Takahashi, J.S. (1981) Neural and endocrine
components of circadian clocks in birds. In: Follett, B.K. and Follett, D.E. (eds)
Biological Clocks in Seasonal Reproductive Cycles. Wright, Bristol, UK,
pp. 171–183.
Patterson, P.H. (1997) The relationship of oviposition time and egg characteristics
to the daily light:dark cycle. Journal of Applied Poultry Research 6, 381–390.
Robinson, F.E., Renema, R.A., Oosterhoff, H.H., Zuidhof, M.J. and Wilson, J.L.
(2001) Carcass traits, ovarian morphology and egg laying characteristics in
early versus late maturing strains of commercial egg-type hens. Poultry Science
80, 37–46.
Romanoff, A.L. and Romanoff, A.J. (1949) The Avian Egg. John Wiley, New York.
Siopes, T.D. and Wilson, W.O. (1980) A circadian rhythm in photosensitivity as the
basis for the testicular response of Japanese quail to intermittent light. Poultry
Science 59, 868–873.
Sturkie, P.D. (1955) Absorption of egg yolk in the body cavity of the hen. Poultry
Science 34, 736–737.
Tilly, J.L., Kowalski, K.I. and Johnson, A.L. (1991) Stage of ovarian follicular
development associated with the initiation of steroidogenic competence in
avian granulosa cells. Biology of Reproduction 44, 305–314.
Turek, F.W., Penev, P., Zhang, Y., Van Reeth, O., Takahashi, J.S. and Zee, P. (1995)
Alterations in the circadian system in advanced age. In: Chadwick, D.J. and
Ackrill, K. (eds) Circadian Clocks and Their Adjustment. Wiley, Chichester, UK,
pp. 212–234.
van Middelkoop, J.H. (1972) The relationship between ovulation interval of White
Plymouth Rock pullets and the laying of abnormal eggs. Archiv für
Geflügelkunde 6, 223–230.
Warren, D.C. and Scott, H.M. (1935) The time factor in egg formation. Poultry
Science 14, 195–207.
Williams, J.B. and Sharp, P.J. (1978) Control of the preovulatory surge of
luteinizing hormone in the hen (Gallus domesticus): the role of progesterone
and androgens. Journal of Endocrinology 77, 57–65.
Zakaria, A.H. (1999) Ovarian follicular growth in laying hens in relation to
sequence length and egg position in various sequence lengths. Archiv für
Geflügelkunde 63, 264–269.
Zuidhof, M.J., Bignell, D. and Robinson, F.E. (1999) Egg Production and Sequence
Analyzer Software. Alberta Poultry Research Centre, Alberta Agriculture, Food
and Rural Development, University of Alberta, Canada.
13 Comparison of Pig Growth Models
– the Genetic Point of View
P. LUITING AND P.W. KNAP
PIC International Group, Ratsteich 31, D-24837 Schleswig, Germany
pieter.knap@pic.com
Introduction
Application models of growth in pigs were pioneered by Whittemore and

Fawcett (1976) and further developed by Moughan and Smith (1984),
Black et al. (1986), Emmans (1988, 1997), Pomar et al. (1991), TMV
(1991), De Greef (1992), Ferguson et al. (1994), Whittemore (1995) and
many others. All these models are dynamic (describing the growth
process over time), deterministic (non-stochastic), and semi-mechanistic
(more or less based on biological growth mechanisms, focusing on the
accretion of several body components). Many of these biological
mechanisms are driven by one or more parameters that are specific to
the genotype (breed, strain, etc.) to be simulated. Proper simulation of
that genotype then requires quantification of those parameters, which
requires experimental effort. Therefore, these parameters are of primary
interest for animal breeders who use the model for their particular
purposes.
The above mentioned publications describe the mechanisms that
form the core of their models with different notation and
argumentation; as a consequence, the exact differences and similarities
between these approaches are often unclear and we attempt here to
clarify some of these issues. Section 1 of the present text describes the
main common mechanisms of the various models and their genotype-
specific aspects. Section 2 describes the details of implementation of
these mechanisms, and attempts to derive a common denominator.
Section 3 discusses the differences between the various models from the
point of view of adequacy and complexity of the description of the
genotype.
Given the large amount of water in the mammalian body, the
prediction of body water mass is perhaps the most critical factor of pig

Comparison of Pig Growth Models 261
growth simulation. Emmans and Kyriazakis (1995) focus on this issue in

quantitative detail, and this is the only such treatment of the matter that we
are aware of; we do not deal with it any further.
Main Common Mechanisms
All the models discussed here distinguish three processes that require
energy: protein deposition (PD), body maintenance, and lipid deposition
(LD), all three on a daily basis. The driving factor is PD; it is predicted
first, after which it determines all other factors in the simulation. The
metabolizable energy (ME) requirements for body maintenance (MEm) are
predicted from metabolic body weight or from body protein mass: MEm =
α × BW or MEm = α × P, where α and are constants. It follows that at
known energy intake (MEI), LD is predicted as the item that balances the
requirements of the law of conservation of energy:
MEI = MEm + PD × EP / kP + LD × EL / kL →
MEI - MEm − (PD × EP / kP ) (13.1)
LD =
EL / kL
where LD = lipid deposition (kg per day), MEI = ME intake (MJ per day),
MEm = ME required for maintenance (MJ per day), PD = protein
deposition (kg per day), EP and EL = combustion energy contents of body
protein and lipid, respectively (MJ per kg), and kP and kL = efficiencies of
use of ME for PD and LD, respectively.
All models assume EP, EL, kP, kL, α and to be constant, although they
vary somewhat with regard to their values. Genetic variation in these
parameters is presumed absent.
All models calculate water and ash deposition from PD, using empirical
equations with constant parameter values. Empty body weight gain (BWG)
is calculated by summation of the four components; adjustment for a
certain amount of gut fill leads to daily growth:
empty BWG = PD + LD + water deposition + ashdeposition

(13.2)
BWG = (empty BWG) × (1 + gutfill correction factor)
All models define an initial status with the starting values for body
protein, lipid, ash and water mass. The predicted deposition values are
added to the current values of body component mass, and next day’s
calculation is performed. This process is repeated until the desired end
status is reached. The models vary somewhat with regard to the water-
and ash-related empirical equations and/or their parameter values, with
regard to the magnitude of gut fill, and with regard to the initial body
component mass assumptions. Genetic variation is presumed absent
here.
262 P. Luiting and P.W. Knap
Most models treat feed intake as a known input parameter. However,

some models predict voluntary feed intake as an output parameter based
on MEm, desired PD, desired LD and Eqn 13.1. The prediction of
desired PD and desired LD is described, respectively, in the two sections
below.
Prediction of Protein Deposition
The main feature of PD is that it is sourced from the digestible crude

protein intake (DCPI). Like MEm, protein requirements for body
maintenance (Pm) are predicted from metabolic body weight or from body
protein mass: Pm = γ × BWδ or Pm = γ × Pδ, where γ and δ are constants. It
follows that at known DCPI, PD can be predicted as:
PD = eP × (DCPI − Pm ) (13.3)
where eP = material efficiency with which protein is utilized above
maintenance, DCPI = digestible crude protein intake (kg per day), and Pm
= protein required for body maintenance (kg per day).
The material efficiency with which protein can be utilized above
maintenance (eP) is assumed to have a constant maximum value (maxep) of
about 0.83, see Fig. 13.1. The various models vary somewhat with regard
to the magnitude of this maximum value for eP, and of γ and δ. Genetic
variation is again presumed absent.
All models described here recognize two situations where eP is lower
than its maximum value, i.e. where part of the protein supply is used as a
source of energy rather than amino acids: (i) PD is at its maximum (section 1);
and (ii) ME supply is insufficient (section 2).
PD
ep = maxep
Pm DCPI
Fig. 13.1. Protein deposition (PD) in relation to digestible crude protein intake (DCPI). Pm
denotes the protein requirement for body maintenance, eP is the material efficiency of protein
utilization. In this range of DCPI, eP attains its maximum value maxeP.
Maximum protein deposition
All models assume a genetically determined upper limit to a pig’s rate of

protein deposition: maxPD. This is the protein deposition that the pig
seeks to realize: maxPD equals the pig’s desired level of PD. By analogy,
the DCPI level that is required to just achieve maxPD is the desired DCPI:
output PD maxPD
eP = = = = maxeP
input DCPI − Pm desired DCPI − Pm
When realized DCPI exceeds desired DCPI, the surplus of digested

protein is used as an energy source. This reduces the material efficiency eP,
calculated as:
output PD maxPD
eP = = = < max eP , for DCPI > desired DCPI (13.4)
input DCPI − Pm DCPI − Pm
Adding this relation to Fig. 13.1 results in Fig. 13.2. The various models
use different definitions for the genetic factor(s) that describe maxPD.
The obtained value for maxPD is used to determine desired DCPI =
maxPD
+ Pm, which is compared to DCPI. If DCPI < desired DCPI,
maxeP
PD is estimated with Eqn 13.3; otherwise, PD = maxPD.
Insufficient energy supply
All models assume implicitly an intrinsic lower limit to a pig’s ME

requirements in order to reach maxPD: minMEImaxPD. From the law of
conservation of energy, all models then assume an intrinsic lower limit to a
pig’s LD coinciding with maxPD:
PD
maxPD
ep < maxep
ep = maxep
Pm DCPI
Fig. 13.2. An extension of Fig. 13.1. PD has an upper limit (maxPD), reached at DCPI =
desired DCPI. At higher DCPI levels, eP is lower than its maximum value.
minMEImaxPD = MEm + (maxPD × EP / kP ) + (minLDmaxPD × EL / kL ) (13.5)
where minLDmaxPD is the lower limit to LD coinciding with maxPD.

The relation of PD and LD with MEI is shown in Fig. 13.3 . At MEI ≤
minMEImaxPD, LD must compete for resources with PD so its realized level
is an indication of the genotype’s priorities for this resource allocation. At
MEI > minMEImaxPD all surplus energy goes to LD, and this is genotype-
independent. So we have a genetically controlled part for LD below
minMEImaxPD and a genetically controlled part for PD above it. It follows
that the point of intersection represents: (i) the pig’s ambition for protein
deposition; and (ii) its genetic priorities for resource allocation between PD
and LD, and minMEImaxPD must then be the lower limit to the energy
intake that the pig seeks to realize.
Another consequence is that PD is not the only genetically determined
factor that drives the system: below minMEImaxPD, the desired LD level is
just as important. Thus a genotype must be characterized by a desired PD
(= maxPD) and a desired LD coinciding with it (= minLDmaxPD).
Dependent on environmental conditions, it may well be that neither of
them ever gets phenotypically expressed.
Whereas at MEI > minMEImaxPD the surplus energy is fully used for
LD, the energy deficit at MEI ≤ minMEImaxPD is not fully accounted for by
reducing LD but also by reducing PD. Part of DCPI is then used as an
energy source and not for protein deposition. This leads to PD < maxPD,
and it reduces eP to:
output PD PD
eP = = = < max eP for PD < maxPD (13.6)
input DCPI − Pm desired DCPI − Pm
Adding this relation to Fig. 13.2 results in Fig. 13.4.
PD LD
maxPD
LD
minLD maxPD
PD
b1
b0 minMEI maxPD MEI
Fig. 13.3. Protein (PD) and lipid deposition (LD) in relation to ME intake (MEI). For PD <
maxPD, PD = f(MEI) = b1 × (MEI – b0). Above that level (i.e. above MEI = minMEImaxPD and
LD = minLDmaxPD), all surplus energy is used for LD.
PD
maxPD
PD =
f(MEI)
eP < maxeP
eP < maxeP
eP = maxeP
Pm DCPI
Fig. 13.4. An extension of Fig. 13.2. At insufficient ME supply, PD has an upper limit that
depends on ME intake (MEI) and eP is lower than its maximum value.
The various models use different definitions of the function f(MEI):

PD = b1 × (MEI – b0) that describes the reduction of PD at MEI <
minMEImaxPD, see the section on ‘Insufficient energy supply’ in ‘Model
comparison’ above.
The value for PD that was obtained from DCPI (in the section on
‘Maximum protein deposition’ in ‘Main common mechanisms’ above)
PD (from DCPI)
is used to determine desired MEI = + b0, which is
b1
compared to MEI. If MEI < desired MEI, PD is estimated via f(MEI);
otherwise, PD from DCPI remains unchanged.
Model Comparison
The main issue from the above section is: body protein deposition (PD)
takes place at a low material efficiency when it is limited either by the
genetic maximum to PD or by insufficient energy supply. The question is
then how the various models parameterize these relations. We deal next
with maximum protein deposition, with insufficient energy supply, and
give a brief comparison of how some of these models describe voluntary
food intake.
Three different methods are used for the definition of maxPD.

The simplest method makes the assumption that maxPD has a constant
value throughout the growth period, at least up to commercial slaughter
weights (Whittemore and Fawcett, 1976). Hence, different genotypes can

be characterized by different values for this single parameter. The other
two methods make maxPD depend on body weight or stage of body
development, describing a pattern where maxPD increases rapidly early in
life, plateaus during the growth period and then decreases towards zero at
maturity. Two types of functions are used to model this trend, as shown in
Fig. 13.5.
First, hyperbolic or parabolic equations of maxPD in terms of BW, in
both cases with two genetic parameters.
Walker and Young (1993) used a skew-hyperbolic relation with BW:
maxPD = BW e(c + d BW) (13.7)
Thorbek (1975) and Whittemore and Fawcett (1976) used a quadratic

relation with metabolic body weight:
maxPD = f MBW2 – g MBW (13.8)
Second, derivatives of sigmoid curves for potential protein growth, based

upon the idea that the effect of BW on maxPD reflects the effect of
physiological development. Emmans (1988) used the Gompertz function;
this equation has two genetic parameters, the rate parameter BGomp and
the mature body protein mass Pmature:
P ⎛ P ⎞
maxPD =BGomp × P × ln(Pmature /P) = BGomp × Pmature × × ln⎜1 / ⎟ (13.9)
Pmature ⎝ Pmature ⎠
Black (1988) used the Richards function, modified to replace one of the
two instances of protein mass by BW:
⎛ P ⎞
maxPD = k × (BW+S)a × [(Pmature − P) / Pmature ]= k × (BW + S)a × ⎜1 − ⎟ (13.10)
⎝ Pmature ⎠
This equation has a high information requirement for characterization of

the genotype: four genetic parameters (k, S, a and Pmature).
maxPD maxPD
Thorbek (1975)
eqn (8)
Emmans (1988) Black (1988)

Walker & Young (1993) eqn (9) eqn (10)
eqn (7)
BW P/Pmat
Fig. 13.5. Various functions to describe the course of maxPD in relation to body weight (BW)
or body protein mass (P) as a proportion of its mature value (Pmat).
All models described here use a function f(MEI) that describes the
reduction of PD for MEI < minMEImaxPD. This function is not always
explicit, but may follow implicitly from other, explicitly defined, relations.
Table 13.1 introduces these structures. Method 1 explicitly defines a value
for minLD/PD, and implicitly describes f(MEI) as a linear regression line
of PD on MEI with b0 = MEm. Methods 2 and 3 define f(MEI) explicitly
but differently. Methods 4 and 5 explicitly define a function for eP and a
value for minL/P, respectively, leading implicitly to a function PD =
f(MEI).
Method 1 was published earliest, and some subsequent methods give,
in addition to their explicit element f(MEI), also their own implicit value
for minLD/PD, for comparison purposes. We include minLD/PD in Table
13.1 for the same reason, although it is not required for a comparison of
these five methods when f(MEI) has already been described.
The general form of f(MEI) is PD = b1 × (MEI – b0), where b0 is an x-
intercept and b1 represents the reduction of PD at MEI < minMEImaxPD.
The general form of minLD/PD is derived as follows. Substituting f(MEI)
into Eqn 13.1 gives:
MEI − MEm − (PD × EP / kP ) MEI − MEm − [b1 × (MEI − b0 ) × EP / kP ]

LD = =
EL / kL EL / kL
1 − b1 × EP / kP ⎛ MEm − b0 × b1 × EP / kP ⎞
= × ⎜ MEI − ⎟
EL / kL ⎝ 1 − b1 × EP / kP ⎠
so that
Table 13.1. Methods for the reduction of protein deposition at MEI < minMEImaxPD.
Method minLD/PD PD = f(MEI) Third function
Whittemore explicit definition implicit description
1
Moughan constant → Lin(MEI)
Black implicit description explicit definition
2
De Greef nonLin(MEI) ← Lin(MEI)
3 Van Milgen implicit description ← explicit definition
nonLin(MEI) Quad(MEI)
Emmans and implicit description ← implicit description ← explicit definition
4
Kyriazakis nonLin(MEI, DCPI) Lin (MEI), nonLin(DCPI) eP = Lin(MEI/DCPI)
5 De Lange implicit description ← implicit description ← explicit definition
nonLin(MEI) Lin(MEI, L, P) minL/P = constant
Lin(): linear function. nonLin(): nonlinear function. Quad(): quadratic function.
References in the text.
1 − b1 × EP / kP ⎛ MEm − b 0 × b1 × EP / kP ⎞
× ⎜ MEI − ⎟
EL / kL ⎝ 1 − b1 × EP / kP ⎠
minLD / PD = =
b1 × (MEI − b0 )
(13.11a)
MEm − b 0 × b1 × EP / kP
MEI −
1 − b1 × EP / kP 1 − b1 × EP / kP
= ×
b1 × EL / kL MEI − b0
which can usefully be rearranged as
1 − b1 × EP / kP ⎛ MEm − b 0 1 ⎞
minLD / PD = × ⎜1 − × ⎟ (13.11b)
b1 × EL / kL ⎝ 1 − b1 × EP / kP MEI − b0 ⎠
Method 2 introduced the parameter ‘marginal minLD/PD’ (margminLD/PD):

the ratio between the regression coefficients of LD and PD on MEI, at MEI <
minMEImaxPD. We show in Appendix 13.1 that this parameter equals the first
1 − b1 × EP / kP
term in Eqn 13.11b: margminLD/PD = . We present
b1 × EL / kL
the implicit values of margminLD/PD as well as minLD/PD for the purpose of
comparison between the models.
The methods are discussed in detail in the following five sections and
in Table 13.2.
Table 13.2a. The forms of regression coefficientsa b0 and b1 for the methods in Table 13.1.
The forms of margminLD/PD and minLD/PD are in Table 13.2b.
Method b0 b1
Whittemore MEm 1
1
Moughan EP / kP + minLD / PD × EL / k L
Black
2 g1 g2
De Greef
g2 × (MEI − MEm − g1) =
3 Van Milgen MEm −4 × maxPD
= × (MEI − MEm − g1)
g12
Emmans and DCPI − Pm

4 0 ×
Kyriazakis DCPI
1
5 De Lange MEm + (minL/P × P – L) × EL / kL
EP / kP + minL / P × EL / k L
aPD = b1 × (MEI – b0).

