THE UNIVERSITY OF CALGARY

Simulation Models of Phyllotaxis

and Morphogenesis in Plants

by

Richard S. Smith

A THESIS

SUBMITTED TO THE FACULTY OF GRADUATE STUDIES

IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE

DEGREE OF DOCTOR OF PHILOSOPHY

DEPARTMENT OF COMPUTER SCIENCE

CALGARY, ALBERTA

September, 2007

c Richard S. Smith 2007
 Abstract

The research presented here is focused on the construction of simulation models intended

to improve our understanding of phyllotaxis and related patterning mechanisms in plants.

Most mechanistic models of phyllotaxis proposed in the literature are based on the idea

that existing primordia inhibit the formation of new primordia nearby. This idea is ex-

plored at two different levels of abstraction. At a more abstract level, a simulation is

presented in which the inhibiting effect is modeled as a simple mathematical function

of distance and primordium age. The resulting model can generate a wide variety of

the phyllotactic patterns observed in nature, and is more robust than previous models

in which primordium age did not influence the inhibition directly. A second simulation

model is presented that offers a explanation for phyllotaxis in molecular terms. A the-

ory inspired by recent experimental work is explored, that suggests that self-enhancing

transport of the morphogen auxin could be responsible for phyllotaxis patterning. This

is fundamentally different from traditional reaction-diffusion models often proposed for

patterning in animals.

Although visually there is little similarity between phyllotaxis and leaf venation, the

transport-based mechanism proposed over 25 years ago for leaf venation is believed to

involve the same molecular components. The same morphogen, and the same transport

machinery is hypothesized to be the basis for both patterning models. Simulation results

presented here and in previous work show that the dramatic difference in the patterns

created by the two mechanisms could be the result of changing the way the transport

machinery obtains its polarity. However, almost nothing is known about how this polarity

is achieved. Furthermore, some experimental evidence suggest that both mechanisms may

actually be operating within the same cells, perhaps even at the same time. Exploring

this possibility, several models of early leaf venation are investigated, using a similar

i
cellular surface and the same molecular components as the phyllotaxis model. These

models provide some initial inroads into the still open problem of integrating the two

patterning mechanisms.

ii
 Acknowledgments

I would first like to thank my supervisor, Przemyslaw Prusinkiewicz, for sharing with me

his extraordinary insight, and for his expert guidance, providing me with focus without

ever limiting my freedom to explore. Thanks to Cris Kuhlemeier for his inspiring passion

in the search for the answer to phyllotaxis, his encouragement, and for sharing with me

his wealth of knowledge and the latest results from his lab. I would also like to thank

Faramarz Samavati, Angus Murphy, Michael Surette, and Christian Jacob for serving on

my supervisory and examination committees. Many thanks to the folks in the lab, both

in the graphics jungle at the University of Calgary as well as at the Institute of Plant

Sciences in Bern. A very special thanks to my wife Ola, and to Nina and Anna who

had to go without dad for a time. Last but not least, I would like to thank the Natural

Sciences and Engineering Research Council of Canada (NSERC), the Alberta Informatics

Circle of Research Excellence (iCore), and the University of Calgary for their generous

scholarship support.

iii
 Table of Contents

Abstract i

Acknowledgments iii

Table of Contents iv

1 Introduction 1
1.1 Morphogenesis in plants . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Why build simulation models of biological systems? . . . . . . . . . . . . 3
1.3 Organization of thesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3

2 Phyllotaxis 5
2.1 A brief history of the study of phyllotaxis . . . . . . . . . . . . . . . . . . 5
2.2 Phyllotaxis terminology . . . . . . . . . . . . . . . . . . . . . . . . . . . 15

3 Modeling the Shoot Apex 18

3.1 The structure of the shoot apex . . . . . . . . . . . . . . . . . . . . . . . 18
3.2 A survey of shoot apex models used in phyllotaxis simulations . . . . . . 19
3.3 A shoot apex model with descriptive growth . . . . . . . . . . . . . . . . 21
3.4 Cells and cell division . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
3.5 VV and the L-Studio modeling environment . . . . . . . . . . . . . . . . 29

4 A Model of Phyllotaxis Based on Inhibition Fields 32

4.1 Field models of phyllotaxis . . . . . . . . . . . . . . . . . . . . . . . . . . 32
4.2 Organization of computation . . . . . . . . . . . . . . . . . . . . . . . . . 34
4.3 Single inhibition function . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
4.3.1 Pattern generation de novo . . . . . . . . . . . . . . . . . . . . . 37
4.3.2 Pattern propagation . . . . . . . . . . . . . . . . . . . . . . . . . 40
4.4 Two inhibition functions . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
4.4.1 Pattern generation de novo . . . . . . . . . . . . . . . . . . . . . 47
4.4.2 Pattern propagation . . . . . . . . . . . . . . . . . . . . . . . . . 49
4.5 Robustness of the models . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
4.6 Summary of results of the inhibition field model of phyllotaxis . . . . . . 53

5 A Transport-Based Model of Phyllotaxis 57

5.1 Mechanistic theories of phyllotaxis . . . . . . . . . . . . . . . . . . . . . 57
5.2 Molecular basis for a transport-based model of phyllotaxis . . . . . . . . 59
5.3 The transport-based patterning mechanism . . . . . . . . . . . . . . . . . 63
5.4 Phyllotaxis model on a cylindrical apex structure . . . . . . . . . . . . . 68
5.5 Phyllotaxis model on a cellular apex structure . . . . . . . . . . . . . . . 73
5.5.1 An essential role for differentiating primordia . . . . . . . . . . . 77

iv
 5.5.2 Phyllotaxis patterns produced by the cellular apex model . . . . . 79
5.5.3 Comparison to experimental observation in Arabidopsis . . . . . . 82
5.6 Sensitivity analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
5.6.1 Fixed PIN1 concentration . . . . . . . . . . . . . . . . . . . . . . 86
5.6.2 Removal of IAA production saturation term . . . . . . . . . . . . 86
5.6.3 Reduction of transport exponent . . . . . . . . . . . . . . . . . . 87
5.6.4 Increase in PIN1 allocation exponent . . . . . . . . . . . . . . . . 88
5.7 Summary of results of the transport-based phyllotaxis model . . . . . . . 88

6 Phyllotaxis Models on an Apex with Physically Based Growth 92

6.1 Physically-based models of the shoot apex . . . . . . . . . . . . . . . . . 92
6.2 A mass-spring model of the shoot apex with a cellular structure . . . . . 94
6.2.1 Local determination of apex and primordium centers . . . . . . . 96
6.2.2 Calculation of distance on the apex surface . . . . . . . . . . . . . 97
6.2.3 Physics simulation . . . . . . . . . . . . . . . . . . . . . . . . . . 98
6.3 Inhibition field model of phyllotaxis on a physically-based model of the
shoot apex . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
6.4 Transport model of phyllotaxis on a physically-based model of the shoot
apex . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
6.5 Summary of results of the phyllotaxis models combined with physically-
based shoot apex models . . . . . . . . . . . . . . . . . . . . . . . . . . . 102

7 Optimality in Phyllotaxis: A Light Capture Study 105

7.1 Phyllotaxis, light capture, and evolution . . . . . . . . . . . . . . . . . . 105
7.2 A simulation model of light capture . . . . . . . . . . . . . . . . . . . . . 105
7.3 Light capture simulation model results . . . . . . . . . . . . . . . . . . . 108
7.3.1 Light capture as a function of leaf width and divergence angle . . 109
7.3.2 Discretization error . . . . . . . . . . . . . . . . . . . . . . . . . . 110
7.3.3 The effects of the angle of incoming light . . . . . . . . . . . . . . 112
7.3.4 The effects of reduced numbers of leaves . . . . . . . . . . . . . . 113
7.4 Summary of results of the light capture models . . . . . . . . . . . . . . 116

8 Mechanistic Models of Leaf Venation 119

8.1 Models of leaf venation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
8.2 A growing leaf surface . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
8.3 Canalization on a cellular leaf structure . . . . . . . . . . . . . . . . . . . 124
8.3.1 Canalization on a growing leaf with discrete sources . . . . . . . . 126
8.3.2 Canalization on a growing leaf with emergent sources . . . . . . . 128
8.4 A combined model of auxin transport . . . . . . . . . . . . . . . . . . . . 129
8.5 Summary of results of leaf venation models . . . . . . . . . . . . . . . . . 132

9 Numerical Methods in VV 135

9.1 Types of problems arising in developmental biology . . . . . . . . . . . . 135
9.2 The VV programming environment . . . . . . . . . . . . . . . . . . . . . 136

v
 9.3 Forward Euler explicit method . . . . . . . . . . . . . . . . . . . . . . . . 137
9.4 Backward Euler and the Crank-Nicholson implicit methods . . . . . . . . 138
9.5 The steady-state solution of a morphogenic gradient . . . . . . . . . . . . 139
9.5.1 Solving for the steady-state . . . . . . . . . . . . . . . . . . . . . 140
9.5.2 Solving a linear system in VV with Gaussian elimination . . . . . 142
9.6 A reaction-diffusion system on a grid of cells . . . . . . . . . . . . . . . . 144
9.6.1 Specifying the model equations . . . . . . . . . . . . . . . . . . . 146
9.6.2 The Crank-Nicholson method combined with the Newton-Raphson
method for non-linear systems . . . . . . . . . . . . . . . . . . . . 147
9.6.3 An implementation of Newtow-Raphson in VV . . . . . . . . . . . 149
9.6.4 A banded linear system solver in VV . . . . . . . . . . . . . . . . 150
9.6.5 A biconjugate-gradient linear system solver in VV . . . . . . . . . 152
9.7 When are implicit methods worth it? . . . . . . . . . . . . . . . . . . . . 155

10 Conclusion 158
10.1 Discussion of research contributions . . . . . . . . . . . . . . . . . . . . . 158
10.1.1 Shoot apex model . . . . . . . . . . . . . . . . . . . . . . . . . . . 158
10.1.2 Model of a tissue with dividing cells . . . . . . . . . . . . . . . . . 159
10.1.3 Inhibition field model of phyllotaxis . . . . . . . . . . . . . . . . . 159
10.1.4 A transport-based patterning mechanism in plants . . . . . . . . . 160
10.1.5 A mechanistic model of phyllotaxis . . . . . . . . . . . . . . . . . 161
10.1.6 A physically-based apex model as a platform for the study of mor-
phogenesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162
10.1.7 A light capture study of phyllotaxis . . . . . . . . . . . . . . . . . 162
10.1.8 Exploration of mechanistic models of leaf venation . . . . . . . . . 163
10.1.9 Implementation of numerical algorithms in VV . . . . . . . . . . . 164
10.2 Future work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 164
10.3 Closing remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167

vi
 List of Tables

4.1 Inhibition threshold minimums. . . . . . . . . . . . . . . . . . . . . . . . 43

4.2 Details of the phyllotaxis types started with different initial conditions . 56

5.1 Divergence angles for the Fibonacci simulation on the cellular apex structure 80
5.2 Divergence angles for sensitivity analysis simulations . . . . . . . . . . . 86

9.1 List of common VV language constructs. . . . . . . . . . . . . . . . . . . 137

vii
 List of Figures

2.1 Relationship between visible parastichies and divergence angle for spiral
phyllotaxis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.2 Relationship between visible parastichies and divergence angle for Fi-
bonacci spiral phyllotaxis . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.3 Sample phyllotactic patterns. . . . . . . . . . . . . . . . . . . . . . . . . 17

3.1 Cylindrical model of a plant apex. . . . . . . . . . . . . . . . . . . . . . . 20

3.2 A schematic representation of the apex. . . . . . . . . . . . . . . . . . . . 22
3.3 Sample RERG functions. . . . . . . . . . . . . . . . . . . . . . . . . . . 23
3.4 Wireframe apex with the topology of a cylindrical grid. . . . . . . . . . . 24
3.5 Cell division in a tissue with cells represented by Voronoi regions . . . . . 26
3.6 Cell growth and division on the shoot apex. . . . . . . . . . . . . . . . . 27
3.7 Details of the model of cell division. . . . . . . . . . . . . . . . . . . . . . 28
3.8 Comparison of the effect of different cell division rules. . . . . . . . . . . 29

4.1 Diagram of inhibition. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35

4.2 Dynamics of the inhibiting field. . . . . . . . . . . . . . . . . . . . . . . . 36
4.3 Effects of changes in inhibition threshold. . . . . . . . . . . . . . . . . . . 38
4.4 Self-starting spiral Fibonacci pattern. . . . . . . . . . . . . . . . . . . . . 39
4.5 Comparison of divergence angles produced by the inhibition field simula-
tion model to divergence angles measured in Arabidopsis . . . . . . . . . 40
4.6 Different patterns propagated using the same parameter values. . . . . . 42
4.7 Examples of bijugate patterns. . . . . . . . . . . . . . . . . . . . . . . . . 44
4.8 Decussate patterns are not sustained. . . . . . . . . . . . . . . . . . . . . 45
4.9 Examples of phyllotactic patterns. . . . . . . . . . . . . . . . . . . . . . . 48
4.10 Examples of transitions in phyllotactic patterns. . . . . . . . . . . . . . . 49
4.11 Decussate pattern simulated in the presence of noise. . . . . . . . . . . . 51
4.12 Phyllotactic patterns propagated with identical model parameters in the
two-inhibitor model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55

5.1 Conceptual model for the regulation of phyllotaxis. . . . . . . . . . . . . 58

5.2 Time lapse model of transport on a line of cells . . . . . . . . . . . . . . 65
5.3 Pattern dependence on model parameters . . . . . . . . . . . . . . . . . . 66
5.4 Time lapse model of transport on a regular grid of cells . . . . . . . . . . 67
5.5 Model of a Lucas phyllotaxis pattern on a cylindrical apex . . . . . . . . 69
5.6 Phyllotactic patterns produced by the transport model. . . . . . . . . . . 73
5.7 Patterning on a cellular structure in the plane . . . . . . . . . . . . . . . 76
5.8 Apex with growing primordia . . . . . . . . . . . . . . . . . . . . . . . . 78
5.9 Types of phyllotaxis produced by the transport-based model on the cellular
apex structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
5.10 Superimposed image of the model and a real Arabidopsis shoot apex . . . 82

viii
5.11 Comparison of divergence angles produced by the transport-based simu-
lation model to divergence angles measured in Arabidopsis . . . . . . . . 83
5.12 Simulation of application of IAA to pin1 mutant apex . . . . . . . . . . 84
5.13 Simulation of cell ablation. . . . . . . . . . . . . . . . . . . . . . . . . . . 85
5.14 Fibonacci spiral phyllotaxis simulation on cellular apex structure . . . . . 89

6.1 Inhibition field model of phyllotaxis combined with a physically-based

model of a shoot apex . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
6.2 Transport model of phyllotaxis combined with a physically-based model
of a shoot apex . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103

7.1 Top and side views of a plant model. . . . . . . . . . . . . . . . . . . . . 106

7.2 Rendering of plant models used for light capture calculations . . . . . . . 108
7.3 Examples of plant models with different leaf width to length ratios. . . . 109
7.4 Light capture fitness landscape relating leaf width to divergence angle . . 110
7.5 Light capture as a function of divergence angle . . . . . . . . . . . . . . . 111
7.6 Light capture fitness landscape with incoming light at 15◦ . . . . . . . . 113
7.7 Light capture fitness landscape with incoming light at 30◦ . . . . . . . . 114
7.8 Light capture fitness landscape with incoming light at 45◦ . . . . . . . . 115
7.9 Light capture as a function of divergence angle for different numbers of
leaves . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
7.10 Light capture averaged over different numbers of leaves . . . . . . . . . . 117

8.1 Modeling a growing leaf by using Bezier surfaces . . . . . . . . . . . . . . 122

8.2 Cell division on a leaf modeled with Bezier surfaces . . . . . . . . . . . . 123
8.3 Favoritism of canalization towards shorter walls . . . . . . . . . . . . . . 125
8.4 Simulation of canalization on a growing leaf . . . . . . . . . . . . . . . . 127
8.5 Simulation of loop formation on a growing leaf . . . . . . . . . . . . . . . 128
8.6 Simulation of vein formation on a leaf with emergent auxin sources . . . 129
8.7 Dual transport model on a rectangular grid of cells . . . . . . . . . . . . 134

9.1 A gradient of morphogen in a line of cells . . . . . . . . . . . . . . . . . . 143

9.2 Reaction-diffusion simulation on a grid of cells . . . . . . . . . . . . . . . 156

ix
 Chapter 1

Introduction

1.1 Morphogenesis in plants

Morphogenesis is one of the most fundamental aspects of developmental biology. How

does a multi-celled organism create organized structures and differentiated, specialized

tissues with an identical set of instructions in every cell? A mouse embryo divided at the

two-cell stage can give rise to two identical adults, but yet these same two cells, if not

separated, are able to collectively coordinate their behavior, each becoming a founder

cell for only a portion of the adult organism. The concept of totipotence is even more

dramatic in plants, where an entire plant can be regenerated from a single cell [Takebe,

1971]. As this single cell divides, how does its progeny collectively decide which cells will

be shoot and which cells will be root? Furthermore, if explants containing more than

one cell are taken, how is it that these cells “know” about each other, and are able to

organize into the various specialized structures of a complete plant? Recent advances

in molecular biology have contributed greatly to our understanding of processes such

as gene transcription, protein synthesis, and cell to cell signaling, but exactly how these

processes result in the observed forms of the developing organism is still largely unknown.

Plants provide excellent systems in which to study morphogenesis. Unlike animals,

plants maintain embryonic cells throughout their lifetime, at the shoot tip, in structures

called shoot apical meristems [Lyndon, 1998]. The embryonic cells in the central zone of

the meristem are founder cells from which all of the aerial structures of the plant develop.

One of the first steps in the transformation of these plant “stem cells” into specialized

tissue, is the differentiation of cells into flower or leaf primordia in the peripheral zone

1
2

of the meristem. Since plant organs undergo little radial repositioning after initiation,

the locations specified at this early stage play an important role in the definition of

plant form. Although the entire peripheral zone of the shoot apex is competent to

initiate organs [Steeves and Sussex, 1989; Lyndon, 1998; Reinhardt et al., 2000], specific

locations are chosen in a predictable fashion, resulting in regular patterns of plant organ

positioning known as phyllotaxis. Striking examples of such patterns can be seen in the

intersecting rows of spirals of florets in a sunflower head, or scales of a pine cone.

Plants display a wide diversity in phyllotactic patterning, often within a single species

or even within the same individual. Some plants produce organs one a time with a

divergence angle between successive organs of 180◦ , as in most grasses, and others produce

spirals. Many plants start out producing decussate patterns, with pairs of horizontally

opposed organs appearing at a rotation of 90◦ from the previous pair. These plants often

switch to a spiral pattern later on in the vegetative stage or during the transition to

flowering.

Phyllotactic patterns have fascinating geometrical properties, such as the numbers of

intersecting visible spirals (parastichies) often being consecutive numbers of the Fibonacci

sequence. As the number of visible parastichies increases, the divergence angle between

successive primordia converges to 137.5◦ , the golden section of a circle. Mathematical

analysis has shown that higher order patterns, like those found in a sunflower head,

require very precise radial positioning [Adler, 1974], even down to a fraction of a degree.

How does the plant accomplish this?

The exploration, through computer simulation, of these and other questions in plant

morphogenesis is the focus of the work presented in this thesis.

 3

1.2 Why build simulation models of biological systems?

“You understand something only if you can program it.” - Gregory Chaitin [2006]

Computer simulation models can play an important role in the understanding of phyl-

lotaxis and other biological phenomena. Causal relationships governing plant develop-

ment are often not directly apparent through observation; we must rely on hypotheses

that are consistent with experimental results. Simulation models can be used to state

conceptual models precisely, and determine if they are plausible. Often the conceptual

model is incomplete, which becomes apparent during simulation model construction. The

simulation model requires all of the details to be addressed, raising questions which may

suggest directions for future experimentation in planta. Models can also be easily modi-

fied to test multiple variants of hypothesis, and parameter exploration may demonstrate

unforeseen properties of the proposed mechanisms.

1.3 Organization of thesis

A significant portion of this thesis is devoted to the use of simulation models to un-

derstand the phenomena of phyllotaxis. This begins in Chapter 2 with a brief history

of phyllotaxis and the introduction of terminology that will be used throughout the re-

mainder of this thesis. Chapter 3 deals with the structure of the shoot apex, and the

development of a simulation model for the apex that will be used by several of the phyl-

lotaxis models that follow. Chapter 4 presents the first simulation model of phyllotaxis,

a model that operates at a geometric level, using simple functions of distance and time

to abstract from the actual mechanism of organ positioning. In Chapter 5 a molecular-

level model of phyllotaxis is presented, that explains phyllotaxis in the context of recent

experimental work. Chapter 6 extends these models to a shoot apex with physically
4

based growth, and Chapter 7 explores phyllotaxis from an evolutionary perspective, with

light capture as the criteria for fitness. Since many of the molecular-level components

involved in organ initiation have also been implicated in the formation of procambial

tissue, mechanistic models of leaf venation are explored in Chapter 8. Numerical meth-

ods for biological simulations is the subject of Chapter 9 which is intended for the more

technical reader and requires a modest computer science or mathematical background.

This is followed by concluding remarks and a discussion of the research contribution of

this work in Chapter 10.

 Chapter 2

Phyllotaxis

2.1 A brief history of the study of phyllotaxis

Few patterns in nature are more conspicuous than the intersecting spirals of florets in a

sunflower head, or scales on a pine cone. Thus it is not surprising that phyllotaxis has

fascinated scientists from a variety of backgrounds for centuries. A considerable body

of literature exists in the study of phyllotaxis, with contributions from biologists, math-

ematicians, physicists, and computer scientists. Although much of the early literature

was limited to a descriptive characterization of phyllotactic patterns, the first hypotheses

on the underlying developmental mechanisms were proposed almost 140 years ago by

Hofmeister [1868]. Schwabe [1984] presents a concise review. With the increased avail-

ability of the digital computer in the 1970s, simulation models of phyllotaxis began to

appear.

According to Adler et al. [1997], systematic study of the geometry of phyllotaxis

began in the 1830s by Schimper, Braun, and the Bravais brothers. Schimper introduced

the notion of the genetic spiral, which traces the creation of all of the leaves sequentially

in time, although other spirals, called parastichies, are usually more visible. He used the

genetic spiral to define the divergence angle by following the path this spiral traces as it

winds around the stem when going from a leaf to the next leaf positioned directly above

it. The divergence angle is calculated as the number of revolutions this path makes

around the stem, divided by the number of successive leaves along the path. In this

way the divergence angle, measured in fractions of a circle, is always rational. Schimper

noticed that this fraction was usually of the form Fn /Fn+2 , where Fn is the nth term

5
6

in the Fibonacci sequence or some other sequence created by the same summation rule

with different initial terms. Working with pine cones, Braun added the notion of the

conspicuous parastichy pair, the two most visible spirals running in opposite directions.

He defined these to be the two sets of spirals formed by connecting each scale with its

closest left and right neighbors. Braun also noticed that Schimper’s rational divergence

angles were convergents1 of the continued fraction expansion for the golden ratio φ−1 ,

whose terms are all 1, marking the first appearance of the golden ratio in phyllotaxis

literature. The Bravais brothers extended this work by showing that Fibonacci numbers

of parastichy pairs are always generated if the divergence angle is close to the golden

angle. They also introduced the still often used idea of representing phyllotaxis as a

point lattice on a cylinder.

The problem of how to determine the divergence angle and vertical displacement of

successive leaves from the number of visible parastichies was first solved by van Iterson

[1907]. Erickson [1983] presents a comprehensive review of this work, with the essential

components presented by Prusinkiewicz and Lindenmayer [1990]. Figure 2.1 gives a

concise visual summary of this relationship for a cylindrical lattice. As the number

of visible parastichies increases, the range for possible values for the divergence angle

becomes progressively smaller. For example, (2,3) Fibonacci spiral phyllotaxis, which

has 2 visible parastichies running in one direction and 3 in the other, occurs within

a range of divergence angles of 1/6th of a circle. As the number of visible parastichies

increases, the range of allowable divergence angles is confined to a nested series of smaller

and smaller intervals about the golden angle φ (Figure 2.2).

1
A convergent of an infinite continued fraction is its expansion limited to a finite number of terms.
For example successive convergents of the infinite continued fraction:
1 1 2 3 5 8
φ−1 = 1 are (1, , , , , , . . . ).
1+ 1
1+ 1+...
2 3 5 8 13

.
 7

Figure 2.1: Relationship between visible parastichies (in brackets), the divergence angle
θ, and vertical displacement h for spiral phyllotaxis. Figure adapted from Prusinkiewicz
and Lindenmayer [1990].

Further work lead to the so-called Fundamental Theorem of Phyllotaxis, in which the

relationship between divergence angle and visible parastichies is shown to hold in both

directions [Adler, 1974; Jean, 1986, 1994]. This theorem states that a parastichy pair will

be visible and opposed if and only if the divergence angle of the system is within some

specified range given in the theorem.

Other authors who were interested in the characterization of phyllotactic patterns

include Church [1904] who defined the bulk ratio, and Richards [1948] who defined a

similar but more commonly used characterization called the plastochron ratio. This is

the ratio of distances of two successive leaves from the center of the apex. Richards [1951]
8

Visible
Divergence angle
parastichies

(2,3) [1 / 3 1 / 2]
(3,5) [1 / 3 2 / 5]
(5,8) [3 / 8 2 / 5]
(8,13) [3 / 8 5 / 13]
(13,21) [8 / 21 5 / 13]
φ

Figure 2.2: Relationship between visible parastichies and the divergence angle for Fi-
bonacci spiral phyllotaxis. As parastichy numbers increase, the divergence angle (shown
in fractions of a circle) is confined to a nested series of smaller and smaller intervals about
φ, the ideal angle.

proposed that only three parameters are required for a complete geometric description

of a phyllotactic pattern; the divergence angle, the plastochron ratio, and the angle of a

cone tangent to the apex at the organ initiating zone.

The most recent descriptive work involves the creation of computer models whose

primary focus is phyllotaxis pattern generation. Vogel [1979] proposed a method that

combined the explicitly assumed golden divergence angle of 137.5◦ with a formula which

he derived for the lateral displacement of successive organs. His formula works by relating

the area consumed on the disk-shaped receptacle by successive lateral displacements to

the average area occupied by a single organ. Ridley [1986] generalized this formula to

arbitrary surfaces of revolution and variable organ size. Fowler [1992] proposed another

model that also uses a specified divergence angle on an arbitrary surface of revolution. In

this model it is the collisions of primordia, rather than the average space they occupy on

the receptacle, that is used to determine the displacement along the profile curve used to

generate the surface. Both of these methods are suitable for the generation of a variety

of densely packed biological structures such as cacti, pineapples, flower capitula, pine
 9

cones, and spike inflorescences. They are useful for the construction of realistic computer

visualizations of plants, but they do not provide an explanation of the biological processes

that lead to the pattern formation.

Mechanistic theory of phyllotaxis starts with the work of Hofmeister [1868], who pro-

posed that new primordia appear periodically in time and are located as far as possible

from the boundaries of existing primordia. He proposed that this location is the point

of greatest elasticity in the shoot apex, and therefore the location where the next pri-

mordium will form. Snow and Snow [1932] suggested a modified version of Hofmeister’s

hypothesis, according to which new primordia appear when and where there is enough

available space. In support of their theory, the Snows did surgical interference experi-

ments on shoot meristems of Lupinus albus, some of which have recently been reproduced

in tomato by Reinhardt et al. [2005]. In these experiments the isolation of a primordium

from the rest of the meristem caused a change in the divergence angle between the next

two primordia. This shows that preexisting primordia have an influence the location of

subsequent primordia.

Schwendener [1878] proposed that contact pressure between adjacent organs causes

repositioning after their initiation. This idea was further developed by Adler [1974,

1977] who gave a proof that under certain assumptions, movement of organs due to

contact pressure can result in a jump to a higher order of phyllotaxis. This model is

further explored with simulation studies by Ridley [1982] and Hellwig et al. [2006] which

are able to show good convergence to the golden angle when primordia are initially

placed at a non-Fibonacci angle, or near the Fibonacci angle, but with considerable noise.

Experimental support for the idea of a “fine-tuning” mechanism for primordia positioning

after initiation is given by Reinhardt et al. [2005], although the actual mechanism may

be chemical in nature.

Schüepp [1916] proposed that even the initial primordia formation could be due to
10

physical forces in the form of tension in the apex caused by excessive growth of tunica.

Green [1980] also proposed a biophysical theory of primordia initiation due to surface

buckling from physical stresses in the apex. He created several simulation models [Green,

1992; Green et al., 1996] but was only able to achieve whorled phyllotaxis. A variant of

Greens model proposed by Shipman and Newell [2005] is able to generate spiral patterns.

Some experimental support for a phyllotaxis theory based on biophysical forces is

provided by experiments that induce changes in phyllotactic patterning by physically

constraining growing plant apices [Hernandez and Green, 1993]. Further support was

provided by Fleming et al. [1997] who were able to induce organ formation in tomato

by applying expansin, thought to increase cell wall extensibility, and thus locally change

the mechanical properties of the apex. In contrast, the work of Reinhardt et al. [2003a]

does not support a theory based on biophysical forces. They performed microsurgical

experiments in which the central zone or parts of the apical tunica were removed. They

suggested that such drastic operations would be expected to significantly change the stress

patterns in the tunica, however primordia were still initiated normally in the portions of

the meristem that were left intact.

To explain the initial positioning of organs missing from Schwendener’s theory, Schoute

[1913] proposed that a chemical substance released from the apex tip and from existing

primordia inhibits the formation of new primordia nearby. When the concentration of

this morphogen drops below a critical threshold in the peripheral zone, a primordium

is initiated. Simulation models based on this theory have successfully reproduced most

of the phyllotactic patterns observed in nature. Examples of these models are given by

Thornley [1975], Mitchison [1977], Young [1978], Schwabe and Clewer [1984], Chapman

and Perry [1987], Douady and Couder [1992, 1996a,b,c], and Yotsumoto [1993]. All of

these models use an inhibition function based on a steady-state approximation to diffu-

sion and decay of an inhibiting substance released by primordia. This is justified by the
 11

assumption that the production, diffusion, and decay of the inhibitor occurs at a much

quicker rate than growth. In some of these models, primordia are created at equal time

intervals representing the average plastochron, at the location of lowest inhibition. The

remaining models position primordia when and where the total inhibition from previous

primordia, and possibly the apex tip, drops below a threshold. This is analogous to the

difference between the theories of Hofmeister and the Snows, with a mechanism based

on an inhibiting substance determining space between organs.

Hellendoorn and Lindenmayer [1974] and Veen and Lindenmayer [1977] did not use

the steady-state assumption, but instead developed molecular-level models that simulate

the production, diffusion, and decay of an inhibitor. Their simulations were performed on

a cylindrical grid of cells, rather than in the more abstract continuous space used by the

models discussed in the previous paragraph. Both primordia, and the top row of cells that

represents the apex tip, produce the inhibitor and cells differentiate into primordia when

their inhibitor concentration drops below a threshold. Growth is modeled by adding rows

of new cells at the top of the cylinder. Under certain initial conditions, theses models

are able to propagate both whorled and spiral phyllotactic patterns.

The inhibition models discussed thus far are based on the assumption that cells dif-

ferentiate when an inhibitor drops below a threshold concentration, or that primordia

appear at equal time intervals at the location with the minimum concentration of the

inhibitor. These models also require the placement of one or more primordia as an initial

condition, to generate an initial profile of inhibitor concentration. No mechanism is given

for these processes, they are simply assumed. Turing [1952] proposed a mechanism called

reaction-diffusion that can explain both of these phenomena in what he envisioned as

a general paradigm for morphogenesis. Turing showed through a purely mathematical

analysis that a pattern of peaks in concentration of two or more chemical substances,

which he termed morphogens, could arise spontaneously in a ring of cells. Starting with
12

homogenous initial conditions, and using the same rules for changing morphogen concen-

tration in each cell, slight perturbations in the initial conditions due to noise could upset

an unstable equilibrium and lead to patterning.

Meinhardt and Gierer [1974] proposed reaction-diffusion equations which are more bi-

ologically plausible than those considered by Turing. One of their models is based on two

substances and is called an activator-inhibitor model. A substance called the activator

enhances its own production, as well as that of another substance, termed the inhibitor.

The inhibitor inhibits production of the activator. Such a system is easy to envision

as a feedback loop in a genetic regulatory network. Small local maxima in activator

concentration due to random variation lead to a local increase in production of both the

activator and the inhibitor. The inhibitor diffuses away more quickly than the activator,

reducing its effect on local activator self-enhancement, while suppressing activator self-

enhancement nearby. In a system of identical cells, each operating with identical rules,

this destabilization can lead to a spatial pattern of peaks in activator concentration,

which can trigger selective differentiation leading to patterning. By using interactions

of multiple substances, combined with multiple cascading interactions, this basic idea

has been used to account for a wide variety of patterning processes observed in nature

[Meinhardt, 1982; Meinhardt et al., 1998; Meinhardt, 2003a]. Meinhardt has proposed

this process as a general mechanism for both the establishment and interpretation of

Wolpert’s notion of positional information2 [Wolpert, 1969]. In the context of a growing

plant apex, the reaction-diffusion mechanism can explain both the emergence of organs

de novo as well as the thresholding mechanism of the inhibition model of phyllotaxis

[Meinhardt, 1982; Meinhardt et al., 1998; Meinhardt, 2003b].

2
Wolpert proposed that cells in an organism can determine their relative position within the organ-
ism with mechanisms such as gradients of morphogens. Reaction-diffusion systems give a mechanistic
explanation for how these gradients may be established. A reaction-diffusion system that establishes
such a gradient can also be coupled with one or more additional systems which can convert the gradient
into sharp boundaries. Thus reaction-diffusion systems provide a mechanistic explanation for both the
establishment and interpretation of positional information.
 13

As pointed out by Jönsson et al. [2006], all that is really needed for phyllotaxis is a

mechanism to explain the appearance of regularly spaced peaks of some organ-initiating

morphogen, in the blank canvas of undifferentiated cells provided by the central zone of

growing plant tip. Despite the fact that an explanation of phyllotaxis was a major mo-

tivation behind the development of the reaction-diffusion model by Turing, the work of

Reinhardt et al. [2000, 2003b] suggests that a different mechanism is involved. Reinhardt

et al. [2000] showed that the phytohormone indole-3-acetic acid (IAA, an auxin) could

initiate plant organs in the peripheral zone of a growing shoot apex, and was therefore

a likely candidate as the activator. This activator could provide the initial signal in the

cascade of reactions that lead to organ identity. In addition, the putative IAA efflux

carrier PIN1 was shown to localize within cells towards the sites of primordium initiation

[Reinhardt et al., 2003b], suggesting directed transport of IAA to these locations. This

led to a hypothesis according to which the peak in activator concentration that results

in organ initiation is due to transport of IAA from surrounding tissue, rather than local

self-enhanced production. Reinhardt et al. concluded their article by proposing a con-

ceptual model for phyllotaxis in which transport of IAA plays a key role in patterning.

Comparing with the activator-inhibitor model, activation is caused by the concentration

of the activator at organ initiation sites by the directional transport, rather than local

production. The inhibiting effect is caused by the draining of the activator from adjacent

tissue, rather than the production of an inhibitor.

A secondary role for physical forces is consistent with a transport-based model of

phyllotaxis. The vesicle trafficking mechanism involved in the relocation of PIN1 pro-

teins within a cell uses cytoskeletal components which are likely to be affected by physical

forces. This could provide a mechanism to reinforce the orientation of PIN1 proteins

towards the sites of newly forming primordia, whose rapid outgrowth would create con-

siderable strain in the cytoskeleton. In addition, after organ initiation, contact pressure
14

between adjacent primordia could be involved in fine-tuning organ positions. This seems

especially likely in densely packed structures, such as flower capitula, that contain higher

numbers of visible spirals which are known to require significant accuracy in the place-

ment of organs [Jean, 1994]. Even in the more loosely-packed meristems of tomato,

Reinhardt et al. [2005] gives experimental support that it is the nearest neighbors in

space, as well as time, that have an instructive role in organ positioning. However their

work does not answer the question as to whether the interaction between a primordium

and its nearest neighbors is physical or chemical in nature.

All of the models considered thus far assume that that phyllotactic patterning is the

result of processes occurring within the meristem itself. A different school of thought

proposes that phyllotaxis results from processes external to the meristem, and that pat-

tering is directed by the location of the vasculature below the meristem [Priestly et al.,

1935; Esau, 1942; Larson, 1975, 1983]. Several authors have reported experimental work

which does not support this hypothesis [Wardlaw, 1943; Snow and Snow, 1947]. Other

authors suggest that the causal relationship is in the other direction, and the phyllotac-

tic pattern influences the positioning of the vasculature [Wardlaw, 1950]. Either way,

there is convincing evidence that patterning of the stem vasculature is correlated to the

phyllotaxis in some species [Girolami, 1953; Kang et al., 2003].

Sachs [1991] linked many developmental events to a plant’s ability to transport sub-

stances that control differentiation, and there is experimental support for his hypothesis

that procambium tissue is induced by the canalization of preferred routes of auxin flux

[Sachs, 1981]. Sachs proposed that the flow of auxin through a cell increases the cells abil-

ity to transport auxin, thus drawing more auxin towards a preferred, self-enhancing path-

way or “canal”. The canalization hypothesis has been supported by several simulation

models [Mitchison, 1980, 1981; Feugier et al., 2005; Rolland-Lagan and Prusinkiewicz,

2005] that are able to select files of cells for differentiation into procambium from a
 15

homogenous tissue. Further experimental support for a directive role for auxin in vas-

culature development comes from experiments in which endogenous application of auxin

has been shown to induce additional procambium tissue [Sachs, 1981; Scarpella et al.,

2006]. [Scarpella et al., 2006] and Reinhardt et al. [2003b] showed that the auxin con-

centration peak that triggers organ initiation also induces formation of the midvein in

young primordia. In addition, the putative auxin efflux transport protein PIN1 is the

first known molecular marker for both plant organ primordia and pre-procambial tissue.

Thus phyllotaxis and vascular patterning are very closely related at the molecular level,

providing further support for the idea that the two processes may have directive roles on

each other.

It seems likely that the primary mechanism of organ positioning in the shoot meristem

is based on chemical morphogens, and is local to the meristem. Nevertheless, influences

by stem vasculature or biophysical forces could play secondary roles in reinforcing or

fine-tuning primordia position.

2.2 Phyllotaxis terminology

The terminology used to describe the arrangement of organ primordia on the growing

surface of the apex is based on the terminology used by Jean [1994] (Figure 2.3). Whorl

size or jugacy, denoted j, is the number of primordia emerging simultaneously. If pri-

mordia are issued one at a time (j = 1), the pattern is unijugate. For unijugate patterns,

the divergence angle is defined as the angle between consecutive primordia. The ver-

tex of this angle lies on the longitudinal axis of the apex. If the divergence angle is

equal to 180◦ , the phyllotactic pattern is distichous, otherwise it is spiral. The latter

term reflects the shape of conspicuous lines, or parastichies, formed by neighboring or-

gans. Typically, there are two sets of intersecting spiral parastichies, running in opposite
16

directions. The number of parastichies is a distinctive feature of the pattern. In Fi-

bonacci patterns, the numbers of opposite parastichies are consecutive elements of the

Fibonacci sequence < 1, 2, 3, 5, 8, 13, 21, . . . >. The first two elements of this sequence

are equal to 1, and each successive element is the sum of two previous elements. In the

more general case of Fibonacci-like summation sequences, each successive element is still

equal to the sum of previous ones, but the first two elements are different. A sequence

< 1, p, p + 1, 2p + 1, 3p + 2, 5p + 3, . . . > with p > 2 is termed the (p − 2)th accessory

sequence. The first accessory sequence < 1, 3, 4, 7, 11, . . . > is also called the Lucas se-

quence. An anomalous sequence is of the form < 2, 2q + 1, 2q + 3, 4q + 4, 6q + 7, . . . >,

where q > 1. Summation sequences beginning with integers greater than 2 are rare in

biology, and are collectively referred to as other sequences.

In multijugate patterns, the numbers of parastichies are multiples of those found in

the underlying unijugate spiral patterns. For example, in a bijugate Fibonacci pattern

(j = 2), the numbers of opposite parastichies are consecutive elements of the sequence

2× < 1, 2, 3, 5, 8, 13, 21, . . . >. Patterns with the parastichy numbers being consecutive

elements of the same sequence are of the same type. Within a type, a pattern with a

higher number of parastichies is said to have a higher order. In contrast to the multijugate

patterns, in whorled patterns new primordia appear in the centers of the spaces between

primordia of the previous whorl. Whorled patterns with the whorl size equal to 2 or 3 are

termed decussate and tricussate patterns, respectively. In the case of multijugate and

whorled patterns, the divergence angle is defined as the smallest angle between primordia

in successive whorls.

Each phyllotactic pattern can only occur if the divergence angle lies within some

interval, which is called the allowable interval for this pattern. For each family of patterns

(e.g., Fibonacci or Lucas patterns), the bounds of this interval tend to a limit value, called

the limit divergence angle, as the order of the pattern increases. Formulas for the bounds
 17

Figure 2.3: Sample phyllotactic patterns, arranged by jugacy j and divergence angle
θ. For spiral patterns, numbers of opposed parastichies are given in parentheses. Inte-
gers p and q characterize families of parastichies, for normal and anomalous phyllotaxis
respectively, corresponding to Fibonacci-like sequences defined in the text.

of intervals and limit divergence angles for typical phyllotactic patterns are presented by

Jean [1994](pp. 36-38).

 Chapter 3

Modeling the Shoot Apex

3.1 The structure of the shoot apex

The shoot apical meristem consists of one or more external layers of cells, called the

tunica, which surround the inner tissue known as the corpus. The outermost layer of

cells of the tunica, called the L1 layer, is the where the molecular processes leading to

phyllotaxis patterning are thought to occur [Reinhardt et al., 2003b; Smith et al., 2006a].

In dicotyledonous plants such as Arabidopsis, this layer of cells can be distinguished from

inner layers in that cell divisions are almost exclusively perpendicular to the surface

[Szymkowiak and Sussex, 1996]. For the purposes of simulation, this layer can be treated

as a curved surface of negligible thickness, which allows the problem of morphogenesis

to be simplified from three dimensions to two [Kramer, 2002].

The central zone of the shoot apical meristem consists of undifferentiated founder

cells, surrounded by a relatively narrow band of cells called the peripheral zone [Steeves

and Sussex, 1989; Lyndon, 1998; Kuhlemeier, 2007]. As the plant develops, the periph-

eral zone maintains an approximately constant distance from the tip of the apex for a

particular developmental stage of the plant. Only cells within the peripheral zone are

competent to initiate organs. In most of the simulations presented in this thesis, the

growth of cells in the central zone is assumed to be slower than in the peripheral zone,

as suggested by experimental data [Lyndon, 1998]. In some cases the notion of an active

ring is used as a convenient abstraction of the peripheral zone. This is defined to be

the line on the apex surface that encircles the apex and is located at the center of the

peripheral zone (Figure 3.2).

18
 19

Phyllotaxis is the result of the interaction between existing plant organ primordia

and the blank canvas of cells supplied by the central zone at the tip of the shoot apex.

This necessitates the construction of a dynamic model of a growing shoot apex in order

to study phyllotaxis through simulation.

3.2 A survey of shoot apex models used in phyllotaxis simula-

tions

Several topologies have previously been proposed to model the shoot apex in phyllotaxis

studies. Thornley [1975] models only the active ring as an abstraction of the peripheral

zone of the shoot apex. The effect of vertical displacement of primordia away from the

peripheral zone due to growth are simulated by introducing a decay of inhibitor concen-

tration with each plastochron. [Meinhardt, 2003b] employs a variation on this theme

where the destabilization of established peaks of activator concentration inherent in his

equations mimics this effect. Such a model can be considered as one-dimensional repre-

sentations of the apex, has the advantage of being simple to implement, and lends itself

well to mathematical analysis [Koch et al., 1994; Kunz, 2003]. d’Ovidio and Mosekilde

[2000] suggest that by reducing the problem of phyllotaxis to the unit circle and eliminat-

ing the expansion of the structure of the system caused by growth, the problem becomes

solvable within ordinary dynamical systems theory.

Most authors use two-dimensional models composed of simple geometric surfaces

of revolution, such as cylinders, disks, or cones, as an abstraction of the shoot apex

[Hellendoorn and Lindenmayer, 1974; Mitchison, 1977; Veen and Lindenmayer, 1977;

Young, 1978; Vogel, 1979; Meinhardt, 1982; Schwabe, 1984; Chapman and Perry, 1987;

Douady and Couder, 1992; Yotsumoto, 1993; Douady and Couder, 1996a,b,c; Meinhardt

et al., 1998; Hellwig et al., 2006]. These surfaces can capture the effects of growth more
20

Figure 3.1: (a) Cylindrical model of a plant apex often used in phyllotaxis simulations.
The cylinder is usually unrolled and presented as a grid with the left and right edges
connected. (b) In the case of cell-based models [Hellendoorn and Lindenmayer, 1974;
Veen and Lindenmayer, 1977; Meinhardt, 1982], growth is simulated by adding rows of
new cells at the top.

realistically, either by displacing primordia on the surface over time, or by adding rows

of cells at the top of the structure (Figure 3.1). Jönsson et al. [2006] uses a variation of

this approach by combining a cylinder and a hemisphere, which for cell-based models,

allows the possibility of communication across the top of the apex, a feature not present

in the models of Hellendoorn and Lindenmayer [1974], Veen and Lindenmayer [1977],

and Meinhardt [1982].

Ridley [1986] and Fowler [1992] present descriptive models that operate on arbitrary

surfaces of revolution. Growth is not considered in these models, which are motivated

largely from a pattern visualization perspective. These models are useful for creating

realistic computer images of flower capitula, fruit bodies, pine cones, and other similar

plant structures containing densely-packed phyllotaxis.

Physically-based apex models have also been used in phyllotaxis simulations by Jönsson

et al. [2006] and Smith [2006]. Both of these models have been implemented by using

mass-spring systems. The apex model of Jönsson et al. [2006] uses physically-based

growth, however all of the defining vertices of the apex are confined to a reference sur-
 21

face. In the model of Smith [2006], parameters controlling the physically-based growth

simulation determine both the shape of the apex and the primordia. In this sense pri-

mordia and apex shape are emergent properties of the model. Physically-based models

are often computationally expensive, as each step of growth involves solving a physical

system, such as a mass-spring system that may contain hundreds or even thousands of

springs. These systems often cannot be solved directly, and must be iterated to conver-

gence. Since shape is an emergent property, it can be difficult to find the correct growth

parameters leading to the desired shape of the both primordia and the apex itself. In

addition, simple geometric properties, such as the location of the center of the apex or a

primordium, can become quite difficult to determine. A more complete discussion of the

issues involved with physically-based models of the shoot apex is presented in Chapter

6.

Despite their drawbacks, physically-based models represent a significant step towards

a more mechanistic model of phyllotaxis and the shoot apex. Physically-based models

include the possibility of simulating the molecular mechanisms involved in the growth

process itself, which might be essential if phyllotactic patterning involves feedback with

growth. In addition, some authors propose a role for physical forces as both a primary

patterning mechanism [Green, 1992; Shipman and Newell, 2005] and as a means to fine-

tune primordia location after initiation [Ridley, 1986; Adler et al., 1997].

3.3 A shoot apex model with descriptive growth

With the exception of the physically-based model presented in Chapter 6, all of the

phyllotaxis simulations presented in this thesis share a common model for the shoot apex

and growth. This model can viewed as an extension of the model of Nakielski [2000] to

arbitrary surfaces of revolution.

22

Figure 3.2: A schematic representation of the apex. The apex shape is defined by a
B-spline generating curve (inset, shown with the control polygon), rotated around the
longitudinal axis of the apex. Sample point P specified by coordinates (θ, a) moves away
from the apex tip with velocity v(a).

The shoot apex is modeled as a surface of revolution, generated by rotating a planar

curve around the longitudinal axis of the apex (Figure 3.2). This planar curve is a B-

spline [Foley et al., 1990] defined interactively using a graphical editor, and can easily

be changed to model apices with various profiles of their central longitudinal sections. A

point P on the apex surface is represented by two coordinates (θ, a) where θ is the angle

of rotation around the axis of the apex, measured with respect to a reference direction,

and a is the distance from the apex tip, measured along the generating curve on the apex

surface. In Cartesian coordinates, the position of point P is thus given by:

(x, y, z) = (x(a) cos(θ), y(a), x(a) sin(θ)). (3.1)

 23

Figure 3.3: Sample plots defining relative elemental rate of growth RERG as functions of
distance a from the apex tip. The functions are defined graphically, using an interactive
B-spline editor. (a) Constant growth function. (b) Function with an increased growth
rate at the active ring.

Individual points move away from the apex tip as a result of growth, while the over-

all shape of the apex remains unchanged. This motion is characterized by a function

RERG(a), [Richards and Kavanagh, 1943; Erickson and Sax, 1956; Silk and Erickson,

1979; Hejnowicz and Romberger, 1984] which defines the relative elemental rate of growth

in the longitudinal direction (along the generating curve) at a distance a from the apex

tip [Hejnowicz et al., 1984; Nakielski, 2000]. The velocity with which a point P = (θ, a)

moves away from the apex tip along the generating curve is then given by the integral:
Z a
v(a) = RERG(a)da. (3.2)
0

Similar to the generating curve, the growth function RERG(a) in the model de-

scribed here is defined graphically, which makes it easy to specify various distributions

of growth on the surface of the apex. Following experimental observation [Lyndon, 1998;

Kwiatkowska and Dumais, 2003; Reddy et al., 2004], the RERG distribution for most

of the simulations presented in this thesis is chosen such that the growth in the central

zone is slower than in the peripheral zone (Figure 3.3).

24

Figure 3.4: Wireframe apex with the topology of a cylindrical grid. All of the vertices
on the top row of the cylinder occupy the same location in space, at the apex tip.

3.4 Cells and cell division

For some models of phyllotaxis, such as the inhibition field model described in Chapter

4, any type of mesh can be used to represent the apex surface, since it is only required

for visualization purposes. However, models operating at the level of cells require a

representation capable of simulating both cell to cell communication and a convenient

method of cell division.

In the simplest case, as in the models of Hellendoorn and Lindenmayer [1974], Veen

and Lindenmayer [1977], and Meinhardt [1982], cells are modeled as squares in a grid

with the left and right edges of the grid connected to form the topology of a cylinder. For

such a simple topology, any convenient data structure can be used. Cells do not divide

per se, but are simply added at the top of the mesh to simulate growth (Figure 3.1).

A variation on this theme can be used to model cells on the reference surface described

in Section 3.3. A regular square grid is embedded in the surface giving the appearance of

a realistic apex shape, but maintaining the topology of a cylinder (Figure 3.4). All of the

points in the top ring of the cylindrical grid share the same spatial coordinates, that is
 25

(θ, a) = (0, 0) for all of the vertices in the top ring. Beginning with a pair of initial rings

of vertices, growth causes all of the vertices except those in the top ring to move down

the apex surface, with a new ring or vertices inserted when the distance between rings

becomes too large. Although this type of subdivision surface is not ideal for cell-based

simulations, it does provide a suitable mesh for apex visualization.

Map L-Systems [Lindenmayer and Rozenberg, 1979] and cell systems [de Boer et al.,

1992] are extensions of L-Systems to two dimensional structures that have been used

to model sheets of cells. Both cell systems and the implementation of map L-Systems

presented by Prusinkiewicz and Lindenmayer [1990] rely on physically-based mass-spring

simulations to produce realistically shaped cells, and are thus not directly applicable to

the purely descriptive growth model described in Section 3.3.

Voronoi diagrams have also been used to model cells [Honda, 1978, 1983] because

of their cell-like appearance. They are not ideal for modeling growing plant tissues,

since the rearrangement of Voronoi regions that inevitably occurs during cell division

creates unrealistic motions of cell walls (Figure 3.5). An example where this problem

occurs can be seen in the phyllotaxis simulation of Jönsson et al. [2006](see supplemental

video). Such motions are not possible in plants due to the rigid extra-cellular matrix

that surrounds plant cells.

Nakielski [2000] presents a model of both cell division and growth of the shoot apex.

A projection of the shoot apex onto the plane makes it possible to consider the apex

as an expanding disk, with all points except the center moving radially outward due to

growth. Cells are modeled as polygons, with the position of cell vertices changing over

time as a result of apex growth. Cell division occurs when the cell size (polygon area)

reaches a threshold value. Two methods are presented for how to determine the dividing

wall. In the first method, cells are divided parallel to one of two principal directions

of growth. In the second method, the dividing wall is chosen so that it passes through
26

Figure 3.5: Cell division in a tissue with cells represented by Voronoi regions. Unrealistic
changes in the cells’ geometry are introduced during cell division. Cell before division
(left), and adjusted regions after division (right). Note the shortening and stretching of
some of the neighbor cell walls, and the occupation of some of the space of the neighbor
cells by the newly created daughter cells.

the center of the cell and is perpendicular to the closest wall to the center. The result is

then projected back onto the parabolic apex surface, producing a realistic looking cellular

apex. Nakielski’s model uses radially uniform growth, which produces good results for

the apices of spruce seedlings that are analyzed in detail in the final section of Nakielski

[2000].

The model of cells and cell division used for the simulations presented in this thesis

is based on an implementation of Nakielski [2000] described by Smith [2006]. The model

presented by Smith [2006] was extended as follows:

• The model was extended to arbitrary surfaces of revolution. This allows the mod-

eling of apices of arbitrary shape.

• The ability for the user to specify a growth function was added to the model. This

is required to simulate the Arabidopsis apex, where the relative elemental rate of

growth changes with distance from the tip [Kwiatkowska and Dumais, 2003].
 27

underlying polygon mesh

visualized surface

Figure 3.6: Cell growth and division on the shoot apex. Growth causes points to move
down the reference surface and polygonal cells to enlarge. Cell division occurs when cells
reach a threshold area.

• The model was extended to three dimensions, with all calculations done on the apex

surface directly, rather than on a disk and then projected back onto the surface.

This eliminates distortions in cell shapes at the periphery of the apex due to the

projection.

• The rules to avoid four-way junctions described in Nakielski [2000] were not imple-

mented in the model described by Smith [2006]. These were added to the model.

• The model by Smith [2006] divided cells by using one of the approaches described

by Nakielski [2000]. The algorithm finds the closest point to the center on any wall.

The cell is then divided at the line passing through this point and the center. These

rules were modified to divide the cell at the shortest wall that passes through the

center [Errera, 1888]. This produces more realistic cell shapes, and in particular
28

Figure 3.7: Details of the model of cell division. (a) A mother cell before division. (b)
The tentative dividing wall is the shortest wall passing through the centroid of the mother
cell. (c) The endpoints of the tentative wall are displaced from any preexisting vertices
of the mother cell in order to avoid 4-way junctions. (d) The form of daughter cells is
adjusted by shortening the dividing wall.

produces fewer “sliver” shaped cells that have one or more vertices with strongly

acute angles (Figure 3.8).

Cells are defined as polygons embedded in the reference surface described in Section

3.3. As the vertices defining the polygons move down the apex surface due to growth,

the polygons expand (Figure 3.6). When a threshold area is reached the cells divide.

By default, the dividing wall follows Errera’s rule and divides the cell at the location

of the shortest wall that passes through the centroid of the cell polygon [Errera, 1888].

The position of the dividing wall may be adjusted to avoid four-way junctions, which are

unusual in plant tissue [Flanders et al., 1990; Goodbody et al., 1991; Dumais, 2007]. To

produce more realistic cell shapes, the cell is “pinched” slightly by moving the vertices

of the dividing wall slightly towards each other (Figure 3.7d).

The dynamics of cell division and the resulting cellular patterns (Figure 3.8b) are

similar to those observed in the Arabidopsis meristem [Kwiatkowska and Dumais, 2003;

Reddy et al., 2004]. Thus, the model is provides an adequate structural support for

modeling phyllotaxis at the molecular level.

 29

Figure 3.8: Comparison of the effect of different cell division rules. (a) Apex with cell
division using the closest wall algorithm. (b) Apex with shortest wall algorithm produces
more realistic cells with less “sliver” shaped cells.

3.5 VV and the L-Studio modeling environment

All of the simulations in this thesis were programmed in C++ using the VV modeling

environment [Smith et al., 2004]. VV allows the local, index-free specification of prob-

lems that can be represented as graphs with the geometry of a two-dimensional surface

embedded in three-dimensional space1 . In a very general sense, VV can be seen as an

extension of L-Systems [Lindenmayer, 1968a,b; Prusinkiewicz and Lindenmayer, 1990],

which are suitable for modeling one-dimensional branching structures, such as plants.

Like L-Systems, the index-free nature of VV allows it to handle problems in which both

the state and the structure of the system are changing over time. Such systems are called

dynamic systems with dynamic structure [Giavitto and Michel, 2001].

Although the models described in this thesis could have been implemented with a

C++ graph library, or by using one of the data structures commonly used in computer

graphics applications for handling polygonal meshes, VV offers several advantages:

• The VV language extensions to C++ offer a much more convenient way to specify
1
In VV it is possible to represent arbitrary graphs, however it has an especially convenient represen-
tation for orientable two-dimensional surfaces.
30

meshes, vertices, and vertex neighborhood relationships than using class libraries.

In addition to this convenience, all of the functionality of C++ is available.

• Everything within VV is specified in local terms, as there is no mechanism for global

access to mesh vertices. This requires the programmer to think in local terms, and

to specify vertex and neighborhood information locally.

• The VV system comes with a built-in viewer and support libraries that simplify

many graphics operations while still allowing full access to OpenGL. Event man-

agement, such as the handling of mouse clicks and movement, is taken care of by

the VV simulator.

• Some very useful high-level commands specific to the VV language are available that

operate on vertices and meshes. An example of this is the “synchronise” command

which allows access to the previous state of vertex data variables and connectivity

information. This can be very useful during mesh subdivision, or when modeling

coupled systems of differential equations.

• VV can operate within the L-Studio [Prusinkiewicz, 2004] modeling environment,

and is able to use all of L-Studio’s standard graphical editing tools, such as the

function editor, the contour editor, the Bezier surface editor, and the materials and

color palette editors. L-Studio also contains an object manager that provides a nice

graphical interface to organize model files.

VV statements are embedded within C++ code and the VV translator converts them

to standard C++ . Thus VV can be seen as an embedding of VV language statements

within C++ . This is contrast to the L+C language [Karwowski and Prusinkiewicz, 2003],

which allows C++ statements to be embedded within the declarative language of L-

Systems. Despite their differences, they do share a common modeling environment,

 31

L-Studio [Prusinkiewicz, 2004] on WindowsTM , or Vlab on LinuxTM . In both systems the

code of the simulator is fully compiled, and the model is loaded as a dynamic library. This

methodology not only reduces compile time, but it also maintains a very clear distinction

between the program code of the model and the simulator.

 Chapter 4

A Model of Phyllotaxis Based on Inhibition Fields1

4.1 Field models of phyllotaxis

The postulate that existing primordia inhibit the formation of new primordia nearby is

fundamental to most mechanistic theories of phyllotaxis. The mechanisms that have been

proposed include contact pressure [Schwendener, 1878], diffusion of an inhibitor [Schoute,

1913], reaction-diffusion [Turing, 1952; Meinhardt, 1982], surface buckling [Green, 1992],

and the polar transport and depletion of an activator [Reinhardt et al., 2003b]. De-

spite the diversity of these mechanisms, they all share the notion of an inhibition field

surrounding existing primordia. This general notion raises two types of questions: (a)

What molecular-level processes may produce such inhibition fields? and (b) What is

the relationship between spatial and temporal properties of the inhibition fields and the

generated patterns?

The simulation model presented below was constructed to address the second ques-

tion, with the objective of finding the best inhibition field functions according to the

following criteria:

• The model can generate a wide variety of phyllotactic patterns.

• Patterns can start de novo, in an empty peripheral zone, or from one or two cotyle-

dons.

• The model can capture transitions in phyllotaxis, such as the often observed tran-

sition from decussate to spiral patterns [Wardlaw, 1968].

1
A substantial portion of this chapter is based on Smith et al. [2006b].

32
 33

• Phyllotactic patterns are initiated and propagated in a robust manner. This im-

plies, in particular: (a) low sensitivity to changes in the model parameter values,

(b) tolerance to low cell counts in the peripheral zone, limiting the precision with

which individual primordia can be placed, and (c) tolerance to random factors

(noise) that may affect the placement of individual primordia.

It is also desirable that the inhibition functions depend on plausible parameters, such

as distance and time, although no hypothesis is made regarding the specific molecular-

level processes that may yield these functions. The investigation is limited to models in

which the angular position of primordia does not change over time. This is consistent

with the current experimentally-based view of phyllotaxis [Reinhardt et al., 2003b].

Previous simulation models have been focused on formulas derived from an assumed

mechanism, the diffusion of an inhibiting substance being the most common [Hellendoorn

and Lindenmayer, 1974; Thornley, 1975; Veen and Lindenmayer, 1977; Mitchison, 1977;

Young, 1978; Meinhardt, 1982; Schwabe and Clewer, 1984; Chapman and Perry, 1987;

Yotsumoto, 1993; Meinhardt, 2003a]. In a steady-state approximation of the concentra-

tion of a diffusing and decaying substance [Thornley, 1975; Mitchison, 1977; Young, 1978;

Yotsumoto, 1993], the inhibiting influence of each primordium decreases exponentially

with distance. In a system containing n primordia, the combined effect h on a sampling

point S can be calculated as the sum:

n
X
h(S) = e−b·d(Pi ,S) (4.1)
i=1

where d(Pi , S) is the distance between primordium Pi and a sampling point S on the apex

surface, and b controls the rate of exponential decrease in inhibition with the distance

from the primordium.

A different formula, motivated by physical experiments in which phyllotactic patterns

were generated in a magnetic field, was proposed by Douady and Couder [1992, 1996a,b,c]
34

With this formula, inhibition decreases with the distance from the source according to

the power function:

n
X 1
h(S) = (4.2)
i=1
d(Pi , S)b

where d(Pi , S) is defined as in Equation (4.1), and b controls the rate of inhibition decrease

with the distance from a primordium.

Formulas (4.1) and (4.2) can generate phyllotactic patterns, however in simulations

conducted as part of this work they showed limitations. Parameters could not be found

to produce some patterns, such as higher-order accessory patterns (p > 6), while other

patterns and their transitions only occurred in very narrow ranges of parameter values.

These limitations are difficult to quantify because of the number of parameters involved.

The results of simulations depend not only on the inhibition fields under investigation,

but also on the assumed shape of the apex, its growth pattern, the position of the

peripheral zone, the initial distribution of primordia, and the manner in which all these

characteristics and parameters change over time. Nevertheless, in a series of interactive

experiments with simulation models it was found that robust de novo generation of a

variety of phyllotactic patterns and their transitions was much easier when the inhibition

fields depended both on the spatial arrangement of existing primordia and on their age.

4.2 Organization of computation

Like previous models of phyllotaxis based on inhibition fields, the simulations presented

in this chapter are based on the idea that the existing primordia exert an inhibiting

influence on the incipient primordia (Figure 4.1). The combined influence of all primordia

constitutes the inhibiting field. The field values are calculated at equally spaced sampling

points on the active ring, unless calculation of the field on the entire apex surface is

required for visualization purposes (Figure 4.2). The number of sampling points can be
 35

Figure 4.1: Diagram of inhibition. (a) The older primordium has a smaller inhibiting
effect on the active ring than the newer primordium. Arrows indicate minima of inhibi-
tion, where a new primordium can appear. (b) The upper minimum of inhibition was
chosen at random as the location of the third primordium.

chosen to approximate the number of cells in the peripheral zone of a particular species,

or can be made much larger to simulate pattern formation in continuous space. The

position of sampling points can be randomly perturbed in both in the circumferential and

the longitudinal directions, which is useful when studying the robustness of phyllotactic

pattern formation.

Simulation of phyllotaxis proceeds in a sequence of time steps. At each step, points

on the apical surface, including existing primordia, are moved away from the apex tip

according to their velocities (Equation 3.2). The inhibition from previous primordia is

then calculated for all sampling points on the active ring. If the field value at one or

more of these points drops below the inhibition threshold, a new primordium is inserted

at the sampling point with the lowest inhibition. If there are two or more points with the

same minimum inhibition, one of them is chosen at random. The inhibition values on the

active ring are then recalculated taking the influence of the newly created primordium

into account. If one or more sampling points are still below the inhibition threshold,

an additional primordium is inserted at a point of minimum inhibition, as described

before. The process is repeated until all sampling points on the active ring are above the
36

(a) (b) (c) (d)

(e) (f) (g) (h)

Figure 4.2: Dynamics of interaction between the inhibiting field and phyllotactic pattern
formation. The field was calculated using Equation (4.3). (a) The field generated by
two symmetrically placed initial primordia. (b) The field shortly before the insertion
of the third primordium. Arrows indicate minima of the inhibition field. The location
indicated by black arrow is chosen at random over the location indicated by grey arrow
as the location of the third primordium. (c-e) The field before the insertion of the fourth,
fifth, and sixth primordium. (f-h) The steady-state dynamics of the inhibiting field. (f)
The field immediately after the insertion of primordium n. (g) The field immediately
before the insertion of primordium n + 1. At the time of insertion of primordium n + 1
(longest arrow), approximate positions of incipient primordia n + 2 and n + 3 are already
visible as smaller local minima of the inhibition field (shorter arrows). (h) The field
immediately after the insertion of primordium n + 1. Note that figures (f) and (h) differ
only by rotation of approximately 137.5◦ .

inhibition threshold. In the subsequent simulation steps, all primordia move away from

the active ring as a result of the apex growth. This movement, combined with a possible

decrease of the inhibiting influences of primordia with their age, reduces the inhibiting

field strength on the active ring. Over time, this inhibition drops sufficiently at some

location to allow for the formation of another primordium, and the process repeats.
 37

4.3 Single inhibition function

The first simulation model uses a single inhibition function. In an apex with n previously

formed primordia, the total inhibiting effect h(S) of previous primordia is calculated as

the sum:
n
X 1
h(S) = e−bti (4.3)
i=1
d(Pi , S)

where d(Pi , S) is defined as in Equation (4.1), ti is the age of primordium i, and b controls

the rate of exponential decrease in inhibition over time. The essential feature of this

formula is the explicit dependence of the field on both the distance from the primordia and

their age. By experimenting with various decreasing functions of distance in simulations,

it was found that the specific form of the distance dependence is not critical, although

the inversely proportional function shown by Equation (4.2) leads to slightly more robust

results than the exponential dependence given in Equation (4.1). Note that, according

to Equation (4.3), the unit value of inhibition h is defined as that produced by a newly

formed primordium (ti = 0) at the unit distance from the primordium center (d(Pi , S) =

1).

The pattern-generating capability of the model that uses the inhibition function of

Equation (4.3) was examined from two perspectives: generation of patterns de novo

(including generation of sequences of patterns, with transitions caused by changes in

parameter values), and perpetuation (maintenance) of patterns initiated with a sequence

of pre-placed primordia.

4.3.1 Pattern generation de novo

In these simulations, no primordia, and thus no inhibition field, were present at the be-

ginning of the simulation. The generating algorithm located the first primordium at a

random position on the active ring (simulations with a predefined position of one ini-
38

Figure 4.3: Effects of changes in inhibition threshold. At low inhibition thresholds, a

self-starting distichous phyllotaxis is produced. Increases in the threshold values cause a
switch to spiro-distichous, and then to spiral Fibonacci patterns.

tial primordium produced the same results). Simulations with high initial values of the

inhibition threshold often produced apparently unorganized arrangements of primordia,

before settling into a pattern. In order to prevent these unorganized arrangements from

occurring, while focusing on the more biologically relevant situation where patterns begin

with the arrangement of cotyledons, the inhibition threshold value was phased in grad-

ually, increasing from an initial value of zero to the final value used in the simulation.

This increase led to transitions in the phyllotaxis type or order.

For example, in the simulation shown in Figure 4.3, the initial value of the threshold

for primordium differentiation was relatively low, yielding a distichous pattern. After an

increase in the threshold value, a spiro-distichous pattern emerged. Continued increases

in the threshold value caused a switch to Fibonacci spiral phyllotaxis. A similar pro-
 39

Figure 4.4: Self-starting spiral Fibonacci pattern created by using an active ring with
38 sampling points. (a) The initial primordium is arbitrarily placed by the model. (b)
The second primordium appears in the area of least inhibition, at a divergence angle of
180◦ from the first. At this point there are two minima of inhibition on the active ring,
both of which are closer to the older of the two existing primordia (arrows). (c) The
third primordium appears at one of these minima, the choice determining the direction
of the spiral. The divergence angle is equal to 114◦ . (d-f) Successive divergence angles of
161◦ , 133◦ , and 142◦ ensue. The model will then continue at 142◦ generating (3, 5) spiral
phyllotaxis.

gression of patterns could be obtained when changing the size of the apex while keeping

the threshold value constant. This is biologically relevant, since changes in the apex size

are known to correlate with changes in phyllotaxis patterns [Kwiatkowska and Florek-

Marwitz, 1999; Jackson and Hake, 1999].