Comparison of Pig Growth Models
Table 13.2b. Continuation of Table 13.2a: the forms of margminLD/PD and minLD/PD for the methods in Table 13.1.
1− b1 × EP / kP ⎛ MEm − b0 1 ⎞
Method margminLD / PD = minLD / PD = margminLD / PD × ⎜1− × ⎟
b1 × EL / kL ⎝ 1 − b1 × EP / k P MEI − b 0⎠
1 = minLD/PD g1
1− g2 × EP / kP ⎛ MEm − g1 1 ⎞
2 margminLD / PD × ⎜1− × ⎟
g2 × EL / kL ⎝ 1− g2 × EP / kP MEI − g1 ⎠
⎛ ⎞
1 ⎜ 1 ⎟
3 ×⎜ − EP / kP ⎟ margminLD/PD
−4 × maxPD
EL / kL ⎜ × (MEI − MEm − g1) ⎟
⎝ g12 ⎠
⎛ ⎞
1 ⎛1 DCPI ⎞ ⎜ MEm 1 ⎟
4 × × − EP / kP ⎟ margminLD / PD × ⎜1− × ⎟
EL / kL ⎜⎝ (DCPI − Pm ) ⎠ ⎜ 1− × (DCPI − Pm ) × E / k MEI ⎟
⎝ DCPI
P P
⎠
margminLD / PD ×
5 minL/P ⎛ (minL / P × P − L) × (EP / kP + minL / P × EL / kL ) 1 ⎞

⎜1+ ×
) L L ⎟⎠
⎝ minL / P MEI − [MEm (
+ minL / P × P − L × E / k ]
269
1. Whittemore and Fawcett (1976)

These authors assumed that an intrinsic minimum amount of lipid
deposition must accompany each unit of protein deposition, implying that
there is an intrinsic lower limit to the ratio between lipid and protein
deposition (minLD/PD), which is defined as a genetically determined
constant throughout the growing period (g1 in Table 13.2).
Substitution of minLD/PD × PD for LD into Eqn 13.1 shows that
f(MEI) is assumed to be a linear regression of PD on MEI:
MEI = MEm + (PD × EP / kP ) + ( minLD / PD × PD × EL / kL ) →

1 (13.12)
PD = × (MEI − MEm )
EP / kP + minLD / PD × EL / kL
It follows that b0 = MEm, b1 = 1/(EP/kP + minLD/PD × EL/kL). The linear

regression coefficient b1 is therefore also assumed to be a genetically
determined constant throughout the growth period.
2. Black et al. (1986)

Black explicitly assumed f(MEI) to be a linear regression of PD on MEI:
PD = b1 × (MEI – b0), and b0 and b1 are assumed to be genetically
determined constants throughout the growth period (g1 and g2 in Table
13.2). In contrast to Whittemore’s model (Eqn 13.12), the x-intercept is not
assumed to be in MEm; all its estimated values are lower than MEm. The
argumentation for b0 < MEm is that when MEI = MEm, protein deposition
is often positive and lipid deposition is often negative (i.e. lipid is
catabolized). Moreover, Black’s model does not require an intrinsic
minimum amount of lipid deposition that must accompany each unit of
protein deposition. Hence minLD/PD is not a constant; it decreases when
MEI decreases, see Table 13.2. It is genetically determined through b0 and
b1. By contrast, margminLD/PD is a constant in this model.
The entity margminLD/PD (the ratio between the regression
coefficients of LD and PD on MEI) was introduced by De Greef (1992) as a
genetically determined constant throughout the growth period. We show in
Appendix 13.1 that, because of Eqn 13.1, the regression of LD on MEI can
be fully expressed in terms of the one for PD, which removes the need for
De Greef ’s two additional genetic parameters (e.g. c0 and c1 in LD = c1 ×
[MEI – c0]). This makes De Greef ’s model collapse entirely to Black’s one;
the only novelty is that it makes the margminLD/PD parameter explicit.
3. Van Milgen et al. (2000)

Like Black, these authors explicitly assumed a regression function for
f(MEI): PD = b1 × (MEI – b0), with an x-intercept b0 = MEm just as in
Whittemore’s model. Their regression coefficient b1 is in itself a linear
regression on (MEI – MEm) and can be written, in our notation, as b1 =
g2 × (MEI – MEm – g1). This way, f(MEI) has the form of a quadratic
regression of PD on (MEI – MEm):
PD = b1 × (MEI − b0 ) = g 2 × (MEI − MEm − g1 ) × (MEI − MEm ) =

(13.13)
= g 2 × (MEI − MEm )2 − g1 × g 2 × (MEI − MEm )
Parameters g1 and g2 are defined as genetically determined constants. Eqn
13.13 is defined as having its maximum value at maxPD; this makes one of
the two parameters (g1, g2) redundant, as follows.
Solving Eqn 13.13 for dPD / d(MEI – MEm) = 0 gives minMEImaxPD:
dPD
= 2 × g 2 × (MEI − MEm ) − g1 × g 2 = 0
d(MEI − MEm )
g × g2
→ minMEImaxPD − MEm = 1 = g1 / 2 (13.14)
2 × g2
→ minMEImaxPD = MEm + g1 / 2
After substitution of minMEImaxPD from equation 13.14 for MEI in
equation 13.13, maxPD can be solved for in order to express b1 in terms of
maxPD and g1 (instead of g1 and g2):
maxPD = g 2 × (minMEImaxPD − MEm )2 − g1 × g 2 × (minMEImaxPD − MEm ) =

−4 × maxPD
= g 2 × (g1 / 2)2 − g1 × g 2 × (g1 / 2) = −0.25 × g 2 × g12 → g 2 = →
g12
b1 =
−4 × maxPD
g12
(
× MEI − MEm − g1 ) (13.15)
It follows that b1 is not a constant but a function of MEI. It is genetically

determined through maxPD and g1.
Like b1, minLD/PD is not a constant but a function of MEI; it is
genetically determined through maxPD and g1 (see Table 13.2). Note that
in this model, margminLD/PD equals minLD/PD and is therefore not a
constant either, in contrast to model 2.
4. Kyriazakis and Emmans (1992)

The reduction in eP with decreasing MEI given a certain amount of DCPI
was represented by these authors in terms of a linear regression of eP on
the ratio between the food ME content and food digestible protein content
(MEC / DCPC, which equals MEI / DCPI), forced through the origin:
MEI PD MEI DCPI − Pm
eP = × → =× → PD = × × MEI (13.16)
DCPI DCPI − Pm DCPI DCPI
The parameter µ is explicitly assumed to be a constant, not genetically

determined. Following Eqn 13.16, f(MEI) has the form PD = b1 × (MEI –
DCPI − Pm
b0) with b0 = 0 and b1 =× . The regression coefficient b1 is
DCPI
then a non-linear function of DCPI, not genetically determined. Table 13.2
shows that minLD/PD is not a constant, but a non-linear function of MEI

and DCPI, not genetically determined. It also shows that margminLD/PD
is not a constant either, but a non-linear function of DCPI, not genetically
determined.
5. De Lange (1995)
De Lange assumed that an intrinsic minimum amount of lipid mass (L)
must accompany each unit of protein mass (P), hence that there is an
intrinsic minimum lipid to protein ratio in the body (minL/P), which is
genetically determined. Therefore, minLD can be written as minLD =
minL/P × (P + PD) – L. The equivalent of Eqn 13.12 is then:
PD =
1
EP / kP + minL / P × EL / kL
(
× MEI − [MEm + (minL / P × P − L) × EL / kL ]) (13.17)
1
So b0 =MEm +(minL / P × P – L) × EL / kL, and b1 =
EP / kP + minL / P × EL / kL
It follows that f(MEI) is implicitly defined as a linear regression with a

regression coefficient b1 that is a genetically determined constant, just as
minL/P. Its x-intercept b0 is a function of P and L, and it is genetically
determined through minL/P. When the actual amount of body lipid exceeds
its intrinsic lower limit, then (minL/P × P – L) 0, so that b0 MEm.
Table 13.2 shows that minLD/PD is a function of P and L, and (non-
linearly) of MEI; it is genetically determined through minL/P;
margminLD/PD is a constant, this time equal to minL/P.
Prediction of voluntary food intake
Two of these models predict voluntary food intake, and they do so from
the combination of maintenance requirements and desired PD and LD.
Desired LD is defined in both cases via an additional equation. The forms
of these equations (and their difference between the two models) are
similar to the associated definitions of maxPD, as follows.
Emmans (1988) used again the derivative of the Gompertz growth
function to describe desired LD. This equation has two genetic parameters:
the rate parameter BGomp (equal to the one for maxPD in Eqn 13.9), and
mature body lipid mass Lmature.
desired LD = BGomp L ln(Lmature / L) (13.18)
Black (1988) used again the derivative of a modified Richards growth
function to describe the desired increase of net energy (NE) in the body:
desired NE = k × (BW + S)a × [(NEmature – NE)/NEmature]
which gives
desiredNE − maxPD × EP
desiredLD = (13.19)
EL
Discussion and Conclusions
Apart from several empirical relations, the models described here contain
three truly mechanistic elements in the form of Eqns 13.1, 13.2 and 13.3.
All genetic factors enter this mechanistic system through eP in Eqn 13.3:
maximum protein deposition (maxPD) and protein deposition at
insufficient energy supply.
The differences and similarities between the various models with
regard to these genetic factors are summarized and discussed below.
The genetic aspects of the five discussed methods are summarized in

Table 13.3.
Black et al. (1986), Moughan and Verstegen (1988) and Emmans and
Kyriazakis (1997) refer to several studies that show an effect of body weight
on maxPD. Although measurement of maxPD seems to be difficult and not
very reliable, especially in the early increasing phase, the trend of
‘increasing rapidly early in life, plateauing during the growing-fattening
stages and then decreasing towards zero at maturity’ is generally accepted.
Hence inclusion of this effect of body weight on maxPD makes the model
valid over a wider growth trajectory than with the assumption of a single
constant maxPD value as in method 1 in Table 13.3.
The hyperbolic, parabolic and sigmoid-derived equations are all
empirical, and it is probably possible to find functions from all these three
classes that fit the observed pattern very well. However, the derivatives of
the sigmoid functions have a more general character and useful properties
like allometry. Disadvantages of Black’s modified Richards curve (method
5) are the large number of genetic parameters and the unclarity of his
modification. The inclusion of BW has strange consequences; for example,
if a pig has a higher BW at a given P because it has more fat in its body,
then its predicted maxPD will be higher. The derivative of the Gompertz
Table 13.3. Methods to describe maximum protein deposition.

Method na maxPD Type
Whittemore
1 1 g1 constant
Moughan
Whittemore
2 2 f (g1, g2) hyperbola
Thorbek
3 Walker and Young 2 f (g1, g2) skewed parabola
4 Emmans 2 f (BGomp, Pmature) Gompertz
5 Black 4 f (k, S, a, Pmature) modified Richards
aNumber of genetic parameters.
curve with two genetic parameters (method 4) seems to be the preferred

model. An additional advantage is the smooth connection with Taylor’s
genetic size scaling principles, see Emmans (1988, 1997).
The genetic aspects of the five discussed methods are summarized in Table
13.4 and Fig. 13.6.
De Greef and Verstegen (1992) showed that minLD/PD decreases with
decreasing MEI. This makes method 1 (with a single constant value for the
genetic parameter minLD/PD, g1 in Table 13.4) unsuitable.
Table 13.4. Methods to describe protein deposition at MEI < minMEImaxPD.

Method na b0 b1 margminLD/PD minLD/PD
Whittemore
1 1 MEm g1 f(g1) g1
Moughan
Black
2 2 g1 f(g2) g2 f(g1, g2, MEI)
De Greef
3 Van Milgen 1 MEm f(g1, maxPD, MEI) f(g1, maxPD, MEI) f(g1, maxPD, MEI)
Emmans and
4 0 0 f(DCPI) f(DCPI) f(DCPI, MEI)
Kyriazakis
5 De Lange 1 f(g1, P, L, MEm) f(g1) f(g1) f(g1, MEI)
aNumber of genetic parameters.
PD PD
high L/P
maxPD maxPD
3 4
1
L/P > minL/P
high
DCPI
low
1 DCPI
3
4
2 L/P = minL/P
0 MEm MEI 0 MEm MEI

b0 < MEm b0 < MEm
Fig. 13.6. Methods to describe protein deposition with insufficient ME supply, as in Table 13.3.
Left: (1) Whittemore, Moughan; (2) Black, De Greef; (3) Van Milgen and Noblet; (4) Emmans
and Kyriazakis. Right: De Lange. References and explanation of symbols in the text.
Table 13.2 shows that method 2 makes minLD/PD dependent on MEI,

for which it requires two genetic parameters (g1 and g2 in Tables 13.2 and
13.4). The reason given by Black et al. (1986) for this is that when ME
intake is just sufficient to achieve zero energy retention (maintenance
energy intake, MEm), protein deposition is positive and lipid deposition is
negative (i.e. lipid is catabolized). This feature has been emphasized by
many authors (e.g. ARC, 1981; Close and Fowler, 1982; Walker and Young,
1993; NRC, 1996; Van Milgen and Noblet, 1999). Thus, the line describing
PD as a linear function of MEI has an x-intercept lower than MEm (b0 <
MEm): the assumption that the ME requirement at zero PD equals MEm (b0
= MEm) is therefore invalid.
Method 3 incorporates the effect of MEI on minLD/PD in a different
way. It requires two genetic parameters, one of which was shown, in Eqn
13.15, to collapse to maxPD; therefore Table 13.1 gives only g1. This is an
advantage over method 2 with its two required genetic parameters.
However, f(MEI) is a quadratic regression of PD on MEI with an x-
intercept equal to MEm (see line (3) in Fig. 13.6), which is (like for method
1) in conflict with the above mentioned finding of b0 < MEm. This makes
method 3 also unsuitable.
Method 4 does not include any genetic parameters, which makes it an
inconvenient method to model animals that differ in their genetically
desired lipid deposition at MEI < minMEImaxPD. This may be caused by
the fact that this method was based on data collected on very young pigs
(4, 6 or 8 weeks after 12 kg body weight, up to 20–50 kg end weight).
Emmans (1997) notices that ‘all genotypes are thin when they are very
immature’ and that this means that the body lipid to protein ratio shows
little variation at a low degree of maturity. This should make it very
difficult to detect genetic variation in this trait at that stage.
Method 4 presents two more complications. First, when an efficiency
parameter (i.e. output/input) such as eP is expressed in terms of MEI, it
would make sense to regress eP on MEI / (DCPI – Pm) where the
denominator properly represents ‘input’. However, the regression was
done on MEC / DCPC, which equals MEI / DCPI. Of course, Pm is small in
comparison to DCPI (especially in very young animals) so the error is
probably small.
Second, the above regression was forced through the origin, defining
b0 = 0. This approach has one parameter less to be estimated, an
important advantage when analysing a small dataset. However, such an
assumption about an intercept (with notoriously large standard errors of
the regression estimates) is very difficult to falsify, and the default
approach should be that the intercept is non-trivial (e.g. Neter et al., 1985,
pp. 163–164). At least there seems to be no valid biological reason to set b0
to zero. Inclusion of Pm as suggested above, and allowance for a non-zero
intercept in the regression, would make method 4 collapse to method 2. All
this makes method 4 unsuitable.
Method 5 has several advantages over the other methods: only a single
genetic parameter is involved and, just as in method 2, minLD/PD depends
on MEI and is not constant, and b0 MEm. There are two more reasons
why this method is attractive.
First, b0 depends on P and L in such a way that b0 approaches MEm
when [minL/P × P – L] approaches zero, i.e. when L/P approaches minL/P.
This is illustrated in Fig. 13.7. When a pig that used to be fed at a high
MEI level (so that it is on a high PD metabolism, and has a high L/P level)
is suddenly brought to a low MEI level (for example, MEI = MEm as in
Fig. 13.7), it will not directly reduce its PD metabolism to zero, but will
catabolize body lipid to support its protein retention. In Fig. 13.6 this
initial situation is represented by the base of the arrow at the upper
regression line with a high L/P level (b0 < MEm). When this low MEI level
continues, the resulting lipid catabolism reduces L/P until it has reached
minL/P (through subsequently lower regression lines in Fig. 13.6), and PD
metabolism follows the arrow to approach zero as represented by the
regression line with b0 = MEm. Black et al. (1986) describe a similar effect
of b0 approaching MEm during a long period (100 days) of insufficient
energy supply. But they had to add yet another empirical equation to
method 2 to quantify this effect; method 5 does not require that.
It also follows that methods that set b0 = MEm (methods 1 and 3 in
Table 13.4) implicitly assume a steady state situation where metabolism has
been adapted to a low MEI level, another reason why they are less suitable.
Second, the concept of minL/P in method 5 is more elegant (because more
mechanistic) than the empirical approach of the linear regression
parameters in method 2. It operates on the same level as Emmans’s (1988,
1997) system of potential body protein mass and desired body lipid mass,
PD LD
maxPD
L/P > minL/P
L/P = minL/P
B
⎫
⎬minLDmaxPD
PD A⎭
LD
0 MEm A B C MEI
⎭
⎬
⎫
minMEI maxPD
Fig. 13.7. An extension of Fig. 13.3. The PD system is the same as in the right-hand plot of
Fig. 13.6. Each PD line has an associated LD line which changes to a steeper slope in the
point minMEImaxPD where PD and LD reach maxPD and minLDmaxPD, respectively. See the
text for scenarios A, B and C.
which presumes that both entities follow Gompertz functions in relation to

age, characterized by three genetic parameters: mature body protein mass
(Pmature), mature body lipid mass (Lmature), and a common specific growth
rate parameter BGomp.
In Emmans’s system as extended by Ferguson et al. (1994), the pig’s
desired lipid deposition at any point in time makes up for the deficit
between desired lipid mass (which follows the above genetically
determined Gompertz pattern) and actual lipid mass (which may deviate
from it for a variety of reasons). Hence for a given level of desired lipid
mass, desired lipid deposition will be lower for higher values of actual lipid
mass. Similarly, in method 5, minLDmaxPD equals [minL/P × (P + maxPD) –
L] which for a given level of minL/P is lower for higher values of L, and
hence for higher values of L/P.
In the example in Fig. 13.7, the pig with the highest actual L/P level
(farthest removed from the minL/P line) reaches its maxPD at MEI level A,
where its LD = minLDmaxPD level is lower than for the pig with the lower
actual L/P which reaches its maxPD and minLDmaxPD values at MEI level B.
The connection between Emmans’s system and method 5 can be
extended as follows.
−1
desired L L mature ⎛ P ⎞
minL / P ≤ = × ⎜ ⎟ (13.20)
desired P Pmature ⎝ Pmature ⎠
0.23
⎛L ⎞
with = 1.46 × ⎜ mature ⎟ as an ‘auxiliary variable’ (Emmans, 1997).
⎝ Pmature ⎠
More generally, the right-hand term in Eqn 13.20 is a non-linear function

of (P/Pmature) with genetic parameter (Lmature /Pmature). Like maxPD it varies
throughout the growth period.
It follows from the above that the two terms in inequality (Eqn 13.20)
behave the same way, and that one serves as a boundary value for the
other. This would suggest a functional relationship between the two entities
desired L
minL/P and , which merits further study.
desired P
As a consequence, minLDmaxPD must be seen as the lower limit to
desired LD. In Fig. 13.7 this leads to the situation where voluntary MEI >
minMEImaxPD so that desired LD > minLDmaxPD, indicated by the circles
for PD (at maxPD) and LD (at desired LD) when the genotype with
minMEImaxPD = B realizes a voluntary MEI = C. In that case we can write:
voluntary MEI = MEm + maxPD × EP/kP + desired LD × EL/kL =
= MEm + maxPD × EP/kP + minLDmaxPD × EL/kL + extra LD × EL/kL
(13.21)
The last term (extra LD) must then be due to either: (i) a drive towards an
LD (desired LD) higher than minLDmaxPD; or (ii) a drive towards a higher
MEI than minMEImaxPD. Both would lead the pig to realize a datapoint (as in
Fig. 13.7) beyond the maxPD level; Campbell et al. (1983, 1985), Campbell
and Taverner (1988) and Dunkin and Black (1987) have shown that this
actually happens. Option (ii) can more appropriately be expressed as:
voluntary MEI − (MEm + maxPD × EP / kP + minLDmaxPD × EL / kL )
extra LD =
EL / kL
(13.22)
It follows that whereas the trajectory for MEI < minMEImaxPD is controlled by
a genetic drive for lipid deposition (expressed as minL/P), the trajectory for
MEI minMEImaxPD is controlled by a genetic drive for protein deposition
(expressed as maxPD) and a genetic drive for either lipid deposition or
‘luxury’ ME intake (which would lead to extra LD through Eqn 13.22).
Final remarks
MEm is quantitatively quite large, and it depends on many environmental

factors. Hence, errors in its estimation may have considerable effects on
model predictions. All models described here assume that MEm is not
subject to genetic variation. This assumption is probably false (see Knap,
2000, p. 164) and MEm must be characterized for the genotype to be
simulated. The system of protein deposition, lipid deposition and
maintenance can be kept internally consistent if MEm and Pm are then
expressed in terms of P/Pmature (e.g. for growing animals, MEm = α ×
P
P =
× 0.73
× Pmature according to Emmans, 1997), similar to maxPD
Pmature
and desired LD (Eqns 13.9 and 13.18). Apart from the models mentioned
in the section on ‘Prediction of voluntary food intake’ in ‘Model
comparison’ above, most models do not simulate voluntary food intake but
treat it as an input parameter. For pig breeding applications, where
voluntary food intake is of true interest, this is unproductive; this again
requires inclusion of MEm as a genotype characteristic in the model, on the
same level as maxPD and desired LD.
Finally, all models assume EP, EL, kP and kL to be constants without
genetic variation; evidence to support or falsify this assumption is virtually
absent (see Luiting, 1990, for an overview in poultry, and Emmans, 1997).
Conclusions
For the purposes of animal breeding, desirable features of a growth model

are: (i) that it is consistent with real-life data; (ii) that it is internally
consistent; (iii) that it has a small number of genotype-specific parameters;
and (iv) that it is able to predict voluntary food intake. Given the shortage
of reliable data that span a sufficiently wide range of P/Pmature on modern
genotypes, item (i) will always be hard to meet, and not necessarily as a
shortcoming of the models. For the description of maxPD, method 4 of
Table 13.3 seems the most appropriate method to meet the other criteria.
With respect to feature (ii) it is best combined with method 5 of Table 13.4
for the description of PD at MEI < minMEImaxPD. An additional advantage
of this combination of methods is that it provides a smooth and consistent
connection with Taylor’s genetic size scaling principles.
References
ARC (1981) The Nutrient Requirements of Pigs. Agricultural Research Council,
Commonwealth Agricultural Bureaux, Slough, UK.
Black, J.L. (1988) Animal growth and its regulation. Journal of Animal Science 66
(Suppl. 3), 1–22.
Black, J.L., Campbell, R.G., Williams, I.H., James, K.J. and Davies, G.T (1986)
Campbell, R.G. and Taverner, M.R (1988) Genotype and sex effects on the
relationship between energy intake and protein deposition in growing pigs.
Campbell, R.G., Taverner, M.R. and Curic, D.M (1983) The influence of feeding
level from 20 to 45 kg liveweight on the performance and body composition of
female and entire male pigs. Animal Production 36, 193–199.
Campbell, R.G., Taverner, M.R. and Curic, D.M (1985) Effects of sex and energy
intake between 48 and 90 kg liveweight on protein deposition in growing pigs.
Close, W.H. and Fowler, V.R (1982) Energy requirements of pigs. In: Haresign, W.
(ed.) Recent Advances in Animal Nutrition. Butterworths, London, UK, pp. 1–16.
De Greef, K.H. (1992) Prediction of production: nutrition induced tissue
partitioning in growing pigs. PhD dissertation, Wageningen Agricultural
University, Wageningen, Netherlands.
De Greef, K.H. and Verstegen, M.W.A. (1992) Partitioning of protein and lipid
deposition in the body of growing pigs. Livestock Production Science 35, 317–328.
De Lange, C.F.M. (1995) Framework for a simplified model to demonstrate
Wageningen Pers, Wageningen, Netherlands, pp. 71–85.
Dunkin, A.C. and Black, J.L. (1987) The relationship between energy intake and
nitrogen balance in the growing pig. In: Moe, P.W., Tyrrell, H.F. and Reynolds,
P.J. (eds) Energy Metabolism in Farm Animals. EAAP publication 32. Rowman and
Littlefield, Totowa, New Jersey, pp. 110–113.
Emmans, G.C. (1988) Genetic components of potential and actual growth. In:
Land, R.B., Bulfield, G. and Hill, W.G. (eds) Animal Breeding Opportunities.
BSAP occasional publication 12, British Society for Animal Production,
Edinburgh, UK, pp. 153–181.
Emmans, G.C. (1997) A method to predict the food intake of domestic animals from
birth to maturity as a function of time. Journal of Theoretical Biology 186, 189–199.
Emmans, G.C. and Kyriazakis, I. (1995) A general method for predicting the
weight of water in the empty bodies of pigs. Animal Science 61, 103–108.
Emmans, G.C. and Kyriazakis, I. (1997) Models of pig growth: problems and
proposed solutions. Livestock Production Science 51, 119–129.
Knap, P.W. (2000) Variation in maintenance requirements of growing pigs in
relation to body composition; a simulation study. PhD dissertation,
Wageningen Agricultural University, Wageningen, Netherlands.
Kyriazakis, I. and Emmans, G.C. (1992) The effects of varying protein and energy
intakes on the growth and body composition of pigs. 2: The effects of varying
both energy and protein intake. British Journal of Nutrition 68, 615–625.
Luiting, P. (1990) Genetic variation of energy partitioning in laying hens: causes of
variation in residual feed consumption. World’s Poultry Science Journal 46, 133–152.
Moughan, P.J. and Smith, W.C. (1984) Prediction of dietary protein quality based
on a model of the digestion and metabolism of nitrogen in the growing pig.
New Zealand Journal of Agricultural Research 27, 501–507.
Moughan, P.J. and Verstegen, M.W.A. (1988) The modelling of growth in the pig.
Netherlands Journal of Agricultural Research 36, 145–166.
Neter, J., Wasserman, W. and Kutner, M. (1985) Applied Linear Statistical Models.
Irwin, Homewood, Illinois.
NRC (1996) Nutrient Requirements of Beef Cattle. National Academy Press,
Washington, DC.
Pomar, C., Harris, D.L. and Minvielle, F. (1991) Computer simulation of swine
production systems. 1: Modeling the growth of young pigs. Journal of Animal
Science 69, 1468–1488.
Thorbek, G. (1975) Studies on energy metabolism in growing pigs. 2: Protein and fat gain in
growing pigs fed different feed compounds. Efficiency of utilization of metabolizable energy
for growth. Report 424, Statens husdyrbrugsforsøg, Copenhagen, Denmark.
TMV (1991) Technisch model varkensvoeding: informatiemodel. Research report P1.66,
Research Institute for Pig Husbandry, Rosmalen, Netherlands.
Van Milgen, J. and Noblet, J. (1999) Energy partitioning in growing pigs: the use of
a multivariate model as an alternative for the factorial analysis. Journal of
Van Milgen, J., Quiniou, N. and Noblet, J. (2000) Modelling the relation between
71, 119–130.
Walker, B. and Young, B.A. (1993) Prediction of protein accretion, support costs
and lipid accretion in the growing female pig and dry sow. Agricultural Systems
42, 343–358.
Whittemore, C.T. (1995) Modelling the requirement of the young growing pig for
dietary protein. Agricultural Systems 47, 415–425.
Appendix 13.1
In addition to a linear regression of PD on MEI with linear regression

coefficient r and x-intercept s, De Greef (1992) assumes a ‘constant’ linear
regression of LD on MEI with linear regression coefficient t and x-intercept v:
PD = r × (MEI − s)
(13.A1)
LD = t × (MEI − v)
He defines the ratio between the two linear regression coefficients as the
marginal minLD/PD ratio and assumes that this parameter is a genetically
determined constant, which makes minLD/PD dependent on MEI:
t × (MEI − v) t MEI − v MEI − v (13.A2)
minLD / PD = = × = margminLD / PD ×
r × (MEI − s) r MEI − s MEI − s
However, because of the law of conservation of energy, the linear
regression of LD on MEI can be written in terms of the parameters of the
linear regression of PD on MEI; De Greef does not mention this:
MEI − MEm − (PD × E / kP )
LD = P =
EL / kL
MEI − MEm − (r × (MEI − s) × EP / kP )
= =
EL / kL (13.A3)
1- r × EP / kP ⎛ MEm − r × s × EP / kP ⎞
= × ⎜ MEI − ⎟=
EL / kL ⎝ 1- r × EP / kP ⎠
= t × (MEI − v)
This shows that t is a function of r, and that v is a function of r, s and MEm,
which automatically makes both t and v constants too. As a consequence,
the ratio between t and r (i.e. marg minLD/PD) is a constant with value (1 –
r × EP / kP) / (r × EL / kL).
Hence given that r = b1 and s = b0, substitution in Eqn 13.A2 shows
that De Greef ’s minLD/PD equals the one of Black (see Table 13.2):
MEI − v
minLD / PD = margminLD / PD × =
MEI − s
MEm − r × s × EP / kP
MEI −
1 − r × EP / kP 1 − r × EP / kP
= × =
r × EL / kL MEI − s
MEm − b0 × b1 × EP / kP
MEI −
1 − b1 × EP / kP 1 − b1 × EP / kP
= × =
b1 × EL / kL MEI − b 0
⎛ MEm − b0 × b1 × EP / kP ⎞
MEI − b0 + b0 − ⎜ ⎟
1 − b1 × EP / kP ⎝ 1 − b1 × EP / kP ⎠
= × =
b1 × EL / kL MEI − b0
⎛ b0 − b0 × b1 × EP / kP − MEm + b0 × b1 × EP / kP ⎞
⎜
1 − b1 × EP / kP ⎜ 1 − b1 × EP / kP ⎟
= × 1+ ⎟=
b1 × EL / kL ⎜ MEI − b0 ⎟
⎜ ⎟
⎝ ⎠
1 − b1 × EP / kP ⎛ MEm − b0 1 ⎞
= × ⎜1 − × ⎟
b1 × EL / kL ⎝ 1 − b1 × EP / kP MEI − b0 ⎠ (13.A4)
14 Mechanistic Modelling at the
Metabolic Level: a Model of
Metabolism in the Sow as an
Example
J.P. MCNAMARA
Department of Animal Sciences, Washington State University
PO Box 646351, Pullman WA 99164-6351, USA
mcnamara@wsu.edu
Introduction
Models are representations of reality. The fields of nutrition, metabolism
and biomedicine have used models to aid in research and its application
since before World War II (see Black et al., 1986; Baldwin, 1995; NRC,
1998, 2001). A model may also be defined as an ordered way of describing
knowledge of some real system. Much research into nutrition of farm
animals since the early 1900s has been used, directly or ‘by default’, to
build, evaluate and improve models of nutrient requirements. An example
would be the Net Energy System, first fully described by Lofgreen and
Garrett (1968) as applied to beef cattle nutrition. Since then, this model
system and others derived from it has been used to determine energy
requirements of many animal species, and to determine energy values of
foods and feeds (NRC, 1998, 2001).
An early model in energy metabolism was the linear equation of fasting
heat production = 0 + 70 kcal/kg BW0.75. Energeticists such as Brody,
Kleiber and Blaxter asked questions about the commonality of energy use
by organisms. This model was derived from empirical data and spurred
decades of discussion, experimentation and further more detailed and
mechanistic model building, encompassing new information as it became
available. These lines of research led to practical empirical models used in
food production and human nutrition as well as to mechanistic research
into the control of energy metabolism in animals. It was the effort of
describing increasingly complex chemical knowledge in strict mathematical
terms, in an iterative fashion as more knowledge became available, and

Mechanistic Modelling at the Metabolic Level 283
applying the knowledge in the real world as appropriate, that has lead to
great efficiencies and improvements in nutrition.
Billions of US dollars, represented as conservation of resources and
decreased costs of raising feed for livestock and of the labour for feeding of
livestock, have been saved by application of these nutritional models (NRC,
1998, 2001). If we assume that application of practical models and
resultant changes in ration formulation and feeding has gained a cost
saving of $0.50 per market pig, and an average of 100 million pigs killed in
the US each year for the last 20 years, that is a saving of about US$1
billion. If we make a liberal estimate of 30 states with two pig nutritionists
per state at a salary of $40,000 and each had a total of $100,000 in total
research support for 20 years, that is about $168 million. Not a bad return
on investment. And this is only for pigs.
Our ability to describe metabolic transactions, and their resultant effect
on nutrient requirements is critical to raising food-producing animals in
efficient ways around the globe. As our knowledge of variation in genetic
and environmental situations continues to increase, this author takes the
philosophy that it is only through continuing to develop models of
increasing complexity, ever grounded in validated research data, that we
will continue to improve our true knowledge, wisdom and their
application to feeding the world.
Another use of such models is in teaching. The more complex models
have great utility for demonstrating to students how the complex system
works. In addition, when used in a teaching programme which also
exposes students to practical models used in production or to daily
nutrient requirements, the natural comparison between the research
models and the practical models can help them to see the direct
connections between the specific biological mechanisms and the practical
application of them.
One final set of reasons to continue to expand research in models has
already been given by one much more capable than myself, so I will close
the introduction with a list of reasons for using research models quoted
directly from Chapter 1 of Modelling Ruminant Digestion and Metabolism
(Baldwin, 1995):
Objectives in research modelling:
(a) Integration of existing concepts and data in a format compatible with
quantitative and dynamic analyses.
(b) Reduction of conceptual difficulties in analyses of interactions among elements
of complex systems.
(c) Evaluation of concepts and data for adequacy in both the quantitative and
dynamic domains.
(d) Evaluation of alternative hypotheses for probable adequacy when current
concepts are found to be inadequate, and identification of critical experiments
and measurements.
(e) Estimation of parameter values not directly measurable and interpretation of
new data.
284 J.P. McNamara
Throughout this chapter I will follow a philosophy and approach

initiated and developed by Lee Baldwin, quoted above, and use primarily
aggregate level pathway biochemistry, such as protein and fat synthesis,
and carbohydrate, fat and amino acid oxidation. Then, we can ask
questions about how ‘genetics’, as exemplified by transcription of certain
genes, including endocrine and neural control, can affect the ‘kinetics’ of
flux through the pathway, and how some specific environments can
attenuate or enhance the genetic drive through substrate supply and
resultant hormonal response. Finally, we will present the first description
of a model of reproductive control by nutrient use and endocrine systems.
I was asked to provide a general overview, and thus I shall, with some
examples from an existing model of metabolism in the pig (Pettigrew et al.,
1992a,b). The newer reader may be interested in several other reviews and
research papers on this topic in dairy cattle and pigs (Black et al., 1986;
McNamara, 1998, 2000, 2003, 2004, 2005; McNamara and Boyd, 1998;
McNamara et al., 2005). The growing knowledge from genomic research
now allows us to expand our models of metabolic control, and thus of
efficiency of food production into more and more ‘realistic representations
of reality’.
The Example of Lactation
Lactation provides a challenge not only to the female, but to the modeller
as well. Key challenges are the interaction of the complex of organs
involved in the adaptation and sustaining of metabolism to the new
physiological state, and the range and speed of change in early lactation.
One goal for the continued improvement of detail and accuracy is to
improve our quantitative understanding of control of metabolic systems.
Pregnancy and early lactation are times of metabolic stress met with a
coordinated response from the hormonal and neural systems (McNamara,
1994). In order to manage the changes in nutrient flows and
interconversions in the animal there is a complex and redundant system of
control factors, better known as hormones and neurotransmitters
(McNamara, 1998; McNamara and Boyd, 1998). Adipose tissue, as an
energy storage reservoir and, as we now understand, an endocrine organ
(Mohamed-Ali et al., 1998), adapts to support fetal growth and lactation
(McNamara et al., 1985; McNamara and Pettigrew, 1994, 2002a,b; Parmley
and McNamara, 1996). These systems are inextricably linked and excellent
reductionist experiments have identified key elements of each subsystem.
However, experiments that describe equation forms and parameter values
for the control exerted on flux are more limited in availability, primarily
because of the cost and complexity of conducting such experiments. A
more detailed analysis of quantifying metabolic control is given in
McNamara and Boyd (1998).
It is not always easy or affordable to do research with sufficient
repeated measures to estimate such rapidly changing fluxes, so some
compromises must be made. However, even with these limitations, using a

modelling approach can help us to design reasonable experiments to help
us improve. Yet another problem is in experimental focus and design.
Many of our nutrition experiments are purposefully done in mid- to late-
lactation to reduce the variance among animals and are usually short term
due to cost. We, as scientists, need to study the metabolic regulation during
the lactation cycle more often and in more detail. There has been a lot of
good new research in genetic control (Bastianelli et al., 1996; Hurley et al.,
2000; Lovatto and Sauvant, 2003), but we need to understand how these
controls relate to metabolic flux. Also, in much nutrition research, animals
are often grouped in genetically similar units instead of designing studies
to determine the interaction between genetics (with a measure admittedly
as crude as previous production) and diet.
A final limitation to having more detailed and accurate models is
(obviously) the complexity of the system itself. Frankly, this author thinks
that proper experiments are often not done because too many scientists
simply either do not appreciate the true complexity of the system, or they
do but are unwilling or unable to actually study it. In order to meet this
goal of describing complexity, we need to focus more strongly on the
endocrine and neural regulation of gluconeogenesis, lipolysis and
lipogenesis, amino acid interconversions and of feed intake. An excellent
way to do this is in the continued development, testing, evaluation, and
challenging with real data, of dynamic, mechanistic, metabolic models of
metabolism. We need to change the philosophy of single-investigator, small
study funding priorities into one of more team approaches and integrative
biology. We need to have funding proposals and panels who can
understand and fund whole animal integrative biology.
Brief Description of Dynamic, Mechanistic, Metabolic Models
A key characteristic of a model is how it describes change over time. A

model that describes a process at one time, usually through an empirical
equation is static. This is true even if the time frame studied was over
several months. A dynamic model integrates change over time. Both are
useful; however, only a dynamic model will help us truly improve our
understanding, first because that is reality; secondly because the
requirements at one time or over a short period are always partially a
function of what has come before. Dynamic models using differential
equations over time, and continuous non-linear functions (e.g. Michaelis-
Menten) can describe the turnover functions that can be ascribed either to
maintenance or to heat increment, such as ion transport, protein turnover
in the muscle and viscera, and triglyceride turnover in the adipose tissue.
When one actually studies these functions, one gets an appreciation for just
how important and variable they can be. For example, a change in muscle
protein turnover of 10%, which is at least a minimum increase in lactation
(Pettigrew et al., 1992a,b; Overton et al., 1998; Drackley, 1999; McNamara
286 J.P. McNamara
and Baldwin, 2000; Phillips et al., 2003) would increase energy for
maintenance by about 8.4 MJ/day (see Baldwin, 1995 for calculations and
stoichiometry). Over 100 days, that is an error of 840 MJ or about 28 kg of
adipose tissue. In a lactating sow, then, we may err by 170 MJ or over 5 kg
of body fat in predicting energy use.
Integrating Genetic Elements into Metabolic Models
There has been a wealth of work on genomics, metabolic control theory,

flux control and the like (Roehe, 1999; Hirooka et al., 2001; Quintanilla
et al., 2002). Simplistically, all kinetics is genetically controlled, even
though many environmental, including nutritional, effects may be larger
(e.g. an animal with a greater genetic drive for muscle growth may
accrete less muscle than a genetically inferior animal if the nutritional
inputs differ sufficiently). There has been significant work done on
structural genetics and gene expression in pigs (Roehe, 1999; Hirooka et
al., 2001; Quintanilla et al., 2002). There is even work beginning on
breeding for traits important for animal welfare (Kanis et al., 2005); it is
past time to start working elements of genetic control into metabolic flux
models.
There should be certain objectives to doing this, such as:
1. Develop a model that integrates transcription of genes coding for key
metabolic enzymes in growing, pregnant and lactating pigs to understand
underlying patterns of control.
2. Develop a model relating genetic variance among lines of pigs to
differences in metabolic patterns in order to identify variation in efficiency
of food use.
3. Use models of metabolism and metabolic control to identify which
genes would exert the majority of control over these flux rates.
4. Integrate models of structural genomics with metabolic models to identify
key patterns and sequences that are common to the most efficient animals.
5. Integrate kinetic flux models with those describing secretions and
mechanisms of endocrine action to identify the quantitative importance of
key control elements.
6. Develop a model of the interactions between nutrient use patterns,
endocrine secretions and actions, and downstream effects on gene
transcription to improve the understanding of key response elements.
Hargrove (2004) has given some examples of modelling gene expression,
which in turn regulate kinetic flux. These models may be used as teaching
tools, such as described by Collins (2004), or as hypothetical models to help
direct mechanistic research trials. We also need to find ways to connect
‘pure gene level’ models of genetic expression (Jiang and Gibson, 1999a,b;
Hirooka et al., 2001) with kinetic models, as the first exists to control the
second! With agricultural animals, we have sufficient genetic records that
we have been incorporating them into models for at least 40 years.
Enzymes control the rates of chemical reactions. Key enzymes are

controlled in turn by substrate and product concentration, allosteric
control by other molecules (such as ATP or messengers from hormonal
action); and hormonal control of enzyme synthesis or destruction. A simple
graphic (Fig. 14.1) summarizes these well-known control elements
(McNamara, 2005). Some simple examples of the types of control are given
in Table 14.1. These varied and redundant mechanisms have direct control
on flux rates through the key pathways of glucose synthesis and oxidation,
and fat and protein synthesis. Sometimes control is simply exerted by
determining the presence or absence of a process or pathway. Other
enzymes, primarily those we study in nutrition, are actively regulated in
both short term (minute-to-minute) to longer term (hours, days) to direct
the use of nutrients by cells, and more pertinent to the animal, by organs.
Most of these processes have several levels of control, from substrates and
products to hormones. Substrate concentration is often the greatest
effector of flux, but in the example of lactation, a sow might be consuming
all the starch she can, but not as much of that absorbed glucose will go to
body fat as it would in the non-lactating state because the hormones of
lactation have altered several genetic level control points (transcription of
acetyl CoA carboxylase for one).
Thus we must model these different levels of control, following
objectives such as those laid out by Baldwin, or those above, or others, to
help detail the specific role of each type of regulation. Some potential
mechanisms will be confirmed as the critical initiative events or responses,
DNA transcription site *

Receptor*
Nucleotides
cAMP mRNA
Hormone* or
Nutrient
*K1
Substrate
Amino
Initial acids
substrate(s) enzyme
>>> Pathway >>>
Vm,Ks*
*K2
product
End Product <<< Pathway <<<
Fig. 14.1. Conceptual model of control of flux through a reaction, indicating types of genetic
control over flux. Conceptual flow of genetic mechanisms of control of flux. The asterisk
indicates states (proteins, nutrients) that are genetically controlled and may have a
measurable heritability. Each state (box) and rate (arrow) can be measured to determine
appropriate parameters. Large dashes indicate a mechanistic state (with components), small
dashes indicate a lower level of organization to the main objective.
288 J.P. McNamara
Table 14.1. Examples of control of enzyme-catalysed reactions.