A simulation that generates a self-staring Fibonacci pattern is shown in Figure 4.4.

This simulation was performed using 38 sampling points on the active ring, a number

chosen to be comparable to the number of cells around the peripheral zone of an Ara-

bidopsis vegetative shoot apex (cf. Figure 2 in Kwiatkowska [2006]). The increase in
40

Figure 4.5: Comparison of divergence angles produced by the inhibition field simulation
model to divergence angles measured in Arabidopsis [Smith et al., 2006a] for the first
10 primordia. The active ring has 38 sampling points, which limits the accuracy of the
placement of primordia to 9.5◦ increments.

the inhibition threshold was faster than in the previous example, and spiral phyllotaxis

is established quickly. Under these conditions, the model produced a sequence of diver-

gence angles that are within the standard error of angles measured during the initial

establishment of a spiral pattern in Arabidopsis (Figure 4.5).

4.3.2 Pattern propagation

Due to various factors, an initial pattern of primordia positions may be different than

that obtained in self-starting simulations described above. These factors include the in-

fluence of cotyledons on the initial state of the simulation, limited accuracy of primordium

placement caused by the relatively small number of cells around the circumference of the
 41

peripheral zone, and other random factors (noise).

Consequently, perpetuation of patterns is also of biological interest. The perpetuation

of patterns was investigated by initiating the simulation with a number of primordia

placed at the limit divergence angle characteristic to a given pattern type.

The model based on Equation (4.3) robustly perpetuated all of the phyllotactic pat-

terns commonly observed in nature, in which primordia appear one at a time. This

includes distichous and spiro-distichous, as well as Fibonacci, Lucas, anomalous, and

other spiral patterns. In all simulations, the opposing parastichy numbers were consec-

utive numbers of some Fibonacci-like summation sequence. High-order patterns, such

as (89, 144) Fibonacci spirals, could be perpetuated if the number of sampling points in

the active ring was sufficiently large to represent a divergence angle within the allowable

interval for the given pattern (the size of this interval decreases as the order of the pat-

tern increases). Occasionally, however, regular phyllotactic patterns could be generated

even when the number of sampling points was smaller; in these cases, the divergence

angle would oscillate and its average value would lie within the bounds (see Section 4.5

discussing pattern robustness for an example).

In a series of simulation experiments it was observed that Fibonacci spiral patterns

were perpetuated under the widest range of model parameter values, followed by Lucas

patterns, and patterns from other Fibonacci type sequences. This is consistent with

observations in nature, where most single spiral patterns come from the Fibonacci, the

Lucas, or the < 2, 5, 7, 12 . . . > anomalous sequence [Zagórska-Marek, 1985; Jean, 1994].

Table 4.1 shows the minimum values of the inhibition threshold for which these patterns

can be perpetuated, assuming constant values of other parameters. There is no maximum

value of this threshold, as it can be raised indefinitely, yielding a sequence of patterns of

increasingly high orders.

Table 4.1 indicates that, if the inhibition threshold is greater than or equal to 128,
42

Figure 4.6: Different patterns propagated using the same parameter values. These simula-
tions differ only by the divergence angle θ between the 20 initial primordia. (a) θ = 77.96◦
yields a 2nd accessory (9, 14) spiral pattern. (b) θ = 99.50◦ yields a Lucas (7, 11) spiral
pattern. (c) θ = 106.4◦ yields a (7, 10) spiral (other) pattern. (d) θ = 137.51◦ yields a Fi-
bonacci (8, 13) spiral pattern. (e) θ = 151.14◦ yields an anomalous (7, 12) spiral pattern.
(f) θ = 68.76◦ and jugacy j = 2 yields a bijugate Fibonacci (6, 10) spiral pattern.

six different patterns can be propagated depending on the placement of initial primordia.

These patterns are illustrated in Figure 4.6. The possibility of propagating such a wide

variety of patterns for the same parameter values was an unexpected result of simula-

tions. It may explain why different types of phyllotaxis may occur in the same species: if

the initial placement of primordia is affected by random factors, different patterns may

emerge and be perpetuated. For example, causal observation of pine cones collected from

the trees on the University of Calgary campus shows that they usually have Fibonacci

spiral phyllotaxis, but many of them have Lucas spiral phyllotaxis, and the two types

often coexist on the same tree. Flowers and inflorescences with a wide range of phyl-
 43

Phyllotaxis Divergence Minimum

Jugacy Sequence
type angle (◦ ) threshold

Fibonacci 1 1, 2, 3, 5, 8, 13, . . . 137.51 4

Lucas 1 1, 3, 4, 7, 11, 18, . . . 99.50 22
Bijugate 2 2, 4, 6, 10, 16, . . . 68.76 31
Anomalous 1 2, 5, 7,12, 19, 31, . . . 151.14 37
Other 1 3, 7, 10, 17, 27, . . . 106.45 65
2nd accessory 1 1, 4, 5, 9, 14, 23, . . . 77.86 128

Table 4.1: Minimum values of the inhibition threshold needed to maintain selected phyl-
lotactic patterns.

lotactic patterns were observed, for example, in Magnolia by Zagórska-Marek [1994], in

Helianthus annuus by Couder [1998], and in Araceae by Jean and Barabé [2001].

The number of initials required to start a pattern varied considerably. While patterns

from the main Fibonacci sequence were easily generated de novo, Lucas patterns required

at least three initials. The anomalous (7, 12) pattern could be started with as few as six

initials placed at the angle 151.14◦ . Patterns derived from accessory sequences with

n > 3 could also be propagated, but required a higher number of carefully placed initial

primordia. For example, 40 initials placed at 64.08◦ were required to start and propagate

(11, 17) phyllotaxis derived from the third accessory sequence (n = 5). These high

numbers of initial primordia indicate that the inhibiting fields required to propagate the

corresponding patterns may emerge in nature (for example, due to a confluence of random

factors), but the probability is small. This is consistent with Jean’s [1994] summary of

frequency data, indicating that phyllotactic patterns derived from various accessory and

anomalous sequences do occur in nature, but are rare.

Once a low-order spiral pattern is established, an increase in the inhibition thresh-

old can result in the pattern switching to higher orders of phyllotaxis. This occurs for

Fibonacci spiral patterns, as well as patterns from accessory, anomalous, and other se-
44

Figure 4.7: Examples of bijugate 2 × (5, 8) patterns. (a) The pattern generated using
the single inhibitor model. Although parastichies are clearly visible, the positioning of
individual primordia is slightly irregular due to the delay in the production of the second
primordium in each pair. (b) The regular pattern generated by the variant of the single
inhibitor model, with explicitly imposed jugacy j = 2. The pattern generated by the
two-inhibitor model is similar.

quences, provided that the increase is not sudden and the number of sampling points

on the active ring is large enough to support the higher order pattern. Sharp increases

in inhibition threshold in accessory, anomalous, and other spiral patterns produce less

predictable results. In some cases the pattern would become disorganized, or switch to

a Fibonacci spiral phyllotaxis was observed. Stable, although not very regular, bijugate

patterns were often generated as well. Parastichy numbers produced by these patterns

were always a multiple (×2) of consecutive elements of the Fibonacci sequence.

In bijugate patterns, pairs of primordia are expected to appear simultaneously, yet

in the above simulations this is not the case. The first primordium of each pair has an

immediate inhibiting effect on the entire ring, which delays the appearance of the second

primordium. As a result of this delay, the bijugate patterns are slightly irregular (Figure

4.7).
 45

Figure 4.8: Decussate patterns are not sustained because of the immediate inhibiting ef-
fect of new primordia. Arrows indicate minima of inhibition, where the third primordium
can appear. (a-d) Consecutive stages of pattern formation.

4.4 Two inhibition functions

In general, the single-inhibitor model defined by Equation (4.3) is not suitable for gen-

erating patterns in which two or more primordia appear simultaneously. Figure 4.8

illustrates this limitation by using a decussate pattern as an example. Suppose the first

two primordia have been placed 180◦ apart, and the inhibition levels at the entire active

ring are too high for additional primordia to appear. Over time, the inhibiting effect of

these initial primordia decreases as a result of apex growth, and two identical inhibition

minima appear in a perpendicular orientation on the active ring (Figure 4.8a). When

the inhibition level at these minima falls below the threshold, one minimum is selected

by the placement algorithm as the location of the third primordium (Figure 4.8b). The

immediate inhibiting influence of this primordium pushes the other minimum above the

threshold, delaying the appearance of the fourth primordium until the apex grows further

(Figure 4.8c). As a result, the third primordium has a weaker inhibiting effect on the

active ring than the more recently formed fourth primordium, and the fifth primordium

is positioned asymmetrically (Figure 4.8d), causing a switch to spiral phyllotaxis.

A variant of the single inhibition model was investigated, in which jugacy j is spec-

ified as a model parameter. Inhibition is calculated with Equation (4.3), considering

46

possible positions of j evenly spaced locations on the active ring. When the sum of in-

hibitions at these locations drops below the inhibition threshold, j primordia are placed

simultaneously. With this strategy, the model is able to robustly produce decussate,

whorled, bijugate, and multijugate patterns. However, the assumed placement algorithm

is difficult to justify in mechanistic terms.

The inability of the model summarized by Equation (4.3) to create patterns in which

two or more primordia appear at once is a consequence of two factors: the spatially

unlimited inhibiting influence of each primordium, and the immediate effect of each

primordium on the entire active ring. In reality, it is unlikely that the full inhibiting

effect of a primordium will be felt the instant a primordium appears. This is especially

true for sampling points located some distance away from the new primordium. In order

to delay the inhibiting effect, two modifications to the above model were introduced.

First, the inhibiting influence of a primordium on the active ring is phased in gradually,

using the following modification of Equation (4.3):

n
−bd ti

X 1
hl (S) = 1.0 − e · e−bl ti . (4.4)
i=1
d (Pi , S)

As in Equation (4.3), bl is the coefficient for exponential decay of inhibition, d(Pi , S) is

the distance between sampling point S and primordium Pi , and ti is the age of primordium

Pi . The additional parameter bd controls the rate at which the inhibiting influence of the

primordium is phased in. According to Equation (4.4), a newly placed primordium does

not immediately increase the inhibition levels at other locations, making it possible for

primordia at other minima to appear concurrently. This sets the stage for the formation

of whorled and multijugate patterns.

The problem with the above mechanism is that the inhibition threshold can be reached

at a series of neighboring locations in quick succession, leading to the formation of closely

spaced primordia. To prevent this from happening, a second inhibiting function was
 47

introduced as follows:
n
X 1
hs (S) = e−bs ti (4.5)
i=1
d (Pi , S)

where parameters d(Pi , S) and ti are as above, and bs controls the rate of exponen-

tial decay of the second inhibition over time. A new primordium is formed when both

inhibiting influences are below predefined threshold values. Equation (4.4) is used to

express long-range inhibition that determines the phyllotactic patterning of primordia,

and Equation (4.5) is used to express short-range inhibition that prevents creation of

multiple primordia at adjacent locations on the active ring. The short-range inhibition

decreases over time faster than the long-range inhibition (bs > bl ), and thus does not

interfere with the phyllotactic positioning of primordia, which remains governed by the

long-range inhibition.

4.4.1 Pattern generation de novo

The model with two inhibition functions (two-inhibitor model for short) can create de

novo the distichous, spiro-distichous and Fibonacci spiral patterns produced by the sin-

gle inhibitor model, as well as decussate, spiro-decussate, and bijugate patterns (Figure

4.9). Experimenting with model parameters showed that Fibonacci spiral patterns are

generated most often, as in the case of the single-inhibitor model. In contrast to the

single-inhibitor model, however, the second most easily produced spiral patterns are bi-

jugate patterns, with parastichy numbers being two-fold multiples of the main Fibonacci

sequence, 2× < 1, 2, 3, 5, 8, 13, . . . >. The next most frequently observed spiral patterns

correspond to the Lucas sequence. These frequencies are consistent with the experimental

data presented by Zagórska-Marek [1985] and Jean [1994](pp. 148-151).

Many plants undergo changes in phyllotaxis during development. Transitions often

occur quickly, so that a new pattern is established within the span of a few plastochrons.
48

Figure 4.9: Some examples of phyllotactic patterns produced by the two-inhibitor model
model: (a) distichous, (b) decussate, (c) spiro-decussate, (d) whorled, (e) Fibonacci
spiral, (f) Lucas spiral, (g) bijugate. Patterns (a-c), (e) and (g) have been generated de
novo, patterns (d) and (f) have been propagated by the model from a predefined initial
pattern.

These transitions can be simulated using the two-inhibitor model by manipulating the

long-range and short-range inhibition thresholds during simulation. For example, Figure

4.10 illustrates the transition from decussate to spiral phyllotaxis, frequently observed

in nature [Wardlaw, 1968; Carpenter et al., 1995; Kwiatkowska, 1995, 1997], and the

transition from decussate to bijugate phyllotaxis. The model can also simulate other

transitions, including distichous to Fibonacci spiral, whorled to Fibonacci or Lucas spiral,

whorled to multijugate, and multijugate to Fibonacci spiral.

 49

Figure 4.10: Examples of transitions in phyllotactic patterns generated by the two-in-

hibitor model. (a) Decussate to spiral. (b) Decussate to bijugate.

4.4.2 Pattern propagation

The model with two inhibition functions can propagate the same patterns as the single

inhibition model, as well as decussate, spiro-decussate, bijugate, multijugate, and whorled

patterns (Figure 4.9 and 4.12). As in the case of the single-inhibitor model, different

patterns can be propagated by the model using the same parameter values. The range of

patterns that can be propagated this way is surprisingly large and includes, for example,

the entire inventory of single and multijugate spiral patterns reported by Zagórska-Marek

[1985]. Figure 4.12 shows the simulation of 25 different patterns propagated by using the

same parameters, with the details of the phyllotaxis types given in Table 4.2. These

are by no means all of the patterns that can be produced by the model; other patterns

can easily be produced with different parameters. For example, decussate and tricussate

patterns are missing, although they are not difficult to produce with the model. These
50

lower-order whorled patterns require a higher threshold for the short-range inhibition

function, so that only 2 or 3 primordia are produced at the same time, and a lower

long-range inhibition threshold to give sufficient spacing between whorls.

4.5 Robustness of the models

Although exceptions exist [Zagórska-Marek, 1994], phyllotactic pattern formation in na-

ture is usually very robust, which means that patterns are formed or maintained under a

wide range of conditions. Plausible models of phyllotaxis should therefore generate stable

patterns or sequences of patterns (including transitions) over a wide range of parameter

values.

An important parameter of the models presented above is the inhibition threshold

at which new primordia are formed. Both the single-inhibitor and two-inhibitor mod-

els produce identifiable phyllotactic patterns over a wide range of long-range inhibition

thresholds. Variation of this parameter during simulation induces transitions in phyl-

lotaxis type or order (Figures 4.3 and 4.10).

The two-inhibitor model also includes a threshold for short-range inhibition. Values

of this parameter are more critical then those of the long-range inhibition threshold.

For example, in the case of decussate patterns, the short-range threshold values can in

some cases only be changed by approximately ±2% without affecting the pattern. This

sensitivity can be understood in the context of the role of the short-range inhibition

in multijugate pattern formation: the threshold values must be low enough to effect

suppression of adjacent primordia, yet high enough to allow for proper positioning of

other members of the whorl. The parameters bd and bs , which control delay of long-range

inhibition and decrease of short-range inhibition over time, can be manipulated within

ranges exceeding ±20% of the mean value.

 51

Figure 4.11: Side and top views of a decussate pattern simulated in the presence of noise.
Sampling points on the active ring were perturbed at random both in the longitudinal
and circumferential directions. Increasing the amplitude of the noise would cause a switch
to spiral phyllotaxis.

The parameter b (bl in the two-inhibitor model), which controls the exponential decay

of inhibition with the age of primordia, also has a significant effect on pattern formation.

Non-zero values of this parameter make it easier to initiate patterns de novo, allow for

faster transitions between patterns, and make it possible to maintain patterns for wider

ranges of apex shapes and other parameter values. On the other hand, the models are

capable of generating regular phyllotactic patterns even as b or bl approach zero. In this

case, the inhibition exercised by a primordium (Equation 4.3 or 4.4) no longer depends

on its age, but decreases only with distance as in the simulations performed by Douady

and Couder [1992].

Recognizable phyllotaxis (determined by examination of divergence angles) is gener-

ated by the models for various apex shapes, such as disk, cylinder, cone, hemisphere, or

arbitrary surfaces of revolution. Both the single-inhibitor and the two-inhibitor models

are also robust with respect to changes in the RERG function. Nevertheless, the shape

of the apex, its growth rate, the size and shape of primordia, and the number of sampling
52

points on the active ring can have drastic effects on which parastichy pairs, if any, are

clearly visible [Schwabe, 1984](p 410).

Both models are robust with respect to noise. For example, decussate phyllotaxis,

which has a strong tendency to break symmetry and switch to spiral phyllotaxis, is

nevertheless stable in the presence of noise (Figure 4.11).

The last parameter considered in this sensitivity analysis is the number of sampling

points on the active ring. With n sampling points, the divergence angles can be repre-

sented with the resolution of 360◦ /n, which limits the precision with which primordia can

be placed on the active ring. Low-order phyllotactic patterns can easily be generated even

for small numbers of sampling points, in the range of 10 to 20. Higher-order patterns,

or patterns from less common accessory sequences, require larger numbers of sampling

points. Nevertheless, the model is robust enough to generate some high-order patterns

even for relatively small numbers of sampling points. For example, an active ring with

19 sampling points makes it possible to represent divergence angles with the resolution

of 360◦ /19 ≈ 18.9◦ . The closest two angles to the limit Fibonacci angle of 137.5◦ are

7 × (360◦ /19) ≈ 132.6◦ and 8 × (360◦ /19) ≈ 151.6◦ . Both of these angles are outside the

allowable interval [135◦ , 144◦ ] of divergence angles for (5, 8) spiral phyllotaxis. Nonethe-

less, the model can create a rough (5, 8) spiral pattern by alternating between 132.6◦ and

151.6◦ as follows: 151.6◦ − 132.6◦ − 132.6◦ − 151.6◦ − 132.6◦ − 132.6◦ − 151.6◦ . . . The

average divergence angle over many primordia is 138.9◦ , which lies within the interval

[135◦ , 144◦ ].
 53

4.6 Summary of results of the inhibition field model of phyl-

lotaxis

Phyllotaxis is an example of an emergent phenomenon, in which properties of the whole

pattern result from interactions between individual elements of the pattern, yet the causal

link between these interactions and the pattern is not obvious. A simulation model was

built to examine the inhibiting effect that the existing primordia may have on the place-

ment of new primordia, both in generation of phyllotactic patterns and their transitions.

It was found that:

• A wide class of patterns can be robustly generated de novo or indefinitely propa-

gated if the inhibiting effect of an existing primordium depends both on the age

of a primordium and on the distance between the primordium and a point on the

apex.

• Different phyllotactic patterns can be propagated by models using the same simu-

lation parameters.

• Multijugate and whorled patterns, in which two or more primordia appear at once,

can be robustly generated if the inhibiting effect of each primordium is not imme-

diate, but gradually phased in.

• An additional short-range inhibiting mechanism is then needed to prevent formation

of series of primordia in neighboring locations.

The cumulative impact of existing primordia on the apex was summarized in terms of

a field that assigns a value of the inhibition to each point on the apex. The inhibition field

represents a useful level of abstraction, which makes it possible to analyze and visualize

the dynamics of interactions between primordia without knowing the specific mechanism
54

by which these interactions are implemented in nature. Nonetheless, general properties

of the inhibition field point to the properties that a molecular-level mechanism may have

in order to produce these patterns. For example, simulations suggest that an optimal

contribution of a primordium to the field decreases with the inverse of distance. Such a

decrease does not need to be of a diffusion-decay type postulated in several earlier models

[Thornley, 1975; Mitchison, 1977; Young, 1978; Yotsumoto, 1993]. This is consistent with

the current view that directed transport of auxin plays an important role in the formation

of phyllotactic patterns [Reinhardt et al., 2003b; Barbier de Reuille et al., 2006; Jönsson

et al., 2006; Smith et al., 2006a].

It was also observed that the simulation models initiate patterns more easily, are

capable of effecting more rapid transitions, and are generally more robust, when the in-

hibiting field depends on the age of primordia. This is consistent with the idea that the

differentiation and growth of primordia may have a direct impact on the phyllotactic pat-

terns. Furthermore, short-range inhibition is required to form multijugate and whorled

patterns in the model. Such a mechanism may be related to the establishment of organ

boundaries and organ separation in nature. These properties of the model are consistent

with the molecular-level model of phyllotaxis proposed by Smith et al. [2006a] described

in the next chapter, which required the introduction of primordium differentiation and

organ boundaries in order to simulate the observed phyllotactic patterns.

 55

Figure 4.12: A large variety of phyllotactic patterns are propagated with identical model
parameters using the two-inhibitor model. The most conspicuous parastichy pairs are
indicated in brackets. See Table 4.2 for details of the type of phyllotaxis, divergence
angle, and jugacy.
56

Phyllotaxis Divergence Initial Figure

Jugacy Sequence
type angle (◦ ) whorls 4.12

6th accessory 1 1,8,9,17,26,. . . 41.77 40 (8,9)

5th accessory 1 1,7,8,15,23,. . . 47.26 40 (7,8)
3rd accessory 1 1,5,6,11,17,. . . 64.08 40 (6,11)
Other (5,11) 1 5,11,16,27,. . . 66.89 40 (5,11)
2nd accessory 1 1,4,5,9,14,23,. . . 77.96 20 (5,9)
Other (4,9) 1 4,9,13,22,35,. . . 82.15 40 (9,13)
1st accessory (Lucas) 1 1,3,4,7,11,18,. . . 99.50 20 (7,11)
Other (3,7) 1 3,7,10,17,27,. . . 106.45 20 (7,10)
Other (3,8) 1 3,8,11,19,30,. . . 132.18 20 (8,11)
Main (Fibonacci) 1 1,2,3,5,8,13,. . . 137.51 20 (8,13)
1st anomalous 1 2,5,7,12,19,. . . 151.14 30 (7,12)
2nd anomalous 1 2,7,9,16,25,. . . 158.14 20 (7,9)
3rd anomalous 1 2,9,11,20,31,. . . 162.42 20 (9,11)

Bijugate 2nd accessory 2 2,8,10,18,28,. . . 38.98 10 (8,10)

Bijugate Lucas 2 2,6,8,14,22,. . . 49.75 10 (6,8)
Bijugate Fibonacci 2 2,4,6,10,16,26,. . . 68.75 10 (6,10)
Bijugate 1st anomalous 2 4,10,14,24,38,. . . 75.57 10 (4,10)

Trijugate Lucas 3 3,9,12,21,33,. . . 33.17 10 (9,12)

Trijugate Fibonacci 3 3,6,9,15,24,. . . 45.84 10 (6,9)

Tetrajugate Fibonacci 4 4,8,12,20,32,. . . 34.38 5 (8,12)

Pentajugate Fibonacci 5 5,10,15,25,40,. . . 27.50 5 (5,10)

Six-jugate Fibonacci 6 6,12,18,30,48,. . . 22.92 5 (6,12)

Whorled (6) 6 6,6 30.00 7 (6,6)

Whorled (8) 8 8,8 22.50 4 (8,8)
Whorled (10) 10 10,10 18.00 2 (10,10)

Table 4.2: Details of the phyllotaxis types shown in Figure 4.12. All simulations used
identical model parameters, but were started with different initial conditions. Initial
primordia were placed using the specified jugacy and divergence angle. The total number
of initial primordia is given by jugacy×initial whorls. For each simulation shown in Figure
4.12, the total number of initials is less than one half of the primordia shown.
 Chapter 5

A Transport-Based Model of Phyllotaxis1

5.1 Mechanistic theories of phyllotaxis

Although several mechanistic theories for phyllotaxis were proposed prior to Reinhardt

et al. [2003b], experimental evidence to support these theories has been lacking. The idea

that diffusing inhibitors operate to suppress organ formation in the otherwise competent

tissue of the peripheral zone is almost 100 years old, but yet no such inhibitors have been

found. Likewise, experimental work casts doubt on theories based on physical forces as

the primary positioning mechanism. Arabidopsis meristems that have had a substantial

part of their tunica removed, or the central zone ablated, are still able to initiate organs

in the remaining portions [Reinhardt et al., 2003a]. Such operations are bound to cause

substantial changes in the stress patterns in the tunica, yet they do not seem to have

significant influence on organ positioning.

Reaction-diffusion systems have also been proposed as a mechanism for many pattern-

ing processes in both plants and animals [Turing, 1952; Meinhardt, 1982]. An example of

such a system is Meinhardt’s activator-inhibitor model of phyllotaxis [Meinhardt et al.,

1998] which is able to create the evenly spaced peaks in activator concentration required

to produce phyllotactic patterns on a growing apex. Although at first auxin may appear

to be good candidate for the activator, it is unlikely that the peaks in auxin concentration

that initiate organ formation are a result of local self-enhanced production as specified in

the activator-inhibitor model. Experimental results suggest that it is the concentration of

auxin from surrounding tissue due to transport, rather than local production, that creates
1
A substantial portion of this chapter is based on Smith et al. [2006a].

57
58

Figure 5.1: Conceptual model for the regulation of phyllotaxis by polar auxin fluxes in
the shoot meristem [Reinhardt et al., 2003b]. (a) PIN1 orientation directs auxin fluxes
(arrows) in the L1 layer, leading to accumulation of auxin (red color) at the initiation
site (I1) in the peripheral zone of the meristem, and eventually inducing organ initiation.
(b) Later, basipetal PIN1 polarization inside the bulging primordium (P1) drains auxin
into inner layers, depleting the neighboring L1 cells. As a consequence, another auxin
maximum is created at a far position (I1) in the peripheral zone. Figure courtesy of
Soazig Guyomarc’h.

developmentally instructive auxin maxima. When growing shoot apices are cultivated in

the presence of auxin transport inhibitors, the induction of lateral organs is blocked, and

the apices grow as radially symmetric structures [Reinhardt et al., 2000]. Application of

auxin to such pin-shaped meristems induces lateral primordia, with the size and position

depending on the concentration and the position of the applied auxin. Furthermore, the

cellular distribution and subcellular localization of the auxin efflux transport facilitator

protein PIN1 is consistent with a role in organ positioning [Reinhardt et al., 2003b].

On the basis of experimental data, Reinhardt et al. [2003b] proposed a conceptual

model of phyllotaxis (Figure 5.1) based on the transport of the organ initiating activator,

auxin. According to this model, auxin is transported acropetally towards the meristem,

where it is redirected to the primordia, that function as sinks. As a result, auxin is

depleted from the surroundings of the primordia, and reaches the organogenetic periph-
 59

eral zone only at a certain minimal distance from the two youngest primordia (P1 and

P2). Auxin accumulates at this position, where it induces a new primordium (incipient

primordium I1) which, in the course of the plastochron, grows out and becomes a sink

itself. The phyllotactic pattern thus results from the dynamics of interaction between

existing and incipient primordia in a growing apex, mediated by the directional transport

of auxin.

The mechanism proposed by Reinhardt et al. [2003b] is plausible in the sense that it

is consistent with the available molecular data and captures qualitatively the inhibitory

effect of the existing primordia on an incipient primordium. However, the question of

whether it is indeed capable of generating the highly constrained geometry of spiral phyl-

lotactic patterns was open. To answer this question, a simulation model was constructed

based on data concerning the induction of primordia by high auxin concentrations and

the polar localization of the auxin transport facilitator PIN1 in the surface layer of the

apex. The model shows that the molecular mechanisms identified by Reinhardt et al.

[2003b] are indeed plausible and can lead to the formation of the phyllotactic patterns

observed in nature.

5.2 Molecular basis for a transport-based model of phyllotaxis

The mechanistic model of phyllotaxis presented here is based on the following hypotheses

consistent with experimental results:

• Only the peripheral zone is competent to induce lateral organs. Auxin

applied to the peripheral zone of pin-like inflorescence meristems induces organ formation

in a dose-dependent fashion [Reinhardt et al., 2000, 2003b]. This happens in the pin-like

meristems of the pin1 and pid mutants, as well as in pin-like meristems created by using

auxin transport inhibitors, such as N-1-naphthylphthalamic acid (NPA). Application of

60

auxin to pin-like meristems above or below the peripheral zone either has no effect, or

induces organ formation in the peripheral zone, close to where the auxin was applied

[Reinhardt et al., 2000, 2003b].

• Organ initiation occurs at sites of high auxin concentration. Various micro-

application experiments in which auxin is applied to meristems of Arabidopsis show that

auxin is capable of initiating lateral organs in both shoots and roots [Reinhardt et al.,

2000, 2003b; Benková et al., 2003]. Studies performed using Arabidopsis plants containing

the DR5::GFP auxin reporter construct, which is thought to reflect endogenous auxin

concentrations [Benková et al., 2003; Casimiro et al., 2001], suggest that convergence

points of high auxin concentration in the meristem are correlated with the sites of organ

initiation [Smith et al., 2006a].

• Peaks in auxin concentration in the meristem result from auxin redistri-

bution by PIN1-mediated transport. Arabidopsis inflorescence meristems contain-

ing the DR5::GFP reporter construct treated with the auxin transport inhibitor NPA do

not show a significant reduction of DR5::GFP expression. However, the phyllotactic pat-

tern of DR5::GFP peaks observed in non-treated meristems disappears and is replaced

by a ring of expression [Smith et al., 2006a]. This suggests that the auxin peaks which

initiate organs are formed as a result of PIN1-mediated transport. Further evidence that

PIN1-mediated transport of auxin is required to create discrete points of auxin conver-

gence is suggested by the work of Reinhardt et al. [2003b]. Auxin applied to the central

zone of pid mutants, which contain PIN1 proteins, results in a groups of 3 or 4 separated

organs in the peripheral zone. On the other hand, auxin applied to pin1 mutants under

the same conditions results in the formation of a ring-shaped organ. A ring-shaped organ

also induced by the application of 2,4-D, thought to be a poor substrate for the PIN1

auxin efflux carrier [Delbarre et al., 1996], to pid mutant meristems. Taken together

these results point to a necessary role for PIN1 in the transport of auxin to discreet
 61

convergence points in the shoot apex.

• Auxin is readily available throughout the meristem. According to the Rein-

hardt et al. [2003b], auxin is not produced in the meristem itself, but imported from more

basal tissues. Transgenic plants containing the DR5::GFP auxin reporter construct in the

pin1 mutant background show extremely low expression of DR5::GFP in the Arabidopsis

inflorescence meristem [Smith et al., 2006a]. This depletion of auxin is likely due to two

factors: (a) the reduction of auxin transport into the meristem due to the lack of PIN1

proteins in the pin1 mutant, and (b) a reduction in auxin production in tissues basal to

the meristem because of a lack of lateral organs, the presumptive sites of auxin synthesis

[Reinhardt et al., 2003b]. As mentioned previously, NPA-treated plants show a ring of

DR5::GFP expression instead of the peaks of expression observed at the sites of incipient

primordia in untreated plants. NPA appears to be able to reduce transport sufficiently

to suppress the formation of auxin peaks in the meristem without, at least in the short

term, significantly reducing the overall supply [Smith et al., 2006a]. Note that the ring

of DR5::GFP expression is relatively uniform in NPA-treated apices. This suggests that

the mechanism providing auxin to the meristem does so in a radially symmetric fashion,

and likely does not influence patterning to any great extent.

• Phyllotactic patterning occurs in the L1. The PIN1 protein is located primar-

ily, although not exclusively, in the external L1 layer [Benková et al., 2003; Reinhardt

et al., 2003b; Smith et al., 2006a] of the shoot apex. In addition, studies on plants

containing DR5::GFP also indicate that auxin is predominately located in the L1 layer

[Smith et al., 2006a]. This suggests that phyllotactic patterns may essentially be formed

on the surface layer of cells in the shoot apical meristem. It is possible that auxin is

retained in the L1 by auxin import carriers, most importantly AUX1, which is strictly

localized to the L1 layer [Reinhardt et al., 2003b].

• PIN1 expression is upregulated by auxin. [Vieten et al., 2005] showed that

62

auxin regulates the expression of PIN proteins in the Arabidopsis root. This suggests

that auxin may regulate the expression of PIN1 in the shoot meristem as well [Peer et al.,

2004]. Analysis of transversal sections of shoot meristems from Arabidopsis shows that

DR5::GFP expression peaks in incipient primordia [Smith et al., 2006a]. In contrast, after

treatment with sirtinol, thought to bypass auxin transport [Dai et al., 2005], all cells of

the L1, including those in the central zone of the meristem, express DR5::GFP at equal

levels [Smith et al., 2006a]. This suggests that all cells in the L1 are auxin-responsive

and that within the L1, the DR5::GFP expression levels are likely reflecting endogenous

auxin concentrations. Since the maxima of PIN1 expression are also located at the sites

of incipient primordia [Reinhardt et al., 2003b; Benková et al., 2003], it is assumed that

auxin positively regulates PIN1 expression in the meristem.

• PIN1 transport facilitator proteins are polarized by auxin. Given the role of

PIN proteins as facilitators of polar transport, their localization in the cell is of primary

importance. In the L1, PIN1 is polarized towards the incipient primordia [Reinhardt

et al., 2003b], in which auxin concentration appears to reach a maximum. On this basis,

in the simulation model that follows, it is hypothesized that PIN1 in the L1 is prefer-

entially polarized towards the neighboring cells with the highest auxin concentration.

Although this assumption has no direct experimental foundation, the molecular mech-

anism is likely to involve two elements: (a) the auxin-modulated endocytosis of PIN1

[Paciorek et al., 2005], and (b) a mechanism that informs cells about auxin concentration

in their neighbors, be it through measuring extracellular auxin or via a receptor-ligand

system.

• Incipient primordia acquire new developmental identity. Cells in the pe-

ripheral zone differentiate to become either primordium cells or intervening tissue. In

the context of the simulation model, the primordia cells may have different parameters

compared to the intervening regions.