Catalytic level: Control of enzyme activity; does not require protein synthesis or
breakdown, usually for the short term; response time in seconds, effect can be quickly
reversed or removed.
Level and type of control Controllers Example reactions
Substrate activation Glucose Glucokinase,

Phosphofructokinase
Product inhibition Fructose-6-phosphate Phosphofructokinase
Fatty acids Fatty acid synthetase
Acetyl CoA Acetyl CoA carboxylase
Allosteric activation and inhibition
Hormonal: Insulin, glucagon, corticoids, usually via:
Phosphorylation Pyruvate dehydrogenase
and dephosphorylation Hormone sensitive lipase
Metabolic: ATP, citrate, glucose, and other metabolites:
Glucose oxidation PFK, FBP-ace
Fatty acids Ac CoA carboxlyase, FAS
Concentration level (control of enzyme presence or concentration); usually requires
protein synthesis or breakdown; fast to slow (few hours or days; may take longer for
response to be reversed or removed).
Enzyme destruction (increasing proteolysis and removal of enzyme)
Enzyme synthesis (increasing synthesis of enzyme, through classic transcription and
translation mechanisms)
Insulin, glucagon, thyroid hormone, growth hormone,
Prolactin, estrogen, progesterone, testosterone
Glucokinase
PDH, AcCoA Carb.
FAS
others may be identified to be important, but downstream, or secondary in

terms of quantitative control. Others may be identified as redundant,
acting only when other systems are not operating. More than one control
system may operate on an enzyme to control flux, but the net effect of
three such controllers does not differ from that of two or one (and any one
of the three can do the job). An example might be pyruvate
dehydrogenase (PDH), which is under allosteric control by hormones and
their second messengers, substrate, products and downstream products
such as ATP. Yet one or all of these controllers might reduce PDH activity
to negligible rates by themselves.
The enzymes of primary importance in glucose and fatty acid
metabolism are controlled at multiple levels from enzyme activation/
deactivation to DNA transcription and RNA translation. Those under the
strictest control would include glucokinase, phosphofructokinase, fructose
bisphosphatase, pyruvate dehydrogenase, acetyl CoA carboxylase, fatty acid

synthetase, phosphoenylpyruvate carboxykinase and pyruvate carboxylase
(Tepperman and Tepperman, 1970; Girard et al., 1997). For mammals it is
the maintenance of a relatively constant supply of blood glucose to the brain
that provides a large part of the metabolic control in the body. In addition,
homeorhetic controls in different physiological states then provide further
regulation. Certainly no nutritional model is pertinent which does not
include some elements of these control points on flux. In turn, the state and
rates of glucose use directly affect the rates of fatty acid and amino acid
metabolism. So it is important that models do an adequate job of describing
this control.
All enzymes and hormones are genetically regulated, from immediate
gene transcription and translation, to heritability of variations in hormone
and enzyme synthesis and secretion. Some examples may be found in
Girard et al. (1997), Bosch et al. (1999) and Daniel et al. (1999). Much
genetic variation may have its initial effect in the central nervous system or
in endocrine organs such as the thyroid or pancreas, but the end effect is
on flux. Years of genetic selection and growth and lactation trials in pigs,
chickens and cattle have proved the point (McNamara, 2000; and
McNamara and Boyd, 1998 for examples). It is also pertinent to describe
the shorter-term genetic controls such as gene transcription and
translation in models. The objective of a model dictates (or should) the
model components. If an objective is to model metabolic flux in any one
species of animal, allowing for description of variation among animals,
then genetic control by definition must be included.
Included in genetic effects on flux would be the presence or absence of
a transcribed and translated gene for an enzyme, the number of copies
(concentration) of the enzyme, as well as the enzyme catalytic ability itself.
While the latter can change through various isozymes and there are some
species differences, within a species this is not thought to be a major
control point. It is the transcription of the gene and translation of the
mRNA that alters the concentration of an enzyme, as well as the hormonal
and neural control of enzyme activity, that are the primary genetic control.
For example, lipoprotein lipase activity will change under the control of
hormones of pregnancy and lactation, in both the mammary gland and the
adipose tissue. In this instance, the change is in opposite directions
(decreasing flux in adipose tissue and increasing flux in mammary;
McNamara, 1998). This is certainly a genetic effect, and is probably
heritable. Adipose tissue enzyme activity is heritable in relation to body fat
amount in agricultural animals (McNamara and Martin, 1982; Hausman
and Hausman, 1993; McNamara, 1998; McNamara and Boyd, 1998),
albeit the actual causative gene or genes (whether in the adipose or due to
hormonal differences) is not clearly known as yet.
Thus, a valid argument is that the initial conditions in growth,
initiation of pregnancy or lactation, are genetic effects that set the
optimum for the ensuing states and rates, which of course can be
attenuated or enhanced by the environment. In the example of milk
290 J.P. McNamara
production (as most phenotypes), genetic control is a pleiotropic effect,

with many genes involved, including those controlling cellular
differentiation and senescence as well as those genes for milk protein,
lactose and fat synthesis. This may be a phenotypic single trait (milk
production), but many genes are involved. Many potential effectors
aggregate to control a few key pathways (protein synthesis, proteolysis, fat
synthesis, esterification, lipolysis, lactose synthesis, glycogen synthesis and
glycogenolysis, gluconeogenesis). Thus a model may initially incorporate
the concentrations of ATP, acetyl CoA, acyl CoA, free CoA, citrate, insulin,
glucagon and translation rates of the gene encoding for acetyl CoA
carboxlyase, on the flux through this enzyme.
So, back to the basic objective of a model. A model with a primary
objective to describe the control of flux of carbon through the acetyl CoA
carboxlyase reaction would include certain pools, equation forms and
parameters at a very specific chemical level. Such conceptual models have
been constructed (Girard et al., 1997 and references therein). However, a
model with the primary objective of describing the metabolism of absorbed
energy-containing nutrients, including amino acids, related to long-term
feeding strategies in the lactation phase of the reproductive cycle of sows
(Pettigrew et al., 1992a,b) would not likely have the same level of detail. Yet
it would contain genetic control elements to describe the effects of genetic
selection and nutrient supply on fatty acid synthesis, in adipose tissue and
the mammary gland. By aggregation of the total sum of reactions into the
lipogenesis pathway, for example, we can describe the reactions at our
desired level of biological organization with pertinent descriptions of
genetic control.
In this type of model – pathway biochemistry, we can sum up control
of an aggregated pathway in a simple and scientifically correct fashion.
Continuing to use lipogenesis as an example, kinetic flux can be described
through the two Michaelis-Menten parameters of maximal velocity (Vmax)
and substrate sensitivity (Km), example shown in Fig. 14.2. Relevant
nutritional models with an objective of describing input/output
relationships at the organism level should be dynamic and aggregated at
the level of pathway biochemistry in each organ. That really is what the
animal does. Higher level (whole body) or lower level (individual cells,
pathways, or enzymes) models are useful, but have a different objective.
This type of description can provide simple possibilities for
construction and testing of models integrating genetic elements. The
equation below is used in Pettigrew’s model of metabolism in lactating sows
to describe glucose conversion to body fat (Pettigrew et al., 1992a):
UGlTs = vGlTs /(1 + (MGlTs /cGl)), and (14.1)

MGlTs = MAGlTs * (( cGlr/cGl)tAGlTs); (14.2)
where MGlTs is the substrate sensitivity constant for glucose, and is

controlled by the concentration of glucose (reflecting the hormone insulin)
such that as glucose concentration rises (insulin increases), the sensitivity
10
V = Vmax/(1 + (Km/[S])Θ)
9
6
Rate of Reaction
Vm = 10, Km = 3
3
Vm = 15, Km = 3
Vm = 10, Km = 1
2
Theta = 3
0
1 1.8 2.6 3.4 4.2 5 5.8
Substrate Concentration
Fig. 14.2. Example of genetic control over substrate saturation (Michaelis-Menten) reactions.
A typical example of a reaction with a maximal velocity and substrate sensitivity. In an
aggregated pathway model, the Vmax is the genetic component representing total enzyme
concentration, cell number, or tissue mass or protein. The substrate sensitivity constant is
also genetically controlled and may also represent enzyme concentration. The theta value on
the (Km/[S]) function indicates sensitivity to substrate, which is also genetically controlled.
This may represent sensitivity to hormonal or substrate/product allosteric control.
constant becomes smaller and reaction rate would increase. The

representation of insulin (cGlr/cGl) is raised to a theta value that can alter
the sensitivity of the reaction.
Let us explore the genetic elements in this equation. The Vmax
represents the total amount of catalytic activity available, in this case, in the
sum of body adipose tissue (at other levels of aggregation, this may be in a
specific organ, cell or single enzyme or receptor; Figs 14.2, 14.3). This is
controlled genetically, inherited from the parents. The Vmax itself may be
variable, decreased by periods of energy deficit that decrease the total mass
of adipose tissue. The value of this parameter for a population of animals
may be determined in studies combining direct measures of enzyme
292 J.P. McNamara
100.0
90.0
80.0
70.0
Percent bound
60.0
50.0
40.0 Normal
Hi M
30.0
Lo K
20.0
Hi Theta
10.0
0.0 2
8
4
5
1
2.
2.
3.
3.
4.
4.
5.
5.
1.
1.
Concentration of Hormone
Fig. 14.3. Example of genetic control over hormone-receptor binding reactions. A typical
example of a hormone-receptor reaction with a maximal binding of receptor [R] and
dissociation constant [K] in response to hormone [H] concentration. The maximal binding of
hormone receptor is the genetic component, as is the dissociation constant. The hormone
concentration is also genetic, but would be represented with another model showing
hormone synthesis and degradation rates. Changing either [R] or K by genetic selection will
change hormone-receptor binding and, possibly, the end result in flux in the cell (by altering
allosteric control of the reaction given in Fig. 14.2, for example).
activity and measures of total body adipose protein content. The K

variable, substrate sensitivity, is also inherited (but may not have a high
heritability, as the catalytic sensitivity of this enzyme, as for most, is a
function of the molecule, not concentration of the molecule). The
aggregate pathway sensitivity is partly controlled by environmental
influences that alter insulin concentrations. The K for glucose in this
reaction reflects the availability of glucose to provide energy. The equation
also introduces the concept of ‘anabolic hormone’, or insulin, which is
calculated from (cGl/rcGl), such that as glucose concentration increases
against the normal or reference glucose concentration, the effect of glucose
increases, decreasing the denominator and increasing reaction rate. The
sensitivity of insulin to glucose availability may be changed to reflect
genetic differences in insulin function if data are available to justify such a
change.
The genetic elements of any metabolic reaction can be incorporated
into flux control. In Fig. 14.2, equation forms for substrate saturation,
which includes most controlled reactions in metabolic pathways, and in

Fig. 14.3, per cent receptor binding for hormonal control are shown.
Through these equations, genetic elements are integrated into flux
models. Maximal rates and substrate sensitivities are genetically inherited
and, in some cases, have a measurable heritability. Substrate sensitivity is
more a function of the molecule itself than of the number of molecules, but
in an aggregate pathway, changes in substrate sensitivity can be measured
(McNamara, 1998; McNamara and Baldwin, 2000). Binding maxima and
sensitivity also are so controlled. We can also envision the Vmax varying
during the life cycle or by hormones related to environmental or
physiological state. An example is available in the Pettigrew model for the
control of lipid storage in adipose tissue:
dQTsdt = PFaTs + PGlTs –UtsFa (14.3)
UFaTs = vFaTs / (1 + (MFaTs / CFa) + (M6FaTs / CGl)) (14.4)
UTsFa = (vTsFa / (1 + (MTsFa / Clh))) * (Qts**0.67)*
(14.5)
(1–((QsTs/Qts)**thTsFa))
In Eqn 14.3, we see the accumulation of triacylglycerols (Qts) on a daily
basis as a function of synthesis from fatty acids (re-esterification, PFaTs),
de novo lipogenesis (PGlTs, given in Eqn 14.1) and lipolysis (UTsFa). We
saw how lipogenesis can be modelled to include genetic and
environmental control in Eqns 14.1 and 14.2. The same is true for re-
esterification (Eqn 14.4), which is controlled by a maximal velocity (vFaTs),
and a sensitivity constant for fatty acids (MFaTs) as well as for glucose
(M6FaTs). These can be considered to be defined by genetics (presence or
absence of genes; transcription of genes in adipose depots). Obviously,
then the environmental control of gene transcription can also be
modelled, if we can measure the effects of nutrients or hormones on gene
transcription. The control by substrate availability is self-evident.
Triacylglycerol lipolysis (UTsFa, Eqn 14.5) is controlled by a maximal rate,
as well as by ‘lactation hormone’ (Clh), which is a function of a basic
lactation curve:
Clh = MYF *(1.0 + 4.02E – 2* (t – T2) – 7.65E – 4 * (t – T2)** 2)
(14.6)
*(step(T2) – step(T3)) + 0.00001
Equation 14.5 (lipolysis) is completed by scaling to metabolic body size (a
step that, in this author’s opinion, is not warranted, but it was the
prevailing thought of the day). This equation is obviously simplistic and
aggregated, but provides a clear and concise view of the flux. We can
model more control as available data warrant it. Genetic control elements
can be included in simple (or complex) descriptions of flux. A theoretical
model describing upstream regulation (transcription control), and
inclusion of other hormones (catecholamines, thyroid) can be built up from
here.
Lipolysis is controlled by catecholamine binding to beta-adrenergic
receptors. There are data in cell systems or cell free systems on binding
294 J.P. McNamara
kinetics of this system, and there are data in adipose tissue on the response
of cAMP and lipolysis to concentrations of catecholamines (McNamara et
al., 1992; McNeel and Mersmann, 1999 and references therein).
Expanding on Eqn 14.5 may yield:
UTsFa = (vTsFa /(1 + (MTsFa /Clh)))* (Qts**0.67)*
(14.6)
(1((QsTs/Qts)**thTsFa))
VTsFa = v1TsFa*(cAMP/rcAMP) (14.7)
cAMP = KcAMP*[B2RecNE] (14.8)
[B2RecNE] = MB2Rec /(1 +(Kd/[NE]θ)) (14.9)
such that the maximal velocity of lipolysis (VTsFa; let us say this represents
total active hormone sensitive lipase) is a function of cAMP concentration
(cAMP) in relation to a reference concentration (rcAMP); cAMP con-
centration is a function of per cent beta-2-adrenergic receptor bound
([B2RecNE]); and bound receptor is in turn a function of maximal binding
(MB2Rec), dissociation constant (Kd) and norepinephrine concentration
([NE]θ), raised to a power (sigmoidal function, as in Fig. 14.3). Further
homeorhetic control (Bauman and Vernon, 1993; McNamara and Boyd,
1998) on the beta-adrenergic response may be exerted by lactational
hormone to increase per cent bound during lactation:
MB2Rec = MB2Rec * (Chl/rCHl) (14.10)
This treatment is basically the same as that proposed in the cow model of
Baldwin (Baldwin, 1995; Baldwin et al., 1987a,b,c), but with more points of
control exerted on the final pathway flux, as in reality. A point to note is
that in fact in any given condition (nutritional status, day of lactation), the
actual pathway flux described in the new model may be exactly the same as
in the old model, but with a fuller description of control.
Another example of descriptions of control within this model of flux is
that of glucose utilization. It is glucose around which the major regulatory
processes of the body have evolved. The brain and central nervous system
require glucose. Although the mass amount of glucose required by the
brain is small compared to that used by the mammary gland or other
organs, the regulatory mechanisms invoked as glucose availability changes
has a major affect on metabolic rates in other organs. Recent work in mice
shows that altering glucose use by food restriction, or by gene insertion for
proteins such as the insulin dependent glucose transporter (GLUT4) result
in changes in transcription of thousands of genes (Fu et al., 2004).
Carbohydrate in the body which is metabolized for energy (or to make
fat or lactose) eventually is converted to triose phosphates or glucose, or is
used through the same metabolic pathways so, for simplicities sake we can
aggregate a lot of this. So changes in glucose concentration or pool size
(Gl) in the body are summed as:
dQGldt = PAaGl + PAbGl + PPaGl + P6TsFa – UGlTs – U6FaTm
(14.11)
– U6FaTs – UGlCd – UGlGc – U6GlTs – UGlLm – UglTm
We sum the uptake of glucose, gluconeogenesis from amino acids (PAaGl),

absorbed glucose (PAbGl), glycerol from lipolysis (P6TsFa), and subtract
the use of glucose for milk and body fat synthesis (UGlTm, UGlTs); use for
glycerol in TAG (U6FaTm, U6FaTs), oxidation to carbon dioxide (UGlCd),
glycogen (UGlGc), lactose (Lm). In this summative equation, all genetic
effects are included in the equations describing each pathway as
exemplified above (Fig. 14.4).
The use of glucose has several dozen if not hundreds of possible control
points throughout the body. This ‘obvious statement’ quite often gets
forgotten for its simple obviousness, resulting in interpretation of research
results far beyond their true importance. For example, one might study the
effect of glucose use in muscle, and discover a novel pathway or regulatory
point and a ‘big deal’ is made of that. That is fine. However, quite often we
do not then make the connection that the glucose use in the muscle affects
and is affected by every single other use of glucose in the body. For example, the
specific process in the muscle may, in fact, have a major effect on glucose
dynamics, or might in fact be so overwhelmed or attenuated by other
processes in other organs that the true physiological significance is minor.
This becomes truly obvious only when we start to construct models that
must make the connection. An example is that one animal or set of animals
will have genetically controlled different maxima for gluconeogenesis than
others, and this definitely will affect their glucose and amino acid use.
Thus, glucose used is a function of the genetic maxima and substrate
sensitivity, and this is enhanced or attenuated by environment. It is
difficult, perhaps, conceptually as well as quantitatively, to determine the
exact mechanism of this genetic control. It is known that the total number
of udder cells is a correlate to milk production, but certainly does not
account for the majority of genetic variance. Flux control is exerted in
large part by the number of proteins translated (enzyme concentration, or
maximal velocity). Other anatomical or physiological factors also come into
play. A greater vascularity increases the maximal substrate supply. This is
certainly inherited, but may be difficult to assign a parameter value to.
Genetic responses in other tissues to the hormones of lactation also affect
substrate supply.
The Modelling Process: Experimentation, Hypothesis Setting and

Improvement
Unfortunately, this particular model has been relatively ignored by
scientists studying lactational biology of the pig, and I think that ignorance
has limited our progress in this area. There is a lot of biology we still do
not know about the pig, including metabolic regulation in the mammary
gland, muscle and adipose tissue. There are many reasons why more
nutritionists or those studying metabolic control do not work in more
integrated approaches; however, the advances in genomic knowledge and
control of gene transcription are starting to reverse the trend. As we learn
296 J.P. McNamara
0.016
Simulated blood glucose concentration, M

0.014
0.012
0.01
0.008
0.006
0.004
0.002
0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
Day of Lactation
0.8
0.7
Gluconeogenesis, moles amino acid /d
0.6
0.5
0.4
0.3
0.2
0.1
0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
Day of Lactation
Fig. 14.4. Simulations of glucose use in a model of metabolism in the sow. (a) Shows the
concentrations of blood glucose during lactation and (b) shows rates of gluconeogenesis
from amino acids. The different lines indicate alterations in either intake or in demand by
mammary gland (McNamara and Pettigrew, 2002a,b). This demonstrates the behaviour and
sensitivity of the model. Further work will then link glucose concentrations and downstream
endocrine signals to reproductive functions (Fig. 14.6).
more about the genome and regulation of transcription, we are coming

back to why we started to study it in the first place: to learn more about
control of metabolism, nutrient use and practical farm animal efficiency.
Our laboratory conducted a number of challenges of this model over
the last 13 years to help refine it and learn more about metabolic
regulation. They are too extensive to describe here, but they focused on a
few key areas: how do energy intake and energy demand (milk production
driven by litter size) alter use of nutrients in the non-mammary tissues
(McNamara and Pettigrew, 2002a,b)? How do our direct estimates of
parameters of maximal velocity and substrate sensitivity for lipid
metabolism in the adipose tissue match those first proposed in the model
and, if they do not, would altering them in the model improve per-
formance (Parmley et al., 1996)? What parameters can easily be identified
and linked to genetic control, and how then do we go about getting better
estimates of their values (McNamara, 1998; McNamara and Boyd, 1998)?
During lactation, the increase in feed intake and in milk synthesis
causes major increases in metabolic rate in the liver (Pettigrew et al.,
1992a,b; Reynolds et al., 2003). These processes all require an increase in
energy need (usually categorized into increased heat increment in Net
Energy schemes). In the most recent version of the sow model, ATP use is
on demand for the metabolic reactions that occur (McNamara and
Pettigrew, 2002a,b) and the increase in demand in turn requires increased
oxidation (irreversible loss) of a mixture of glucose and fatty acids. If, for
example, there are systematic errors in the model parameters which set
the rate of muscle protein turnover and metabolic rates in the liver (rates
that have genetic components) such that these energetic costs are too low,
then the error will actually show up in an over-accumulation of fat in the
model sow.
The utility of modelling has been recently demonstrated with such an
example in the cow model of Baldwin (1995). Others and I have been
working with this model, challenging its performance with experimental
data and refining the model as we go (McNamara and Baldwin, 2000;
McNamara, 2005; Hanigan et al., 2005). One result obtained in all
situations: the model cow accumulated too much fat. Many measurements
were made to determine possible reasons for this error, and much novel
information on regulatory biology was found. Yet, recently, it has been
proposed that in fact the yield of ATP from electron transport is likely
lower than the three ATP/NADH2 and two ATP/FADH2 incorporated into
the model. If these figures are in fact 2.5 and 1.5, then the demand for
ATP is not met as efficiently as we thought. In fact, changing these
parameter values in the model thus increased the obligatory oxidation of
glucose and fatty acids and then fed back on the uses of glucose, acetate
and fatty acid and the over-accumulation of fat was primarily repaired
(McNamara, 2005). This is an example of the modelling approach working
in a research programme: a properly constructed model founded in
proper biological knowledge can help to point out specific errors in our
knowledge that otherwise would be quite hard to uncover.
298 J.P. McNamara
Using this information, we easily went back to test the effects in the sow
model described herein. We had known that this model also over-
predicted the accumulation of body fat on low-fat, but not high-fat diets
(McNamara and Pettigrew, 2002a,b), and wondered whether correcting
the ATP yield could prevent this in the sow model. Fig. 14.5 shows that the
answer is ‘It depends’. Reducing the ATP yield by 16% corrected the error
on the low fat diet, but then did not describe enough fat on the high fat
diet. This indicates that there are other errors remaining. Certainly some
of these are related to the parameters for glucose use and body fat
synthesis, which are genetically regulated. Construction of more refined
models including genetic regulation will help us define and correct further
errors.
A question that can now be asked is: ‘Can we select for, or by other
means change, the genetic control of body fat synthesis to improve
efficiency of milk synthesis (without reducing the health of the cow)?’
Complex mechanistic dynamic models, integrating ‘genetics’ into ‘kinetics’,
can help us to predict the potential outcomes of such decisions. These
simple examples given here just touch on the complexity of genetic
changes for ‘practical traits’ at the organism level. It is one thing to do a
‘knockout’ to determine what happens if a protein is not made. But it is a
30
Hi F Obs = 29.4
29
28
Body Fat, kg
27
LoFnorm
26 HiFnorm
LoFHiPO Lo F Obs = 26.3
25
HiFHiPO
24
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
Day of Lactation
Fig. 14.5. Effect of altering ATP yield on use of body fat in lactation of sows. If the current
newest estimates of ATP yield from NADH2 and FADH2 of 2.5 and 1.5 are correct, this will
help explain the over-prediction of body fat. The sow model over-predicts the accumulation of
body fat in situations of high fat intake. However, the error does not seem to be in the
stoichiometry of the adipose tissue equations (McNamara and Pettigrew, 2002b). If instead,
the original older estimates of ATP yield were too high, then less glucose and fat would be
oxidized to supply the ATP demand, and more body fat would accumulate. An example of
how a properly constructed model can help to point out errors in knowledge, which can then
be corrected.
different problem to describe and predict what may happen due to more
practical genetic changes. That is our charge, however.
Differences in milk production, body fat and body protein are genetic
and inherited. Rates of amino acid, fatty acid and glucose use in the major
organs are also genetically controlled. Empirical applied data (feed intake,
litter growth) can be collected and compared to a ‘theoretical’ model to ask
such questions as: ‘What would we expect the range of milk production to
be if this number were used as the Vmax for lactose synthesis (or body
protein synthesis)?’ Here again, our quantitative knowledge of
biochemistry allows us to ask such questions. What factors control gene
transcription in the mammary gland? Conceptual and quantitative models
will be of great assistance here to codify our knowledge and set hypotheses.
How About a Model of Reproduction?

In 1992, Pettigrew et al. (1992a) wrote:
The mechanisms connecting the diet to reproductive performance are
presently unknown but may include variations in voluntary feed intake,
digestion, absorption, metabolism of absorbed nutrients, and endocrine effects.
Clear understanding and manipulation of this connection to optimize long-
term sow herd performance requires ability to track, systematically and
quantitatively, dietary effects through the various processes to reproductive
performance. The objective of the present work is to begin to develop the
required systematic, quantitative understanding of that connection.
In the 12 years since that article was published, a large body of work has
been done on the control of reproduction in sows, and even more in dairy
cattle. Yet no one, to my knowledge, has published a true mechanistic
model of nutrient metabolism and reproduction. At the same time, several
companies are promising great results in reproductive improvement with
various nutritional products or management. So, if the results are so
predictable, why don’t we have a model? It should have been easy! I have
alluded to several of the political, scientific and monetary reasons in the
paragraphs above. Those won’t be solved here or soon. Thus, I would like
to present the possibility of the beginnings of such a model, to continue the
work proposed by my colleagues long ago.
We know growth and ovulation of viable eggs is under the regulation
of follicle stimulating hormone (FSH) and luteinizing hormone (LH).
Further, much is understood about the relationship of glucose flux to
ovulation, effects of fats and IGF on the ovary, and potential negative
effects of ammonia or urea on the uterine environment (Garcia-Bojalil et
al., 1998; Staples et al., 1990; Mattos et al., 2002; Westwood et al., 2002;
Farmer and Palin, 2005). A flow diagram connecting the essential elements
of the Pettigrew sow model and Baldwin cow model with reproduction is
proposed in Fig. 14.6. Examples of equations necessary to describe part of
the system may include control of FSH and LH secretion and degradation,
and their effects on follicular development and ovulation.
300 J.P. McNamara
Hypothalamus
Leptin
Oestrogen
Adipose Pituitary
Glucose
IGF1 FSH LH Progesterone
Fatty Acids Dominant

Follicle
(CLA) Developing
follicles
NH3 Ovulated
Amino Egg(s)
Acids
Fetus(es)
Implantation Corpus
Body Luteum
Protein
Fig. 14.6. Description of a model of nutrient control of reproduction in the sow. Flux rates of
several nutrients, and secondary hormonal responses, have direct effects on follicular
development, ovulation and uterine environment. Why do we not yet have a mechanistic
model of nutrient/reproduction interactions, given the wealth of information we have? Here is
presented a flow diagram of one such model. Equation forms will need to be derived and
parameters estimated. The model then becomes a framework for identifying inadequacies in
our knowledge.
dLH = secLH – degLH (14.12)

secLH = pLH * kGlLH * kLpLH * kEsLH (14.13)
Deg LH = KdegLH * [LH] (14.14)
sFSH = pFSH * kGlFSH *kPGFSH – KEsFSH (14.15)
Deg FSH = KdFSH * [FSH] (14.16)
V RFoll to DevFoll = VReDe * (KfollFSH * [FSH]) (14.17)
V DevFoll to Domfoll = VDvDo * (KfollFSH * [FSH]) (14.18)
Equation 14.12 shows the change in LH as a function of secretion and
degradation, secLH (Eqn 14.13) is a function of synthesis, glucose,
luteinizing hormone releasing hormone and oestrogen. Degradation of
LH (Eqn 14.14) is some constant (but it does not have to be a constant), the
same is true for FSH (Eqns 14.15 and 14.16). Eqns 14.17 and 14.18
describe the development of developing follicles from resting follicles and
on to dominant follicles. I would contend that there is enough data in the
literature to begin to construct these equation forms and parameter values.
Further data would need to be found or collected to relate glucose
concentrations to these functions, and on to all the equations depicted in

the flow diagram of Fig. 14.6. From there we can begin more specific
experiments into both applied models and models of genetic (genomic,
transcriptional, translational) control of these processes. Why not?
Summary and Implications
When one stands on the shoulders of giants, one is humbled and grateful
that the horizon is closer in view. While the objective of eventually
understanding everything that happens in an animal may be far beyond
the horizon, progress is being made and our true purpose of feeding the
world efficiently throughout many geographical areas and for many
centuries is becoming closer. A quantitative and systems approach is the
only way to continue upon this path. Reductionistic research, especially
into the true functionality of the genome and physiological control of
metabolism will be essential to improvement. In addition, continued efforts
in simple, economical and effective chemical methodology for defining
feedstuffs is a must. But over-riding all, is the integration of knowledge,
data, and concepts, into an organized representation of reality: a model
that will achieve our goal. Research can help us overcome politics and
other human frailties with a system that can ensure adequate food
supply for all. That is our ultimate goal, and societal, industrial and
governmental support for the modelling approach to research is critical to
achieve it.
References
Baldwin, R.L. (1995) Modelling Ruminant Digestion and Metabolism. Chapman & Hall,
New York, pp. 469–518.
Baldwin, R.L., France, J. and Gill, M. (1987a) Metabolism of the lactating cow. I.
Animal elements of a mechanistic model. Journal of Dairy Research 54, 74–105.
Baldwin, R.L., France, J. and Gill, M. (1987b) Metabolism of the lactating cow. II.
Rumen elements of a mechanistic model. Journal of Dairy Research 54, 106–132.
Baldwin, R.L., France, J., Beever, D.E., Gill, M. and Thornley, J.H.M. (1987c)
Metabolism of the lactating cow. III. Properties of mechanistic models suitable
for evaluation of energetic relationships and factors involved in the partition of
nutrients. Journal of Dairy Research 54, 133–145.
Bastianelli, D., Sauvant, D. and Rerat, A. (1996) Mathematical modeling of
digestion and nutrient absorption in pigs. Journal of Animal Science 74,
1873–1887.
Bauman, D.E and Vernon, R.G. (1993) Effects of exogenous somatotropin on
lactation. Annual Review of Nutrition 13, 437–461.
Simulation of energy and amino acid utilization in the pig. Research and
Bosch, F., Pujol, A. and Valera, A. (1999) Transgenic mice in the analysis of
metabolic regulation. Annual Review of Nutrition 18, 207–232.
302 J.P. McNamara
Collins, J.C., Harris, A., Linsenmeier, R.A. and Olds, S. (2005) Testing hypotheses
of learning at the VaNTH bioengineering education center: applications to
nutrition and metabolism. Chapter 14. In Hargrove, J.L. and Berdanier, C.D.
(eds) Mathematical Modeling in Nutrition and Toxicology. Mathematical Biology
Press, Athens, Georgia, pp. 235–244.
Daniel, P.B., Walker, W.H. and Habener, J.F. (1999) Cyclic AMP signaling and gene
regulation. Annual Review of Nutrition 18, 353–383.
Drackley, J.K. (1999) Biology of dairy cows during the transition period: the final
frontier. Journal of Dairy Science 82, 2259–2273.
Farmer, C. and Palin, M.F. (2005) Exogenous prolactin stimulates mammary
development and alters expression of prolactin-related genes in prepubertal
gilts. Journal of Animal Science 83, 825–832.
Fu, C., Xi, L., Wu, Y., McCarter, R., Richardson, A., Hickey, M. and Han, E.S.
(2004) Hepatic genes altered in expression by food restriction are not
influenced by the low plasma glucose level in young male GLUT4 transgenic
mice. Journal of Nutrition 134, 2965–2974.
Garcia-Bojalil, C.M., Staples, C.R., Risco, C.A., Savio, J.D. and Thatcher, W.W.
(1998) Protein degradability and calcium salts of long-chain fatty acids in the
diets of lactating dairy cows: reproductive responses. Journal of Dairy Science 81,
1385–1395.
Girard, J., Ferre, P. and Foufelle, F. (1997) Mechanism by which carbohydrates
regulate expression of genes for glycolytic and lipogenic enzymes. Annual
Review of Nutrition 17, 325–352.
Hanigan, M.D., Bateman, H.G., Fadel, J.G. and McNamara, J.P. (2005) Metabolic
models of ruminant metabolism: recent improvements and current status.
Journal of Dairy Science 88 (in press).
Hargrove, J.L. (2004) Kinetic basis for modeling gene expression: a historical
overview. In: Hargrove, J.L. and Berdanier, C.D. (eds) Mathematical Modeling
in Nutrition and Toxicology. Mathematical Biology Press, Athens, Georgia,
pp. 265–280.
Hausman, G.J. and Hausman, D.B. (1993) Endocrine regulation of porcine adipose
tissue development: cellular and metabolic aspects. In: Hollis, G.R. (ed.) Growth
of the Pig. CAB International, Wallingford, UK, pp. 49–73.
Hirooka, H., de Koning, D.J., Harlizius, B., van Arendonk, J.A., Rattink, A.P.,
Groenen, M.A., Brascamp, E.W. and Bovenhuis, H. (2001) A whole-genome
scan for quantitative trait loci affecting teat number in pigs. Journal of Animal
Science 79, 2320–2326.
Hurley, W.L., Wang, H., Bryson, J.M. and Shennan. D.B. (2000) Lysine uptake by
mammary gland tissue from lactating sows. Journal of Animal Science 78,
391–395.
Jiang, Z.H. and Gibson, J.P. (1999a) Genetic polymorphisms in the leptin gene and
their association with fatness in four pig breeds. Mammalian Genome 10,
191–193.
Jiang, Z.H. and Gibson, J.P. (1999b) Bi-PASA genotyping of a new polymorphism in
the APOB gene shows no evidence for an association with fatness in pigs.
Animal Genetics 30, 54–56.
Kanis, E., DeGreef, K.H., Hiemstra, A. and Van Arendonk, J.A.M. (2005) Breeding
for societally important traits in pigs. Journal of Animal Science 83, 948–957.
Lofgreen, G.P. and Garrett, W.N. (1968) A system for expressing net energy
requirements and feed values for growing and finishing beef cattle. Journal of
Lovatto, P.A. and Sauvant, D. (2003) Modeling homeorhetic and homeostatic

controls of pig growth. Journal of Animal Science 81, 683–696.
Mattos, R., Staples, C.R., Williams, J., Amorocho, A., McGuire, M.A. and Thatcher,
W.W. (2002) Uterine, ovarian, and production responses of lactating dairy
cows to increasing dietary concentrations of menhaden fish meal. Journal of
Dairy Science 85, 755–764.
McNamara, J.P. (1998) Interaction of glucose and amino acid metabolism in
lactating sows: estimating internal parameters of a model of metabolism. In:
McCracken, K.J., Unsworth, E.F. and Wylie, A.R.G. (eds) Energy Metabolism of
Farm Animals: Proceedings of the 14th Symposium. CAB International, Wallingford,
UK, pp. 32–35.
McNamara, J.P. (2000) Integrating the effects of genotype and nutrition on
utilization of body reserves during lactation of dairy cattle. In: Cronje, P.B.
(ed.) Ruminant Physiology: Digestion, Metabolism, Growth and Reproduction. CAB
McNamara, J.P. (2003) A perspective on research and future of metabolic models of
farm animals. In: Souffrant, W.B. and Metges, C.C. (eds) Progress in Research on
Energy and Protein Metabolism. EAAP Publication No.109, pp. 99–112.
McNamara, J.P. (2004) Research, improvement and application of mechanistic,
biochemical, dynamic models of metabolism in lactating dairy cattle. Animal
Feed Science and Technology 112, 155–176.
McNamara, J.P. (2005) Research, improvement and application of mechanistic,
biochemical, dynamic models: from genetics to kinetics. In: Hargrove, J.L. and
Berdanier, C.D. (eds) Mathematical Modeling in Nutrition and Toxicology.
Mathematical Biology Press, Athens, Georgia, pp. 87–110.
McNamara, J.P. and Baldwin, R.L. (2000) Estimation of parameters describing lipid
metabolism in lactation: challenge of existing knowledge described in a model
of metabolism. Journal of Dairy Science 83, 128–143.
McNamara, J.P. and Boyd, D.E. (1998) Hormones as quantitative controllers. In:
Kyriazakis, I. (ed.) A Quantitative Biology of the Pig. CAB International,
McNamara, J.P. and Martin, R.J. (1982) Muscle and adipose tissue lipoprotein
lipase in fetal and neonatal swine as affected by genetic selection for high or
low backfat. Journal of Animal Science 55, 1057–1061.
McNamara, J.P. and Pettigrew, J.E. (1994) Lipid metabolism in lactating swine
consuming different amounts of energy: a challenge of current knowledge as
represented in a model of sow metabolism. In: Aguilera, J.F. (ed.) Energy
Metabolism of Farm Animals: Proceedings of the 13th Symposium. CAB International,
McNamara, J.P. and Pettigrew, J.E. (2002a) Protein and energy intake in lactating
sows. 1: Effects on milk production and body composition. Journal of Animal
Science 80, 2442–2451.
McNamara, J.P. and Pettigrew, J.E. (2002b). Protein and energy intake in lactating
sows. 2. Challenging parameters of a model of metabolism. Journal of Animal
Science 80, 2452–2460.
McNamara, J.P., Dehoff, M.H., Bazer, F.W. and Collier, R.J. (1985) Adipose tissue
fatty acid metabolism changes during pregnancy in swine. Journal of Animal
Science 61, 410–415.
McNamara, J.P., Becker-Khaleel, B. and Parmley, K. (1992) Quantitative relationships
between cyclic adenosine-3’,5’-monophosphate and lipolysis in adipose tissue
during the peripartum period. Journal of Dairy Science 75, 1901–1910.
304 J.P. McNamara
McNamara, J.P., Phillips, G.J., Citron, T.L. and Sage, J. S. (2006) Challenge and
improvement of a model of dairy cattle metabolism to describe changes in
metabolism in cows in early lactation. In: International Workshop on Modelling
Nutrient Utilization. Proceedings of 6th International Workshop on Modelling in Farm
Animals. CAB International, Wallingford, UK (in press).
McNeel, R.L. and Mersmann, H.J. (1999) Distribution and quantification of beta1-,
beta2-, and beta3-adrenergic receptor subtype transcripts in porcine tissues.
Mohamed-Ali, V., Pinkney, J.H. and Coppack, S.W. (1998) Adipose tissue as an
endocrine and paracrine organ. International Journal of Obesity and Related
Metabolic Disorders 22, 1145–1158.
National Research Council (1998) Nutrient Requirements of Swine, 10th revised edn.
National Academy Press, Washington, DC.
National Research Council (2001) Nutrient Requirements of Dairy Cattle, 7th revised
edn. National Academy Press, Washington, DC.
Overton, T.R. and Waldron, M.R. (2004) Nutritional management of transition
dairy cows: strategies to optimize metabolic health. Journal of Dairy Science 87,
E105–E119.
Parmley, K.L.S. and McNamara, J.P. (1996) Lipid metabolism in adipose tissue of
pigs fed varying amounts of energy. Journal of Nutrition 126, 1644–1656.
Pettigrew, J.E., Gill, M., France, J. and Close, W.H. (1992a) A mathematical
integration of energy and amino acid metabolism. Journal of Animal Science 70,
3742–3761.
Pettigrew, J.E., Gill, M., France, J. and Close, W.H. (1992b) Evaluation of a
mathematical model of lactating sow metabolism. Journal of Animal Science 70,
3762–3773.
Phillips, G.J., Citron, T.L., Sage, J.S., Cummins, K.A., Cecava, M.J. and McNamara,
J.P. (2003) Adaptations in body muscle and fat in transition dairy cattle fed
differing amounts of protein and methionine hydroxy analog. Journal of Dairy
Science 86, 3634–3647.
Quintanilla, R., Milan, D. and Bidanel, J.P. (2002) A further look at quantitative
trait loci affecting growth and fatness in a cross between Meishan and Large
White pig populations. Genetic Selection and Evolution 34, 193–210.
Reynolds, C.K., Aikman, P.C., Lupoli, B., Humphries, D.J. and Beever, D.E. (2003)
Splanchnic metabolism of dairy cows during the transition from late gestation
through early lactation. Journal of Dairy Science 86, 1201–1217.
Roehe, R. (1999) Genetic determination of individual birth weight and its
association with sow productivity traits using Bayesian analyses. Journal of
Staples, C.R., Thatcher, W.W. and Clark, J.H. (1990) Relationship between ovarian
activity and energy status during the early postpartum period of high
producing dairy cows. Journal of Dairy Science 73, 938–947.
Tepperman, J.and Tepperman, H.M. (1970) Gluconeogenesis, lipogenesis and the
Sherringtonian metaphor. Federation Proceedings 20, 1284–1293.
Westwood, T., Lean, I.J. and Garvin, J.K. (2002) Factors influencing fertility of
Holstein dairy cows: a multivariate description. Journal of Dairy Science 85,
3225–3237.
15 The Place of Models in the New
Technologies of Production
Systems
D.M. GREEN1 AND D.J. PARSONS2
1Universityof Oxford, Department of Zoology, South Parks Road, Oxford,
OX1 3PS, UK; 2Cranfield University, Silsoe, Bedford, MK45 4HS, UK.
darren.green@zoo.ox.ac.uk
Introduction
The problem
Farmers aim to produce livestock within the tight targets imposed on them
by their customers (increasingly, supermarkets) and to do so at a profit.
Both the consumer and the government, however, are increasingly being
influenced by concerns other than price alone, including issues such as
environmental damage and pollution, human and animal health and
animal welfare (Fox and Sachs, 2003). The role of science and technology
in agriculture is to enable the farmer both to increase profit and to address
the further concerns of the consumer.
Integrated management systems
Most parts of the animal production process have been subject to

automation. Equipment is available to monitor the animals, to feed them,
and to control their environment. Integrated management systems (IMS)
represent an attempt not only to combine these various subsystems but
also to delegate control of them to an automatic controller (Frost, 2001).
The potential benefits of IMS are multiple (Whittemore et al., 2001a).
Precise control, while freeing staff time, will allow precise formulation
and rationing of the diet to improve economic and environmental
efficiency of livestock production, with a product of high quality giving a
high margin, while also addressing health, welfare and environmental
concerns.