 63

• Auxin is exported from primordia to subepidermal layers. Strong PIN1

expression in the primordia [Reinhardt et al., 2003b] coincides with the strands of high

expression of the DR5::GFP reporter construct in the center of primordia, which persist

during organ outgrowth and become connected to the central vasculature [Smith et al.,

2006a; Mattsson et al., 2003; Kang and Dengler, 2002]. This suggests that primordia act

as sites of auxin export from the L1 to the central vasculature. The capacity of AUX1 to

retain auxin in the L1 may be exceeded at I1, where it ”escapes” into internal layers and

induces organ identity and organ outgrowth [Reinhardt et al., 2003b]. Thus primordia

are assumed to act as sinks of auxin in the L1 in spite of the high concentration and,

possibly, production of auxin in the primordia.

These experimental results form the basis for a mechanism generating phyllotaxis

where transport plays the central role. This is fundamentally different from reaction-

diffusion models [Turing, 1952; Meinhardt et al., 1998] where local self-enhanced pro-

duction of a morphogen, rather than directed transport, is responsible for the peaks in

morphogen concentration that lead to tissue patterning.

5.3 The transport-based patterning mechanism

In order to study the transport-based patterning mechanism it is convenient to consider

a topology less complex than a growing plant apex, with as few assumptions as necessary

to capture the essence of the mechanism. A simple example is a single line of cells,

connected at the ends to form a ring. This is one of the topologies originally considered

by Turing [1952] in his analysis of reaction-diffusion systems. If growth is not considered,

then even production and decay of IAA and PIN1 can be ignored, as only the dynamics

of the transport mechanism itself are required for patterning.

The first simulation model of transport-based patterning considered is implemented

64

on a ring of cells, with no growth or cell division, and with all cells are equal in size.

Each cell stores an IAA concentration, a PIN1 concentration, and the orientation of PIN1

transport facilitator proteins to its left and right neighbors. The simulation begins with

each cell containing an initial IAA concentration, which changes over time due to diffusion

and PIN1-mediated transport to and from neighbor cells. There is no production or decay

of IAA. There is also no production or decay of PIN1, with each cell being assigned an

initial concentration that does not change throughout the simulation. On the other hand

PIN1 orientation, and thus IAA transport, does change. PIN1 proteins are assumed

to be preferentially located on the membranes facing neighbor cells with higher IAA

concentration according to the formula:

[IAA]j
[P IN ]i→j = [P IN ]i P (5.1)
[IAA]k
k∈Ni

where [P IN ]i→j is the number of PIN1 proteins located in the membrane of cell i facing

neighboring cell j, [IAA]j is the concentration of IAA in neighbor cell j, [P IN ]i is the

total concentration of PIN1 proteins in cell i, and Ni are the neighbors of cell i. The

orientation of PIN1 towards neighbor cells given by Equation (5.1) is a linear function of

IAA concentration, although higher powers of concentration or an exponential function of

concentration will also produce patterning on a line of cells. In the phyllotaxis simulations

considered in later sections, an exponential function of concentration is used.

Active transport depends on the orientation of PIN1 proteins at the cell membranes.

The effect of PIN1 on the efflux of auxin from cell i to a neighboring cell j is modeled

using the formula:

[IAA]i
transporti→j = T [P IN ]i→j (5.2)
1 + κT [IAA]j

where T is the transport coefficient, and κT is the transport saturation coefficient. For

a given number of PIN1 molecules near the wall separating cell i from cell j, the flux of
 65

Figure 5.2: Pattern emergence on a line of 50 cells, with wrap-around boundary condi-
tions (the leftmost and rightmost cells are considered neighbors). Taller bars (brighter
green) indicate higher IAA concentration. A small amount of noise present in the ini-
tial distribution is required to break symmetry. Relatively evenly spaced peaks in IAA
concentration are formed.

auxin from i to j is thus assumed to increase with the concentration of auxin in cell i,

and saturate with the increasing concentration of auxin in cell j. In this case a linear

power of concentration is used, however the phyllotaxis simulations described later use a

quadratic dependence of the flux on concentration.

The simulation proceeds in a sequence of time steps, with PIN1 polarity calculated

at the beginning of each time step. The entire contents of PIN1 in a cell is allocated

to the cell’s membrane, with Equation (5.1) specifying the portioning out of available

PIN1 to sections of the membrane facing each of the neighbor cells. The change in IAA

concentration in each cell i is modeled by the following equation:

d[IAA]i
= diffusion + transport
dt
X
= D([IAA]j − [IAA]i ) +
j∈Ni
X [IAA]j [IAA]i

T [P IN ]j→i − T [P IN ]i→j . (5.3)
j∈Ni
1 + κT [IAA]i 1 + κT [IAA]j

Diffusion is assumed to take place directly between neighboring cells (extracellular

space is ignored), with the cell membranes representing the main obstacle to diffusion.

This is also the case for transport, since the auxin transported out of a cell is deposited

directly into the neighbor cell, bypassing extracellular space.

If the model is started in the absence of noise, all cells will begin with the same
66

Figure 5.3: Pattern dependence on model parameters. Model parameters affect how many
peaks a given number of cells will create. Higher values for the transport coefficient will
create more peaks. Similarly, higher values for the diffusion coefficient will reduce the
number of peaks. If the diffusion coefficient is too high, or transport coefficient too low,
no peaks will form at all. Parameters: (center) D = 1, T = 70; (top left) D = .5, T = 70;
(bottom left) D = 1.5, T = 70; (top right) D = 1, T = 40; (bottom right) D = 1, T = 180.

initial concentration, and therefore will have the same PIN1 orientation, transport, and

diffusion. The concentration of IAA in all of the cells will remain equal, and no pattern

will form. This is similar to what happens in reaction-diffusion systems in the absence

of noise that is needed to “break symmetry”. If even a small amount of noise is added

to the model, for example by perturbing initial cell concentrations, a relatively evenly-

spaced pattern of peaks in IAA concentration will form (Figure 5.2). Cells with small

local maxima due to the initial conditions cause the PIN1 proteins in neighbor to cells

to be oriented preferentially towards them (very slightly) which causes the concentration

of IAA at these maxima to increase. A positive feedback loop is then formed which will

recruit even more PIN1 proteins over time. The depletion of IAA in the cells surrounding

a peak prevents the formation of other peaks nearby.

The amplitude and spacing between these peaks can be controlled by manipulating

model parameters (Figure 5.3), with decreased transport leading to fewer peaks, and
 67

Figure 5.4: Pattern emergence on a 30 × 30 grid of cells without wraparound bound-

ary conditions. Color (brighter green) indicates higher IAA concentration. Boundary
conditions break symmetry causing the pattern to form from the outside in.

increased transport leading to more peaks. A reverse relationship is observed when

manipulating the diffusion coefficient as the spacing of the peaks is a balance between

diffusion and transport. If the diffusion coefficient is too high, or the transport coefficient

is too low, no peaks will form at all. For a given cell count and parameter set, the same

number of peaks usually forms, regardless of the pattern of initial noise. However, the

locations of the peaks will vary. If the simulation is performed with fewer cells, at some

point a smaller number of peaks will form. On the boundary of this transition, it is not

uncommon for the larger number of peaks to form initially, and than a pair of peaks will

coalesce to form the smaller number.

Similar results are obtained on a grid of cells. The simulation depicted in Figure 5.4

was performed with the same equations as the simulation on a line of cells, only the

PIN1 proteins are now portioned among four neighbor cells instead of two. The removal

of wraparound conditions is enough to break symmetry and thus no noise is required to

start the pattern. Note that this causes the pattern to form from the outside in.
68

5.4 Phyllotaxis model on a cylindrical apex structure

The first transport-based model of phyllotaxis described here uses an apex with the topol-

ogy of a regular cylindrical grid (Figure 3.4). Vertex points defining the grid represent

cells, and are embedded in the apex surface model described in Section 3.3. All cells

have four neighbors and are assumed to be the same size, even though for visualization

purposes the cylinder is drawn to approximate the shape of a plant apex.

In order to accommodate a growing structure, a source of auxin must be considered,

as pattern formation can no longer be seen as simply redistributing auxin from some

almost uniform initial state. Experimental evidence suggests that most of the auxin

available in the shoot apex is supplied there by directed transport rather than local

production [Reinhardt et al., 2003b; Smith et al., 2006a]. In order to separate out the

transport processes involved in supplying auxin to the meristem from those responsible

for phyllotaxis patterning, the following assumption is used: a uniform supply of auxin,

modeled as local production, is provided to all cells of the apex which are outside of the

central zone. These are defined to be all of the cells which are situated at a distance

greater than a given distance from the tip of the apex (Figure 5.5). This distance,

which also corresponds to the beginning of the peripheral zone, is provided as a model

parameter and remains fixed throughout the simulation. The formula used to model

auxin production and decay is given in Equation (5.6).

Since no auxin is produced to the central zone, this area is relatively low in auxin

and does not participate in patterning to any great extent. It is for this reason that the

RERG function described in Section 3.3 is chosen such that all of the growth, and hence

insertion of new vertices, occurs within the central zone of the apex. A more biologically

plausible growth function for a shoot apex would specify a similar or increased growth

rate in the peripheral zone over that of the central zone. However, the insertion of new
 69

Figure 5.5: Model of a Lucas phyllotaxis pattern on a cylindrical apex with 40 vertices
around. (a) Actual topology of the apex with rectangular cells. All cell are considered
to be the same size and have 4 neighbors. Green color indicates auxin concentration.
The auxin production zone begins at a fixed distance from the tip with all the cells
below this distance producing auxin. This distance remains unchanged throughout the
simulation. (b-c) Top and side views of the same simulation rendered with peaks in auxin
concentration extruded from the surface.

rings of vertices in the area of the apex where the pattern is forming is likely to to have

an unrealistic influence on the resulting pattern. In a real apex, the cells are not regular,

and cell divisions are spread out, both spatially and temporally. A geometry with more

realistically shaped cells and more realistic cell division is presented in Section 5.5.

The allocation of PIN1 transport facilitator proteins to cell membrane sections in

the phyllotaxis model on the cylindrical grid of cells is no longer linear, but instead is

modeled by the following equation:

b[IAA]j
[P IN ]i→j = [P IN ]i P [IAA] (5.4)
b k

k∈Ni

where the exponentiation base b > 1 controls the extent to which PIN1 protein distri-

bution is affected by the concentration of IAA in neighboring cells. Note that the only

difference between this equation and Equation (5.1) is that an exponential relation is

used for PIN1 orientation, instead of a linear one.

The total amount of PIN1 proteins in a cell is no longer fixed, but depends on pro-
70

duction and decay modeled by the formula:

d[P IN ]i
= production − decay
dt
ρP IN [IAA]i
= − µP IN [P IN ]i (5.5)
1 + κP IN [P IN ]i

where the coefficient ρP IN controls IAA dependent PIN1 production and κP IN is the PIN1

production saturation coefficient. µP IN represents PIN1 decay, which is proportional to

PIN1 concentration.

IAA concentration in a cell is the result of auxin supply (modeled as production) and

decay combined with the effects of cell to cell diffusion and transport. Combining these

factors, the change in IAA concentration in cell i is given by:

d[IAA]i
= production − decay + diffusion + transport
dt
ρIAA X
= − µIAA [IAA]i + D([IAA]j − [IAA]i )
1 + κIAA [IAA]i j∈Ni
X [IAA]2j [IAA]2i

+ T [P IN ]j→i − T [P IN ]i→j (5.6)
j∈N
1 + κT [IAA]2i 1 + κT [IAA]2j
i

where ρIAA controls the production of IAA with saturation coefficient κIAA , and µIAA

controls the decay of IAA which is dependent on the IAA concentration. IAA transport

is modeled as in Equation (5.2) for the line model, only the terms for transport and

transport saturation are quadratic instead of linear. In phyllotaxis simulations, it was

found that the exponential relationship for PIN1 orientation defined in Equation (5.4)

combined with quadratic terms for transport in Equation (5.6), was able to generate a

wider variety of phyllotaxis patterns over a wider range of model parameters than linear

relations.

The simulation begins with an initial pair of rings, both containing the same number

of vertices evenly space around the apex surface. If the shape of the profile generating
 71

curve intersects the y axis (Figure 3.2) at the tip, then all of the vertices of the top ring

will share the same position in space. Otherwise, the apex will have a hole in the center at

the tip. The number of vertices around the apex is the same for all rings and is provided

as a model parameter. This number remains unchanged throughout the simulation, and

can be chosen to approximate number of cells around the apex at the peripheral zone.

In general, distichous patterns patterns require a lower number of vertices than spiral or

decussate patterns. Whorled patterns, in which 3 or more primordia appear at the same

time, require higher vertex counts.

As the simulation progresses, the bottom ring of vertices moves away from the apex

tip due to growth. When a specified distance between the rings is reached, a new ring

of vertices is inserted in-between. Vertices are only inserted between the top two rings,

and all of the growth of the apex surface occurs within the boundary of the central zone.

Since all the cells in the model are considered to be the same size, growth only affects

the simulation when a new row of cells is added at the tip, or when a ring of cells moves

outside the central zone and begins to produce IAA. Despite the visual appearance of the

apex in the simulations, the model more closely resembles the cylindrical apex depicted

in Figure 3.1. This type of apex structure is very similar to the cell-based simulations

of Hellendoorn and Lindenmayer [1974], Veen and Lindenmayer [1977], and Meinhardt

[1982].

In the initial stages of the simulation all vertices are within the central zone where

no auxin in produced, and thus no patterning occurs. Auxin production begins when

growth causes some of the cells to move outside of the central zone. At this point,

depending on parameters, several different patterning events can happen. In the most

frequent case, one or more peaks in IAA concentration (indicating primordium initiation

sites) will appear evenly spaced about the apex at an equal distance from the tip. This is

quite similar to what happens in the simulation of a line of cells discussed in Section 5.3,
72

where a ring of cells is divided into relatively evenly spaced peaks. This initial whorl of

primordia is located a few cells outside the boundary of the central zone. After some time

the whorl of primordia moves down the apex surface, away from the central zone, and the

next whorl of primordia appears in the centers of the spaces between the previous whorl.

The new whorl is positioned at roughly the same distance from the apex tip as the first

whorl, and this distance does not change throughout the simulation. This is the most

common outcome, and yields distichous, decussate, or whorled phyllotaxis, depending on

how many primordia appear at once. For a given set of model parameters, this number is

an increasing function of the number of vertices around the apex in the cylindrical mesh

of cells. If the number of vertices around the apex is a small odd number, this initial

symmetry will sometimes be broken and a spiral and spiro-distichous pattern will result.

The initial symmetry can also be broken deliberately, by placing some initial peaks of

IAA in a predefined pattern. The PIN1 protein orientation in adjacent cells towards

these cells of higher auxin concentration will maintain the peaks. The incorporation of

an initial pattern of auxin peaks into the model can result in the propagation of various

types of spiral or spiro-distichous patterns, in addition to the distichous, decussate, or

whorled patterns that arise de novo (Figure 5.6).

The cylindrical apex model had a strong tendency towards producing decussate and

whorled patterns. In many cases a pre-pattern of auxin peaks in a spiral arrangement

eventually reverts to decussate or whorled phyllotaxis. Once established, these patterns

are very stable and will not return to spiral. A decussate or whorled pattern started on

an apex with enough vertices around to support a higher order whorled pattern will often

eventually switch to that higher order pattern. Interestingly, one such transition, from

decussate to tricussate, occurs in tricussate plants that start from a pair of cotyledons

[Gómez-Campo, 1974].

The tendency towards decussate and whorled patterns may be the result of the topol-
 73

Figure 5.6: Phyllotactic patterns produced by the transport model on an apex with
cylindrical topology. (a) Fibonacci spiral phyllotaxis on an apex with v = 33 vertices
around. (b) Distichous phyllotaxis (v = 13). (c) Lucas spiral phyllotaxis (v = 40). (d)
Decussate phyllotaxis (v = 27). (e) Spiro-distichous phyllotaxis (v = 15). (f) Tricussate
phyllotaxis (v = 38).

ogy created by the cylindrical representation of the shoot apex. In a real plant, there

are fewer cells around the apex closer to the tip than in the peripheral zone. This may

interfere with cell-to-cell communication, artificially increasing the distances between pri-

mordia in all but the radial direction. In addition, the regularity of the cylindrical grid

apex structure may have an influence on patterning.

5.5 Phyllotaxis model on a cellular apex structure

The model of phyllotaxis discussed in Chapter 4 is based on the notion that phyllotaxis

is the result of a field produced by existing primordia that inhibits the formation of new
74

primordia nearby. The phyllotaxis simulation model of Section 5.4 suggests a mecha-

nism, based on directed transport of IAA, by which this inhibition may be achieved.

In this transport-based model, the contents of each cell are considered to be uniformly

distributed, and it is the cell-to-cell transport and diffusion that determines the profile

of the auxin concentration gradient. Thus, as far as inhibition is concerned, distance is

more accurately measured in number of cells rather than the actual Euclidean distance

along the surface. In this sense, the cylindrical apex model does not provide a realistic

distance measure for a signaling process where the level of signal depends directly on

distance.

To solve this problem, an apex cell structure is required in which the number of cells

around the apex can vary with the distance from the tip. This is especially important in

the area between the tip of the apex and the outer border of the peripheral zone, where

the signaling processes that cause patterning are most active. The apex model with the

cellular structure based on the model by Nakielski [2000], as described in Section 3.4, fits

this criterion. This model also produces cells with shapes very similar to those found in

a real Arabidopsis meristem [Reddy et al., 2004], with a realistic variation in cell size,

eliminating the regularity of the cylindrical grid. In addition, cell divisions do not occur

synchronously, as happens when inserting rings of cells in the cylindrical model, which

minimizes any impact the synchronous divisions might have on patterning.

In order to accommodate the cellular structure, the formula for PIN1 protein alloca-

tion to membrane sections must be modified to account for the differences in the lengths

of cell walls. Since only the surface layer of cells is modeled, and it is assumed that all

cells have the same thickness, quantities normally expressed as volumes can be expressed

by areas, and areas can be expressed by lengths. Adding the effect of variable cell well
 75

length to Equation (5.4) gives:

li→j b[IAA]j
[P IN ]i→j = [P IN ]i P (5.7)
li→k b[IAA]k
k∈Ni

where li→j is the length of the interface between cell i and cell j, with the other symbols

defined as before.

An additional term is also added to the PIN1 production equation, so that a only

a portion of production is dependent on IAA concentration ρP IN and the remaining

production is constant or default production ρP IN0 . Since PIN1 proteins are too large to

diffuse and are not subject to transport themselves, they do not move from cell to cell.

In addition, any effects on the concentration of PIN1 from cell expansion due to growth

or the “pinching” of cells that occurs during cell division are ignored. Thus the change

in concentration of PIN1 proteins within a cell i does not depend on cell geometry and

is given by:

d[P IN ]i
= production − decay
dt
ρP IN0 + ρP IN [IAA]i
= − µP IN [P IN ]i . (5.8)
1 + κP IN [P IN ]i

Since the model has cells with varying size, a cell’s area is taken into account when

calculating the change in IAA concentration. Flux due to diffusion or transport will

affect the cell’s concentration by an amount inversely proportional to the cell’s area.

When calculating diffusion, the length of the interface between cells is also considered.

Note that in the case of IAA flux due to transport facilitated by PIN1 proteins, the length

of cell walls is already accounted for in Equation (5.7). Combining these modifications,
76

Figure 5.7: Patterning on a cellular structure in the plane with auxin visualized in green
and PIN1 proteins visualized in red. (Left) Sheet of cells covered with approximately
evenly spaced peaks in auxin concentration. (Right) Close up around auxin peak show-
ing polar PIN1 localization on sections of cell membranes facing cells of higher auxin
concentration.

the equation to model the change in the IAA concentration of a cell i becomes:

d[IAA]i
= production − decay + diffusion + transport
dt
ρIAA D X
= − µIAA [IAA]i + li→j ([IAA]j − [IAA]i )
1 + κIAA [IAA]i Ai j∈N
i
2
[IAA] [IAA]2i
 
T X j
+ [P IN ]j→i − [P IN ]i→j (5.9)
Ai j∈N 1 + κT [IAA]2i 1 + κT [IAA]2j
i

where li→j is the length of the interface between cell i and cell j, and Ai is the area of cell

i. Again no adjustment is made for increasing cell area due to growth, or the changes in

cell area that occur as a result of the pinching of cells during cell division.

The patterning model of Equations (5.7−5.9) applied to a sheet of cells in the plane

is depicted in Figure 5.7. As with the transport model on a regular grid of cells, a

self-organizing pattern of relatively evenly spaced peaks in auxin concentration emerges.

 77

5.5.1 An essential role for differentiating primordia

The phyllotaxis model on the cylindrical apex structure (Section 5.4) includes the im-

plicit hypothesis that phyllotaxis in Arabidopsis is determined directly by the transport-

based patterning mechanism operating on the growing surface of the apical meristem. In

simulations on the cellular apex, however, the topology proved too irregular to obtain

sustained spiral phyllotactic patterns using this mechanism alone. Patterns would appear

to produce primordia in a spiral arrangement for for a short time, and then switch to

decussate. After a few whorls of decussate, a spiral arrangement would reappear, some-

times in the opposite direction. This observation was upheld by numerous simulations, in

which diverse parameter values and different formulas for polarizing PIN1 were used. A

similar phenomena was reported by Jönsson et al. [2006] in their phyllotaxis simulations.

It was thus concluded that additional factors, in particular the effects of differentiating

primordia, must play an important role in the generation of phyllotaxis patterns. To test

this hypothesis, the transport-based patterning model was extended with the following

elements:

• The surface of the apex is divided into three zones: the central zone, the pe-

ripheral zone, and the proximal zone (located below the peripheral zone). Auxin

production depends on the zone in which a cell is located. No auxin is produced

in the central zone. In the peripheral and proximal zones, auxin is produced us-

ing different production coefficients, with ρIAA (peripheral zone) being greater than

ρIAA (proximal zone).

• The peripheral zone is divided into primordia and intervening regions between the

primordia. A primordium is initiated when the IAA concentration in two adjacent

cells reaches a predefined threshold Th. The primordium center is located at the

midpoint of the centroids of these two cells. Radius r and height h (Figure 5.8),
78

Figure 5.8: Apex with growing primordia. Growth causes the primordia to move down
the apex surface, and increase in size (radius r and height h).

which define the size of a primordium, increase with time according to a predefined

function that is provided to the model as an input parameter. Primordia cells are

differentiated, and thus behave differently from the intervening regions of the shoot

apical meristem.

• The differentiated cells produce auxin at a relatively higher rate:

 
ρ (primordium) di
additional production = IAA 1− (5.10)
1 + κIAA [IAA]i r

where di is the distance of the centroid of a primordium cell i from the primordium

center.
 79

• The concentration of auxin in a primordium is capped at a maximum level [IAA]max ,

with the excess amount of auxin assumed to flow into the inner tissues of the

primordium and initiate the vasculature.r.

• Polarization of PIN1 proteins towards the primordium center is augmented, com-

pared to the intervening region. This is achieved using an auxin-concentration

equivalent IAA0 , instead of the current auxin concentration, as the polarizing fac-

tor. This factor is defined using the formula:

 
0 di
[IAA ]i = max [IAA]i , [IAA]max · (1 − ) (5.11)
r

where [IAA]max is the model maximum IAA concentration.

A possible molecular mechanism for the resulting bias of PIN1 protein polarization

towards the primordium center might be a diffusing substance released by cells located at

the center. This substance would be prevented from spreading outside the primordium

by boundary cells. Another possibility might be a mechanical induction of PIN1 polar-

ization, caused by stresses or strains in the primordium.

5.5.2 Phyllotaxis patterns produced by the cellular apex model

The phyllotaxis model on the cellular apex structure is able to produce the following

phyllotactic patterns: distichous, decussate, sprial, and tricussate (Figure 5.9). These

different phyllotaxis types were found by making changes to multiple parameters and it

is difficult to attribute particular parameters to specific changes in phyllotaxis. This is

in contrast to the inhibition field model described in Chapter 4 that is able to model the

simple transition from distichous to spiral by changing a single parameter, the inhibition

threshold. In mechanistic terms, this effect could be accomplished in several ways. The

size of the apex could be increased, so that a primordium would have a smaller relative
80

Primordium 1 2 3 4 5 6 7 8 9 10
Time 2.8 3.0 6.8 7.2 10.5 11.1 14.0 14.9 17.0 18.5
Divergence Angle 0.0 162.7 96.3 155.9 123.4 143.2 139.3 128.6 142.1 124.8

Primordium 11 12 13 14 15 16 17 18 19 20
Time 19.8 21.5 22.9 24.3 26.4 27.4 29.5 30.9 32.7 34.0
Divergence Angle 144.2 131.9 125.4 141.2 132.3 131.6 125.0 137.3 135.4 124.1

Primordium 21 22 23 24 25 26 27 28 29 30
Tim 35.6 37.1 38.7 40.2 42.0 43.7 45.6 47.1 49.3 50.1
Divergence Angle 123.0 144.8 114.6 128.9 133.1 135.0 143.0 150.0 130.9 123.5

Primordium 31 32 33 34 35 36 37 38 39 40
Time 52.0 53.5 55.0 57.1 58.4 60.0 61.3 63.0 64.6 66.6
Divergence Angle 132.8 126.7 136.4 130.1 134.6 130.9 135.2 125.4 153.2 126.1

Table 5.1: Divergence angles in degrees of the first 40 primordia for the Fibonacci spiral
phyllotaxis simulation on the cellular apex structure.

range of influence measured as a fraction of the apex size. Another possibility would be to

lower diffusion, so that the peaks in auxin concentration can form closer together. Closer

peak spacing could also be achieved by increasing transport, either by increasing PIN1

production, or the transport coefficient itself. Even an increase in IAA production, which

could also be done in several ways, would be expected to produce more densely packed

primordia. In a real plant, the individual components of the patterning mechanism are

likely coupled. For example, if the auxin level in the meristem was artificially increased,

would it change the apex size or the transport rate? It is therefore reasonable to expect

that a change from one stable phyllotaxis type to another might involve changes to

multiple parameters simultaneously. This may also be part of the reason that there are

very few mutants that change one stable phyllotactic pattern into another (for discussion

see Kuhlemeier and Reinhardt [2001]).

One conclusion that can be drawn from the simulations is that there is a relation-

ship in non-spiral patterns between apex size and the number of primordia the apex

is able to produce at one time. This is not surprising and fits nicely with theories of

phyllotaxis based on the availability of space [Snow and Snow, 1932]. Starting with dis-
 81

Figure 5.9: Types of phyllotaxis produced by the transport-based model on the cellular
apex structure: (a) distichous; (b) decussate; (c) tricussate; (d) Fibonacci spiral.

tichous patterns (j = 1), the apex size progressively increases as the jugacy increases in

the simulations for decussate (j = 2) and tricussate (j = 3) patterns. This is consistent

with observations of the abphyl1 mutant in maize, that has an enlarged meristem and

produces decussate phyllotaxis instead of the distichous phyllotaxis normally produced

by the wild type [Jackson and Hake, 1999].

As expected, there is also a relationship between primordia density and the amount
82

Figure 5.10: Superimposed image of the model and a real Arabidopsis shoot apex. (left)
Arabidopsis shoot apex photograph taken with a light microscope. (center) Simulation
model with all but the 8 most recent primordia removed. (right) Superimposition of light
microscope photograph and simulation model.

of auxin supplied to the meristem. Patterns in which primordia are more densely packed,

such as Fibonacci spiral or tricussate, have higher IAA production coefficients than the

distichous or decussate patterns, which are considerably more sparse.

5.5.3 Comparison to experimental observation in Arabidopsis

In dicotyledonous plants such as Arabidopsis, the embryo produces two cotyledons that

arise near-simultaneously at angles close to 180◦ . From this decussate starting situation,

a spiral phyllotaxis with divergence angles of approximately 137.5◦ gradually emerges

[Mündermann et al., 2005; Callos, 1994]. The model recreates this pattern de novo within

the standard error of measured divergence angles (Figure 5.11). The spiral phyllotactic

pattern is perpetuated if the simulation is continued (Table 5.1). A visual comparison of

the model to a real Arabidopsis apex can be seen in Figure 5.10.

Further support for the model is obtained by simulating mutant phenotypes and the

outcomes of experimental manipulations made to real Arabidopsis apices. Specifically,

the pin1 phenotype is characterized by the absence of primordia [Okada et al., 1991]
 83

Figure 5.11: Comparison of divergence angles produced by the transport-based simulation

model. Divergence angles measured in Arabidopsis with standard error bars in green and
divergence angles generated by the spiral phyllotaxis model in blue. See Smith et al.
[2006a] for details of divergence angle measurement.

and low DR5::GFP expression [Smith et al., 2006a]. Removing production of PIN1 in

the model reproduces the inhibition of organ formation (Figure 5.12a). Note that this is

not the only possible outcome of disabling the transport mechanism. Since the supply of

auxin (modeled as local production) to the meristem is not reduced in this simulation,

another possibility would be that the auxin would build up uniformly until it exceeds

the primordia differentiation threshold. In this case a ring shaped primordia would be

expected. In fact the concentration of auxin does appear to be higher in the simulation

than in a real pin1 mutant and more similar to an NPA-treated apex (Figure 2C and

2D in [Smith et al., 2006a]). This is likely because the auxin that is believed to be

transported from below in a real plant is reduced in the pin1 mutant, and there is no

compensation for this in the simulation.

84

Figure 5.12: Simulation of application of IAA to pin1 mutant apex. (a) Simulated pin1
mutant apex. (b) Primordium formed in the pin1 mutant after localized application
of auxin in the peripheral zone. (c) Primordium ring formed in the pin1 mutant after
localized application of auxin at the tip of the apex. Arrows indicate the site of auxin
application.

Local application of auxin in the peripheral zone of a pin1 mutant produces an isolated

primordium (Figure 3A in [Reinhardt et al., 2003b]), whereas local application of auxin

at the tip of the apex results in the formation of a ring-shaped primordium (Figure 3C

in [Reinhardt et al., 2003b]). These experimental results are reproduced by the model

(Figure 5.12b and 5.12c). In contrast, the model does not capture the dependence of

the primordium size on the amount of auxin; furthermore, the simulated ring consists

of many separate primordia, as observed in the pinoid mutant (Figure 3F in [Reinhardt

et al., 2003b]). These shortcomings are due to the simplistic nature of the geometric

algorithm defining position and scope of primordia in the model, rather than failure of

the postulated molecular patterning mechanism.

Laser ablation of incipient primordium cells causes the emergence of a displaced pri-

mordium in the vicinity of the ablated site (Figure 3 in [Reinhardt et al., 2005]). This

effect is reproduced in the model (Figure 5.13).

 85

Figure 5.13: Simulation of cell ablation. (a) Control apex. (b) The same apex in which
five cells shown in black have been removed. Note the shift in the position of the pri-
mordium initial (cells with high auxin concentration, shown in bright green) near the
ablation site.

5.6 Sensitivity analysis

In order to provide insight into the form and role of parameters in the model equations,

a sensitivity analysis was performed. Starting with the parameter values used for the

Fibonacci spiral phyllotaxis simulation as a baseline, selected equations or parameters

of the model were changed, and simulations run to determine if spiral phyllotaxis could

be sustained. This invariably required the adjustment of additional parameters, but an

attempt was made to minimize the number of parameters changed. For this analysis the

focus was on on pattern maintenance rather than pattern initiation. Consequently, the

simulations were started by placing 10 initial primordia in a Fibonacci spiral pattern,

at a divergence angle of 137.5◦ , spaced temporally at the average spacing for primordia

11 − 20 in the Fibonacci spiral phyllotaxis simulation (see Table 5.1). Four test cases

were considered, and the divergence angles of primordia 11-20 recorded (Table 5.2).
86

Primordium
11 12 13 14 14 16 17 18 19 20

Fixed PIN1 145.1 143.8 139.5 133.8 143.3 134.7 130.0 125.7 137.4 144.3
IAA Saturation 137.3 128.6 138.9 122.2 143.4 139.3 122.6 143.5 126.9 136.0
Transport Power 127.7 145.4 125.2 136.3 130.9 126.8 135.5 124.2 148.9 122.8
Allocation Base 144.0 131.2 134.9 119.8 145.1 137.2 125.0 126.2 128.1 133.7

Table 5.2: Divergence angles in degrees for sensitivity analysis simulations initiated with
10 primordia placed at 137.5◦ . The angles of the next 10 primordia produced by the
model are shown.

5.6.1 Fixed PIN1 concentration

The equation for PIN1 production and decay (Equation 5.8) was removed from the

model, and PIN1 concentration was fixed at 0.94, the approximate value of PIN1 in non-

primordia cells in the Fibonacci spiral phyllotaxis simulation. Pattern maintenance con-

ditions were found by lowering IAA proximal zone production ρIAA (proximal zone), from

3.0 to 2.5, and changing additional IAA production within primordia ρIAA (primordium)

from 3.5 to 2.5.

After fixing PIN1 concentration, it was not difficult to find the new parameter values

that resulted in pattern maintenance (“Fix PIN1” simulation in Table 5.2). This sug-

gests that the exact form of the PIN1 production and decay equations is not critical to

phyllotactic pattern formation.

5.6.2 Removal of IAA production saturation term

The two equations that have IAA production terms (Equations 5.9 and 5.10) both have

a saturation component based on IAA concentration. Here the case is considered where

this saturation is removed by setting the IAA production saturation coefficient κIAA equal

to zero. It was found that spiral phyllotaxis can be maintained if IAA proximal zone

production ρIAA (proximal zone) is changed from 3.0 to 1.0, additional IAA production

in the peripheral zone ρIAA (peripheral) is changed from 3.0 to 0.9, and additional IAA

production within primordia ρIAA (primordium) is changed from 3.5 to 1.0.

 87

It was expected that the removal of IAA production saturation based on IAA concen-

tration would require a reduction in IAA production coefficients. It was not difficult to

find values that lead to pattern maintenance (“IAA Saturation” simulation in Table 5.2).

Although it appears that these saturation terms are not essential to pattern formation,

subsequent investigation revealed that in this condition the model was no longer able to

reproduce the pin1 mutant phenotype. With κIAA set to zero, the IAA concentration

in the model would build up until primordia formed continuously, completely filling the

peripheral zone.