306 D.M. Green and D.J. Parsons
At the core of an IMS lies a model, which provides two benefits. First, it
allows forward predictions to be made of the impact of management
decisions: the best course of action can then be automatically determined
by the system. Secondly, IMS involves considerable automated monitoring
of livestock and its environment therefore the model provides a powerful
diagnostic tool to detect when livestock performance deviates from that
expected and to alert the stockperson to the potential problem.
There are few examples in the literature of fully functioning IMS
systems, but various authors have made steps in this direction. Of these,
automated feeding and weighing systems will be discussed below.
Halachmi et al. (1998) discuss an automatic feeding system for cows, which
allows a substantial collection of data, but this is not integrated with a
model or control system. Pietersma et al. (2001) developed a decision
support system to analyse lactation data for cattle and to examine
individual performance, but again without a model or control system.
Many models of pig and poultry nutrition exist, ranging in scope from
those describing the nutrition of a single animal up to those describing the
whole farm and its economy. But aside from the examples described
below, these models have not yet been integrated into on-line livestock
production systems.
The general approach shared by the IMS systems described below is
that of model-based predictive control (IMS Pigs1; IMS Poultry2; Aerts et al.,
2003a,b). This is a two-stage process: first, continuously collected data are
fed back into an adaptive (growth) model; second, forward predictions of
this model are used to identify the management regime required to satisfy
the control objectives. How the various components of a generic IMS
system interact is shown in Fig. 15.1. The goal of IMS is to produce a
system that is entirely on-line and operates in real-time without human
intervention (in normal operation: problems would require intervention
when detected). The systems currently in development achieve this to
varying degrees.
IMS brings together many of the subjects that have been discussed
elsewhere in this book. Different approaches can be used to model feed
intake, nutrient and energy usage, growth and feed intake: approaches
from detailed modelling at the metabolic level (e.g. Green and
Whittemore, 2003) through to data-based ‘black box’ approaches (e.g.
Aerts et al., 2003a,b). Regardless of the approach used, in practice no
animal will ever perform exactly as predicted by a model, due to
consequences of the social, disease and thermal environment (see
Wellock et al., Chapter 4, this volume) and genetic effects. These
unknown parameters must be estimated from the data before the model
can be used for accurate prediction. Identifying these parameters is a
form of model inversion problem (see Doeschl-Wilson et al., Chapter 9,
this volume).
This chapter addresses the components and concepts of IMS systems.
First, the various pressures and objectives of IMS are addressed, then the
equipment and data sources available on the farm. The models themselves
Models in the New Technologies of Production Systems 307
Previous
status
Diet and
Run model
environment
Predict
status now
Review and
amend model
Measure Diet and
status now environment
Amend diet
Run model
and environment
Predict Future
future status control targets
Fig. 15.1. Flowchart showing design of an abstract integrated management system.
are then discussed: modelling strategies used for IMS, methods for
optimizing models in response to the collected data, and how model
predictions can be used to make control decisions. Finally, some IMS
systems already developed and in development are discussed, before
considering likely future developments.
Market Forces and Control Objectives
The IMS approach is adaptable for a variety of objectives, either alone or

in combination. These can be categorized as follows:
1. Economic objectives. In most cases, the primary objectives of producers
will be to maximize profit. The best course of action here may be to rear
livestock as quickly as possible, although this is not necessarily the case. For
example, faster growth might be at the expense of meat quality, and if this
adversely affects the price achieved then slower growth may be desired. A
full model of not only growth and nutrition, but also of economics at the
farm level, will be of benefit here in choosing the optimal strategy. The
market may impose constraints on the farmer: animals at or below a given
fatness may be required, or delivery of animals of a specified size at a
specified date. Multiple routes to profit may exist which an IMS system
could identify: e.g. either, on the one hand, high levels of production with
low margin per animal sold or, on the other, a lower level of production of
meat of higher value.
2. Environmental objectives. In England, over 70% of nitrates and 40% of
phosphates in water are derived from agricultural land (DEFRA, 2002).
The increasing concern within the EU has led to the designation of nitrate
vulnerable zones and legislation to minimize nitrogen emissions. IMS will
be able to assist here as emissions targets, along with production targets,
can be embedded in the systems. A typical target might be to minimize
nitrogen emissions per animal produced. A model is required to solve this
problem: feed excess dietary protein and growth will be fast, but wasteful
of nitrogen; feed too little dietary protein and growth will be slow, with the
animal contributing to pollution over a longer timescale.
3. Health and welfare objectives. It is not anticipated that automated
systems will ever replace the experienced stockperson. However, where
time is limited, the IMS system will provide a useful extra tool to alert the
stockperson to health and welfare issues. Indeed, where animals are
continuously monitored in terms of growth rate and feed intake, an IMS
system may alert the farmer to a problem before any clinical signs appear.
For example, an animal with low intake and with both low predicted and
low observed growth would be categorized as suffering from low appetite.
Alternatively, an animal with normal intake and normal predicted growth,
but low observed growth (i.e. a high feed conversion ratio) would be
designated a sick animal.
Multiple objectives. The optimum for control is much more difficult to
determine where there are multiple control objectives. A discussion of the
complexities of multiple criteria decision making is beyond the scope of this
chapter. It may, however, be simpler to consider additional objectives as
constraints on the system; this may be appropriate where, for example, there
are government-imposed limits on emissions associated with production.
The IMS systems IMS Pigs and IMS Poultry and that of Aerts et al.
(2003a,b) were not equipped with full economic models of production,
having specific growth targets already set, in order to test the ability of
control to achieve such targets. For the poultry systems, the target was
weight at a particular age; for pigs, as well as growth rate or final size, an
additional objective was specified: level of fatness at the end of the trial.
This is an important target in a species where payment by the
slaughterhouse may be based on carcass quality.
Equipment
An important step towards integrated management systems has been the

growth in sensor systems to measure performance and related data (Frost
et al., 1997), and in particular, a growth in affordable sensor systems.
Without a commercially viable means of collecting data with sufficient
frequency, accuracy, and at a suitable level (individual, pen, etc.), the
affordability of modern computer systems is of little benefit. For both

sensor systems and control systems (environmental controllers, feeders,
etc.), there are the following considerations.
1. Cost. The necessity for equipment to be affordable has already been
mentioned. This applies not only to installation costs but also to operating
costs in terms of both money and personnel time.
2. Maintenance. Maintenance requirements must be low, especially where
a large number of pieces of equipment are installed. Devices with moving
parts tend to become clogged with dust and require cleaning and
recalibration. Devices at ground level will be soiled and require cleaning.
In the case of pigs, there is the likelihood that any novel object within
reach will attract their special attention.
3. Invasiveness. Equipment must be non-invasive, and not cause
deleterious change to animal behaviour. Any requirement for frequent
maintenance will, of course, lead to more animal disturbance. Addressing
both of these last two points, a device that can operate remotely from the
animal is of more benefit than one in contact with the animal.
All of the above criteria are satisfied by the recently developed ‘Visual
Image Analysis’ (VIA) system (Marchant et al., 1999; Schofield et al., 1999),
which is commercially available [Osborne (Europe) Ltd]. This novel system
allows for measurement of the plan-view area of an (indoor housed) pig, by
video camera mounted overhead. These cameras are inexpensive
compared with computer hardware and software costs, easy to install and
calibrate, and easy to maintain, since maintenance extends only to lens
cleaning, there being no moving parts. A number of cameras can share one
computer. If pigs are fitted with radio-frequency transponders, then
individual pigs can be measured; without these, estimates of pen-average
pig weight can be obtained.
A series of images is continuously collected while an animal is in view.
From these area and length measurements (Fig. 15.2), live weight can be
determined with accuracy similar to that of weigh scales (White et al.,
2004), and there is tantalizing evidence showing that other useful indices
of pig shape and size, correlated with composition and conformation, can
be measured too (Doeschl et al., 2004; Doeschl-Wilson et al., 2005). Live
weight itself is, in effect, a proxy for ‘amount of meat’: the potential of
VIA is that it may measure this directly, rather than doing so via live
weight.
VIA offers promise as a means of monitoring size for animals with a
clear, relatively invariable outline; similar approaches have been used for
fish (Tillett et al., 2000), and work is also in progress to determine the
three-dimensional shape of live pigs (Wu et al., 2004). However, this
requires camera equipment that is unlikely to become affordable for
commercial use and the placement of cameras at pig level is also
problematic. Fur and feathers provide more of a challenge for VIA systems
because the animals can present more variety of apparent sizes to the
camera. Nevertheless, progress has been made on using the VIA approach
Fig. 15.2. Screenshot from a running VIA system showing identification of three body regions.
for poultry (de Wet et al., 2003). VIA analysis of pigs from alternative
viewing angles (side and rear views) is also currently under investigation.
Such data may provide measurement of condition, which is particularly
important for sows (Schofield, 2004): but note again the concern over
camera placement.
Automatic (conventional) weighing systems have long been available
for both pigs and poultry. For example, Williams et al. (1996) report on the
effectiveness of weighing pigs through automatic forelegs-only weighing
platforms associated with feeders. For pigs, electronic tagging allows
collection of weight data for individual animals, whereas for poultry,
average weights of the birds visiting the weigher are obtained.
Pig fatness is generally determined by backfat depth at the P2 site (in
the UK). Though such measurements would be valuable input into IMS
systems for pigs, providing an added dimension of data for model
adaptation, its manual measurement can cause distress and is labour
intensive. Ultrasonic P2 measurement remains currently the most practical
means of easily determining carcass composition, amongst a field including
ultrasound (Newcom et al., 2002) and alternatives such as CT scanning
(Kolstad, 2001) and MRI (Mitchell et al., 2001), whose use is restricted to
research.
Feeding equipment is at once both a potential sensor and a controller.
Systems exist for controlling and measuring feed intake and blend for both
pigs and poultry (Filmer, 2001; Ellis and Hyun, 2005), though whether
feeding other than ad libitum is possible will depend upon housing
conditions.
Models and Model Adaptation
Modelling strategies
Modelling represents an attempt to abstract important elements of reality,

in order to study the behaviour of a system and make predictions. Any
such model, however, cannot describe every detail of the system.
Depending upon how much a priori information is encompassed by the
model, two extremes of modelling approach exist: empirical and
mechanistic. The empirical model abstracts from the observables. This is a
simple approach, requiring few assumptions as to the nature of the system,
though data must be collected in order to build the model. Such models
are generally quick to compute, and so can be used in time-critical
situations. However, they cannot safely be used to make predictions outside
the range of data that they were built from.
At the other end of the scale are mechanistic models. These models are
more complex, based on more assumptions regarding the underlying
system and then tested against the observables. They are slower to
compute. However, their predictive power is greater and not so limited by
the range of data already collected, though model predictions must always
be used with caution, and more so with highly complex models. They can
also be used predictively early in a trial, where an empirical model could
not, as insufficient data would be collected up to that point.
A second modelling distinction can be made between stochastic and
deterministic models: deterministic models always give the same results
each time the model is run; stochastic models give a range of likely outputs
from each set of inputs. Deterministic models are quick to run and
repeatable, but only stochastic models give the user not only a prediction,
but also a level of confidence in that prediction. Stochastic models are
useful where it is necessary to model populations of animals: the variation
in weight around the mean at the end of a trial may be of as much interest
as agreement between the mean and the target weight. For example, the
value of carcasses might be considerably reduced above a certain level of
fatness. In this case, targeting just under this level of fatness will be
acceptable where the variation in fatness is predicted to be small, but
unacceptable where large.
The literature concerning pig and poultry growth and nutrition
models is extensive, and a full review of the different approaches used is
beyond the scope of this chapter. It is necessary, however, to draw attention
to those models, and their sources, that have been used in the
development of the IMS systems exampled here. In modelling the growing
pig (IMS Pigs), Green and Whittemore (2003, 2005) used a detailed
mechanistic approach, considering the growth and nutrient use of pigs
with different genetic merit and in social, thermal, and disease
environments with different levels of challenge. The justification for this
approach was largely drawn from the reviews of Whittemore et al.
(2001b,c,d). A mechanistic approach was chosen in order to allow the IMS
system to make control decisions at the beginning of the trials, before data
had been collected.
The IMS Poultry programme (Frost et al., 2003; Stacey et al., 2004) uses
a semi-mechanistic model of poultry growth: a compromise between a
complex, slow, mechanistic model, with extensive a priori assumptions
concerning growth built in, and a quick-running empirical model, which
would be of limited use early in a trial. This model was built from the
principles established by Emmans (1981, 1987, 1989, 1994), Emmans and
Fisher (1986) and Hancock et al. (1995). Also for poultry, Aerts et al.
(2003a) chose an empirical approach, and used a recursive linear
regression relationship to relate cumulative feed intake to mass, using a
limited number of past measurements to predict future weight; they found
a time-window of 5 days to give good forward prediction (Aerts et al.,
2003b).
Model adaptation
Environmental, health, and genetic effects will mean that at run time, the
model will produce output that deviates to a greater or lesser degree from
the observed data. Therefore, the model is continuously updated by
reparameterization from these data. Model parameterization can be
considered as a function minimization problem, where the function to be
minimized is a loss function describing the poorness of fit of the model;
minimization by least squares or negative log-likelihood are the two main
approaches here. Strategies of function minimization are many and varied
(Press et al., 1992), and which is the most appropriate depends upon the
nature of the model response surface (in however many dimensions there
are parameters to be optimized). The surface can have single or multiple
minima, be continuous or discontinuous, differentiable or nondifferentiable
(Fig. 15.3). The most simple and foolproof method is the grid search (in one
dimension, linear search): this method searches through the whole of
parameter space, with a given precision, to find the minimum value.
However, such a brute force approach requires an extraordinarily large
amount of processing time, especially where the number of dimensions
(model parameters) is not small.
Several more ‘intelligent’ approaches exist, which can be divided
generally into methods that only require evaluation of the model/function,
and those that in addition require evaluation of the gradient (Press et al.,
1992). The latter methods are less appropriate for models where output
shows plateaux and discontinuities; such output is possibly more likely
encountered with complex mechanistic models. Gradient-following
approaches include the Newton-Raphson and quasi-Newton methods (Gill et
al., 1981), as well as the Marquardt-Levenberg algorithm (Marquardt, 1963).
The revised simplex method (Nelder and Mead, 1965) does not require
calculation of gradients, but does follow the descent of the model response
surface and so can become trapped at local minima. The methods of
Output variable
Input variable
Fig. 15.3. Smooth and unsmooth model response surfaces.
simulated annealing (Kirkpatrick et al., 1983) and genetic algorithms (Holland,

1975; Goldberg, 1989) both have the benefit of being able to escape from
local minima. However, they tend to take longer to compute. These two
methods are stochastic, in that a different solution will be obtained each
time the algorithm is run. The amount of variation in the parameters
obtained in repeated runs will provide some estimate of the parameter
error. A comparison of the different methods of function minimization is
shown in Table 15.1.
IMS Poultry fitted a single parameter to their model of broiler growth
at run time using a linear search algorithm. This had the advantage of
always finding the minimum, and yet at an acceptable speed. With a larger
number of model parameters fitted (two), IMS Pigs used a variant of the
Table 15.1. Comparison of methods of function minimization or optimization.

Risk of Can fit
Inherently Number of local non-linear
bounded Stochastic parameters minima functions Speed
Linear/grid search yes no low low yes low
Quasi-Newton/
Newton-Raphson/ no no medium high yes high
Marquardt-Levenburg
Revised simplex no no medium high yes high
Simplex (linear yes no medium low no high
programming)
Genetic algorithm yes yes high medium yes medium
Simulated annealing no yes high medium yes medium
revised simplex approach, amending the method to include the

specification of sensible constraints on the model parameters (Parsons,
1992). For this particular problem, with a small but not singular number of
parameters, and a complex model response surface, it was found to be
more robust than the quasi-Newton approach, and faster than the genetic
algorithm, offering a compromise between high speed and high robustness
(Parsons et al., 2004).
Parameter selection for model adaptation
For data-based, empirical modelling approaches, model building and model

parameterization can be equivalent. The model itself may be as simple as a
set of linear equations, though other strategies, such as the use of neural
networks, have been proposed. The interpretation of these models,
however, can become difficult: though externally simple, they can be
internally very complex, and as such not easily understandable. For this
reason, such approaches have been described as ‘black box’ models. The
dangers of using predictions from these models outside the range of the
data collected are greater: this is essentially a problem of model over-fitting.
The more transparent approach used by Aerts et al. (2003a,b) to model
growth response, also empirical, was conceptually more easy to understand,
and more robust. These authors used recursive linear regression to describe
the relationship between growth and cumulative feed intake.
The problem of model over-fitting is not limited to empirical models;
the same issue arises when fitting the parameters of a mechanistic model.
Where too many parameters are fitted, parameter estimates can become
correlated. The model will fit the current data set well; but model
predictions can be unstable outside the possibly narrow range of data used
for parameterization, producing biologically unfeasible output. Under
these conditions, it is as well to combine the correlated parameters into a
single parameter and fit this new parameter.
In the IMS Poultry project (Frost et al., 2003; Stacey et al., 2004), a
single parameter was chosen for adaptation. This parameter controlled the
combined efficiencies of utilization of protein and energy; this was found
more robust than fitting the two efficiencies simultaneously as two
parameters. The growth model of Aerts et al. (2003a,b) and IMS Pigs both
fitted two parameters. In the former case, these represented the slope and
intercepts of the simple linear model. In the latter, mechanistic model, a
number of parameters were available and were considered. Two
parameters existed controlling maximum rate of protein retention and two
controlling efficiency of energy and protein use (both genetic and disease
effects) (Green and Whittemore, 2005). However, a good fit was found,
without generating biologically unreasonable model output, by fitting one
from each pair. Since weight data from the VIA system was available for
individually identified pigs, model parameterization in IMS Pigs was
carried out on the individual animals.
Stacey et al. (2004) draw attention to the question of how much data it
is necessary to collect before model adaptation should be performed.
Though this period depends critically on the confidence level used, a
certain minimal period must elapse before statistically significant changes
in animal state can be detected. Kearney et al. (2004), on automatic
weighing of cattle, recommend collection of data for up to 2 months in
order to quantify growth rate. Stacey et al. (2004) decided upon a period of
14 days for growing chickens; this was chosen as a compromise between
being able to respond correctly and promptly to differences in animal
growth rate, and responding falsely to measurement errors. For pigs,
White et al. (2004) noted that for platform weighers, a period of 4–13 days
is required before changes in pig state can be detected with 95%
confidence; and for VIA systems, 8–10 days. Precisely quantifying growth
rate would take longer. Aerts et al. (2003a,b) chose a window of 5 days of
collected data on which to base forward predictions.
In the IMS Poultry, IMS Pigs and Aerts et al. (2003a,b) growth models, a
single dimension of data, live weight, was available for model optimization.
With VIA (White et al., 2004) or with automatic fat depth measurements
(Frost et al., 2004), there is the potential of extra dimensions of data in
future systems. In this case, there will be greater potential to include more
parameters and preserve model robustness.
Control
Controllable elements
Whittemore (1998) draws attention to four aspects of the production

process that are subject to control.
1. The feed intake ration.
2. The feed intake blend, particularly with respect to the ratio of energy to
amino acids.
3. The choice of the point of slaughter.
4. The choice of breed, its lean:fat ratio, composition, and conformation.
These aspects are as applicable for poultry as they are for pigs. To these
can be added:
5. The social environment of the animal; group size, mixing and lighting.
6. The thermal environment; temperature, floor materials, humidity,
ventilation, etc.
Objectives
An adapted model, once generated, can be used to make forward pre-

dictions. The difference between these forward predictions of production
outputs and their control objectives can be minimized by adjusting the
aforementioned controllable elements in a manner parallel to fitting a

model to collected data. Many of the same principles discussed for model
adaptation apply here. Control objectives can be implemented in a number
of ways, as follows:
1. Single-point targets. Here, the target for control is a single point at the
end of the trial. The control system can devise any route in order to reach
this goal. This is simple to implement but has drawbacks. The system may
‘select’ a route that demands growth performance at the limit of what is
possible for an average animal. This may constitute a welfare problem, and
should the performance demanded exceed that possible for an individual
differing from the population average, control may fail to achieve the target.
2. Trajectories. The problems of single-point targets for growth can be
avoided by setting a trajectory for the animals to grow along through the
whole trial period. This approach was used for poultry projects by both
IMS Poultry and Aerts et al. (2003a,b). IMS Pigs used both trajectory and
end-point approaches, as the control software was adapted for either. The
weighting of deviations of predicted growth from this trajectory need not,
however, be equal through the whole trajectory: it might, for example, be
weighted more heavily at the end, to ensure that early forward predictions
are not constrained to give a good fit to the target trajectory at the expense
of achieving the final target weight. In effect, this is a compromise between
end-point and trajectory approaches.
3. Integrals. Both of the above methods are similar in that the system is
attempting to steer towards a target at a particular time. They differ in that
the latter method specifies more the route to be taken towards this target.
In contrast, there are objectives for which the value is summed over the
whole period of production. The value at a particular time is not of
interest, only its summation over time. Production costs or nitrogen
emission targets are likely to fall into this category.
4. Constraints. Some production outputs, such as nitrogen output, are
best considered as constraints on the system, rather than as absolute
targets. It may be, for example, that nitrogen output below a particular
level is acceptable, not that a particular level is desired. This distinction is
one that can be made for single-point targets, trajectories, and integrals
alike.
Control algorithms and control variable selection
The number of control variables subject to optimization during control is

likely to be larger than the number of parameters fitted during model
adaptation. Moreover, each control variable can be regarded separately for
each period modelled during the remainder of the trial, giving a
potentially large number of dimensions for optimization. The algorithms
available to fit model predictions to control targets are the same as those
available to fit model predictions to previously collected data and the same
principles apply in choosing the best method.
In the IMS Poultry project, both control of feed intake and control of
feed blend (varying dietary protein content between two extremes) were
possible. With each control variable able to be specified for each day of the
trial, the dimensions of the problem were at least 30. With this large
number of dimensions, a genetic algorithm (Holland, 1975; Goldberg,
1989) was found to be robust, and operated at an acceptable speed (Stacey
et al., 2004). The system of Aerts et al. (2003a,b), also studying poultry,
considered control by restricted feed intake alone. These authors fitted a
single control variable – the amount of food to be fed in the succeeding 24
hours – to follow a defined target trajectory for growth.
Rationing of feed was not possible in the IMS Pigs study. But here, pigs
were housed in relatively small groups of 12 and the feed blend could be
varied at the individual feeder level (again, dietary protein content varied
between two extremes). This control variable was divided into four control
periods, as tests showed that little improvement in control would be gained
by using shorter periods (Parsons et al., 2004). The number of dimensions
here being considerably smaller than in Stacey et al. (2004), the revised
simplex algorithm was found to be suitable for optimization, as was also
used for model adaptation in this study.
In the case of blending feed as a control variable, both the IMS Pigs
and IMS Poultry projects considered not the blend itself, but the rate of
change of the blend, constrained within narrow bounds. This prevented
large step changes or quick changes in the blend of the diets, which might
be poorly tolerated.
Modelling feed intake
The accurate prediction of ad libitum feed intake can be considered the

bane of growth modelling. For pigs, food intake in a commercial
environment differs substantially from that in experimental conditions and
there are marked differences between genotypes, premises, and
management practices (Whittemore et al., 2001b). Mechanistic models are
informative in predicting the qualitative effects of a change of environment
(social, disease or thermal) or diet (Wellock et al., 2003a,b; Whittemore et
al., 2003); however, some authors have considered feed intake as being best
obtained by observation (Schinckel and de Lange, 1996) and therefore best
specified as a model input (Green and Whittemore, 2003). Where feed is
rationed below the likely ad libitum intake (and refusals can be assumed
low) as may be the case for individually housed sows, this is reasonable.
Elsewhere however, it is necessary to predict ad libitum intake from
observations.
For both the IMS Pigs and IMS Poultry projects, the experimental set-
up included measurements of feed consumption. Feed intake data from
preceding trials were used to predict ad libitum feed intake for animals of a
specified size for entry into the growth models. However, allowance was
made for individual deviation from the assumed feed intake for short
periods. Where an animal was eating more than an average amount, for
the purposes of future prediction (and therefore control) this behaviour
was expected to continue for a period, after which average feed intake was
then assumed. The problem of feed intake prediction was not an issue for
the growth model of Aerts et al. (2003a,b), who fed specific feed rations to
groups of chickens subject to control.
IMS Systems in Practice
By way of an example showing the interaction of the many components of

IMS systems described above, we will now turn to two systems that have
been developed. The first of these, IMS Pigs, is in development; the
second, IMS Poultry, forms a completed research project now in
commercial use (available as PIMS from David Filmer Ltd).
Integrated management systems for pigs
The IMS Pigs project investigated the performance of a novel system for
pig production and pollution control (Green et al., 2004; Parsons et al.,
2004). This system is work in development and not all components are
online and fully automated. Nevertheless, the scientific proof of principle
has been shown. The main data stream for IMS Pigs was a VIA system
(Marchant et al., 1999; Schofield et al., 1999) which provided automatic live
weight estimates on a daily basis for individual animals. It was
complemented with manual P2 backfat measurements and manual
measurements of live weight. Feeding was ad libitum but the crude protein
content of the feed was varied between 13 and 19% on a per-pen basis
(manually, at the instruction of the IMS controller). Each of the 12 pens
housed 12 pigs during each trial, pairs of pens sharing one of six
controlled-environment rooms at ADAS Terrington, Norfolk, UK.
A revised simplex method (Nelder and Mead, 1965) was used to fit two
parameters of a mechanistic model of pig growth (Green and Whittemore,
2003, 2005) to the data collected from the VIA system. In effect, one
parameter specified the efficiency of growth and the other defined the
partitioning of dietary nutrients between protein and lipid deposition.
Weekly forward predictions from the fitted models were used to determine
the crude protein content to be blended for each pen in order to best
satisfy the control objectives. Fitting of model output to the control
objectives was also carried out using the revised simplex approach.
In the trials, two sets of objective targets were specified: targets for
weight gain (50 or 60 kg gain during the growth period) and for P2
backfat depth at the slaughter point (12 or 16 mm). Results from one of
the trials are shown in Table 15.2. Model fit was good before optimization,
with the means of the final observed and predicted weights differing by
only 6 kg. Optimization reduced this to less than 1 kg, even where
Table 15.2. Growth targets in IMS Pigs. Growth targets for each pen,
with mean final deviation from targets. Standard errors are shown in
brackets.
Target Deviation
Weight gain (kg) (approximate) 50 2.1 (2.4)
50 2.3 (0.9)
60 –5.8 (1.5)
60 2.0 (2.4)
Final P2 backfat depth (mm) 12 –0.9 (0.53)
12 0.2 (0.60)
16 –2.1 (0.72)
16 –2.4 (0.68)
optimization was performed only on the early part of the collected data.
This shows the optimized models to have good predictive power.
Control of fat depth was partially successful. The lower target was
achieved with good accuracy, but the pigs set the higher target ended the
trial 2–2.5 mm thinner than specified. However, reference pigs fed only on
the low protein diet were similarly below target, showing that achievement
of this target was beyond the capability of the system. Weight gain control
was achieved within 3 kg of the targets on three of the four pens. The
fourth pen finished below target by nearly 6 kg. This was unavoidable
because these pigs, although on target until the penultimate week of the
trial, then showed an unexplained reduction in growth rate. Where such
changes, possibly caused by a sub-clinical illness, occur early in the trial,
there is time for the IMS system to take corrective action. In this case,
there was no time remaining on trial for such a severe departure from the
target to be corrected. Nevertheless, for healthy pigs, this study provides
proof of principle of control of pig weight through control of diet on ad
libitum feeding. If rationed feeding is implemented, greater control of both
weight and fatness can be expected.
Integrated management systems for poultry
The IMS Poultry project uses, as does the IMS Pigs project, the IMS
structure shown in Fig. 15.1 (Frost et al., 2003; Stacey et al., 2004).
Compared with the individual animals studied in IMS Pigs the commercial-
sized scale here was much larger: each of the eight houses on the trial farm
held 30,000–40,000 birds. Mean live weight for each house was provided
by automatic ‘Flockman’ weigher systems (http://www.flockman.com), as
with IMS Pigs providing daily weight data. Again, as with IMS Pigs, the
main controller was diet blend: the birds were usually fed ad libitum, but
the composition of the food was varied to give a range of protein content.
Rationed feed intake was, however, available in some trials.
As with IMS Pigs, a mechanistic model of growth and nutrition was

used, albeit one that was simpler and faster to compute. Only a single
model variable was fitted to the data, specifying efficiency of use of
nutrients, and so the robust linear search algorithm was used for model
parameterization, since the more complex algorithms were unnecessary
with a one-dimensional problem. For control, a genetic algorithm
(Holland, 1975) was used, their being many more dimensions in the
optimization, of the order of 30.
The control targets comprised a target weight trajectory, with control
decisions being made by the IMS system thrice weekly. Model output, as
found with the IMS Pigs model, was good, but varied from house to house.
This is of course expected, and is why the model adaptation mechanism is
implemented; following model adaptation, error between the model and
actual weights was reduced from approximately 108 g to approximately
29 g. The IMS system provided a precision of control equal to that
achieved by human managers. Stacey et al. (2004) comment that, with
control of not only feed blend but also feed ration, even better
performance of the system will be found.
Future Prospects
Further progress in IMS systems for pigs and poultry are to some extent
determined by the development of sensor systems. For sows, some
progress has been made in the development of VIA systems (Schofield,
2004). However, the changes during pregnancy in the sow are small
compared with those of the growing pig, being more a matter of shape
than size, and this provides more of a challenge for the VIA system
(Doeschl-Wilson et al., 2005). Side-view VIA is more appropriate for
evaluating the sow, but needs more awkward placement of cameras than
the overhead placement used for viewing the growing pig. Amongst other
uses for VIA is the potential to track animal movements (Tillett et al.,
1997), thereby providing a measure of activity and its concomitant energy
usage.
Another data source that may prove useful in IMS is sound.
Vocalizations are known to be an indicator of welfare in both pigs and
poultry (Manteuffel et al., 2004) and progress has been made in
recognition of vocalization types with neural networks (Chedad et al.,
2001). For reviews of recent developments in the detection of chemicals
relating to health and welfare, nuisance, or pollution see Persaud (2001)
and Frost et al. (1997).
The IMS systems described above are all indoor systems. IMS outdoors
will provide more obstacles. VIA systems would have to be more robust
and contend with more varied lighting conditions than are found indoors.
Modelling of the thermal environment and its effects on energy use would
be considerably more difficult than in the controlled conditions of indoor
housing.
Conclusions
Integrated management systems represent an exciting new use of models,

operating in real time as part of livestock production systems to provide a
basis for management decisions to be made by automatic control systems.
Their use relies on the provision of continuous data streams from automatic
sensor systems, which are growing in affordability and reliability. The
commercial development of such systems is just beginning, but trial studies
show that their performance compares favourably with that of human
managers. It is hoped that IMS will provide benefits to the producer, to the
consumer, to the welfare of livestock, and to the environment.
Acknowledgements
With thanks to Andrea Doeschl-Wilson and Istvan Kiss.
Endnotes
1. Integrated Management Systems for pig nutrition control and pollution reduction.
LINK Sustainable Livestock Production programme funded project; UK
Department of Environment, Food, and Rural Affairs (formerly MAFF)
http://www.defra.gov.uk/science/Link/Agriculture/SLP/
Key references: Parsons et al. (2004); Green and Whittemore (2003,
2005); White et al. (2004).
2. Integrated management systems to enhance efficiency and pollution control in
poultry production. As above.
Key references: Frost et al. (2003); Stacey et al. (2004).
References
Aerts, J.-M., Wathes, C.M. and Berckmans, D. (2003a) Dynamic data-based
modelling of heat production and growth of broiler chickens: development of
an integrated management system. Biosystems Engineering 84, 257–266.
Aerts, J.-M., Lippens, M., De Groote, G., Buyse, J., Decuypere, E., Vranken, E. and
Berckmans, D. (2003b) Recursive prediction of broiler growth response to
food intake by using a time-variant parameter estimation method. Poultry
Science 82, 40–49.
Chedad, A., Moshou, D., Aerts, J.-M., van Hirtum, A., Ramon, H. and Berckmans,
D. (2001) Recognition system for pig cough based on probabilistic neural
networks. Journal of Agricultural Engineering Research 79, 449–457.
DEFRA (2002) Nitrates – Reducing Water Pollution from Agriculture. http://www.defra.
gov.uk/environment/water/quality/nitrate/
de Wet, L., Vranken, E., Chedad, A., Aerts, J.-M., Ceunen, J. and Berckmans, D.
(2003) Computer-assisted image analysis to quantify daily growth rates of
broiler chickens. British Poultry Science 44, 524–532.
Doeschl, A.B., Green, D.M., Whittemore, C.T., Schofield, C.P., Fisher, A.V. and
Knap, P.W. (2004) The relationship between the body shape of living pigs and
their carcass morphology and composition. Animal Science 79, 73–83.
Doeschl-Wilson, A.B., Green, D.M., Fisher, A.V., Carroll, S.M., Schofield, C.P. and
Whittemore, C.T. (2005) The relationship between body dimensions of living
pigs and their carcass composition. Meat Science 70, 229–240.
Ellis, M. and Hyun, Y. (2005) Feed Intake Behavior and Implications for Energy Utilization.
http://www.traill.uiuc.edu/porknet/paperDisplay.cfm?Type=paper&ContentID=63
animals, particularly poultry. In: Hillyer, G.M., Whittemore, C.T. and Gunn,
R.G. (eds) Computers in Animal Production, Occasional Publication No. 5. British
Society of Animal Production, Thames Ditton, UK, pp. 103–110.
Emmans, G.C. (1987) Growth body composition and feed intake. World Poultry
Science 3, 208–227.
Emmans, G.C. (1989) The growth of turkeys. In: Nixey, C. and Grey, T.C. (eds)
Recent Advances in Turkey Science. Butterworths, London, pp. 135–166.
Emmans, G.C. (1994) Effective energy: a concept of energy utilisation applied
and Boorman, K.N. (eds) Nutritional Requirements of Poultry and Nutritional
Filmer, D. (2001) Nutritional management of meat poultry. In: Wathes, C.M., Frost,
A.R., Gordon, F. and Wood, J.D. (eds) Integrated Management Systems for
Livestock. Occasional Publication No. 28. British Society of Animal Science,
Fox, L. and Sachs, J. (2003) The Meatrix http://www.themeatrix.com/
Frost, A.R. (2001) An overview of integrated management systems for sustainable
livestock production. In: Wathes, C.M., Frost, A.R., Gordon, F. and Wood, J.D.
(eds) Integrated Management Systems for Livestock. Occasional Publication No. 28.
British Society of Animal Science, Edinburgh, UK, pp. 45–50.
Frost, A.R., Schofield, C.P., Beaulah, S.A., Mottram, T.T., Lines, J.A. and Wathes,
C.M. (1997) A review of livestock monitoring and the need for integrated
systems. Computers and Electronics in Agriculture 17, 139–159.
Frost, A.R., Parsons, D.J., Stacey, K.F., Robertson, A.P., Welch, S.K., Filmer, D. and
Fothergill, A. (2003) Progress towards the development of an integrated
management system for broiler chicken production. Computers and Electronics in
Frost, A.R., French, A.P., Tillett, R.D., Pridmore, T.P. and Welch, S.K. (2004) A
vision guided robot for tracking a live, loosely constrained pig. Computers and
Electronics in Agriculture 44, 93–106.
Gill, P.E., Murray, W. and Wright, M.H. (1981) Practical Optimization. Academic
Press, New York.
Goldberg, D.E. (1989) Genetic Algorithms in Search, Optimization and Machine Learning.
Addison-Wesley, Reading, MA, USA.
Green, D.M. and Whittemore, C.T. (2003) Architecture of a harmonised model of
77, 113–130.
Green, D.M. and Whittemore, C.T. (2005) Calibration and sensitivity analysis of a
model of the growing pig for weight gain and composition. Agricultural Systems
84, 279–295.
Green, D.M., Parsons, D.J., Schofield, C.P. and Whittemore, C.T. (2004) Real-time
control of pig growth through an integrated management system (IMS).
Proceedings of the British Society of Animal Science, p. 26.
Halachmi, I., Edan, Y., Maltz, E., Peiper, U.M., Moallem, U. and Brukental, I.
(1998) A real-time control system for individual dairy cow food intake.
Computers and Electronics in Agriculture 20, 131–144.
Hancock, C.E., Bradford, G.D., Emmans, G.C. and Gous, R.M. (1995) The
evaluation of the growth parameters of six strains of commercial broiler
chickens. British Poultry Science 36, 247–264.
Holland, J.H. (1975) Adaptation in Natural and Artificial Systems. University of
Michigan Press, Ann Arbor, Michigan.
Kearney, G.A., Knee, B.W., Graham, J.F. and Knott, S.A. (2004) The length of test
required to measure liveweight change when testing for feed efficiency in
cattle. Australian Journal of Experimental Agriculture 44, 411–414.
Kirkpatrick, S., Gelatt, C.D. and Vecchi, M.P. (1983) Optimization by simulated
annealing. Science 220, 671–680.
Kolstad, K. (2001) Fat deposition and distribution measured by computer tomography
in three genetic groups of pigs. Livestock Production Science 67, 281–292.
Manteuffel, G., Puppe, B. and Schön, P.C. (2004) Vocalization of farm animals as a
measure of welfare. Applied Animal Behaviour Science 88, 163–182.
Marchant, J.A., Schofield, C.P. and White, R.P. (1999) Pig growth and conformation
monitoring using image analysis. Animal Science 68, 141–150.
Marquardt, D.W. (1963) An algorithm for least squares estimation of parameters.
Journal of the Society of Industrial and Applied Mathematics 11, 431–441.
Mitchell, A.D., Scholz, A.M., Wange, P.C. and Song, H. (2001) Body composition
analysis of the pig by magnetic resonance imaging. Journal of Animal Science 79,
1800–1813.
Nelder, J.A. and Mead, R. (1965) A simplex method for function minimization.
Computer Journal 7, 308–313.
Newcom, D.W., Bass, T.J. and Lampe, E.F. (2002) Prediction of intramuscular fat
percentage in life swine using real-time ultrasound. Journal of Animal Science
80, 3046–3052.
Parsons, D.J. (1992) Performance of constrained forms of the simplex method for nonlinear
optimization. IED Technical Note TN 003, Silsoe Research Institute, Silsoe, UK.
Parsons, D.J., Schofield, C.P., Green, D.M., Whittemore, C.T., Carroll, S. and Kay,
R. (2004) Real-time control of pig growth through an integrated management system
(IMS). Institute of Agricultural Engineers, Leuven, Belgium.
Persaud, K.C. (2001) Biosensors for livestock production. In: Wathes, C.M., Frost,
A.R., Gordon, F. and Wood, J.D. (eds) Integrated Management Systems for
Livestock. Occasional Publication No. 28. British Society of Animal Science,
Pietersma, D., Lacroix, R., Lefebvre, D., Block, E. and Wade, K.M. (2001) A case-
acquisition and decision-support system for the analysis of group-average
lactation curves. Journal of Dairy Science 84, 730–739.
Press, W.H., Teukolsky, S.A., Vetterling, W.T. and Flannery, B.P. (1992) Numerical
Recipes in C. Cambridge University Press, Cambridge, UK.
Schinckel, A.P. and de Lange, C.F.M. (1996) Characterization of growth parameters
needed as inputs for pig growth models. Journal of Animal Science 74, 2021–2036.
Schofield, C.P. (2004) Integrated management systems for pig nutrition control and
pollution reduction. DEFRA LINK project, technical report. Silsoe Research
Institute, UK.
Schofield, C.P., Marchant, J.A., White, R.P., Brandl, N. and Wilson, M. (1999)
Monitoring pig growth using a prototype imaging system. Journal of
Agricultural Engineering Research 72, 205–210.
Stacey, K.F., Parsons, D.J., Frost, A.R., Fisher, C., Filmer, D. and Fothergill, A.
(2004) An automatic growth and nutrition control system for broiler
production. Biosystems Engineering 89, 363–371.
Tillett, R.D., Onyango, C.M. and Marchant, J.A. (1997) Using model-based image
processing to track animal movements. Computers and Electronics in Agriculture
17, 249–261.
Tillett, R., McFarlane, N. and Lines, J. (2000) Estimating dimensions of free-
swimming fish using 3D point distribution models. Computer Vision and Image
Understanding 79, 123–141.
81, 2995–3007.
White, R.P., Schofield, C.P., Green, D.M., Parsons, D.J. and Whittemore, C.T.
(2004) The effectiveness of a visual image analysis (VIA) system for monitoring
the performance of growing/finishing pigs. Animal Science 78, 409–418.
Whittemore, C.T. (1998) Simulation modelling. In: The Science and Practice of Pig
Production. Blackwell Science, Oxford, UK.
Whittemore, C.T., Green, D.M. and Schofield, C.P. (2001a) Nutrition management
of growing pigs. In: Wathes, C.M., Frost, A.R., Gordon, F. and Wood, J.D.
(eds) Integrated management systems for livestock. Occasional Publication No. 28.
British Society of Animal Science, Edinburgh, UK, pp. 89–95.
Whittemore, C.T., Green, D.M. and Knap, P.W. (2001b) Technical review of the
energy and protein requirements of growing pigs: food intake. Animal Science
73, 3–17.
Whittemore, C.T., Green, D.M. and Knap, P.W. (2001c) Technical review of the
energy and protein requirements of growing pigs: energy. Animal Science 73,
199–215.
Whittemore, C.T., Green, D.M. and Knap, P.W. (2001d) Technical review of the
energy and protein requirements of growing pigs: protein. Animal Science 73,
363–374.
Whittemore, E.C., Emmans, G.C. and Kyriazakis, I. (2003) The problem of
predicting food intake during the period of adaptation to a new food: a model.
Williams, S.R.O., Moore, G.A. and Currie, E. (1996) Automatic weighing of pigs fed
ad libitum. Journal of Agricultural Engineering Research 64, 1–10.
Wu, J.H., Tillett, R., McFarlane, N., Ju, X.Y., Siebert, J.P. and Schofield, P. (2004)
Extracting the three-dimensional shape of live pigs using stereo
photogrammetry. Computers and Electronics in Agriculture 44, 203–222.
Index
ability to cope 65 anabolic hormone 292