5.6.3 Reduction of transport exponent

The transport component of Equation (5.9) has an effect on IAA flux proportional to the

square of the source cells IAA concentration. The effect on flux is inversely proportional

to the square of the destination cells IAA concentration. It was found in early experiments

on models with lines of cells, that the lack of a transport saturation term leads to peaks

in IAA concentration that are only a few cells apart. The addition of a linear saturation

term increased the number of cells between peaks, but it was found that a non-linear

relation provided even greater cell numbers.

In this simulation, these quadratic relations were changed to a power of 1.5. It

was found that spiral phyllotaxis could be maintained if the transport coefficient T was

reduced from 22.5 to 19.0, the primordium differentiation threshold Th from 7.5 to 7.0,

and the peripheral zone size from 2.9 to 2.57.

It was possible to maintain a spiral phyllotactic pattern (“Transport Power” simula-

tion in Table 5.2), but the model was very sensitive to parameter values. This could be

due to the reduced cell count of the smaller peripheral zone, placing greater demands on

the robustness of the pattern forming mechanism.

88

5.6.4 Increase in PIN1 allocation exponent

The PIN1 allocation equation (Equation 5.7) gives an exponential preference to neighbor

cells of higher IAA concentration. Experiments with other formulas for preference such

as linear, or a power of neighbor cells’ IAA concentration, did not yield spiral phyllotactic

patterns.

A simulation was performed with the exponentiation base b in the PIN1 allocation

equation (Equation 5.7) increased from 3.0 to 4.0. Diffusion coefficient D was increased

from 4.0 to 4.5, and transport coefficient T reduced slightly from 22.5 to 22.

Again spiral phyllotaxis was maintained (“Allocation Base” simulation in Table 5.2),

but the simulation was very sensitive to parameter values. Experience with this parame-

ter has shown that decussate and tricussate simulations are tolerant of values as high as

5.0, but Fibonacci spiral simulations were most stable at values around 3.0. Reduction

of this parameter value below 2.5 did not yield spiral phyllotactic patterns.

5.7 Summary of results of the transport-based phyllotaxis model

A simulation model of shoot apical meristems was constructed that reproduces the phyl-

lotaxis of Arabidopsis vegetative shoots (Figure 5.14). The model supports the basic

tenet of the conceptual model proposed by Reinhardt et al. [2003b], according to which

phyllotaxis results from the dynamics of interaction between existing and incipient pri-

mordia in a growing apex, mediated by transported auxin. New primordia emerge in the

areas of high auxin concentration and maintain a distance from each other by depleting

auxin in their proximity.

An important aspect of the model is the transport-based patterning mechanism, which

involves a putative positive feedback loop between the concentration of auxin, the local-

ization of PIN1, and the polar transport of auxin. This mechanism, combined with
 89

Figure 5.14: Top and side view of Fibonacci spiral phyllotaxis simulation showing the
first 20 primordia.

diffusion, can “break symmetry” of an approximately uniform auxin distribution, and

form a spatial pattern of auxin concentrations that is stable over time (Figure 5.2). The

magnitude and spacing of concentration maxima depend on model parameters (Figure

5.3). This patterning mechanism is fundamentally different from reaction-diffusion in

that it postulates a controlled redistribution of an existing morphogen (auxin) in space,

whereas reaction-diffusion postulates a controlled local production of two or more mor-

phogens. The transport-based patterning model presented here is also different from

canalization, which postulates a sustained polar transport of auxin preferentially ori-

ented in the direction of high flux [Mitchison, 1980, 1981; Sachs, 1981; Feugier et al.,

2005; Rolland-Lagan and Prusinkiewicz, 2005]. In addition, the initial events that lead

to patterning in the canalization model rely on the transport of auxin down its concen-

tration gradient. In contrast, the phyllotaxis models presented here specify that auxin is

preferentially pumped towards the neighboring cell with the highest auxin concentration.

Depending on the difference in auxin concentration between the source and target cell,
90

the resulting flux may be with, but more often against the concentration gradient.

Experimental data shows that PIN1 proteins in the L1 layer of a wild type Arabidop-

sis meristem are generally oriented towards the primordium initial, and the position of

primordia coincides with the maxima of auxin concentration. The transport-based pat-

terning mechanism is consistent with this data, and is able to generate stable phyllotactic

patterns on a regular cylindrical grid of cells. However, in simulation experiments on a

cellular apex structure, this mechanism does not generate phyllotactic patterns by itself.

Therefore, an additional factor is postulated: the differentiation of cells within primordia.

The primordia cells are characterized by increased auxin, enhanced polarization of PIN1

proteins in the L1 towards primordium centers, and export of auxin from the L1 towards

the internal layers. All of these have a strong experimental basis (Figure 2 in [Reinhardt

et al., 2003b]).

According to the cellular model, phyllotaxis is thus not governed by a single mecha-

nism, but represents a combined effect of several factors. This complexity may be needed

in nature to generate phyllotactic patterns in the presence of noise. The small number of

cells around the peripheral zone of the Arabidopsis apex limits the precision in primor-

dia placement to approximately 15◦ (360◦ divided by approximately 24 cells). A robust

mechanism is thus needed to initiate and maintain phyllotactic patterns in spite of a

relatively large departure in the placement of individual primordia from their mathemat-

ically ideal positions. The problem of generation of phyllotactic patterns in the context of

an irregular geometry was not considered in previous models, which assumed continuous

or uniformly discretized space.

Assumptions of the model include the localization of auxin sources, and the molecular

mechanism of PIN1 protein localization in a cell. According to the model of Reinhardt

et al. [2003b], auxin is produced outside the shoot apical meristem, and is acropetally

transported into the meristem through the L1. Under these conditions, it was not possible
 91

to create spiral phyllotactic patterns. Instead a uniform production throughout the L1

in the peripheral zone was assumed, with an additional boost in the primordia. Also, the

model postulates localization of PIN1 towards the neighboring cells with the highest auxin

concentration, but leaves open the question of what molecular mechanism may produce

this localization. The answers to these questions may lead to the integration of the

model of phyllotaxis with a model of vein formation in the leaf and stem. Although both

processes are mediated by auxin, and appear to share the same PIN1-mediated transport

machinery, the proposed mechanism of PIN1 polarization in the phyllotaxis model is

almost opposite to the canalization mechanism proposed for veins [Mitchison, 1980, 1981;

Sachs, 1981; Rolland-Lagan and Prusinkiewicz, 2005]. It is thus interesting how these

different mechanisms may be reconciled in the growing plant. The work presented in the

later parts of Chapter 8 presents a very preliminary exploration of this question.

 Chapter 6

Phyllotaxis Models on an Apex with Physically

Based Growth

6.1 Physically-based models of the shoot apex

The phyllotaxis simulations presented in Chapters 4 and 5 use a descriptive model of

growth for the shoot shoot apex. The shapes of the apex, and of the outgrowing pri-

mordia, are specified directly in geometric terms, and are not emergent properties of

the model. This approach is simple and allows the modeler to focus on the patterning

mechanism of phyllotaxis while abstracting from the processes involved in maintaining

the shape of the apex and outgrowing primordia.

In this chapter, phyllotaxis on shoot apices simulated with a physically-based growth

model will be investigated. Unlike descriptive models, physically-based models have the

potential to generate apex and primordium shapes as emergent properties of the model.

This opens up the possibility of modeling the processes that lead to these shapes, and

the interaction between growth and form. For phyllotaxis, physically-based models may

have a special significance, since many authors have suggested that physical forces may

play a direct or indirect role in primordia positioning (for example Hofmeister [1868];

Schwendener [1878]; Green [1980]; Hernandez and Green [1993]; Dumais and Steele [2000]

and Shipman and Newell [2005]).

Several types of physically-based models have been proposed for the shoot apex, all of

which are based on mass-spring systems. In the simplest case, only the cell centers of the

surface (L1) layer of the tunica are defined and connected by a network of springs. Such

a model is used by Jönsson et al. [2006] (also described in Mjolsness [2006]) as a platform

92
 93

for a transport-based model of phyllotaxis. Although their model is physically-based,

the shape of the apex is not an emergent property of the model. The points defining

the cells are only free to move within a predefined two-dimensional surface provided by

the modeler as a parameter to the model. Within the confines of this surface, a mass-

spring system is used to create a growing tissue. The system grows by increasing the

rest length of the springs, causing the cell centers to move farther apart. If the cell

walls need to be visualized, then the Voronoi regions of the cell centers can be used, or

alternatively, the cells can be rendered as intersecting spheres [Jönsson et al., 2006]. This

type of model makes is possible to consider locally induced differences in growth rate,

a property which would be difficult to achieve in a purely descriptive models, such as

the ones used in Nakielski [2000] and in Chapters 4 and 5. However, this possibility is

not exploited in the work of Jönsson et al. [2006], and their model does have have some

significant disadvantages. During cell division, the model creates unrealistic movements

of cells with respect to each other within the tissue, and the recalculation of the Voronoi

neighborhoods creates unrealistic changes in cell geometry (Figure 3.5). This may be the

reason why, in their work, cell geometry is ignored in the transport simulations, with all

of the cells and cell-to-cell interfaces considered to be the same size in the simulation of

auxin transport and diffusion.

Another approach also uses a two-dimensional polygonal mesh or sheet of cells, but

is able to create an apex model with emergent shape. The mesh is not constrained to

a specified surface, and in order to prevent the it from collapsing the authors apply a

force normal to the surface at each vertex [Smith et al., 2004; Smith, 2006; Stoma et al.,

2007]. This force can be viewed as representing the combined effects of turgor pressure

and other forces from within the meristem. Thus the apex surface can be likened to the

skin of a balloon that is holding back pressure from within. In order to anchor the apex

in space, the positions of the vertices at the bottom of the apex are fixed. Growth can be
94

simulated by changing the rest lengths of the springs [Stoma et al., 2007], by subdividing

and adding additional springs [Smith et al., 2004, 2006b], or by a combination of these

methods. Different topologies for the polygonal mesh which defines the surface have been

proposed. Stoma et al. [2007] uses the junctions of the cell walls as the vertices of the

mesh, so that the polygons represents cells, whereas Smith et al. [2004] and Smith [2006]

use a triangular mesh that in general has no correspondence to a cellular structure.

The use of this “balloon” model makes the modeling of primordia bulging from the

surface straightforward. The model of Stoma et al. [2007] simulates this bulging after

the application of auxin to a simulated PIN1 mutant apex by increasing the rest length

of springs based on auxin concentration. Smith et al. [2004] and Smith [2006] use the

inhibition field phyllotaxis model of Thornley [1975] to position primordia, which then

grow out from the apex by the insertion of extra springs due to localized subdivision at

the tips of primordia.

6.2 A mass-spring model of the shoot apex with a cellular struc-

ture

The construction of a mass-spring model of the shoot apex presented here is similar to

that of Stoma et al. [2007], and is an extension of the mass-spring model of a cellular

tissue by Fracchia et al. [1990]. Starting with the cellular surface described in Chapter

3, the walls are considered to be springs that are connected to point masses located at

the junctions of the cell walls. The point masses at all such junctions are considered to

have an identical value of 1.

The force acting on the point mass of a vertex v located at position pv due to springs
 95

is based on Hooke’s law and is calculated as:

X  kpu − pv k

pu − pv
s
Fv = k 1− (6.1)
u∈N
lv→u kpu − pv k
v

where Nv is the neighbors of vertex v, pu is the position of a neighbor vertex u, k is the

spring constant per unit length, and lv→u is the rest length of spring joining vertex v to

neighbor neighbor vertex u. The norm symbol indicates the Euclidean norm or distance

between the points. The expression within the brackets is positive when the distance

between v and its neighbor u less than the rest length, and negative if it is greater. Note

that for a given difference from the rest length, the magnitude of the force is reduced

as the rest length of the spring increases. This magnitude is then multiplied by the

normalized vector (pu − pv )/ kpu − pv k and by k to give the forces. These forces are

then summed over all the neighbors of the vertex and multiplied by the spring constant

which corresponds to the stiffness of the springs.

In addition to the forces on a vertex due to springs, a uniform force representing the

internal turgor pressure Ft is included in the model which acts in the direction of the

surface normal at each vertex. A real plant apex is able to control its growth direction,

enabling a shoot to grow straight, to favor an upward growth direction, or to grow towards

the light. These processes are not modeled directly but are accounted for by a directional

force Fd which is applied to the vertices at the tip of the apex where directional growth

occurs. This directional force is also applied at the tips of growing primordia, although

the strength of the force may be different than that used for the main apex. Combining

the various components, the total force from all source Fv acting on a vertex v is:

Fv = Fsv + Ftv + Fdv . (6.2)

Note that for cells in the apex which are not at the tip of a growing apex or primordia,

the last term Fdv is zero.

96

Once the forces on the vertices are calculated, the steady-state of the system is de-

termined (see Section 6.2.3), and growth is simulated by increasing the rest lengths of

the springs. If a spring is longer than its rest length, then its rest length is increased

proportional to this difference. In the case of cells walls within the extent of an apex or

primordium tip, this increase is larger, since the directional force F d is also applied. This

causes the apex and bulging primordia to grow quickly at the tips, while still increasing

in size at a lesser rate at locations more distal from the tips.

6.2.1 Local determination of apex and primordium centers

In the previous section a directional force was specified for vertices near the tip of the

apex. In general it is not straightforward to determine the center of the apex tip, and

which cells belong to it, in cellular models with emergent growth. The problem is even

more difficult when trying to determine the centers of protruding primordia. In the model

proposed by Smith et al. [2004] a single vertex is designated as the center of the tip of the

apex or primordium, and the subdivision scheme used keeps this vertex at the center as

the simulation proceeds. In cellular models, however, a central cell may divide, causing

the center of the tip of the apex or primordium to shift as one of the new daughter cells

is chosen as the center. In addition, the irregularity of the wall lengths can cause cells to

grow at different rates, shifting the location of a central cell of an apex or primordium

tip.

To overcome this problem, a method is needed so that the position of an apex or

primordium center can be determined on the basis of local information contained in

the cells that make up the primordium itself. Ideally, this would involve modeling the

molecular components that are responsible for determining apex or primordium identity,

extent, and boundary. In the model presented here, a somewhat simpler approach is

taken which nonetheless retains a local character.

 97

For the main apex, the center is specified at the start of the simulation, whereas

primordium centers are determined as the simulation proceeds. Along with the location

of the center, an initial extent of the apex or primordium is specified, which may change

as the apex or primordium ages. All of the vertices within the radius of this initial extent

are considered as part of the growing tip. These vertices are subject to the directional

force Fd discussed previously, and participate in the dynamic process of determining the

center of the tip. At the end of each time step, after growth, the physics simulation,

and any cell divisions have occurred, all of the positions of the vertices belonging to the

apex or primordium center from the previous step are averaged to determine a tentattive

center point for the next iteration. This average is weighted in favor of vertices closer to

the previous center, and is in general not on the apex or primordium surface. A point on

the surface is obtained by projecting this tentative center onto the surface layer of cells,

in the direction of force Fd . All of the points that are within the apex or primordium

extent are then updated with the distance to the new center, and will participate in

determination of the center at next time step.

6.2.2 Calculation of distance on the apex surface

The maintenance of the apex and primordium centers, as well as the determination of

which cells belong to a particular apex or primordium, requires the calculation of their

distances to their respective centers. For the main apex itself, Euclidean distance can

be used, however this can cause problems with protruding primordia. If the extent of

a primordium at initiation reaches to the other side of a tall slender apex, vertices on

the other side of the apex might be included. If the tip of one primordium comes too

close to the flank of another, it can recruit vertices belonging to a completely different

primordium.

This problem is solved by using the distance along the apex surface, instead of the
98

distance through space, when determining how far a vertex is from an apex center.

Although calculating the geodesic would give the most accurate measure of distance on

the surface, the weighted graph distance (using the lengths of the edges), which can be

calculated with Dijkstra’s algorithm, has several advantages. Aside from being simple,

fast, and east to implement, graph distance might actually be a more accurate measure

of distance as far as the diffusion and transport of molecular components is concerned.

In the phyllotaxis simulations that follow, a variant of Dijkstra’s algorithm is used to

find the distance from the center for all vertices belonging to the main apex or primordia

tips. These cells are visualized in blue in Figure 6.1.

6.2.3 Physics simulation

In mass-spring simulations it is often desirable to visualize the dynamic behavior of a

system as it progresses towards a steady state. This would be the case for the simulation

of a bouncing ball, where the dynamic behavior of the physics simulation is of interest.

In the case of the growing apex model presented here, each time step starts with the

simulation at equilibrium. Growth, which changes the rest lengths of the springs, upsets

this equilibrium and the system must be solved for the equilibrium state under these

new conditions. This means that in general, there will be many iterations of the physical

simulation for each growth step of the model.

There are several methods by which the equilibrium state of the system can be deter-

mined. Given the total force Fw on a vertex w as defined earlier, the following equations

describe the change in velocity v and position p over time:

dvw
= (Fw − ςvw )dt (6.3)
dt
dpw
= vw dt (6.4)
dt

where ς is the damping constant.

 99

This assumes that the point masses at all of the vertices in the model are equal to 1.

Using the forward Euler method, the values for velocity and position at time t + ∆t can

be calculated as:

vt+∆t = vt + (F − ςvt )∆t (6.5)

pt+∆t = pt + vt ∆t. (6.6)

Although these formulas give acceptable results, if the velocity at time t + ∆t is

substituted in Equation (6.6) for vt , then the step size can be increased by a factor of 5 or

more in many cases. This is because in this case the resulting update formulas correspond

to the energy-conserving discrete Lagrangian for the system and has been called the

symplectic forward Euler method [Stern and Desbrun, 2006]. With this modification the

update formulas become:

vt+∆t = vt + (F − ςvt )∆t (6.7)

pt+∆t = pt + vt+∆t ∆t. (6.8)

The physics simulation in each growth step is a separate simulation within the main

simulation, which has its own time step and integration routines. This simulation is

iterated until an equilibrium state is reached which is determined by the examining

the maximum total force acting on any vertex. Since the total force will go to zero

when the system is at a steady-state, this can be used to apply a threshold to stop

the iteration. In practice it is convenient to also limit the number of iterations, as this

improves performance and smooths out the simulation when primordia first appear and

tend to “pop out” from the apex surface.

Since only the equilibrium state of the physics simulation is sought at each time step

of growth, a relaxation method can be used that does not consider velocity. Vertices can
100

simply be moved in the direction of the total force by some amount proportional to the

magnitude of the force. The system is then iterated until the forces approach zero. In

the two models that follow both approaches have been used, and both methods produce

similar results with similar execution times.

6.3 Inhibition field model of phyllotaxis on a physically-based

model of the shoot apex

In this section the inhibition field model of phyllotaxis described in Chapter 4 is combined

with the physically-based model of the shoot apex. The inhibition model is used to specify

the locations of new primordia, with the growth of the apex and of the outgrowing

primordia modeled with the mass-spring system described in the previous section.

The inhibition field model of phyllotaxis uses the active ring as an abstraction for the

peripheral zone. A function of distance and primordium age defines the inhibiting effect

that preexisting primordia have on the active ring. When a location on the active ring

drops below the model threshold, a new primordium is initiated.

In order to locate the active ring, the central axis of the apex must be defined. This

axis is located at the line through the center point of the apex tip in the direction of the

directional force Fd , which for the main apex is straight up. The ring is located at a fixed

distance on the surface below the apex tip center. This distance is supplied to the model

as a parameter. When a location on the active ring is chosen for primordium initiation,

it becomes the center point for the new primordium tip. The initial directional force

for primordia is set to the surface normal at this location and gradually shifted towards

vertical, causing primordia to curve upward as they grow.

As in the model of Chapter 4, growth causes the apex tip to grow up and away from

the newly formed primordium. The primordium also grows out from the apex surface.
 101

Figure 6.1: Inhibition field model of phyllotaxis combined with a physically-based model
of a shoot apex with growing primordia. A self-starting Fibonacci spiral pattern is
shown. Cells in blue represent cells which are considered part of the growing apex tip or
the growing tips of primordia.

Time and distance reduce the inhibiting effect that this and other primordia have on the

active ring until another primordium appears and the process repeats. Combining all of

these events, the simulation results in a self-starting Fibonacci spiral phyllotaxis pattern

(Figure 6.1). The physics simulation is implemented with the symplectic forward Euler

method described in the previous section.

102

6.4 Transport model of phyllotaxis on a physically-based model

of the shoot apex

Since the physically-based apex is implemented as a cellular structure, it is straightfor-

ward to incorporate the transport model of phyllotaxis described in Chapter 5. The same

basic model is used, but the rules for the production of auxin are simplified with auxin

only being produced in the peripheral zone. The size and location of the peripheral zone

are given as model parameters, and do not change throughout the simulation. The rules

for diffusion and decay of auxin, as well as for the production, decay, and localization

of PIN1 are the same as in the model of Chapter 5. The rules for primordia initiation

are also the same, with primordia being initiated when a pair of cells exceeds an auxin

concentration threshold.

Figure 6.2 shows a simulation using the transport model on the physically-based apex.

In this model, the relaxation method is used in the physics simulation, but the result

is very similar to the inhibition field model on the same surface which used symplectic

forward Euler.

6.5 Summary of results of the phyllotaxis models combined

with physically-based shoot apex models

Simulations of phyllotaxis implemented on the shoot apex with physically-based growth

require considerably more computation time than the simulations using descriptive growth

described in Chapter 4 and 5. This is because each growth step requires that the physics

simulation be iterated until equilibrium is reached. Implicit methods, such as those dis-

cussed in Chapter 9, do not improve upon this situation because, in the case of forward

Euler, the increase in time step that can be achieved does not make up for the additional
 103

Figure 6.2: Transport model of phyllotaxis combined with a physically-based model of

a shoot apex with growing primordia. The simulation uses a mass-spring model solved
with a relaxation method. Cells that are brighter green have high auxin concentration
which initiates organ formation. Red indicates PIN1 proteins with polar localization that
direct auxin transport towards organ initiation sites.

computational overhead of the more advanced methods. This is also true for the relax-

ation model, where a more sophisticated method to solve the non-linear system takes as

much time to run as the simple iterative approach used here.

Despite this drawback, the physically-based models of the shoot apex and young
104

developing primordia presented here provide a good framework for further work. It is

quite straightforward to add tropisms to developing primordia or to have lateral apices

which themselves produce primordia. In addition, it is possible to model primordia

which are not radially symmetric, and develop into leaves, flowers, or other organs. All

of these would be much more difficult to implement in the descriptive growth models

presented earlier. In addition, the physically-based model of the shoot apex provides

an excellent platform to study the mechanisms that determine primordium extent and

boundary determination, as well the possible influence of these processes may have on

the fine tuning primordia positioning in phyllotaxis models.

 Chapter 7

Optimality in Phyllotaxis: A Light Capture Study

7.1 Phyllotaxis, light capture, and evolution

The frequency of spiral phyllotaxis patterns in nature raises the question as to whether

or not these patterns are optimal in some way that is beneficial to the plant. King et al.

[2004] presents a simulation model of light capture which provides support for the idea

that the Fibonacci angle of 137.5◦ optimizes light harvesting, and has been selected by

evolutionary pressure. In contrast, Niklas [1998] proposes that morphological features

not correlated to the divergence angle, such as leaf shape and internode length, can

significantly affect the fitness landscape. These changes reduce the effect of a non-ideal

angle, and thus the importance of evolutionary pressure on the divergence angle.

In this chapter a simulation model for light capture is presented, and the fitness

landscape is examined over a range of conditions. Although previous models provide

support for the notion that the Fibonacci angle provides optimal light capture, this

study indicates that other noble angles are equally optimal. The model confirms previous

results [Niklas, 1998; King et al., 2004] according to which morphological features of the

plant can reduce the advantage that ideal divergence angles provide. A similar reduction

in advantage of the ideal angles is shown for the effects of varying the angle of incidence

of incoming light.

7.2 A simulation model of light capture

The plant model used for light capture analysis is a simple, fairly compact, monopo-

dial form with no lateral branching (Figure 7.1). It was constructed in the L-Studio

105
106

Figure 7.1: Top and side views of a plant model. Only the leaves are visualized. Leaf
size, angle between the leaf and the stem, internode length, and petiole length change as
a function of the distance from the apex tip.

modeling environment [Prusinkiewicz, 2004]. The plant model has a central axis, with

successive leaves placed at a fixed divergence angle θ around this axis. Only the leaves

are modeled, ignoring the shading and light gathering effects of the stem and petioles.

Leaves are constructed using Bezier patches with the interactive surface editor provided

by L-Studio. Four interactively defined functions are used to specify model parameters

that vary based on their position on the plant stem [Prusinkiewicz et al., 2001]. These

parameters determine various leaf attributes as a function of their distance from the apex

tip. The functions control the following leaf attributes:

• Leaf size.

• Internode length.

• Angle between the leaf axis and the central axis of plant.

• Petiole length (distance between the base of the leaf and the central axis).
 107

Several methods have been used previously to assess the light capture ability of plant

models. King et al. [2004] defines a shadow function for each leaf, which is used to ap-

proximate the shading effects of upper leaves on the lower ones, taking into account a

decrease in the shading effect with distance. Pearcy and Yang [1998], interested in a Red-

wood forest under-story plant Adenocaulon bicolor, use a fairly elaborate method based

on canopy photographs, taking into account diffuse light from different angles combined

with occasional direct light coming from holes in the forest canopy. Niklas [1988, 1998]

uses a simpler approach by considering only direct light of varying angles based on the

position of the sun. In all of these approaches only single plants are considered, without

considering the effects of shading from neighbors.

An approach similar to that used by Niklas [1998] is used here to calculate light

capture. Only direct light is considered, and total light capture is calculated by projecting

the leaves on a plane orthogonal to the direction of the incoming light. The plane is then

divided into pixels, with the pixels that are contained in leaves summed to determine the

total light capture.

These calculations can be done quickly by exploiting the graphics hardware of a

personal computer, using the pixels of the bitmap image in the frame buffer as the

discretization. Flat shading, combined with a surface material with only an ambient

component, gives an extremely fast rendering in which the color of pixels does not depend

on the viewing angle (Figure 7.2). If the view point is considered to be the position of

the light source, then the resulting image gives a two dimensional projection of the

leaves as seen by the light source. Only the leaves are rendered, and the pixel values

for the background are set to zero (black). The total light capture of the plant is then

determined by summing the green channel values of the pixels in the frame buffer. This

method also has the added advantage of visual verification, since the actual projection

used for calculating light capture model can be monitored as the simulation progresses.
108

Figure 7.2: Rendering of plant models used for light capture calculations. By using
flat shading and a material containing only an ambient component, a two dimensional
projection of the portions of the leaves exposed to sunlight can be obtained. A count of
the green pixels gives the total amount of light captured.

Translucency can also be simulated by using blending.

All light capture simulation programs were written in C++ using the VV [Smith et al.,

2004] and L-Studio [Prusinkiewicz, 2004] modeling environments. The images for light

capture calculations were rendered with OpenGL with a frame buffer resolution of 1600

x 1600.

7.3 Light capture simulation model results

In the first series of simulations incoming light came from a fixed angle directly over-

head. The total light capture for plant models containing 30 leaves was calculated for

divergence angles ranging from 60◦ to 180◦ , in increments of 2◦ . To determine how light

capture changes as the leaf width/length ratio changes, 30 series of these simulations

were performed for various leaf width/length ratios ranging from 1/8 to 3/4 (Figure 7.3).
 109

Figure 7.3: Examples of plant models with different leaf width to length ratios. Values
range from 1/8 (left) to 3/4 (right).

7.3.1 Light capture as a function of leaf width and divergence angle

Following the presentation of Niklas [1998], results of the simulation are summarized in a

three dimensional plot, giving a fitness landscape for the two traits: divergence angle and

leaf width (Figure 7.4). It is not surprising that as leaf width increases, so does the overall

light captured. There is a noticeable dip at divergence angles of 90◦ and 120◦ where the

leaves line up in groups of four and three respectively. A similar fitness landscape is

depicted in Figure 5 of Niklas [1998], but it looks simpler because the complexity is

obscured by the clustering of the points about the golden angle of 137.5◦ . In Niklas’

figure, there is only a single global maximum about 137.5◦ , likely because there are not

enough points plotted between 137.5◦ and 180◦ . In contrast, the simulation presented

here (Figure 7.4) shows that there are other local maxima in this area (see also Figure

7.5). In agreement with Niklas results, narrow leaves show an increased sensitivity to the

divergence angle1

Although previous models suggest that a global maximum at 137.5◦ should appear,

this is not the case. The plot in Figure 7.4 shows many local maxima with no clear global

maximum for the entire range of low (1/8) to high (3/4) leaf width/length ratios.

Both Niklas [1998] and King et al. [2004] considered the transparency of leaves in their
1
In Niklas [1998] the leaf area is constant, whereas in the simulations describe here the length of the
leaf is constant. Thus Niklas’ results are scaled to total leaf area so that long narrow leaves become
more fit due to reduced shading from other leaves.
110

Light Capture

0.75 Leaf
width/length
0.125
60 70 80 90 100 110 120 130 140 150 160 170 180
Divergence Angle

Figure 7.4: Light capture fitness landscape relating leaf width/length ratio to divergence
angle for a simulated plant with 30 leaves. Incoming light is directly overhead (parallel
to the stem axis).

models, either directly or indirectly. In this study, a simulation with partially transparent

leaves was performed by using OpenGL blending and a semi-transparent rendering of

the leaves. This produced almost identical results to simulations without transparency.

Although overall light capture was reduced, the shape of the fitness landscape looked

almost identical to the model without transparency and is not presented here.

7.3.2 Discretization error

Any calculation involving discretization is subject to error. If the model is rotated

and light capture recalculated, a slightly different value results. In order to minimize

this effect, a series of simulations at different rotations where performed using a leaf
 111

Light Capture 64 78 99.5 106.5 132 137.5 151 158 162.5165

103 154.5
72 144
90
120

60 70 80 90 100 110 120 130 140 150 160 170 180

Divergence Angle

Figure 7.5: Light capture as a function of divergence angle for medium-width leaves
(width/length = 3/8). Simulated plant has 30 leaves and direct overhead light paral-
lel to stem axis. To reduce the error due to discretization, each point represents the
average over 90 simulation runs, with the initial leaf position varying from 0◦ − 90◦ .
Angles marked without boxes denote convergence angles for Fibonacci-like summation
sequences, and generally correspond to local maxima: 64◦ = 3rd accessory, 78◦ = 2nd
accessory, 99.5◦ = Lucas, 106.5◦ = other, 132◦ = other, 137.5◦ = Fibonacci, 151◦ = 1st
anomalous, 158◦ = 2nd anomalous, 162.5◦ = 3rd anomalous, 165◦ = 4th anomalous. An-
gles indicated in boxes are rational fractions with low denominators, and correspond to
local minima: 72◦ = 360◦ × 1/5, 90◦ = 360◦ × 1/4, 103◦ ≈ 360◦ × 2/7, 120◦ = 360◦ × 1/3,
144◦ = 360◦ × 2/5, 154.5◦ ≈ 360◦ × 3/7.

width/length ratio of 3/8 (medium width). For each divergence angle considered, 90

simulations were performed with the only difference being an initial rotation between 0◦

and 90◦ . Averaged results are shown in Figure 7.5.

Although various Fibonacci-like summation sequences are associated with conver-

gence angles that are near local maxima, they are often not located exactly at the peak

of these maxima. This is likely due to the relatively low leaf count and the width of the

leaves used in the simulations. If a greater number of narrower leafs were used, the noble

angles would become closer to the peaks of the maxima. This happens because irrational

divergence angle prevent the leaves from completely shadowing each other, allowing more

leaves to contribute to total light capture. If extremely narrow leaves are used, then a

larger number of leaves are required to completely fill all of the area that is possible,
112

requiring higher denominators of rational fractions. Thus as the total number of leaves

contributing to light capture increases, the irrational noble numbers become more and

more optimal.

Note that most of the local maxima in Figure 7.5 are similar in magnitude, indicating

that the various forms of spiral phyllotaxis, corresponding to different noble angles, are

equally fit from a light capture perspective. Perhaps more insight can be gained from

Figure 7.5 by looking at the valleys rather then the peaks. All of the rational fractions

with small denominators within this interval are represented by local minima. This

corresponds to locations where a number of leaves are superposed. Fractions such as 1/3

and 1/4 have a large number leaves on top of each other, where as smaller fractions,

such as 2/7 have less. Thus the graph in Figure 7.5 can be viewed as a straight line

with a collection of valleys that grow increasingly smaller as the denominators of their

associated fractions increase.

7.3.3 The effects of the angle of incoming light

Up to this point only overhead light has been considered. In Niklas’ model, the angle

of incoming light was varied to follow the path of the sun throughout the day, with the

total value over the day calculated. Although this may be more biologically realistic, it

is also interesting to examine the fitness landscape at particular angles. This shows how

the angle of incoming light affects the interaction between leaf width/length ratio and

the divergence angle. Figures 7.6−7.8 show fitness landscapes with the incoming light at

angles of 15◦ , 30◦ , and 45◦ from vertical.

The shape of the fitness landscape for incoming light at 15◦ is qualitatively similar

to when the light comes from directly overhead. This is especially true for wider leaves

(compare Figures 7.4 and 7.6). If shading from other plants is considered, considerably

more light will be collected at small angles from vertical, when the sun is higher in the sky.
 113

Light Capture

0.75 Leaf
width/length
0.125
60 70 80 90 100 110 120 130 140 150 160 170 180
Divergence Angle

Figure 7.6: Light capture fitness landscape for a simulation with 30 leaves with the
incoming light at a 15◦ angle to the stem axis.

This provides justification for the approach used by King et al. [2004] who only considers

light coming from directly overhead when investigating the relationship between light

capture and divergence angle.

As the angle of incoming light and the stem axis increases, the data become noisier,

and the divergence angle plays a lesser role in overall fitness. At 45◦ , narrow leaves

show almost no sensitivity to the divergence angle and wider leaves show a considerably

reduced sensitivity (Figure 7.8).

7.3.4 The effects of reduced numbers of leaves

All of the previous simulations were done with a plant model containing 30 leaves. In

this section the investigation focuses on the effects of reduced numbers of leaves on the
114

Light Capture

0.75 Leaf
width/length
0.125
60 70 80 90 100 110 120 130 140 150 160 170 180
Divergence Angle

Figure 7.7: Light capture fitness landscape for a simulation with 30 leaves with the
incoming light at a 30◦ angle to the stem axis.

relationship between light capture and divergence angle. Since plants often do not grow

in isolation, it is common that only the uppermost leaves receive large amounts of direct

light. With this in mind, simulations were performed for a fixed leaf width while varying

the number of leaves on the plant model. A wide leaf width was chosen, as the divergence

angle should have a greater effect on light capture with wide leaves when the plant has

a small number of them. The results for simulations using between 1 and 10 leaves are

shown in Figure 7.9.