acclimation state 215 anabolism 156–157
adipose tissue animal
control of lipid storage in 293 behavioural changes 36–37
energy storage tissue 284 breeding strategies 64
enzyme activity 289 description 24, 46, 48
see also fat disease and health 39, 41, 117, 136
age at first egg, mean, prediction of health 39–42, 117–118, 129–136,
236–237, 255 305, 320
air quality 38, 82–84 homeothermic 188–192
air velocity (and speed) 34, 54, 190–191, thermal environment 210–214, 306
198–200, 213, 215, 221 appendages, bare 190–194, 203
algebraic model inversion 167, appetite 32, 56–57, 71, 308
169–177, 185
artificial neural networks 98, 103, 105
sources of error 174
ash deposition 261
allometric
ATP cost 152–153
coefficients 24
ATP synthesis 149–152, 158
relationships 24, 49, 80
ATP yield 158, 298
amino acids
coefficients 30 automated feeding and weighing
deaminated 29, 145, 148, 152 systems 306, 310
first-limiting 25, 46, 49, 57, 125–126
feed intake 29
immune response 71, 126, 128–135, backfat 56, 153, 171–180, 310,
138–139 318–319
limiting 2,25, 31, 33 basal metabolic rate 153
optimum content 91–93 between-animal variation 63–64, 66, 71
profiles 29–30 behavioural freedom 215
ratio with energy 46, 77 bimodal distribution 237, 243, 256
ratio with protein 46 biological responses, accuracy of
response 79, 132 defining 22
utilization of limiting 2 birds in transit 218
325
326 Index
body digestible energy 45–46, 143, 172

composition 15, 35, 49, 95, digestion 15, 28, 49
154–156, 171, 178–179 Discomfort Index 197
maintenance 169, 261–262 disease
reserves 26, 154 and performance 39–41, 71
temperature 37–38, 189, 193, 202, challenge 23, 38
212, 216–222, 244 clinical 127
see also carcass intestinal 16, 30
boundary layer 191–203 susceptibility 71
breast meat yield 78, 80 diseased growing pigs 117
breeding values, estimated (EBV) 108,
163
broiler economic objectives 307
behaviour 82 effective energy 28, 43–44, 147, 153
feeding programmes 77–82, 85, efficiency of utilization of protein (ep)
91–94 31, 43, 262
bulk constraint 31–32, 38 effect of temperature on 124
egg
double-yolked 231, 236, 242,
calorimetry, indirect 145 245–246, 255–257
carbon dioxide balance 211 soft-shelled 242, 246–247, 255
carcass weight, predicting 230, 249–252,
composition 77, 80, 310 255–256
protein 157 empiricism 4,10
quality 308 energy
value 311 balance 37, 145, 214
catalytic activity 291 content, optimizing 2
chance constrained programming 109 cost 148–153, 172
chaos theory 99, 101 of physical activity 154
circadian rhythm 229–230, 238–241, dietary 2,31, 55, 172
255 gross 143, 172
coccidiosis 16 insufficient supply 267
compensatory responses 26, 34 losses 3,143
controllable elements in production 315 faecal 144
control algorithms 316–317 urinary and gaseous 145, 146
control objectives 306–308, 315–318 maintenance 28, 37, 40, 59,
constraints 316 151–154, 166, 275
integrals 316 metabolizable 29, 119–122, 143,
single-point targets 316 145, 261
trajectories 316 partitioning 155
cost–benefit analysis 84 systems 143, 145, 147–148, 157
costs classical 147
feed 77 digestible 143
processing and transport 77 effective 28, 43–44, 147, 153
metabolizable 143
net 143
decision analysis 111 transactions, stoichiometry of 148
decision support system 306 environment
degree of maturity 25, 275 macro- and micro- 189
diet spatial and temporal variation 188
composition 81, 171 thermal 35, 37, 48, 56, 210–218,
least cost 14, 78 306, 315, 320
Index 327
environmental factors fibre (crude) 27, 31, 144–147

air velocity 4, 54, 190–191, 198–200, floor type 54
213, 215, 221 follicle maturation 230–235
ambient temperature 35, 43, 48, 54, follicular hierarchy 238, 245, 255
170, 213, 220 foods
see also temperature energy limiting 119, 120–123
environmental stressors 54–70, 117, protein-adequate 121
163, 210, 214 protein limiting 122, 125
environmental objectives 308 fractal analysis 212
exposure to pathogens 71, 126–127, fractional anabolic and catabolic rates
129, 136 156
fuzzy logic 98–100, 211
faecal energy losses 144

fasting heat production (FHP) 28, 146, genetic
282 algorithms 103, 107, 313–314, 317
fat factors 273
content 26 parameters 62, 64, 70, 81, 85,
deposition 26, 33–34, 37 87–94, 183, 266–277
growth 25–27, 32, 35 potential 23, 67, 79, 163, 166, 178,
reserves 26 185
synthesis 284, 287, 290, 298 gluconeogenesis 10, 150, 285, 290, 295
see also carcass; lipid glucose 18–19, 144, 148–152, 287–303
growth
fatness, inherent 24
ash 27
fatty acids 18–19, 144, 149–151,
compensatory 26, 34
288–293, 297, 299
fat (lipid) 25–27, 32–35, 178
feather
feather 81, 88
coat 90, 190, 193
homoeorhetic and homoeostatic
density score 85
control 156–157, 289, 294
growth, rate of 81, 85, 88–90
moisture 27
total score 85
parameters 25, 65, 68
feathering rate
potential rate of 23, 25, 31, 165, 169
multiplier 88 protein 24–29, 38, 39, 49, 80–83,
non-linear effect on variation 91 91, 130, 165, 266
feed gut capacity 23, 33, 43, 49
costs 77
formulation 78, 83, 99, 108–110
pelleted 84 health
variation in chemical and physical status 22–23, 27–28, 39–42, 49
nature 84 and welfare objectives 308
see also feeding heat
feeding exchange 188, 190–193, 205,
ad libitum 44–45, 117, 125, 310, 212–216, 219, 222
317–319 conduction 191, 215
choice 46 convection 82, 191, 192, 215
costs 77–78, 92–94 radiation 191
restricted 170, 317 physiology of 188
see also feed increment of feeding 8,146–147,
feed intake 150–152, 158, 285, 297
constrained 31, 89 production 145
desired 28 stress 85, 197, 212, 215, 221
328 Index
heat loss 35–38, 212, 215 lipid:protein ratio at maturity (LPRm)

evaporative (latent) 35–36, 24–26, 43–48, 85, 88–92
192–194, 202–204, 206 lipid
from Dumbo 212 deposition 46, 147–156, 166, 172,
sensible (non-evaporative) 36–37, 261, 270, 275–278
191, 193, 204, 215, 228 growth 26, 34, 178
heritability 70, 85, 165, 179, 181, 287, metabolism 151, 297
292–293 protein ratio, minimum 27, 118,
hierarchy of organizational levels 8 155, 171, 174, 272
homeostatic and homeorhetic controls lipogenesis 151, 285, 290, 293
156 lipolysis 285, 290, 293–295
homeothermic animals 188–192 lipoproteins 289
housing 45, 210–211, 215, 310, 320 litter size 297
hypothermia 212, 221–222 liver 145, 150, 297
livestock houses 211
logistic growth curve 230
ideal protein 130–133, 170
intake 118–119, 125, 131
requirement 30–31, 128 maintenance
immune response 71, 126–139 activity component 28, 37, 40–41,
acquired 126, 129, 135 59–63, 154, 320
variation in 71 requirement 119, 169, 272, 309
immunity 126–135 amino acids 125
maintenance of 129
disease/immunity 39, 127, 129
rate of acquiring 135
energy 38, 40–41, 155
infectious 17, 54, 71
support costs 153
environment 54
mammary gland 18, 289–290,
stressors 71
294–296, 299
initial weight 88
marginal response in PR on energy
innate immunity and repair 127–128
limiting foods
integrated control systems 211
effect of genotype 121
integrated management systems (IMS)
effect of liveweight 120
305–308, 318–321
effect of temperature 121
internal cycle length 237–241,
253–254 mature body protein (Pm) 24–25, 266,
isothermal net radiation 190–204 277
maximum lipid in gain 88
meat quality 22, 307–308
Kalman filter 103–108 metabolic body size 293
metabolism 148, 150–154, 276,
282–299
lactating sow 18, 286, 290 metabolisable energy (ME)
lactation 19, 108, 284–298 content 29, 121, 145
lactation curve 108, 293 intake 119, 120
lactose 17–19, 290, 294–295 requirement for maintenance 119,
synthesis 19, 290, 299 261
laying hen 2,79–80, 92, 104, 111, 257 system 143
see also egg milk 18
linear programming (LP) 2,12, 77, 99, production 295–299
108–110, 313 synthesis 17–19, 290, 295–298
linear search algorithm 313, 320 minerals 84, 104
Index 329
model optimization
application of 2,10, 14, 22, 79, 98, algorithm 176
165, 263 criteria in model inversion 176
Bristol–Reading 237 pig production strategies 64
deterministic 14, 99, 171, 209, 260, optimum (optimal)
311 amino acid content in feed 77–92
dynamic 14, 101, 169, 202, 209, feeding programme 77–82, 93
212, 285, 298 nutrient density 77
economic 2,308 production environment 212
empirical 4,9–10, 15, 20, 97–103, thermal micro-environment
210–213, 216
163, 230, 282, 311–314
ovarian follicle 229, 232
examples 1
oviposition time 240–243, 255–256
inversion 165
ovulation
by optimization 167
internal 236, 244–245, 254–257
laying hen rate 230, 234–243, 254–256
linear multiple regression 12 ovulatory cycle 229–235, 240–241, 254
metabolic flux 286, 289
mixture 104
Reading 79 partitioning
sow 297–299 nutrients 23, 155, 318
of nutrients during disease 126
stochastic 64, 79, 97–101, 109, 168,
of scarce food resources 117
178, 212, 230, 236, 311
immune pigs 129
teleonomic 9
naïve pigs 129
thermal balance 212
pathogen
see also requirement response challenge 117, 126–136
modelling virulence of 129
progress in 19 phase-shift 241
physiological response 217 photoperiodic control 230
mortality 83, 218, 222 piglets 46, 51
pigs per pen 62
pollution control 318
neutral detergent fibre (NDF) 144 populations 71, 79, 92, 126, 311
nitrogen correlated distributions 85
balance techniques 145 Predicted Mean Vote (PMV) 214
prime sequence length 238, 253–254
excretion 145, 147
programming
retention 29, 147
chance constrained 99, 109–110
see also protein
compromise 14
nursery pigs 43, 44
goal 14, 110
nutrient linear 2,12, 77, 83, 99, 108–110
absorption 10 parametric 14
content, variation in 83–84, 109 quadratic 110
density 33, 43, 77–78 separable 14
partitioning 15, 23, 126, 130 protein
requirements 3,22, 71, 109, 118, deaminated 29, 145, 148, 152
257, 283 deposition rate 25, 152, 261
specifications 2 maximum 265
utilization 10, 79, 297, 306, 311 losses
nutritional status 294 endogenous gut 29
objective function 13–14, 78, 184, 210 integument 29, 30
operations research 12, 103–108, 110 maintenance protein 30
330 Index
protein continued stress

synthesis 152–153, 287–290 cold 215, 221–222
turnover 29, 152, 154, 178, 285, 297 effect on performance 55, 61
see also carcass, feedstuff, nitrogen heat 212–215, 220, 224
quadratic programming 110, 209 in group housing 38
subclinical infection 117, 127
substrate sensitivity 290–297
radiation increment 198, 199, 203 sugars 144, 146
radio-telemetric systems 217 support costs for maintenance 153
rate of maturing (B) 24, 25, 39, 80, 81 surface wetting 221
rate:state formalism 12, 14, 16, 19 synthesis
reductionist approach 100 milk fat and lactose 17
regression analysis 103–105 systems approach 159, 301
reproductive performance 210, 299
respiratory tract 204
response temperature
curvilinear-plateau, of population Apparent Equivalent (AET)
66 197–198, 201–202, 218
homeostatic 217 deep body 212, 218–222
immune 71, 126–139 effective 35, 37, 197–200, 204–205,
linear-plateau 66 214, 216
mean population 65 humidity index 216
of a system 9,15 standard operating 216
to an LH surge 232 thermal
to environmental constraints 38 balance models 212
to light 257 comfort 213–223
to stressors 65–72 hygrometric index 197
revised simplex algorithm 312–317 resistance 190–203
Richards function 266 thermoregulation 82, 178, 189193–194
transport
animal 215, 217, 221–223
sensitivity analysis 88, 174 environment 215, 217
sensor systems 308–309, 320–321 feed 77–78, 83–84
sequence length, mean 253–254, 256 nutrient 152–153, 246, 285, 294,
simulation, stochastic 76 297
social environment 23, 38, 49, 62, 64,
71, 315
social dominance 64 ultrasound 310
social stressors 54–72 urea
ability to cope, genetic differences excretion 145, 148
54, 61–72, 133–134, 165 synthesis 152, 153
effect on performance 56–57 uric acid 145
feeder space allowance 54, 59, uterus 246
70–71
group size 54–75, 315
mixing 54, 58, 60–71 variation
space allowance 54, 57–71 between-animal 63–71, 154, 166
starch 121, 144, 146–151 chemical and physical nature of feed
state, current 23, 25, 63, 81, 156 84
stoichiometry 145, 147–153, 158–159, feed composition 76, 83–84
286, 298 genetic 62–62, 76126, 261–262,
stocking density 23, 38–39, 81–82 275, 278, 289
Index 331
immune response 71 weaning, post 40

non-linear effect, feathering rate 91 wind chill factor 197
spatial and temporal, of environment welfare 55, 64, 67–69, 210, 217–223,
188 305, 308, 316, 320
within-house 76 breeding for improved, in pigs 69,
within-population 166 286
ventilation 35, 82, 211, 215, 222–223,
315 yolk
Visual Image Analysis (VIA) system double 231, 236, 242, 245–246,
309 255–257
vocalizations as indicator of welfare 320 resorption 243
weight
water deposition 261 and age 236, 250, 256
water-holding capacity 32, 58, 60 prediction 249, 255

Mechanistic Modelling in Pig and Poultry Production

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Mechanistic Modelling in Pig and Poultry Production

Uploaded by

Copyright:

Available Formats

MECHANISTIC MODELLING

IN PIG AND POULTRY

CABI Head Ofﬁce CABI North American Ofﬁce

Tel: +44 (0)1491 832111 Tel: +1 617 395 4056

© CAB International 2006. All rights reserved. No part of this publication

Typeset by Columns Design Ltd, Reading

List of Contributors vii

9 Evaluating Animal Genotypes through Model Inversion 163

I. Kyriazakis, Animal Nutrition and Health Department, Scottish Agricultural

This volume records the proceedings of a conference held in South Africa

Financial support for this symposium is gratefully acknowledged from:

© CAB International 2006. Mechanistic Modelling in Pig and Poultry Production

‘science is observation encapsulated in equations’. The aim of most

could be shown to ﬁt all the available experimental data for lysine,

Wageningen University, Marijkeweg 40, 6709 PG Wageningen, The

Nature and Progress of Science

© CAB International 2006. Mechanistic Modelling in Pig and Poultry Production

(Popper, 1968), stresses the importance of predictive ability rather than

Role and Practice of Mathematical Modelling

Fig. 2.1. Nature and progress of science.

simple that it can be grasped and controlled without abstraction. Abstraction

scheme for the hierarchy of organizational levels is shown in Table 2.1.

Teleonomic models (see Monod, 1975, for a discussion of teleonomy) are

Table 2.1. Levels of organization.

i+3 Collection of organisms (herd, ﬂock)

Mechanistic models are process-based and seek to understand causation. A

inter-relate the components that make a model mechanistic. Mechanistic

Model evaluation is not a wholly objective process. Models can be

completely non-coercive way, by appealing to reason, by enabling us to see the

Linear multiple regression models pervade applied biology. The mathe-

E(Y X1 , X2 , ..., X q ) = β0 + β1 X1 + β2 X2 + ... + β q X q , (2.2)

Y is known as the dependent variable, X1, X2, …, Xq as the independent

Statistics Operations Applied Numerical Pure

← Empirical modelling →← Mechanistic modelling

Fig. 2.2. Mathematical spectrum.

An LP problem has three quantitative aspects: an objective; alternative

∑ aij X j ≥ or ≤ bi ; i = 1,2, ..., m , [constraints]

X j ≥ 0; j = 1, 2, ..., q, [non-negativity constraints]

coefﬁcients and right-hand-side values, respectively. The paradigm is

rate:state formalism is not as restrictive as ﬁrst appears because any higher-

Total Merozoites Sporozoites,

milk, in the mammary gland of the lactating sow. The biochemical

The ﬁrst step in the application of scientiﬁc precepts to a problem is to

Fatty acids, Glucose,

Fatty acyl CoA,

Milk fat, Milk lactose,

The simulation of animal growth potentially provides a way of predicting

© CAB International 2006. Mechanistic Modelling in Pig and Poultry Production

Actual Intake and Growth

Economics LCF/Optimize Feed Budgets N and P

There is a plethora of mathematical functions describing the pattern of

Potential protein growth

Body protein weight over time is determined from the function:

where Pt = body protein weight at time t (kg)

It has been widely acknowledged that an animal has an inherent potential

Table 3.1. Growth parameters from various literature sources.

Ferguson and Gous (1993)

ideal conditions (Webster, 1993; Schinckel and de Lange, 1996; Moughan,

equal to 0. It is therefore possible to obtain signiﬁcant protein growth rates

Moisture and ash growth

As the allometric coefﬁcient describing the relationship between body

Live weight gains

Prediction of Voluntary Feed Intake

The principle behind predicting voluntary feed intake assumes that an