On the model with only 2 leaves, light capture steadily increases until around 158◦ ,

when the 2 leaves no longer overlap. On a plant with 3 leaves, the maximum light

capture is at 120◦ as expected. As the number of leaves increases, the shape of the

fitness landscape becomes more complex and quickly approaches that shown in Figure
 115

Light Capture

0.75 Leaf
width/length
0.125
60 70 80 90 100 110 120 130 140 150 160 170 180
Divergence Angle

Figure 7.8: Light capture fitness landscape for a simulation with 30 leaves with the
incoming light at a 45◦ angle to the stem axis.

7.5. A similar effect is seen with narrower leaves, although the valleys are deeper and

the graphs of lower numbers of leaves flatten out sooner. It is interesting to note how

few leaves it takes for the divergence angle to have a significant effect on light capture.

Taking the same data as presented in figure 7.9, a second plot was created in which

the light capture is averaged over multiple simulations containing up to 10 leaves (Figure

7.10). The data is first plotted is for 2 leaves, then the averages are taken for 2 and

3 leaves. Next the average light capture is calculated for simulations containing 2-4

and 2-5 leaves. The process is repeated until the average is taken over 9 simulation runs

containing from 2 to 10 leaves. Taking averages in this way, the same basic profile emerges

as in the previous section, although up until the average of 10 leaves the local maximum

at 137.5◦ has a slight advantage over the others. This suggests that the Fibonacci angle
116

Light Capture

10
Leaves
2
0 10 20 30 40 50 60 70 80 90 100 110 120 130 140 150 160 170 180

Divergence Angle

Figure 7.9: Light capture as a function of divergence angle for different numbers of leaves
on a plant model. Simulations were performed with a model plant with very wide leaves
(width/length ratio of 10/7).

might have an advantage in the face of uncertainty for plants that produce few leaves,

or in cases where only a few leaves at the very top of the plant are exposed to direct

light. An argument for the role of uncertainty in the evolutionary selection of plants has

recently been proposed by Prusinkiewicz et al. [2007].

7.4 Summary of results of the light capture models

The graph of light capture as a function of divergence angle can be characterized by Figure

7.5. Although changes in leaf shape, leaf angle, petiole length, transparency, or the angle

of incoming light may add noise or flatten the curve somewhat, the same overall shape is

preserved. Local maxima in light capture are associated with the irrational angles from

the Fibonacci, Lucas, and various other Fibonacci-like summation sequence, whereas

local minima are associated with rational angles with low denominators.

In previous light capture simulation studies some authors have reported a global
 117

10
9
8
7
6
5
Light Capture

4
3
2
Leaves

0 20 40 60 80 100 120 140 160 180

Divergence Angle

Figure 7.10: Light capture as a function of divergence angle averaged over different
numbers of leaves. The bottom line on the graph shows light capture for a plant model
with 2 leaves. The next line shows the average of the light captured by simulations with
2 and 3 leaves. This continues until the line labeled 10 leaves, which averages the light
captured from 9 simulations with from 2 to 10 leaves. Simulations were performed with
a plant model with wide leaves (width/length ratio of 10/7).

maxima in light capture at 137.5◦ [Pearcy and Yang, 1998; Niklas, 1998; King et al., 2004],

although in some cases their data show other local maxima very close in amplitude at the

Lucas angle of 99.5◦ [Pearcy and Yang, 1998; King et al., 2004] as well as the anomalous

sequences near 151◦ and 158◦ [King et al., 2004]. The results presented here show that

there are also other angles associated with Fibonacci-like summation sequences that

provide light capture comparable to the golden angle of 137.5◦ . In addition, changing

morphological features of the model plant or changing the angle of incoming light may

flatten the fitness landscape or make it more noisy, but it does not favor the Fibonacci

angle. Even for small leaf counts, except in extreme cases, there is still a variety of angles
118

that can be chosen for optimal light capture.

Thus despite there being several simulation studies on the subject, the question as

to whether the Fibonacci divergence angle is selected by evolutionary pressure remains

unanswered. Perhaps this is not the right question to ask. Although the exact mechanism

for phyllotaxis is still unknown, it is widely believed that phyllotactic patterns are the

result of a leaf spacing mechanism issuing organs one after another on the growing shoot

tip. The simulation models presented in Chapter 4 demonstrate that such a mechanism

is capable of generating all of the angles for the Fibonacci and Fibonacci-like summation

sequences appearing at the local maxima shown in Figure 7.5. In addition, the phyllotaxis

model of Chapter 4 shows that sequences that appear with higher frequency in nature

correspond to sequences that are more stable, and easier to produce in the model. The

inhibition field model also does not, in general, produce spiral phyllotaxis with angles

that are poor for light capture, such as 90◦ or 120◦ . If the question were instead “Why

spiral phyllotaxis?”, rather than focusing on the particular case of “Why 137.5◦ ?”, then

indeed a strong argument for evolutionary pressure could be made. The mechanism of

phyllotaxis is able to produce reasonably optimal light capture with a simple leaf spacing

mechanism on a growing plant tip. Even though this may lead to several possibilities for

the divergence angle, the outcome is similar, as the various spiral phyllotaxis angles all

have relatively equal light capture. Thus it is spiral phyllotaxis itself, and its mechanism

that is being selected, rather than a particular divergence angle.

 Chapter 8

Mechanistic Models of Leaf Venation

8.1 Models of leaf venation

Spiral phyllotaxis is not the only example of conspicuous patterning in plants. The

intricate venation patterns seen in the leaves of vascular plants are in some cases almost

as striking as spiral phyllotaxis. Although the literature on the modeling of venation

patterns is not as vast as for phyllotaxis, there is still a considerable variety in the

proposed algorithms. Gottlieb [1993] proposed a model operating on a grid with grid

subdivision used to simulate growth. Vein tips grow towards the centers of new grid

cells, which can be seen as the source of some hormone. The model of Rodkaew et al.

[2002] uses a particle system, in which the paths traced by the particles represent the

areas of the leaf that differentiate into veins. Runions et al. [2005] proposes an algorithm

similar in spirit to Gottlieb [1993] where vein tips grow towards discrete sources of auxin

in the leaf lamina. These algorithms are all biologically inspired and model the processes

of leaf venation at a similar level of abstraction as the inhibition field model of phyllotaxis

presented in Chapter 4. Runions et al. [2005] presents a nice review of this class of models

as well as a new algorithm that produces very convincing leaf venation patterns.

The first mechanistic simulation model of leaf venation was proposed by Meinhardt

[1976] with several variants given in Meinhardt [1982]. His reaction-diffusion model

creates peaks of self-enhancing morphogen at growing vein tips. These peaks cause

differentiation of the cells at the tips, which in turn changes the tip cells’ behavior,

pushing the peaks to adjacent cells. This leads to extension of the veins at the tips.

However, as in the case of organ initiation, experimental data points to the transport

119
120

of auxin, rather then the local production of a morphogen, as a key factor in the selection

of cells for differentiation into procambium [Sachs, 1981, 1991; Galweiler et al., 1998;

Reinhardt et al., 2003a; Scarpella et al., 2006]. This data is largely consistent with

the canalization hypothesis for the initiation of vascular systems in plants [Sachs, 1981].

This hypothesis proposes that the canalization of auxin into preferred routes of auxin

flux occurs in much the same way as water carves rivers in soft terrain. A cell’s ability

to transport auxin is assumed to increase with auxin flux, causing any initially dominant

path to be reinforced. As in the case of river formation, as soon as the smallest canal

begins to emerge due to random variation, it will be accentuated and attract even more

flow, causing the preferred path to strengthen. Simulation models of the canalization

hypothesis on grids of cells have shown that this mechanism is indeed capable of selecting

strands from a tissue of undifferentiated cells [Mitchison, 1980, 1981; Rolland-Lagan

and Prusinkiewicz, 2005; Feugier et al., 2005; Feugier and Iwasa, 2006]. These models

simulate the production, decay, diffusion, and transport of a single morphogen, auxin,

with the procambium assumed to differentiate along the paths of high auxin flux. It is

usually assumed that the same PIN1 auxin transport proteins implicated in phyllotaxis

patterning are responsible for the creation of the high throughput paths, and that their

concentration at the wall of a cell is a non-linear, increasing function of the flux of auxin

at that wall.

Thus the hypothesized mechanisms for phyllotaxis and leaf venation both involve

transport-based patterning, with the same morphogen and the same transporter pro-

teins. The main difference between the mechanisms is how the orientation of the PIN1

transporter proteins is determined, a process for which the molecular basis is still largely

unknown. In the case of phyllotaxis, PIN1 proteins are hypothesized to orient pref-

erentially up the auxin gradient, whereas in the canalization model PIN1 proteins are

hypothesized to orient in the direction of greater flux.

 121

There is a strong experimental basis for a close link between the molecular mechanism

of phyllotaxis and procambium initiation. The earliest known marker for both organ

initiation [Reinhardt et al., 2003a] and procambium initiation [Scarpella et al., 2006] is

the dramatic upregulation of the PIN1 transport protein. The convergence point of auxin

in the meristem that leads to organ initiation also leads to the initiation of the midvein

of the leaf. In addition, micro-application of auxin is able to induce organ formation in

the meristems of tomato and Arabidopsis [Reinhardt et al., 2000, 2003a], as well induce

additional convergence points and vascular tissue in the leaf [Sachs, 1981; Scarpella et al.,

2006].

At this point it is worth mentioning that Couder et al. [2002] proposed a very different

mechanistic model for leaf venation. They present a physical model in which a gel film

is dried under controlled conditions. This leads to the creation of patterns that resemble

leaf venation, much like those seen in the cracking of drying mud. They propose that

leaf venation occurs as the result of stresses in the leaf lamina that occur due to growth.

Although it is not widely accepted that the primary mechanism for leaf venation pat-

terning is physically-based, their model does suggest that diverse patterning phenomena

can be unified at a higher level into generalized pattern formation mechanisms.

8.2 A growing leaf surface

Previous mechanistic simulation models of leaf venation consider only pattern formation

on static grids of cells [Meinhardt, 1976; Mitchison, 1980, 1981; Meinhardt, 1982; Feugier

et al., 2005; Rolland-Lagan and Prusinkiewicz, 2005; Feugier and Iwasa, 2006]. In the

section that follows, canalization models are explored on a growing cellular leaf structure.

This provides insight into how the canalization model behaves on a growing structure

with irregular geometry.

122

(u,v) coordinates
1.0

0.0 1.0
u
(x,y,z) positions

t=0 t = .5 t = 1.0

Figure 8.1: Modeling a growing leaf by using Bezier surfaces. A sequence of Bezier
surfaces represents the leaf at various times t during the simulation. Coordinates in the
parameter space (u, v) shown in the top left are mapped into three-dimensional space
(x, y, z) as shown. The position of points at time steps in between the three key frames
is found by linear interpolation.

The leaf structure used in the canalization models presented here is implemented

by defining a sequence of Bezier surfaces that represent the leaf at different times (key

frames) during the simulation (Figure 8.1). These surfaces are defined interactively using

the surface editor provided with L-Studio [Prusinkiewicz, 2004]. Linear interpolation of

the control points of the surfaces is used to produce the leaf at any desired point between

the key frames. By advancing time in small steps, a smoothly growing leaf surface is

produced.

The Bezier surfaces used to model the leaf are two-dimensional parametric surfaces

embedded in three-dimensional space. Points on the surface are represented by coordi-

 123

Figure 8.2: Cell division on a leaf modeled with Bezier surfaces. The model starts with a
single cell covering the entire surface at time t = 0. This cell is subdivided until all cells
are below the threshold area. As the surface grows, cells divide when their area exceeds
the threshold.

nates (u, v) which are mapped by Bezier surface evaluators to positions (x, y, z). As the

simulation progresses and the leaf changes shape, a (u, v) coordinate at the tip of the leaf

will remain at the tip even though its actual position in space might change considerably.

The same cellular structure described in Chapter 3 and used for the phyllotaxis model

in Chapter 5 can be implemented on the growing leaf. The simulation starts with a single

initial cell that covers the entire surface (Figure 8.2). The same cell division algorithm

used in the transport model of phyllotaxis is applied until there are no cells left above

the threshold area. As the simulation proceeds, the size and shape of the leaf changes,

causing the polygons that define the cells to enlarge. When cells reach the threshold

area, they are divided.

The C++ program library that implements the Bezier surface has two main functions;

one to compute the (x, y, z) coordinates of a point on the surface given its (u, v) coordi-

nates, and the other to perform the inverse operation. The evaluator used to compute

the (x, y, z) coordinates was programmed from the equations given in Neider et al. [1994],
124

however in general there is no closed formula to compute the (u, v) coordinate given a

particular (x, y, z) position that may or may not be on the surface. This inverse operation

is required whenever new points are added to the model and is implemented by using a

steepest decent search algorithm on the evaluator function.

8.3 Canalization on a cellular leaf structure

The first model of canalization presented here follows the polar transport model presented

in Rolland-Lagan and Prusinkiewicz [2005], which is based on of Mitchison [1980, 1981].

Since both of these models were implemented on rectangular grids of cells, adjustments

must be made for the variable number of neighbors and the irregular geometry of the

cellular surface. Each cell contains an auxin concentration, as well as the amount of PIN1

transporter proteins located at the interfaces facing each of its neighbor cells. Diffusion,

transport, and decay of auxin are modeled, as well as production at discrete sources

placed programmatically on the leaf surface. Extracellular space is ignored, with auxin

moving directly from cell to cell as the result of transport and diffusion. The following

equation specifies the change in auxin concentration of a cell i:

d[IAA]i D X
= ρ − µ[IAA]i + li→k ([IAA]k − [IAA]i )
dt Ai k∈N
i

T X
+ li→k (Pk→i [IAA]k − Pi→k [IAA]i ) (8.1)
Ai k∈N
i

where [IAA]i is the auxin concentration of cell i, ρ specifies the rate of auxin production,

µ specifies auxin decay, Ni represents the neighbors of cell i, D is the diffusion coefficient,

T is the transport coefficient, li→k is the length of the interface between cell i and cell

k, Pi→k is the number (per unit length) of transport proteins on the interface, and Ai

is the area of cell i. The coefficient ρ is zero in all cells except the auxin sources, and

the coefficient µ is increased at sink cells located at the base of the leaf. These sink cells
 125

Figure 8.3: Favoritism of canalization towards shorter walls. The cell on the bottom left
has a fixed concentration of auxin that is higher than its neighbors, all of which have
the same area. At first diffusion to all the neighbors is equal per unit length of interface,
since the gradient between the cell and its neighbors is the same for all cells. Since longer
interfaces will allow more auxin to diffuse, the two neighbors with the longer interfaces
will increase in auxin concentration quicker than the one with the shorter interface. This
results in a larger gradient across the shorter wall than the other walls. A larger gradient
results in a larger production of PIN1 proteins at this interface.

with increased decay represent the effects of connecting to the vasculature of the rest of

the plant at the petiole.

The production of PIN1 transporter proteins uses an independent production model,

which means the amount of PIN1 at a particular cell interface is independent of the

quantities at the interfaces with other cell neighbors. When flux is less than or equal to

zero, there is no flux based production. There is a maximum PIN1 amount allowed at

any one interface. Let the flux from cell i to cell k be defined as:

φi→k = D([IAA]i − [IAA]k ) + T (Pi→k [IAA]i − Pk→i [IAA]k ) (8.2)

then the equation to model the change in PIN1 proteins at the interface from cell i to

cell k is:
dPi→k p
= α li→k max(0, φi→k )2 + β − γPi→k (8.3)
dt

where α controls the production of PIN1 dependent on auxin flux, β controls background

production of PIN1, and γ controls decay. The rate of production of PIN1 is a quadratic
126

function of flux, with no flux related production when the flux is less than zero. Note

that the flux related production depends on the square root a power of the length of

the cell to cell interface li→k . This is required to remove the bias in the canalization

model towards short walls (Figure 8.3). Previous simulation studies of canalization were

always done on regular grid structures, with all of the interfaces between cells being the

same size, whereas in the cellular leaf model the walls can vary in length. The variation

in results in an unequal diffusion of auxin to neighbors cells, causing the gradient to be

larger for neighbors with shorter interfaces. This will favor PIN1 accumulation at shorter

interfaces. A similar problem occurs with cells of varying areas, as larger cells will have

a smaller change in concentration than smaller ones for the same total flux. Since the

cell division algorithm produces cells that are roughly similar in size, no adjustment is

made to PIN1 production at interfaces based on neighbor cell area.

8.3.1 Canalization on a growing leaf with discrete sources

In the Arabidopsis shoot apex the upregulation of PIN1 at the sites of organ initiation

appears simultaneously with the expression of PIN1 in internal tissues [Reinhardt et al.,

2003a]. In his conceptual model for phyllotaxis Reinhardt suggests that the same con-

vergence points of auxin that initiate leaf primordia are also responsible for the initiation

of the midvein of the leaf. Scarpella et al. [2006] provides support for this assertion and

further suggests that the first lateral veins are also initiated by convergence points of

auxin in the leaf margin.

In the first simulation of canalization presented here, these convergence points are

simulated by designating small groups of cells as sources that produce auxin (ρ > 0 in

Equation 8.1). Cells at the base of the leaf are designated as sinks. These cells have an

increased value for auxin decay (µ in Equation 8.1) and represent the effects of connecting

to the existing vasculature of the plant.

 127

Figure 8.4: Simulation of canalization on a growing leaf. Starting with the initial source
at the tip of the leaf, a canal emerges progressing from source to sink. As new sources
appear, more canals emerge connecting them with existing ones. Dark cells at the bottom
of the leaf are sinks for auxin and represent a connection to the stem vasculature of the
plant.

The simulation begins with sink cells located at the base of the leaf. After a short

time, auxin source cells are placed at the tip of the leaf that initiate the formation of the

midvein. The simulation continues and sources are introduced at the sides of the leaf

which initiate canals that connect to midvein (Figure 8.4). As more sources are added,

new veins are initiated that connect up with neighboring vasculature. Often the canals

are initially comprised of two or more files of cells and gradually narrow to a single file.

If the production at the source cells is increased, the source can support several

canals. In this case loops can form, with the polarity of the cell files changing direction

at the sources (Figure 8.5). Although Scarpella et al. [2006] does report a reversal of cell

polarity towards the upper middle portion of the initial vein loops in Arabidopsis, they

were unable to confirm a convergence point of auxin at this location. They did notice,

however, that this transition point is often marked by a single bipolar cell.
128

Figure 8.5: Simulation of loop formation on a growing leaf. If the sources are strong
enough, several canals connecting with preexisting ones can be formed. This leads to the
emergence of closed venation patterns. The source at the tip of the leaf is removed just
before the sources at the sides appear. This causes the tip to be a strong sink for auxin
helping to attract the upper portions of the loops. The dark cells at the bottom of the
leaf are sinks for auxin and represent the connection to the existing vasculature of the
plant.

8.3.2 Canalization on a growing leaf with emergent sources

The first convergence point that initiates the midvein is created by the same process that

initiates organ formation. In the phyllotaxis model of Chapter 5 this is hypothesized to

be caused by self-emergent behavior of the PIN1 mediated auxin transport system. The

convergence points that lead to the initiation of the second step, the lower loop lateral

veins, are also thought to be created by PIN1 mediated auxin transport Scarpella et al.

[2006] in the margin of the leaf.

This suggests a combination of mechanisms, where PIN1 proteins in the cell margin

are oriented up the auxin concentration gradient, with canalization operating in the

interior cells. In the second model presented, the margin cells use the same equation

as the phyllotaxis model (Equation 5.7) to orient PIN1, with the interior cells using

Equation (8.3) as before. The margin cells are considered to be separate from the inner

layers, with no diffusion or transport between the margin and the inner cells. In addition,

uniform production of auxin is assumed in the margin cells.

 129

Figure 8.6: Simulation of vein formation on a leaf with emergent auxin sources. The
gradient-based model of PIN1 orientation used in the phyllotaxis model operates in the
margin cells. This causes peaks of auxin to appear in the margin. After the auxin builds
up to a threshold concentration, it leaks into internal cells and initiates vein formation.
As the leaf grows, more sources appear causing more veins to be initiated. Dark cells at
the bottom are sinks for auxin and represent a connection to the existing vasculature of
the plant.

The gradient-based transport model operating in the margin cells will cause an initial

peak of auxin to appear at the tip. When the auxin at this location reaches a threshold

concentration, it is allowed to leak into the inner cells. This initiates the formation of

the midvein. As the leaf grows, and more space opens up in the margin, more sources

appear. This causes the initiation of more veins that connect to existing vasculature

(Figure 8.6).

8.4 A combined model of auxin transport

The simulation in the previous section used two different algorithms for orienting PIN1

proteins depending on which cells were involved. Here the case is considered where a
130

combination of orientation methods are used in the same cell. Allocation of PIN1 to a

cell interface is based on both the flux across the interface as well as the concentration

of the neighbor cell on the other side of the interface.

To simplify the exploration of this question, a static rectangular grid of cells is con-

sidered without growth. All cells produce a background amount of auxin with increased

production occurring at discrete sources placed as the simulation proceeds. Equation

(8.1) is used to determine the change in auxin concentration of a cell over time, although

the equation becomes a bit simpler since the wall lengths and areas of the cells are con-

sidered to be 1. The equation to model the allocation of PIN1 to cell interfaces allocates

a portion of PIN1 based on flux, as well as a portion based on the neighbor cell’s auxin

concentration. Each cell is assumed to have a fixed total PIN1 amount which is allocated

to cell interfaces as follows:

dPi→k
= Piavail α max(0, φi→k )2 + δ[IAA]k − γPi→k


(8.4)
dt

where φi→k is the flux from cell i to k as defined by Equation (8.2), α controls PIN1

allocation to the cell interfaces due to flux, δ controls PIN1 allocation to the cell interface

based on the concentration of auxin in neighbor cells, γ represents decay of PIN1 from

the cell interface back to the cytosol, and Piavail is the PIN1 in the cytosol available for

allocation. The PIN1 available for allocation can be found by:

X
Piavail = Pitotal − Pi→k (8.5)
k∈Ni

The total PIN1 in the cell Pitotal is the same for all cells and fixed throughout the

simulation. PIN1 is allocated to cell interfaces by a quadratic function of the flux to

neighbor cells, and a linear function of the concentration of neighbor cells. If a higher

power of concentration is used, or too high a value for δ, then the model has a strong
 131

tendency to make cycles of cells. These cycles are different from the loops observed in

Figure 8.5 as there is no polarity switch within the loop. This causes auxin to be pumped

around in circles, which is unlikely to be of any benefit in a real plant, although Sachs

[1991] does report this as an occasional outcome after wounding.

Figure 8.7 shows a simulation performed with the PIN1 allocation model of Equation

(8.4). The simulation is able to create canals initiated by the auxin sources, as well as

canals initiated by the sink. If no discrete sources are added to the simulation, the sink

driven canals will grow up the center and sides of the grid to form 3 canals that drain

the entire tissue. One of the sink driven canals in the simulation depicted in Figure 8.7

grows up and connects with a source driven canal, creating a loop. Note that the auxin

concentration in both the source and sink driven canals is higher than in surrounding

tissue, however this property is much more pronounced for source driven canals.

The higher concentration of auxin in canals is consistent with the results of Feugier

et al. [2005], who reports this phenomenon in simulations where PIN1 is allocated to the

individual interfaces from a pool of PIN1 in the cytosol. In contrast, the original models

of Mitchison [1980, 1981] produce PIN1 at each interface independently from the other

interfaces, and the concentration of auxin in the canals is always equal to or lower than

in other cells. This is not supported by DR5::GUS data which suggests that the auxin

concentration is higher in the pre-procambium than in surrounding tissue [Scarpella et al.,

2006].

If the component of PIN1 orientation based on gradient is removed from the model

by setting δ = 0 in Equation (8.4), then the resulting flux-based allocation model does

create some canals with higher concentration than surrounding cells, but this happens

only with canals that have several branches feeding into them. In the canals with higher

concentration the difference is also much smaller than in the model with both polarization

factors.
132

Another property of the dual transport model is that the cells adjacent to the canal

have noticeable PIN1 polarization towards the center of the canal, a feature corresponding

to what [Scarpella et al., 2006] refers to as “centrobasal” polarity. Although a very slight

amount of polarization towards canals can be seen in cells adjacent to the canals in the

model with PIN1 polarization based only on flux, again it is much less pronounced than

in the model with two polarization factors.

8.5 Summary of results of leaf venation models

The models presented in this chapter demonstrate that the canalization hypothesis is

indeed capable of explaining vein formation on the irregular geometry of a growing cellular

leaf structure. However, the irregular geometry does raise some questions about the

dynamics of canalization that are not seen in models implemented on regular grids. In

the cellular models, an adjustment was required to compensate for the tendency to favor

short walls. How does the plant do this? Would the inclusion of extra-cellular space

mitigate this to some degree? Is it possible that this favoring of short walls also happens

in the plant? If so, this would suggest that the geometry of individual cells plays a

significant role in the selection of preferred pathways of auxin flow.

Simulations of canalization have been limited in the size of the patterns they can cre-

ate, typically single strands or small branching structures and networks. The simulations

presented here are no exception. If the simulation depicted in Figure 8.6 is continued,

with additional sources appearing as the leaf grows, then the amount of auxin coming

from the sources will eventually overtake the capacity of the midvein. When this hap-

pens, the model will forms cycles, with clumps of cells directing auxin around in circles.

In order to allow the simulations to continue and create more complex patterns, differen-

tiation of vein cells is likely required. The cells in older veins could then have increased
 133

auxin flow, allowing the veins to keep up with the additional sources. Unfortunately it

is not straightforward to maintain differentiated files of cells on a surface with global

descriptive growth. Cells that differentiate at an early stage in the model will consume

too much space at a later stage if their growth is not restricted in at least one direc-

tion. Restricting growth in this fashion is not straightforward if the growth is not being

specified locally.

The simulation depicted in Figure 8.6 uses gradient-based PIN1 polarization in the

margin to create auxin convergence points, and canalization in the interior cells to create

cell files of increased PIN1 expression. This is consistent with the observation of Scarpella

et al. [2006] for the early sections of veins that are initiated by auxin convergence points.

Since the same PIN1 proteins are used for both processes operating in very close proxim-

ity, it suggests that the two different PIN1 orientation strategies might be aspects of the

same mechanism. The last model presented on a rectangular grid of cells demonstrates

that when the two orientation strategies are combined, the model is able to maintain a

higher concentration of auxin in the veins than in surrounding tissue. The model also

has significant lateral PIN1 polarization towards the center of the emerging strand. This

fits better with experimental data than flux-only models of canalization.

134

(a) (b)

(c) (d)

Figure 8.7: Dual transport model on a rectangular grid of cells. Blue color represents
concentration of auxin and red represents PIN1 proteins at the cell to cell interfaces.
Lines from the center of cells represent the direction of auxin flux. Source cells are
outlined in green, and the sink cell is outlined in black. (a) The simulation starts with
an initial source at the top center and a sink at the bottom center of the grid. Since this
source cell has a higher auxin concentration than other cells, there is PIN1 polarization
towards it in its neighbors. (b) After a short time, flux-based PIN1 orientation causes a
canal to form propagating from the source downward as well as from the sink upward.
Sink driven lateral canals also appear. Note that the concentration of auxin in the canals
is higher than in the surrounding tissue. (c) The addition of more sources induces more
canals. (d) More sources are added and the sink driven canal on the left connects to a
source driven canal to form a loop.
 Chapter 9

Numerical Methods in VV

9.1 Types of problems arising in developmental biology

The simulations described in Chapters 5 and 6 are typical examples of problems arising in

developmental biology. They can be characterized by a state vector of all the quantities

in the system that are changing, combined with rules (differential equations) for how

this state vector changes over time. The quantities in the state vector are typically of

two types, those that represent chemical quantities, such as morphogens or proteins, and

those that represent physical quantities, such as forces, position, and velocity. Regardless

of type of quantities, the problem is the same. Given a state vector at some initial time

t, and a system of differential equations that describes how the state vector changes over

time, find the state of the system at time t + ∆t. This is called the initial value problem.

In many of the simulations described in Chapters 5 and 6, it is not only the state of

the system that changes over time, but also the structure. As cells divide, new variables

are added to the state vector, and as cells are removed, variables are deleted. Such

systems are called dynamic systems with dynamic structure [Giavitto and Michel, 2001].

As mentioned in Chapter 3, VV provides a convenient programming environment to

model systems with dynamic structure in a local index-free manner.

Within a single time step of a model, the structure of the system can be considered

as constant, so that standard techniques used to solve systems of differential equations

on static structures can be used. Press et al. [1992] presents a collection of routines

programmed in C and C++ to solve such problems. It would be convenient to use these

routines as is, since they are readily available and well-tested. Unfortunately, they al-

135
136

most invariable depend on data structures that are not index-free (arrays) and are not

compatible with the VV data structure. This means that data must be moved back

and forth between the different structures at the beginning and the end of every time

step. Maintaining a copy of the system’s state in an auxiliary array does not solve this

problem since the changing structure of the system requires that the array be constantly

recreated. This results in considerable inefficiency which can only be overcome by an

implementation of solvers directly on the VV data structure itself.

9.2 The VV programming environment

The examples in this chapter are programmed using the VV language extensions to C++

[Smith et al., 2004; Smith, 2006]. A brief introduction to VV is provided as well as Table

9.1 that lists all of the VV language constructs used in the example programs that follow.

VV implements a graph rotation system, which is a directed graph structure that

stores the neighborhood of each vertex as an ordered circular list. In geometric terms,

this ordering allows the representation of orientable 2-manifolds (two-dimensional sur-

faces) embedded in three-dimensional space. Loosely speaking, the fact that the surface

is orientable means that it has two sides, which can be distinguished at vertices by ex-

amining the ordering of their neighbors.

The graph data structure of VV is global to the program, including the connectivity

between vertices. Sets of vertices are called meshes. In the examples that follow there is

only one mesh S that contains all of the vertices in the graph. The VV language extensions

provide statements to iterate over the vertices in a mesh, as well as the neighborhoods

of vertices. Data elements, such as the system’s state vector and other model variables,

can be stored in both the vertices and the edges of the VV graph.
 137

VV Expression Meaning
vertex v Declare vertex v.
mesh S Declare mesh S (set of vertices).
S.x Data element x stored in mesh S.
v$x Data element x stored in vertex v.
v^n The edge from v to n
(v^n).x The data element x stored in the edge from v to n.
forall v in S {...} Iterate v over all vertices in mesh S.
forall n in v {...} Iterate n over all of the neighbors of vertex v.
nextto n in v Expression for the vertex after n in the neighborhood of vertex v.
prevto n in v Expression for the vertex before n in the neighborhood of vertex v.
splice a after n in v Add vertex a to the neighborhood of vertex v after vertex n.
splice a before n in v Add vertex a to the neighborhood of vertex v before vertex n.
make {a,b,c} nb_of v Replace the neighborhood of v with the vertices a, b, and c.
valence v Expression for the number of vertices in the neighborhood of vertex v.

Table 9.1: List of common VV language constructs.

9.3 Forward Euler explicit method

The forward Euler method is the simplest method for numerical integration of systems

of differential equations and leads to a natural translation of differential equations into

computer program statements. Consider the following ordinary differential equation in

one variable:
dc
= f (c) (9.1)
dt

Given a concentration ct at time t, we would like to determine the concentration at time

t+∆t. If ∆t is sufficiently small, then this value can be approximated by using an explicit

formula based on the value of c at time t:

ct+∆t = ct + f (ct )∆t (9.2)

This is known as the forward Euler explicit method and can be written programmatically

(in C or C++ ) as:

c = c + f(c) * dt;
138

where dt = ∆t.

The main drawback to this method is that it requires the simulation to proceed in

relatively small time steps, as the method is unstable for large steps. In practice this is

often not a problem, since in developmental modeling it is frequently desirable to visualize

the results of the simulation at fairly small steps. For example, in the movies presented

in the supplemental materials of Smith et al. [2006a] there are only 8 integration steps

per frame.

The simplicity of the forward Euler method aids in the process of model validation.

It is often difficult to determine if a complex program is really doing what the author

intended. In the forward Euler scheme, differential equations are translated into program

statements in a very straightforward manner. By using a time step small enough to avoid

stability problems, forward Euler can be used to verify the results of more sophisticated

integration methods.

9.4 Backward Euler and the Crank-Nicholson implicit methods

For some problems, the step size required for the forward Euler method is so small that

it becomes impractical, and it is necessary to consider alternate methods. In Equation

(9.2) (forward Euler) the value of c at time t is used to estimate f (c) over the entire time

step. In the backward Euler method, value of c at time the end of the time step (t + ∆t)

is used as an estimate instead:

ct+∆t = ct + f (ct+∆t )∆t (9.3)

This is called an implicit method because the value for cc+∆t is not explicitly given since

it appears on both the left and right hand sides of the equation. In order to obtain a

solution for Equation (9.3), the equation must be solved for ct+∆t . This method is called
 139

the backward Euler method, and is stable for arbitrarily large values of ∆t. Unfortunately

many problems involve systems of non-linear equations, in many variables, making them

difficult to solve.

The explicit forward Euler method is not stable for large ∆t, however the implicit

backward Euler method is not so accurate in t. A compromise is to use the average of

the values of c calculated at the beginning and the end of the time step. This is called

the Crank-Nicholson method, and is more accurate while still retaining the stability of

backward Euler with respect to large time steps. The Crank-Nicholson method is given

by:
ct + ct+∆t
ct+∆t = ct + f ( )∆t (9.4)
2
Note that this is still an implicit method since again ct+∆t appears on both sides of the

equation.

Although implicit methods are almost always more computationally expensive than

explicit ones, in some cases the step size can be increased by one or more orders of

magnitude, making the additional computational costs worthwhile.

9.5 The steady-state solution of a morphogenic gradient

A simple case which is convenient to illustrate the use of implicit methods in VV consists

of a line of cells with a single morphogen. Suppose we are interested in finding the steady

state of a Wolpert-style morphogenic gradient [Wolpert, 1969] which might be used to

provide positional information to some filamentous organism consisting of a line of cells.

It is assumed that one cell produces the morphogen, that this morphogen diffuses from

cell to cell, and that the morphogen will decay to some extent in every cell. The equations

to describe the change in morphogen concentrations over time in a line of n cells can be
140

written as:

dc1
= ρ + D(c2 − c1 ) − µc1 (9.5)
dt
dci
= D(ci−1 − ci ) + D(ci+1 − ci ) − µci 1<i<n (9.6)
dt
dcn
= D(cn−1 − cn ) − µcn (9.7)
dt

where ci represents the concentration of morphogen in cell i, ρ controls production of the

morphogen in cell 1, µ is the decay coefficient, and D controls cell to cell diffusion.

9.5.1 Solving for the steady-state

The steady-state solution to equations 9.5−9.7 will occur when the concentrations are
dci
not changing, that is when dt
= 0 for 1 ≤ i ≤ n. This gives the following linear system

of equations:

(−µ − D)c1 + Dc2 = −ρ (9.8)

Dci−1 + (−µ − 2D)ci + Dci+1 = 0 1<i<n (9.9)

Dcn−1 + (−µ − D)cn = 0 (9.10)

This is equivalent to setting up the equations for the backward Euler implicit method

and taking the limit as ∆t goes to infinity. Notice that the solution does not depend on

the initial concentrations in any of the cells. The system can be expressed in matrix form

(Ac = b) as:
 141

    
−µ − D D 0 0 0 0 ... c 0 −ρ
  1   
    

 D −µ − 2D D 0 0 0 ... 0   c2

 
  0 

    
0 D −µ − 2D D 0 0 ... 0   c3 0
    
   
  = 
 .. .. ..  .
  ..
  .. 

 . . . 
 
  . 

    

 0 ... 0 0 0 D −µ − 2D D  c
  n−1
 
  0 

    
0 ... 0 0 0 0 D −µ − D cn 0
(9.11)

This system of equations is linear, and if it is not too large, it can be solved by Gaussian

elimination.

In VV a line of cells can be modeled by defining each of the cells as a vertex, and then

connecting them to their left and right neighbors. Since the VV data structure is a graph,

it also provides a graph representation for matrices and can be used to store the matrix

shown in Equation (9.11). The diagonal entries can be stored in vertex variables v$a

for a vertex v, and the remaining non-zero entries in each row can be stored in variables

in the edges connecting the diagonal entry on that row to its neighbors, for example

(v^n).a for a neighbor vertex n of v. The vector on the right hand side of equation

9.11 (b in Ac = b) can also be stored in the vertices as another data element v$b. The

initialization of the system shown in Equation (9.11) is done with the following VV code

segment:

1 forall v in S { // Initialize matrix A and vector b (Ax=b)

2 v$b = 0;
3 if(v == startv) // Production in first cell only
4 v$b = -rho;
5 v$a = -mu - D * valence v; // Decay & diffusion to neighbors
6 forall n in v {
7 (v^n).a = D; // Diffusion from neighbors
8 }
9 }
 142

Line 1 creates an iterator that loops through all of the vertices in S, that is, all of the

vertices in the line of cells. Since there are no indices, the starting vertex is recorded in

the vertex variable startv so that the test on line 3 can be used to ensure that this is

the only vertex producing the morphogen. The variables rho, mu, and D are used to store

the parameters ρ, µ, and D for production, decay, and diffusion of the morphogen as

specified in Equations (9.5−9.7). The loop in lines 6−8 is used to fill in the off-diagonal

entries of the matrix that correspond to diffusion from neighbor cells.

9.5.2 Solving a linear system in VV with Gaussian elimination

Once the matrix of equation 9.11 is initialized, the system can be solved. The following

VV code segment will perform this task using Gaussian elimination:

1 forall v in S { // For all cells (diagonal)

2 double m = v$a;
3 forall n in v {
4 (v^n).a /= m; // Divide row by diagonal entry
5 }
6 v$b /= m; // Divide b vector as well
7 v$a = 1.0;
8 // Now eliminate column
9 forall r in v { // For all rows in column
10 m = -(r^v).a; // Find multiple of row to subtract
11 (r^v).a = 0.0;
12 r$a += m * (v^r).a; // Subtract from r
13 r$b += m * v$b; // Do b as well
14 vertex c = nextto r in v;
15 while(c != r) { // Loop though rest of row r
16 if((v^c).a != 0.0) {
17 if(!(is c in r)) { // Add new edges if required
18 splice c after any in r in r;
19 splice r after any in c in c; // Connect both ways
20 (c^r).a = (r^c).a = 0.0;
21 }
22 (r^c).a += m * (v^c).a; // Subtract entry
23 }
24 c = nextto c in v; // Next column in row
25 }
26 }
27 }
 143

Figure 9.1: A gradient of morphogen in a line of cells. The morphogen is produced in the
leftmost cell, is subject to decay in all cells, and diffuses from cell to cell. Steady state
solution is shown with the height and color of the bars representing concentration of the
morphogen within a cell.

The main loop starting on line 1 steps through the vertices which correspond to the

entries on the diagonal of the matrix. The order in which the vertices are processed is

not important. Lines 2−7 are used to preform the elementary row operation of dividing

the row by a scalar, which is the diagonal entry at this vertex. This results in a 1 at

this position of the diagonal (note the lack of pivoting). Lines 9−26 then use this row to

eliminate all of the other entries this column. Note that the loop on line 9 contains the

implicit assumption that the graph of the matrix has symmetric connectivity, otherwise

it would be difficult to determine which rows contain non-zero entries in this column. In

practice this is not a problem, since it is usually the case that if vertex v is a neighbor

of vertex u, then u is also a neighbor of v, although data items stored in the edges may

differ. Lines 18 and 19 maintain this symmetry, and in some cases missing edges must

be added to the graph as elimination proceeds. The result of the steady state solution of

the line of cells with a diffusing morphogen described by equations 9.5−9.7 is shown in

Figure 9.1.
144

The program performs Gaussian elimination without a global indexing scheme, and

since the matrix under consideration is sparse, the computation required can be consider-

ably less than normally required for a matrix of this size. For lines and rectangular grids

of cells, the matrix will in general form a banded system if arranged in the correct order.

To improve performance, the main iterator on line 1 can be divided into 2 passes, with the

first pass used to eliminate entries below the diagonal. Processing in the reverse order on

the second pass, the entries above the diagonal are eliminated, resulting in a reasonably

efficient banded solver that is also able to solve a general system. This requires that the

main loop beginning on line 1 be performed in a specific order to minimize the number

of edges that must be inserted into the graph as the computation proceeds. The next

section presents the VV program for such a banded solver.

Note that the above program could be written more naturally by using the VV con-

struct “forall c in v” on line 15. Unfortunately the prototype version of VV described

in Smith [2006] does not support nested iterators, requiring the workaround of using an

ordinary C++ while loop.

9.6 A reaction-diffusion system on a grid of cells

The next problem considered is a reaction-diffusion system [Turing, 1952] on a grid

of cells. The equations follow those of Meinhardt and Gierer [1974] for an activator-

inhibitor system that is able to create peaks of morphogen on a rectangular array of

cells. There are two substances, the activator a, and the inhibitor h. The activator

enhances the production of both substances, while the inhibitor inhibits production of

the activator. The inhibitor also diffuses more quickly then the activator. Under these

conditions, relatively uniformly spaced peaks in activator concentration can result from

initially homogenous concentrations of the morphogens. A small amount of noise must

 145

introduced to break the symmetry of the unstable equilibrium that must be present in

the system for patterning to occur.

The simulation begins with a uniform distribution of both substances, with a very

small amount of noise introduced. Small local maxima in activator concentration due

to this noise become larger because of the self-enhancement of the activator. Since the

activator also enhances production of the inhibitor, more inhibitor is also produced at

these local maxima. This addition inhibitor will not suppress the activator peaks, since

it diffuses away more quickly than activator. It will, however, suppress the formation of

additional activator peaks nearby.

The following equations are used to model the change in concentration of the two

morphogens in this system:

da ρ a a2 X
= + σa − µa a + Da (ai − a) (9.12)
dt 1.0 + h i∈N
dh X
= ρh a2 + σh − µh h + Dh (hi − h) (9.13)
dt i∈N

where ρa and ρh control activator-dependent production of the morphogens, σa and σh

control default or background production, µa and µh control decay, and Da and Dh

control the cell to cell diffusion of the morphogens. The expression i ∈ N under the

summation sign indicates that the sum is taken over all of the neighbors of a cell.

Unlike the previous example in Section 9.5, an implicit method yields a non-linear set

of equations. This is complicated by the fact that there is more than one variable per cell

or vertex in the VV graph. The non-linear system can be solved by approximating it with

linear system using the Newton-Raphson method, and then iterating until it converges.

If the time step of the simulation is not too large, the method will converge fairly quickly,

often after only a few iterations.

In order to make it easy to re-use the VV program for various problems, the Crank-

Nicholson method, its linearization with Newton-Raphson, and the solution of the linear
 146

system of equation will be done for the general case.

9.6.1 Specifying the model equations

Suppose a function is coded in VV to find the derivatives of data values in each node of

the VV mesh. In the current case of Equations (9.12) and (9.13), this function would find

the change in concentration of the two morphogens, the activator a, and the inhibitor h.

Such a function can be coded in VV as follows:

1 void FindDerivatives(vertex v) { // Find change in concentrations at a vertex

2 double dA, dH;
3 double A = v$vars[0]; double H = v$vars[1];
4
5 dA = rho_A * A * A/(1.0 + H) + sigma_A; // Production
6 dH = rho_H * A * A + sigma_H;
7
8 dA -= mu_A * A; // Decay
9 dH -= mu_H * H;
10
11 vertex n = any in v; // Loop through all neighbors
12 vertex first = n;
13 do {
14 double nA = n$vars[0]; double nH = n$vars[1];
15
16 dA -= D_A * (A - nA); // Diffusion
17 dH -= D_H * (H - nH);
18
19 n = nextto n in v;
20 } while(n != first);
21
22 v$dvars[0] = dA; v$dvars[1] = dH;
23 }

The data variables of the model (the state vector) are stored in the array vars and

the derivatives are stored in dvars. Note that lines 2, 3, 14, and 22 are not required, but

are used to improve the readability of the program. It would be possible to perform the

operations directly on the vars and dvars variables instead. Unfortunately, since the VV

language parser executes before the C++ preprocessor, the programmer cannot simply

use #define statements to do this task. Nevertheless, the above program segment is a

very straightforward translation of Equations (9.12) and (9.13) into program statements,
 147

and can be used to perform forward Euler integration over a mesh S of vertices with the

following calling program:

1 forall v in S { // Find derivatives

2 FindDerivatives(v);
3 }
4
5 forall v in S { // Update concentrations
6 v$vars += DT * v$dvars;
7 }

The loop on lines 1−3 visits all the cells in the grid to calculate the rate of change of

the morphogens, and then the second loop updates the model variables with the forward

Euler method. This provides a convenient means to verify the methods that follow, as

an implicit method will be developed that calls the same FindDerivatives() function.

This results in an implicit implementation that is as straightforward to specify as the

forward Euler method.

9.6.2 The Crank-Nicholson method combined with the Newton-Raphson

method for non-linear systems

Suppose the following system of coupled first order ordinary differential equations, which

may be non-linear, describes the evolution of the state vector x = (x1 , x2 . . . xn ) of the

system:
dxi
= fi (x1 , x2 . . . xn ) (9.14)
dt

where the xi represent the model variables, with one or more at each vertex in the VV

graph. The Crank-Nicholson method takes the average of forward and backward Euler

methods and is given by:

fi (xt1 , xt2 . . . xtn ) + fi (xt+∆t

1 , xt+∆t
2 . . . xnt+∆t )
xt+∆t
i = xti + ∆t (9.15)
2

rearranging the equation becomes:

148

fi (xt+∆t
1 , xt+∆t
2 . . . xt+∆t
n ) fi (xt1 , xt2 . . . xtn )
∆t − xt+∆t
i + ∆t + xti = 0. (9.16)
2 2

If we let:
fi (x1 , x2 . . . xn ) fi (xt1 , xt2 . . . xtn )
Fi (x) = ∆t − xi + ∆t + xti (9.17)
2 | 2 {z }
Constant

be a set of scalar valued functions on the vector variable x = (xi , x2 , . . . xn ), then the

simultaneous solution to Fi (x) = 0 for all i, corresponds to x = xt+∆t . The function

Fi (x) can be approximated by the first two terms of the multidimensional form of the

Taylor expansion as follows:

n
t
X
t∂
Fi (x) = Fi (x + ∆x) = Fi (x ) + Fi (x)∆xj + O(∆x2 )
j=1
∂x j
n
∼
X ∂
= Fi (xt ) + F (x)i ∆xj (9.18)
j=1
∂xj

Ignoring all but the first two terms we get the second order error term O(∆x2 ), giving a

first order linear approximation to the non-linear system. Setting Fi (x) = 0, in matrix

notation we have:

JF ∆x = −F(xt ) (9.19)

where JF is the Jacobian of the Fi (x) functions. Note that when taking the partial

derivatives, the constant term of Equation (9.17) is removed, so that the expression

becomes:
∂ ∂ ∆t ∂
Fi (x) = fi (x1 , x2 . . . xn ) − xi (9.20)
∂xj ∂xj 2 ∂xj
or: 
∂
f (x , x2 . . . xn ) ∆t −1 i=j

∂ 
∂xj i 1 2
Fi (x) = (9.21)
∂xj ∂
f (x , x2 . . . xn ) ∆t i 6= j

∂xj i 1

2
 149

9.6.3 An implementation of Newtow-Raphson in VV

Returning to the activator-inhibitor system of Equations (9.12) and (9.13), the partial

derivatives in Equation (9.21) can be found by using finite differences and calling the

FindDerivatives function defined earlier. The VV program code to do this is:

1 void FindPartials(vertex v) { // Find Jacobian

2 FindDerivatives(v);
3 Vec ovars = v$vars; // Save original variables and derivatives
4 Vec odvars = v$dvars;
5
6 for(int i = 0; i < VARS; i++) { // For all data variables in this vertex
7 v$vars[i] += DX; // Add dx
8 FindDerivatives(v); // Find new derivative
9 Vec partials = (v$dvars - odvars)/DX * DT/2.0; // Take finite difference
10 partials[i] -= 1.0;
11 v$AT[i] = partials; // Fill in a row of partials
12 v$vars = ovars; // Reset data variables
13 }
14 v$A = v$AT.transpose(); // Partials are tranposed
15
16 forall n in v { // Find partials for neighbors
17 Vec ovars = n$vars;
18 for(int i = 0; i < VARS; i++) { // For all data variables in neighbor vertex
19 n$vars[i] += DX; // Add dx
20 FindDerivatives(v); // Find new derivative
21 Vec partials = (v$dvars - odvars)/DX * DT/2.0; // Take finite difference
22 (v^n).AT[i] = partials; // Fill in a row of partials
23 n$vars = ovars; // Reset data variables
24 }
25 (n^v).A = (v^n).AT.transpose(); // Partials are tranposed
26 }
27 v$dvars = odvars; // Reset derivative variables
28 }

The VARS variable holds the number of data elements in each vertex, in this case 2,

and DX holds the epsilon that is used to take the finite difference. The variables v$A

and v$AT are square matrix variables, with as many rows and columns as there are data

elements at each vertex. Depending on the method chosen to solve the linear system,

the v$AT variables may or may not need to be saved in the vertices and edges.

Now that the routine to fill in the matrix of partials is defined, the Newton-Raphson

iteration can programmed in VV as follows:

 150

1 forall v in S { // Find time 0 values

2 FindDerivatives(v);
3 v$t0 = v$vars + v$dvars * DT/2.0;
4 }
5 double newterr = 0;
6 do { // Main loop for Newton-Raphson iteration
7 forall v in S {
8 FindPartials(v); // Initialize matrix A with partials
9 v$b = v$vars - v$t0 - v$dvars * DT/2.0; // Initialize vector b in Ax=b
10 }

..
.

Solve linear system

..
.

11 newterr = 0;
12 forall v in S { // Find error
13 double sz = v$x.norm();
14 if(sz > newterr)
15 newterr = sz;
16 v$vars += v$x; // Update model variables
17 }
18 } while(newterr > S.NewtTol); // Exit when tolerence reached

The vector variable v$t0 is used to store values needed from the beginning of the

timestep, which is the constant part of F (x) from Equation (9.17). The routine requires

that some type of linear system solver will solve the system and put a result in the v$x

vector variable. This variable is used to both update the model variables v$vars as well

as to determine the size of the maximum correction. The Newton-Raphson iteration

stops when the maximum correction goes below some tolerance.

9.6.4 A banded linear system solver in VV

In Section 9.5 a Gaussian elimination solver was used. The following program implements

the 2 pass variation discussed earlier that will in general be more efficient for banded

systems:

1 for(int pass = 1; pass <= 2; pass++) { // Do in two passes

2 vertex v = 0; vertex endv = 0;
 151

3 if(pass == 1) // Eliminate lower triangle in first pass

4 v = endv = firstv;
5 else // Eliminate upper triangle in second pass
6 v = endv = lastv; // Process second pass in reverse order
7 do {
8 if(pass == 1) {
9 Mat M = v$A.inverse(); // Find inverse to multiply row by
10 v$A.identity(); // Multiply current entry and b vector
11 v$b = M * v$b;
12 forall n in v { // Multiply rest of the row
13 (v^n).A = M * (v^n).A;
14 }
15 }
16 forall r in v { // Now eliminate rest of column
17 if((pass == 1 && r$row > v$row) || (pass == 2 && r$row < v$row)) {
18 Mat M = -(r^v).A; // Find multiples of rows to subtract
19 (r^v).A = 0;
20 r$A += M * (v^r).A; // Subtract current entry
21 r$b += M * v$b;
22 vertex c = nextto r in v;
23 while(c != r) { // Subtract from rest of row
24 if((v^c).A != Zero) {
25 if(!(is c in r)) { // Create entry if required
26 splice c after any in r in r;
27 splice r after any in c in c;
28 (c^r).A = (r^c).A = 0;
29 (c^r).temp = (r^c).temp = true;
30 }
31 (r^c).A += M * (v^c).A; // Subtract entry
32 }
33 c = nextto c in v;
34 }
35 }
36 }
37 if(pass == 1)
38 v = nextto v in sortv; // Forward on first pass
39 else
40 v = prevto v in sortv; // Backward on second
41 } while(v != endv);
42 }
43 forall v in S {
44 v$x = v$b; // Update model variables
45 forall n in v {
46 if((v^n).temp) {
47 erase n from v; // Remove temp edges
48 }
49 }
50 }

The algorithm requires that the vertices are processed in a specific order in both

passes. Since VV has no built-in mechanism to order vertices within a mesh, a special
152

vertex has been created called sortv. This vertex includes all of the vertices in the grid

in its neighborhood placed in the correct order. In addition, the variables firstv and

lastv have been set to the first and last vertices in the grid.

For a grid of cells, the system will be banded, with the width of the band being twice

the width of the grid. In the first pass the matrix is processed from left to right, and

all of the entries below the diagonal will be eliminated. Continuing from the bottom

right, the second pass will eliminate the entries above the diagonal working in the reverse

order. Lines 25−29 insert any extra edges that are required to perform elementary row

operations. In the case of a banded system, these extra edges will be confined to the

interior of the band. The temp variable stored in the edge is used to mark these temporary

edges so that they can be deleted later if required. For a grid of cells where the structure

does not change, the deletion of these edges is optional.

The matrix entries at each vertex are not scalers, but are themselves square matrices.

The type Mat is predefined to represent this data type, which is in this case is a 2 by 2

matrix of floating point values. Even though the variables are not scalars, elementary

row operations can be performed as block operations, and the program looks only slightly

different than in the scalar case.

9.6.5 A biconjugate-gradient linear system solver in VV

It is quite often the case that the matrix that needs to be solved is quite large. For a

50 by 50 grid of cells, the matrix is 2500 by 2500, and takes a considerable amount of

computation to solve. In addition, for non-banded systems, the processing time of the

above solver goes up considerably, making alternative methods more attractive.

One such alternative is the conjugate gradient method for sparse systems. See Press

et al. [1992] for a complete description of the algorithm. The general idea is that the
 153

solution to Ax = b is also the unique minimizer of the function:

1
f (x) = xT Ax − bx (9.22)
2

This function has a minimum when its gradient is zero ∇f = 0. The algorithm finds

a sequence of mutually orthogonal correction vectors pi such that the solution can be

expressed as the linear combination:

x = α1 p1 + α2 p2 + . . . + αn pn (9.23)

At each iteration one of the α scalars and one of the pi vectors is determined. Since

n mutually orthogonal vectors forms a basis for an n dimensional vector space, the

algorithm will theoretically complete in at most n iterations, assuming exact arithmetic.

In practice the algorithm will normally produce a sufficient approximation much quicker

than this, often by an order of magnitude or more.

The conjugate gradient method requires that the matrix in question be both sym-

metric and positive definite. Unfortunately it is rarely the case that a model yields a

symmetric system. One solution is to solve the system AT Ax = AT b instead, since the

matrix AT A is symmetric. However this matrix is often significantly more ill-conditioned

then A. Another solution is to use a related method that works in the non-symmetric

case called the biconjugate gradient method. The following is an algorithm for the bi-

conjugate method:

1 r ← Ax − b
2 r̄ ← r
3 rr ← r̄T r
4 p ← −r
5 p̄ ← −r̄
6 while rr > tolerance {
7 α ← rr/(p̄T Ap)
8 x ← x + αp
 154

9 r ← r + αAp
10 r̄ ← r̄ + αAT p̄
11 rrold ← rr
12 rr ← r̄T r
13 β ← rr/rrold
14 p ← −r + βp
15 p̄ ← −r̄ + β p̄
16 }

In the biconjugate gradient method, two sets of mutually orthogonal vectors pi and

p̄i are constructed instead of one.

Conjugate gradient methods are convenient for use with a VV, since they can be

coded without requiring the programmer to specify the matrix A directly [Press et al.,

1992]. The programmer is only required to calculate the product of the matrix and a

column vector. In case of the biconjugate gradient method, the programmer must also

be able to calculate the product of the transpose of the matrix and a column vector.

In the reaction-diffusion example with two morphogens, it is convenient to store both

the matrix A and its transpose AT at each node in the mesh. Although most problems,

including this one, do not create a symmetric matrix, the matrix does have symmet-

ric connectivity. The following code which implements the biconjugate gradient in VV

exploits this fact:

1 forall v in S {
2 v$x.zero(); // Clear x vector
3 v$r1 = v$r2 = -v$b; // Set intial residuals
4 v$p1 = -v$r1; v$p2 = -v$r2; // Set intial correction vectors
5 rr += v$r2 * v$r1;
6 }
7 do { // Main loop for biconjugate-gradient
8 forall v in S {
9 v$Ap1 = v$A * v$p1; // Multiply A x p1
10 v$ATp2 = v$AT * v$p2; // Multiply AT x p2
11 forall n in v {
12 v$Ap1 += (v^n).A * n$p1;
13 v$ATp2 += (v^n).AT * n$p2;
14 }
15 }
16 double p2Ap1 = 0;
 155

17 forall v in S {
18 p2Ap1 += v$p2 * v$Ap1; // Find p2 x A x p1
19 }
20 double alpha = rr / p2Ap1; // Find alpha
21 oldrr = rr; rr = 0;
22 forall v in S {
23 v$x += alpha * v$p1; // Improve solution vector
24 v$r1 += alpha * v$Ap1;
25 v$r2 += alpha * v$ATp2; // Next residuals
26 rr += v$r2 * v$r1;
27 }
28 double beta = rr/oldrr;
29 forall v in S {
30 v$p1 = beta * v$p1 - v$r1; // Find next p1 and p2
31 v$p2 = beta * v$p2 - v$r2;
32 }
33 } while(fabs(rr)/(double)S.N > S.ConjGradTol); // Exit when tolerence reached

Figure 9.2 show the results of the simulation of the reaction-diffusion system defined

in Equations (9.12) and (9.13) performed on a grid of cells. For this simulation, the

biconjugate-gradient method typically requires 20-60 iterations to converge, which is

much less then the theoretical maximum of 2500 for a 50 by 50 grid of cells.

9.7 When are implicit methods worth it?

All of the simulations in the Chapters 5 and 6 use the forward Euler explicit method.

Although the implicit solution of the activator-inhibitor system in the previous section

allows a time step over 10 times larger, it is not enough to compensate for the increased

processing cost. The implicit solution takes twice as long to run as the explicit one. This

is often the case for equations that are not too stiff. If the coefficient for the diffusion of

the inhibitor is increased 25 fold, then only one or two activator peaks will appear on a 50

by 50 grid. In this case the opposite occurs, and the explicit solution takes approximately

twice as long to execute as the implicit one.

When choosing an integration method, it is necessary to consider the time scale at

which the various model processes are occurring. For example in the models in Chapter
156

Figure 9.2: Steady state of a reaction-diffusion simulation on a 50 by 50 grid of cells

using the activator-inhibitor system defined by Equations (9.12) and (9.13). Bright spots
indicate peaks in activator concentration. A small amount of noise is added initially to
break symmetry. Both simulations use the same parameters except for the time step
dt. The simulation on the left was performed using the forward Euler explicit method
with a time step of dt = .05. The simulation on the right was performed using the
Crank-Nicholson implicit method with a time step of dt = .5. The non-linear set of
equations arising from the implicit method is approximated by using the Newton-Raphson
method. The resulting linear system is then solved with the biconjugate gradient method
implemented directly on the VV data structure. The grid is too large to use the banded
solver described in Section 9.6.4, although a similar result can be produced on smaller
grids.

5, the growth steps are quite small. Even if implicit methods were faster for long time

steps, this might be of no use if the time step required to realize an overall performance

gain exceeds the length of the small time steps required for growth.

Another factor that can affect the choice of integration methods is the nature of

constraints in the model. If the equations describing the model are not differentiable

in places, then the Newton-Raphson approximation of the non-linear system may not

behave well. This is the case in several places in the phyllotaxis model of Chapter 5.

One example is the concentration limits imposed on auxin and the PIN1 protein. These

limits create discontinuities in the derivatives of the differential equations that describe
 157

the change in concentration of these quantities.

Ideally, the modeling environment should make the process of deciding which integra-

tion method to use easier by providing several “canned” methods for the programmer to

use. If a selection of techniques is available within the same model specification frame-

work, then the task of trying different integration methods is greatly simplified.
 Chapter 10

Conclusion

10.1 Discussion of research contributions

The research presented in this thesis is devoted to computer simulation models of mor-

phogenesis, and in particular phyllotaxis. The main objective of this work was to gain

understanding of the mechanisms of plant development. In some cases the models pre-

sented were able to confirm existing hypotheses about morphogenesis, and in other cases

questions raised by the models led to further hypotheses. Not surprisingly, many ques-

tions remain unanswered, and will be the subject of future work.

The main research contributions presented in this thesis are discussed below.

10.1.1 Shoot apex model

The shoot apex model described in Chapter 3 uses a combination of previous work from

various sources to create a flexible, computationally inexpensive model of the shoot apex

suitable as a platform for phyllotaxis simulations. The fact that the same apex model is

used the inhibition field models of Chapter 4 operating in continuous space, as well as

the transport-based phyllotaxis models of Chapter 5 operating on a cellular structure,

demonstrates the utility of this model. This model doesn’t offer a new algorithm or

patterning process, but combines existing elements into a useful tool for phyllotaxis

research.

The transport-based model of phyllotaxis on the cellular apex structure presented in

Chapter 5 extends this shoot apex model by creating a compound surface, in order to

simulate bulging primordia, within the framework of descriptive growth.

158
 159

10.1.2 Model of a tissue with dividing cells

The cellular tissue used in the simulation models in Chapters 5, 6, and 8 is again largely

based on a combination of ideas from previous work from various sources. The con-

tributions in this area include modifications to the existing algorithm to produce more

realistic cell shapes, as well as the generalization of the model to three dimensions, which

enables the same tissue model to be used with a variety of biological structures and

growth models. Within this thesis, the same tissue implementation is used for the shoot

apex model with descriptive growth and bulging primordia, the shoot apex model with

physically-based growth, as well as for the growing leaf structure with growth simulated

using Bezier surface interpolation.

The utility and versatility of the growing tissue model with dividing cells presented

here has resulted in its use in other models of morphogenesis [Barbier de Reuille et al.,

2007] and has inspired an effort currently underway to incorporate it into the libraries

provided with the VV simulation environment (Barbier de Reuille, unpublished).

10.1.3 Inhibition field model of phyllotaxis

The inhibition field model of phyllotaxis presented in Chapter 4 makes a contribution

towards the simulation modeling of phyllotaxis with the discovery of an improved in-

hibition function for field models of phyllotaxis. Previous models have always used a

decreasing function of distance, whereas the model presented in this thesis uses an in-

hibition function based on both distance and primordium age. It was found that this

function is able to produce a greater variety of phyllotaxis patterns more robustly than

previous models. A new algorithm using two inhibition functions was also proposed to

explain decussate, whorled, and multi-jugate patterns.

Previous models of phyllotaxis based on inhibition fields have been presented from the

perspective of trying to simulate a specific mechanism of inhibition, such as the diffusion

160

of a chemical substance. The approach taken in the work presented in this thesis was

to look for a plausible inhibition function based on its ability to create the patterning

observed in nature, and then see what this function suggests about the mechanism. The

best inhibition function found in this investigation does not correspond to what would be

expected for a mechanism based on diffusion. In addition, the dependence of inhibition

on primordium age, and not just primordium location, suggests an important role for

primordium differentiation in phyllotaxis.

It was also discovered that the inhibition field model presented in Chapter 4 was able

to produce a large variety of patterns using the same model parameters. 25 different

phyllotaxis patterns could be obtained by varying only the initial conditions in the two-

inhibitor model (Figure 4.12). The extent of this phenomenon was previously unreported,

and has implications for experimental work in developmental biology.

10.1.4 A transport-based patterning mechanism in plants

Another contribution in this thesis is the discovery made through simulation modeling,

that transport-based patterning can create relatively uniformly spaced peaks of a mor-

phogen in an otherwise homogenous tissue. Such peaks can trigger differentiation and

result in pattern formation, and thus suggest an answer to one of the questions raised in

the introduction to this thesis: How is it that form is created from identical, undifferen-

tiated cells with the same genetic code?

It should be mentioned at this point that different implementations of this same

mechanism were published simultaneously by Jönsson et al. [2006] and Smith et al.

[2006a].

The mechanism, based on self-enhanced transport, provides an alternative to the long-

standing paradigm for patterning, reaction-diffusion, which dates back to Turing [1952].

Although the canalization model of Sachs [1981] for vein formation is also transport-
 161

based, that mechanism is limited to the creation of strands and branching patterns, and

is not able to create evenly spaced peaks of morphogen. The model presented in this thesis

shows that transport-based patterning is able to create a wider range of patterns than

previously thought, simply by changing the rules that control the direction of transport

of the morphogen.

10.1.5 A mechanistic model of phyllotaxis

The transport-based patterning mechanism discussed in the previous section was inspired

by experimental work aimed at understanding the patterning mechanism of phyllotaxis.

This work culminated in a conceptual model of phyllotaxis [Reinhardt et al., 2003b]

that suggested that the transport of auxin, rather than local self-enhanced production,

causes the peaks in auxin concentration that lead to organ initiation. Efforts to build a

simulation model of this theory raised questions, such as what controlled the direction

of auxin transport, that led to the discovery of the patterning mechanism.

The model of phyllotaxis presented in Chapter 5 is perhaps the most plausible model

of phyllotaxis to date, in the sense of being consistent with experimental results. It is

able to produce a variety of phyllotaxis patterns, and is more stable than the model

proposed concurrently by Jönsson et al. [2006], which can only produce phyllotaxis-like

patterns for a few primordia at a time. Jönsson et al. [2006] attribute this shortcoming

to their implementation of a growing apex surface, which disrupts the patterning process

with unrealistic movements of cells during division. This underscores the importance of

having a sound growing apex model and dividing cellular structure as a basis for creating

molecular-level phyllotaxis simulations. Another factor contributing to the difference of

the outcomes of the two models is likely related to the equations for auxin transport an

PIN1 auxin transporter localization. Jönsson et al. [2006] used linear relations motivated

by the current understanding of molecular-level processes, whereas in the work presented

162

in this thesis, plausible equations were selected based on their pattern formation char-

acteristics. Obviously, equations that have a strong experimental basis are preferable,

however the molecular basis for PIN1 orientation, and aspects of auxin transport, is

largely unknown. Instead, equations were chosen that are plausible, and produce the

desired phenomena. As is the case with inhibition function found in Chapter 4, it is now

possible to ask what the equations for auxin transport used in the model suggest about

the mechanism.

10.1.6 A physically-based apex model as a platform for the study of mor-

phogenesis

Despite the popularity of physically-based models in computer graphics and computer

animation, there are surprisingly few such models of the shoot apex. The model presented

in this thesis was inspired by preliminary work by Stoma et al. [2007] in which a cellular

model of shoot apex with a single bulging primordium is presented. The physically-

based apex model presented in Chapter 6 improves upon this model by including both

the inhibition field model of phyllotaxis from Chapter 4 and the molecular-level model of

phyllotaxis from Chapter 5. Although an inhibition field model of phyllotaxis on a shoot

apex with physically-based growth has been presented previously by Smith et al. [2004],

no molecular-level model of phyllotaxis on a physically-based apex with emergent shape

has been previously described.

10.1.7 A light capture study of phyllotaxis

The light capture study of phyllotaxis presented in Chapter 7 was motivated by the fol-

lowing question: Is the frequency of the golden angle of 137.5◦ in nature due to a selective

advantage this divergence angle provides from a light capture perspective? Previous sim-

ulation studies have addressed this question with conflicting conclusions [Niklas, 1998;
 163

King et al., 2004]. The simulation study presented in this thesis shows that spiral phyl-

lotaxis does provide enhanced light capture, however the angles commonly observed in

spiral phyllotaxis are all equally fit. This is interesting when considered in combination

with results from the inhibition field model of phyllotaxis from Chapter 4. In the inhibi-

tion field model, initial conditions determine the choice among the commonly observed

phyllotaxis angles, with the frequency of the various patterns observed in nature corre-

lating with relative stability of the various patterns in the model. This suggests that it is

the mechanism of phyllotaxis that is being selected, rather than a particular divergence

angle. This result comes from the creation of a clearer light capture fitness landscape at a

higher resolution than those described previously, which suggested a different perspective

on the question.

10.1.8 Exploration of mechanistic models of leaf venation

Previous models of leaf venation based on the canalization hypothesis [Sachs, 1981] have

always operated on regular grids of cells. In Chapter 8 it was shown that this mechanism is

capable of vein formation on a growing cellular structure. These models also demonstrate

some interesting dynamics of canalization that do not appear on regular structures, such

as the tendency to favor short cell interfaces, or neighbors with larger area. This raises

some questions about potential mechanisms for canalization and about the effect of cell

geometry on the initiation of procambium.

The molecular basis for the auxin transport-based patterning mechanism of canaliza-

tion is hypothesized to involve the same components as the transport-based patterning

mechanism proposed for phyllotaxis. Chapter 8 presents some preliminary work in uni-

fying these two processes into a common model.

164

10.1.9 Implementation of numerical algorithms in VV

All of the simulations presented in this thesis were programmed in the VV simulation

environment. VV allows the specification of problems in a local, index-free manner.

Chapter 9 was included to demonstrate how numerical methods relevant to biological

modeling could be implemented in this context. All numerical methods implemented

were based on well established algorithms. The contribution of this work lies in the

demonstration that these algorithms can be specified in a local index-free manner without

the use of matrices. The fact that the description of these results was presented using

matrices underscores the present lack of a general mathematical formalism to even discuss

such problems in a local, index-free manner.

10.2 Future work

The discovery that transport-based patterning is able to create a greater variety of pat-

terns then previously thought suggests the need for a more in-depth exploration of this

area. Is this mechanism capable of producing periodic waves similar to what is possible

with reaction-diffusion? What types of patterns can be produced? What about com-

bining mechanisms? Rarely does a signaling molecule act directly on its target. This

suggests that hybrid systems that combine transport-based models and reaction-diffusion

might be useful to explore. Could such a mechanism produce a more robust phyllotaxis

model?

A related direction for future work is suggested by the similarity between the proposed

auxin transport-based patterning mechanism for phyllotaxis, and the transport-based

canalization mechanisms proposed for leaf venation. Are the two patterning processes,

which share the same molecular components, simply manifestations of a single mecha-

nism? The models presented in the second half of Chapter 8 only begin to explore this
 165

question.

It is possible that the answer requires extension of the cellular model to include

extracellular space. There is evidence suggesting that auxin import carriers play an

essential role in both organ positioning [Stieger et al., 2002; Reinhardt et al., 2003b] and

in auxin accumulation in procambium [Kramer, 2004], but including import carriers in

the models requires extending the model to incorporate extracellular space.

Although more difficult to construct, a three-dimensional model of the shoot apex

would provide an excellent platform to study the interaction between transport-based

phyllotaxis and the canalization mechanism. In two dimensions, it is difficult to study

both processes at once, which might prove essential if there is significant feedback between

them. Would the “shortest wall through the centroid” algorithm for cell division in

two dimensions extend nicely to a “smallest area through centroid” algorithm in three

dimensions? VV is capable of modeling two-dimensional surfaces, such as the surface

of a single three-dimensional cell. If VV were modified to handle multiple meshes with

different neighborhoods, could a three-dimensional cellular model be implemented as a

collection of meshes, one for each cell?

A three-dimensional model would also be useful to explore physically-based simula-

tions that can address questions relating to the emergence of shape, such as why leaves

are flat, or how growing apex tips can grow straight. Although it is likely that the

physically-based model described in Chapter 6 can be extended to make some inroads

in these directions, a full three-dimensional model is may be required to obtain complex

emergent shapes (cf. Stoma et al. [2007]).

Other problems are likely within the reach of two-dimensional models with physically-

based growth. A leaf with physically-based growth might provide a solution to the

problems discussed in Chapter 8, in which differentiation of vein cells within a descriptive

growth model became an issue. In a physically-based model, it is much simpler to control

166

growth locally, opening up the possibility of extending the canalization models beyond

single strands or simple branching patterns. In addition, investigating the local growth

requirements required to produce various leaf shapes, or the molecular processes involved

in compound leaves, is possible within the framework of two-dimensional models.

Another direction for future work relates to the software development side of biological

modeling. A recurrent theme in software development is the idea of creating reusable

pieces or components that can be simply used as is, freeing the software developer to

focus on the problem at hand. Biological modeling is no exception.

Consider the model of the shoot apex described in Chapter 3. It was used for both

the inhibition field model and the molecular level model of phyllotaxis. Another surface

with descriptive growth was defined in Chapter 8 to model growing leaves. These surfaces

have much in common, and are in fact programmed with the bulk of the work done by

two basic operations, one to specify the movement of points due to growth, and one to

allow the insertion of new points. These operations are:

• Given a point on the surface at time t, find the location of the point at time t + ∆t.

• Given the (x, y, z) position of a point in space, find the closest surface point.

Although it may prove useful to add additional operations, the experience gained in

crafting the models presented in this thesis suggests that other growth models might also

be accommodated. This could lead to the creation of a generic growth library for the

VV simulation environment, allowing the modeler to select from a variety of predefined

growth and surface models.

This same observation holds true for the model of the tissue with dividing cells, which

is used on the apex with descriptive growth, the apex with physically based growth, as

well as the growing leaf model. The implementation used in this thesis allows a selection

of division algorithms (Figure 3.8). Is it possible to extend this to other situations, such
 167

as the lineage-driven division wall orientation seen in stomata? Can a convenient yet

powerful application interface for potentially complex operations, such as determining

the location of a division wall, be created?

The numerical routines presented in Chapter 9 are also good example of routines

that are useful in many different models. Can they be incorporated into a library in a

convenient manner so that the modeler can simply select which numerical integration

method they wish to use?

The incorporation of such objects into the simulation environment would free the

modeler to focus on the research question, rather then having to program these compo-

nent which are often seen simply as “overhead” in the process of creating development

models. Of course the design of an application interface that is simple, yet flexible enough

to endure, remains a challenge.

10.3 Closing remarks

The period in which we live has been dubbed the information age. Nowhere does this

term seem more appropriate than in the world of molecular biology. The amount of

information gathered on the details of life in the past few decades is staggering, yet

this information rarely translates directly into understanding. The goal of computer

modeling of biological systems is to compress this information into theory that can be

stated precisely, programmatically, capturing the essential features of phenomena while

abstracting from the details [Prusinkiewicz, 1998; Chaitin, 2006]. It is only through the

use of such models can we hope to make sense of the enormous complexity of living

things. It is the author’s hope that the simulation models presented in this thesis have

made a small step on this path, from information to understanding.

 Bibliography

Adler, I. 1974. A model of contact pressure in phyllotaxis. Journal of Theoretical

Biology, 45(1):1–79.

Adler, I. 1977. An application of the contact pressure model of phyllotaxis to the close

packing of spheres around a cylinder in biological fine structure. Journal of Theoretical

Biology, 67(3):447–458.

Adler, I., Barabe, D., and Jean, R. V. 1997. A history of the study of phyllotaxis.

Annals of Botany, 80(3):231–244.

Barbier de Reuille, P., Bohn-Courseau, I., Ljung, K., Morin, H., Carraro, N., Godin, C.,

and Traas, J. 2006. Computer simulations reveal properties of the cell-cell signaling

network at the shoot apex in Arabidopsis. Proceedings of the National Academy of

Science of the USA, 103(5):1627–1632.

Barbier de Reuille, P., Runions, A., Smith, R., Coen, E., and Prusinkiewicz, P. 2007.

Trichome patterning on growing tissue. In Proceedings of the 5th International Work-

shop on Functional-Structural Plant Models, pages 365–370.

Benková, E., Michniewicz, M., Sauer, M., Teichmann, T., Seifertova, D., Jurgens, G.,

and Friml, J. 2003. Local, efflux-dependent auxin gradients as a common module for

plant organ formation. Cell, 115:591–602.

Callos, J. D. 1994. Organ positions and pattern formation in the shoot apex. The Plant

Journal, 6:1–7.

Carpenter, R., Copsey, L., Vincent, C., Doyle, S., Magrath, R., and Coen, E. 1995. Con-

trol of flower development and phyllotaxy by meristem identity genes in antirrhinum.

Plant Cell, 7(12):2001–2011.

168
 169

Casimiro, I., Marchant, A., Bhalerao, R. P., Beeckman, T., Dhooge, S., Swarup, R.,

Graham, N., Inz, D., Sandberg, G., Casero, P. J., and Bennett, M. 2001. Auxin

transport promotes Arabidopsis lateral root initiation. Plant Cell, 13(4):843–852.

Chaitin, G. 2006. Meta Math!: The Quest for Omega. Vintage Books.

Chapman, J. M. and Perry, R. 1987. A diffusion model of phyllotaxis. Annals of Botany,

60:377–389.

Church, A. H. 1904. On the Relation of Phyllotaxis to Mechanical Laws. Williams and

Norgate, London.

Couder, Y., Pauchard, L., Allain, C., Adda-Bedia, M., and Douady, S. 2002. The leaf

venation as formed in a tensorial

eld. European Physical Journal B, 28:13–138.

Couder, Y. 1998. Initial transitions, order and disorder in phyllotactic patterns: the

ontogeny of Hellianthus annuus. A case study. Acta Societatis Botanicorum Poloniae,

67(2):129–150.

Dai, X., Hayashi, K., i., Nozaki, H., Cheng, Y., and Zhao, Y. 2005. Genetic and chemical

analyses of the action mechanisms of sirtinol in Arabidopsis. Proc Natl Acad Sci U S

A, 102(8):3129–3134.

Boer, M., d., Fracchia, F., and Prusinkiewicz, P. 1992. Rozenberg, G. and Salomaa,

A., editors. Lindenmayer systems: Impacts on theoretical computer science, computer

graphics, and developmental biology, chapter A model for cellular development in mor-

phogenetic fields, pages 351–370. Springer-Verlag, Berlin.

Delbarre, A., Muller, P., Imhoff, V., and Guern, J. 1996. Comparison of mechanisms

controlling uptake and accumulation of 2,4-dichlorophenoxy acetic acid, naphthalene-

170

1-acetic acid, and indole-3-acetic acid in suspension-cultured tobacco cells. Planta, 198

(4):532–541.

Douady, S. and Couder, Y. 1992. Phyllotaxis as a physical self-organized growth process.

Physical Review Letters, 68(13):2098–2101.

Douady, S. and Couder, Y. 1996a. Phyllotaxis as a dynamical self organizing process.

Part I: The spiral modes resulting from time-periodic iterations. Journal of Theoretical

Biology, 178:255–274.

Douady, S. and Couder, Y. 1996b. Phyllotaxis as a dynamical self organizing process.

Part II: The spontaneous formation of a periodicity and the coexistence of spiral and

whorled patterns. Journal of Theoretical Biology, 178:255–274.

Douady, S. and Couder, Y. 1996c. Phyllotaxis as a dynamical self organizing process.

Part III: The simulation of the transient regimes of ontogeny. Journal of Theoretical

Biology, 178:255–274.

d’Ovidio, and Mosekilde, . 2000. Dynamical system approach to phyllotaxis. Phys Rev

E Stat Phys Plasmas Fluids Relat Interdiscip Topics, 61(1):354–365.

Dumais, and Steele, . 2000. New evidence for the role of mechanical forces in the shoot

apical meristem. J Plant Growth Regul, 19(1):7–18.

Dumais, J. 2007. Can mechanics control pattern formation in plants? Curr Opin Plant

Biol, 10(1):58–62.

Erickson, R. O. 1983. Dale, J. and Milthorpe, F. L., editors. The growth and functioning

of leaves., chapter The geometry of phyllotaxis., pages 53–88. Cambridge University

Press.
 171

Erickson, R. O. and Sax, K. B. 1956. Elemental growth rate of the primary root of

Zea mays. Philosophical Transactions of the Royal Society of London. B. Biological

Sciences, 100(5):487–498.

Errera, L. 1888. Über zellfromen und seifenblasen. Botanisches Centralblatt, 34:395–398.

Esau, K. 1942. Vascular differentiation in the vegetative shoot of Linum. i.: The

procambium. American Journal of Botany, 29(9):738–747.

Feugier, F. G., Mochizuki, A., and Iwasa, Y. 2005. Self-organization of the vascular

system in plant leaves: inter-dependent dynamics of auxin flux and carrier proteins.

Journal of Theoretical Biology, 236(4):366–375.

Feugier, F. G. and Iwasa, Y. 2006. How canalization can make loops: a new model of

reticulated leaf vascular pattern formation. J Theor Biol, 243(2):235–244.

Flanders, D. J., Rawlins, D. J., Shaw, P. J., and Lloyd, C. W. 1990. Nucleus-associated

microtubules help determine the division plane of plant epidermal cells: avoidance

of four-way junctions and the role of cell geometry. Journal of Cell Biology, 110(4):

1111–1122.

Fleming, A. J., McQueen-Mason, S., T., M., and Kuhlemeier, C. 1997. Induction of leaf

primordia by the cell wall protein expansin. Science, 276:1415–1418.

Foley, J. D., Dam, A., v., Feiner, S. K., and Hughes, J. F. 1990. Computer graphics -

Principles and practice. Addison Wesley, 2nd edition.

Fowler, D. 1992. A collision-based model of spiral phyllotaxis. Computer Graphics

(ACM SIGGRAPH), 26(2):361–368.

Fracchia, D., Prusinkiewicz, P., and Boer, M., d. 1990. Animation of the development

of multicellular structures. In Computer animation ’90.

172

Galweiler, L., Guan, C., Muller, A., Wisman, E., Mendgen, K., Yephremov, A., and

Palme, K. 1998. Regulation of polar auxin transport by AtPIN1 in Arabidopsis

vascular tissue. Science, 282:2226–30.

Giavitto, J.-L. and Michel, O. 2001. MGS: a programming language for the transforma-

tions of topological collections. Technical Report 61–2001, Laboratoire de Méthodes

Informatiques, CNRS – Université d’Evry Val d’Essonne, Evry, France.

Girolami, G. 1953. Relation between phyllotaxis and primary vascular organization in

linum. American Journal of Botany, 40(8):618–625.

Gómez-Campo, C. 1974. Phyllotactic patterns in Bryophyllum tubiflorum harv. Botan-

ical Gazette, 135(1):49–58.

Goodbody, K., Venverloo, C., and Lloyd, C. 1991. Laser microsurgery demonstrates

that cytoplasmic strands anchoring the nucleus accross the vacuole of premitotic plant

cells are under tension, implications for division plane alignment. Development, 113:

931–939.

Gottlieb, M. E. 1993. Garcia-Ruiz, J. M., Louis, E., Meakin, P., and Sander, L. M.,

editors. Growth patterns in physical sciences and biology, chapter Angiogenesis and vas-

cular networks: complex anatomies from deterministic non-linear physiologies., pages

267–276. Plenum Press, New York.

Green, P. B. 1980. Organogenesis a biophysical view. Annual Review of Plant Physi-

ology, 31:51–82.

Green, P. B. 1992. Pattern formation in shoots: a likely role for minimal energy

configurations of the tunica. International Journal of the Plant Sciences, 153(3):S59–

S75.
 173

Green, P. B., Steele, C. S., and Rennich, S. C. 1996. Phyllotactic patterns: a biophysical

mechanism for their origin. Annals of Botany, 77(5):515–528.

Hejnowicz, Z. and Romberger, J. A. 1984. Growth tensor of plant organs. Journal of

Theoretical Biology, 110:93114.

Hejnowicz, Z., Nakielski, J., and Hejnowicz, K. 1984. Modeling of spatial variations

of growth within apical domes by means of the growth tensor. ii. growth specified on

dome surface. Acta Societatis Botanicorum Poloniae, 53:301–316.

Hellendoorn, P. H. and Lindenmayer, A. 1974. Phyllotaxis in bryophyllum tubiflorum:

morphogenetic studies and computer simulations. Acta Botanica Néerlandica, 23(4):

473–492.

Hellwig, H., Engelmann, R., and Deussen, O. 2006. Contact pressure models for spiral

phyllotaxis and their computer simulation. Journal of Theoretical Biology, 240(3):

489–500.

Hernandez, L. F. and Green, P. B. 1993. Transductions for the expression of structural

pattern: Analysis in sunflower. Plant Cell, 5(12):1725–1738.

Hofmeister, W. 1868. Handbuch der Physiologishen Botanik. Engelmann, Leipzig.

Honda, H. 1978. Description of cellular patterns by Dirichlet domains: the two-

dimensional case. Journal of Theoretical Biology, 72:523–543.

Honda, H. 1983. Geometrical models for cells in tissues. International Review of

Cytology, 81:191–248.

Jackson, D. and Hake, S. 1999. Control of phyllotaxy in maize by the abphyl1 gene.

Development, 126(2):315–323.
174

Jean, R. V. 1986. A basic theorem on and a fundamental approach to pattern formation

on plants. Mathematical Biosciences, 79:127–154.

Jean, R. V. 1994. Phyllotaxis: A Systematic Study in Plant Morphogenesis. Cambridge

University Press.

Jean, R. and Barabé, D. 2001. Application of two mathematical models to the araceae,

a family of plants with enigmatic phyllotaxis. Annals of Botany, 88:173–186.

Jönsson, H., Heisler, M., Shapiro, B. E., Meyerowitz, E. M., and Mjolsness, E. 2006.

An auxin-driven polarized transport model for phyllotaxis. Proceedings of the National

Academy of Science of the USA.

Kang, J. and Dengler, N. 2002. Cell cycling frequency and expression of the homeobox

gene athb-8 during leaf vein development in Arabidopsis. Planta, 216(2):212–219.

Kang, J., Tang, J., Donnelly, P., and Dengler, N. 2003. Primary vascular pattern and

expression of ATHB-8 in shoots of Arabidopsis. New Phytologist, 158:443454.

Karwowski, R. and Prusinkiewicz, P. 2003. Design and implementation of the L+C

modeling language. Electronic Notes in Theoretical Computer Science, 86(2):1–19.

King, S., Beck, F., and Lüttge, U. 2004. On the mystery of the golden angle in phyl-

lotaxis. Plant, Cell and Environment, 27:685–695.

Koch, A. J., Guerreiro, J., Bernasconi, G. P., and Sadik, J. 1994. An analytic model of

phyllotaxis. Journal de Physique France, 4:187–207.

Kramer, E. M. 2002. A mathematical model of pattern formation in the vascular

cambium of trees. Journal of Theoretical Biology, 216(2):147–158.

Kramer, E. M. 2004. Pin and aux/lax proteins: their role in auxin accumulation. Trends

in Plant Science, 9(12):578–582.

 175

Kuhlemeier, C. and Reinhardt, D. 2001. Auxin and phyllotaxis. Trends Plant Sci, 6(5):

187–189.

Kuhlemeier, C. 2007. Phyllotaxis. Trends in Plant Science, 12(4):143–150.

Kunz, M. 2003. Dynamical models of phyllotaxis. Physica D, 157:147–165.

Kwiatkowska, D. 1995. Ontogenetic changes of phyllotaxis in Anagallis arvensis L. Acta

Societatis Botanicorum Poloniae, 64(4):319325.

Kwiatkowska, D. 1997. Intraspecific variation of phyllotaxis in Anagallis arvensis. Acta

Societatis Botanicorum Poloniae, 66(3-4):259271.

Kwiatkowska, D. and Florek-Marwitz, J. 1999. Ontogenetic variation of phyllotaxis and

apex geometry in vegetative shoots of Sedum maximum (l.) hoffm. Acta Societatis

Botanicorum Poloniae, 68(2):85–95.

Kwiatkowska, D. 2006. Flower primordium formation at the Arabidopsis shoot apex:

quantitative analysis of surface geometry and growth. Journal of Experimental Botany,

57(3):571–580.

Kwiatkowska, D. and Dumais, J. 2003. Growth and morphogenesis at the vegetative

shoot apex of Anagallis arvensis L. Journal of Experimental Botany, 54(387):1585–

1595.

Larson, P. R. 1975. Development and organization of the primary vascular system

in Populus deltoides according to phyllotaxy. American Journal of Botany, 62(10):

1084–1099.

Larson, P. R. 1983. Dale, J. E. and Milthorpe, F. L., editors. The Growth and Function-

ing of Leaves, chapter Primary vascularization and the siting of primordia. Cambridge

University Press.
176

Lindenmayer, A. and Rozenberg, G. 1979. Claus, V., Ehrig, H., and Rozenberg, G.,

editors. Graph grammars and their application to computer science; First International

Workshop, Lecture Notes in Computer Science 73, chapter Parallel generation of maps:

Developmental systems for cell layers, pages 301–316. Springer-Verlag, Berlin.

Lindenmayer, A. 1968a. Mathematical models for cellular interactions in development.

I. Filaments with one-sided inputs. Journal of Theoretical Biology, 18(3):280–299.

Lindenmayer, A. 1968b. Mathematical models for cellular interactions in development.

II. Simple and branching filaments with two-sided inputs. Journal of Theoretical Bi-

ology, 18(3):300–315.

Lyndon, R. F. 1998. The shoot apical meristem. Cambridge University Press.

Mattsson, J., Ckurshumova, W., and Berleth, T. 2003. Auxin signaling in Arabidopsis

leaf vascular development. Plant Physiol, 131(3):1327–1339.

Meinhardt, H. 1976. Morphogenesis of lines and nets. Differentiation, 6(2):117–123.

Meinhardt, H., J., K. A., and G., B. 1998. Jean, R. V. and Barabé, D., editors. Symmetry

in Plants, chapter Models of pattern formation applied to plant development., pages

723–758. World Scientific Publishing, Singapore.

Meinhardt, H. 1982. Models of Biological Pattern Formation. Academic Press.

Meinhardt, H. 2003a. The algorithmic beauty of sea shells. The Virtual Laboratory.

Springer, 3ème edition.

Meinhardt, H. 2003b. Complex pattern formation by a self-destabilization of established

patterns: chemotactic orientation and phyllotaxis as examples. C. R. Biologies, 326

(2):223–237.
 177

Meinhardt, H. and Gierer, A. 1974. Applications of a theory of biological pattern

formation based on lateral inhibition. Journal of Cell Science, 15(2):321–346.

Mitchison, G. J. 1980. A model for vein formation in higher plants. Philosophical

Transactions of the Royal Society of London. B. Biological Sciences, 207:79–109.

Mitchison, G. J. 1981. The polar transport of auxin and vein patterns in plants.

Philosophical Transactions of the Royal Society of London. B. Biological Sciences, 295:

461–271.

Mitchison, G. 1977. Phyllotaxis and the fibonacci series,. Science, 196:270–275.

Mjolsness, E. 2006. The growth and development of some recent plant models: A

viewpoint. Journal of Plant Growth Regulation, 25:270–277.

Mündermann, L., Erasmus, Y., Lane, B., Coen, E., and Prusinkiewicz, P. 2005. Quan-

titative modeling of Arabidopsis development. Plant Physiology, 139(2):960–968.

Nakielski, J. 2000. Carbone, A., Gromov, M., and Prusinkiewicz, P., editors. Pattern

formation in biology, vision and dynamics, chapter Tensorial model for growth and cell

division in the shoot apex, pages 252–286. World Scientific.

Neider, J., Davis, T., and Woo, M. 1994. OpenGL Programming Guide 9 (The Red

Book). Addison-Wesley.

Niklas, K. J. 1988. The role of phyllotactic pattern as a developmental constraint on

the interception of light by leaf surfaces. Evolution, 42:1–16.

Niklas, K. J. 1998. Jean, R. V. and Barabé, D., editors. Symmetry in Plants., chap-

ter Light harvesting fitness landscapes for vertical shoots with different phyllotactic

patterns, pages 759–773. World Scientific Publishing, Singapore.

178

Okada, K., Ueda, J., Komaki, M. K., Bell, C. J., and Shimura, Y. 1991. Requirement of

the auxin polar transport system in early stages of Arabidopsis floral bud formation.

Plant Cell, 3(7):677–684.

Paciorek, T., Zazmalov, E., Ruthardt, N., Petrsek, J., Stierhof, Y.-D., Kleine-Vehn, J.,

Morris, D. A., Emans, N., Jrgens, G., Geldner, N., and Friml, J. 2005. Auxin inhibits

endocytosis and promotes its own efflux from cells. Nature, 435(7046):1251–1256.

Pearcy, R. W. and Yang, W. 1998. The functional morphology of light capture and car-

bon gain in the Redwood forest understory plant Adenocaulon bicolor Hook. Functional

Ecology, 12:543–552.

Peer, W. A., Bandyopadhyay, A., Blakeslee, J. J., Makam, S. N., Patrick H. Masson,

R. J. C., a., and Murphy, A. S. 2004. Variation in Expression and Protein Localization

of the PIN Family of Auxin Efflux Facilitator Proteins in Flavonoid Mutants with

Altered Auxin Transport in Arabidopsis thaliana. The Plant Cell, 16:1898–1911.

Press, W. H., Flannery, B. P., Teukolsky, S. A., and Vetterling, W. T. 1992. Numerical

recipes: the art of scientific computing. Cambridge University Press, Cambridge (UK)

and New York, 2nd edition.

Priestly, J., Scott, L., and Gillett, E. 1935. The development of the shoot in Alstroemeria

and the unit of shoot growth in monocotyledons. Annals of Botany, 49:161–179.

Prusinkiewicz, P. 1998. Sommerer, C. and Mignonneau, L., editors. Art@Science,

chapter In search of the right abstraction: The synergy between art, science, and

information technology, pages 60–68. Springer, Wien.

Prusinkiewicz, P. 2004. Art and science for life: Designing and growing virtual plants

with l-systems. Acta Hortic, 630:1528.

 179

Prusinkiewicz, P. and Lindenmayer, A. 1990. Algorithmic Beauty of Plants. Springer-

Verlag.

Prusinkiewicz, P., Muendermann, L., Karwowski, R., and Lane, B. 2001. The use of

positional information in the modeling of plants. In Proceedings of SIGGRAPH, pages

289–300.

Prusinkiewicz, P., Erasmus, Y., Lane, B., Harder, L. D., and Coen, E. 2007. Evolution

and development of inflorescence architectures. Science, 316(5830):1452–1456.

Reddy, G. V., Heisler, M. G., Ehrhardt, D. W., and Meyerowitz, E. M. 2004. Real-time

lineage analysis reveals oriented cell divisions associated with morphogenesis at the

shoot apex of Arabidopsis thaliana. Development, 131(17):4225–4237.

Reinhardt, D., Mandel, T., and Kuhlemeier, C. 2000. Auxin regulates the initiation and

radial position of plant lateral organs. Plant Cell, 12(4):507–518.

Reinhardt, D., Frenz, M., Mandel, T., and Kuhlemeier, C. 2003a. Microsurgical and

laser ablation analysis of interactions between the zones and layers of the tomato shoot

apical meristem. Development, 130(17):4073–4083.

Reinhardt, D., Pesce, E.-R., Stieger, P., Mandel, T., Baltensperger, K., Bennett, M.,

Traas, J., Friml, J., and Kuhlemeier, C. 2003b. Regulation of phyllotaxis by polar

auxin transport. Nature, 426(6964):255–260.

Reinhardt, D., Frenz, M., Mandel, T., and Kuhlemeier, C. 2005. Microsurgical and laser

ablation analysis of leaf positioning and dorsoventral patterning in tomato. Develop-

ment, 132(1):15–26.

Richards, F. J. 1948. The geometry of phyllotaxis and its origin. In Symposium of the

Society for Experimental Biology 2., pages 217–245.

180

Richards, F. J. 1951. Phyllotaxis: Its quantitative expression and relation to growth in

the apex.pdf. Philosophical Transactions of the Royal Society of London. B. Biological

Sciences, 235:509–564.

Richards, O. W. and Kavanagh, A. J. 1943. The analysis of the relative growth gradients

and changing form of growing organisms illustrated by the tobacco leaf. American

Naturalist, 77:385–399.

Ridley, J. N. 1982. Packing efficiency in sunflower heads. Mathematical Biosciences, 58:

129–139.

Ridley, J. N. 1986. Ideal phyllotaxis on general surfaces of revolution. Mathematical

Biosciences, 79:1–24.

Rodkaew, Y., Siripant, S., Lursinsap, C., and Chongstitvatana, P. 2002. An algorithm for

generating vein images for realistic modeling of a leaf. Prodeedings of the International

Conference on Computational Mathematics and Modeling, 9:73–78.

Rolland-Lagan, A. G. and Prusinkiewicz, P. 2005. Reviewing models of auxin canaliza-

tion in the context of leaf vein pattern formation in Arabidopsis. The Plant Journal,

44(5):854–865.

Runions, A., Fuhrer, M., Lane, B., Federl, P., Rolland-Lagan, A.-G., and Prusinkiewicz,

P. 2005. Modeling and visualization of leaf venation patterns. ACM Transactions on

Graphics, 24(3):702–711.

Sachs, T. 1981. The control of patterned differentiation of vascular tissues. Advances in

Botanical Research, 9:151–262.

Sachs, T. 1991. Pattern Formation in Plant Tissues. Cambridge University Press.

 181

Scarpella, E., Marcos, D., Friml, J., and Berleth, T. 2006. Control of leaf vascular

patterning by polar auxin transport. Genes and Development, 20(8):1015–1027.

Schoute, J. C. 1913. Beitrage zur blattstellunglehre. i. die theorie. Recueilde Travaux

Botániques Néerlandais., 10:153–339.

Schüepp, O. 1916. Beiträge zur theorie des vegetationspunktes. Berichte der Deutschen

Botanishen, 34:847857.

Schwabe, W. W. 1984. Barlow, P. W. and Carr, D. J., editors. Positional control in

plant development, chapter Phyllotaxis, pages 403–440. Cambridge University Press.

Schwabe, W. and Clewer, A. G. 1984. Phyllotaxisa simple computer model based on the

theory of a polarly-translocated inhibitor. Journal of Theoretical Biology, 109:595–619.

Schwendener, S. 1878. Mechanische Theorie der Blattstellungen. Engelmann.

Shipman, P. D. and Newell, A. C. 2005. Polygonal planforms and phyllotaxis on plants.

Journal of Theoretical Biology, 236(2):154–197.

Silk, W. K. and Erickson, R. O. 1979. Kinematics of plant growth. Journal of Theoretical

Biology, 76:481–501.

Smith, C. 2006. On Vertex-Vertex Systems and their use in geometric and biological

modelling. PhD thesis, The University of Calgary, Department of Computer Science.

Smith, C., Prusinkiewicz, P., and Samavati, F. 2004. Local specification of surface

subdivision algorithms. In Pfaltz, J. L., Nagl, M., and Blen, B., editors, Applications

of Graph Transformations with Industrial Relevance: Second International Workshop,

AGTIVE 2003, pages 313–327. Springer-Verlag GmbH.

182

Smith, R. S., Guyomarc’h, S., Mandel, T., Reinhardt, D., Kuhlemeier, C., and

Prusinkiewicz, P. 2006a. A plausible model of phyllotaxis. Proceedings of the National

Academy of Science of the USA, 103(5):1301–1306.

Smith, R. S., Kuhlemeier, C., and Prusinkiewicz, P. 2006b. Inhibition fields for phyllot

actic pattern formation: a simulation study. Canadian Journal of Botany, 84(11).

Snow, M. and Snow, R. 1932. Experiments on phyllotaxis. i. the effect of isolating a

primordium. Philosophical Transactions of the Royal Society of London. B. Biological

Sciences, 221:1–43.

Snow, M. and Snow, R. 1947. On the determination of leaves. New Phytologist, 46(1):

519.

Steeves, T. A. and Sussex, I. A. 1989. Patterns in plant development. Cambridge

University Press, 2 edition.

Stern, A. and Desbrun, M. 2006. Discrete geometric mechanics for variational time

integrators. In Discrete Differential Geometry: An Applied Introduction, a fullday

course at SIGGRAPH ’06.

Stieger, P. A., Reinhardt, D., and Kuhlemeier, C. 2002. The auxin influx carrier is

essential for correct leaf positioning. The Plant Journal, 32(4):509–517.

Stoma, S., Chopard, J., Godin, C., and Traas, J. 2007. Using mechanics in the mod-

elling of meristem morphogenesis. In Proceedings of the 5th International Workshop

on Functional-Structural Plant Models, pages 365–370.

Szymkowiak, E. J. and Sussex, I. M. 1996. What chimeras can tell us about plant

development. Annual Review of Plant Physiology and Plant Molecular Biology, 47:

351–376.
 183

Takebe, I. 1971. Regeneration of whole plants from isolated mesophyll protoplasts of

tobacco. Naturwissenschaften, 58:318–320.

Thornley, J. H. M. 1975. Phyllotaxis i. a mechanistic model. Annals of Botany, 39:

491–07.

Turing, A. 1952. The chemical basis of morphogenesis. Philosophical Transactions of

the Royal Society of London. B. Biological Sciences, 237:37–52.

van Iterson, G. 1907. Mathematische und Microscopisch Anatomische Studien über

Blattstellungen, nebst Betraschungen über den Schalebau der Miliolinen. Gustav-

FischerVerlag, Iena.

Veen, A. H. and Lindenmayer, A. 1977. Diffusion mechanism for phyllotaxis: theoretical

physico-chemical and computer study. Plant Physiology, 60(1):127–139.

Vieten, A., Vanneste, S., Wisniewska, J., RenBenjamins, E. B., Beeckman, T., Luschnig,

C., and JirFriml, . 2005. Functional redundancy of PIN proteins is accompanied by

auxin-dependent cross-regulation of PIN expression. Development, 132(20):4521–4531.

Vogel, H. 1979. A better way to construct a sunflower head. Mathematical Biosciences,

44:179–189.

Wardlaw, C. W. 1943. Experimental and analytical studies of pteridophytes: I. pre-

liminary observations on the development of buds on the rhizome of the ostrich fern

(matteuccia struthiopteris tod.). Annals of Botany, 7:171184.

Wardlaw, C. W. 1950. The comparative investigation of apices of vascular plants by

experimental methods. Philosophical Transactions of the Royal Society of London. B.

Biological Sciences, 234:583–602.

Wardlaw, C. W. 1968. Morphogenesis in Plants. Methuen & Co., London, U.K.

184

Wolpert, L. 1969. Positional information and the spatial pattern of cellular differentia-

tion. Journal of Theoretical Biology, 25:1–47.

Yotsumoto, A. 1993. A diffusion model for phyllotaxis. Journal of Theoretical Biology,

162:131–151.

Young, D. A. 1978. On the diffusion theory of phyllotaxis. Journal of Theoretical

Biology, 71(3):421–432.

Zagórska-Marek, B. 1985. Phyllotactic patterns and transitions and transitions in Abies

balsamea. Canadian Journal of Botany, 63:1844–1854.

Zagórska-Marek, B. 1994. Phyllotactic diversity in magnolia flowers. Acta Societatis

Botanicorum Poloniae, 63:117–137.