The editors of this compilation declare two goals: to show how computer science can help us to understand biological development, and to show how ideas from developmental biology can stimulate new thinking in computer science. The objectives are neatly symmetrical, but they do not sit comfortably together. Developmental biologists and computer scientists speak different languages and inhabit different worlds. On one side of the cultural divide, the developmental biologists struggle to digest complex experimental findings and make sense of them, tangled in wearisome quantities of data that defy quantitative analysis. On the other side, the computer modellers construct ingenious idealized systems, with a blithe disregard for biological realities and a greater concern for what is possible than for what is actual. A terrible jungle or a ridiculous toyland: in bad moments that seems to be the dichotomy. The question in my mind as I embarked on On Growth, Form and Computers was whether it would show us some happy middle ground, where computer modelling casts clear fresh light on the workings of biological systems, and in particular of real developing organisms.
The book begins with an introductory chapter that is mainly addressed to computer scientists and in which, among other things, the editors rashly undertake to summarise developmental, cell and molecular biology in nine pages for the benefit of those who do not know about these things. The rest of the book consists of 21 heterogeneous essays by different authors, including six pieces that are reprinted from other publications. The first few contributions (Wolpert on development and evolution, Hancock on cell signalling, Bolker on the genotype-phenotype relationship, and Brockes and Kumar on amphibian regeneration) discuss aspects of developmental biology but say little or nothing directly about computer modelling. Developmental biologists will find some things to interest them here, but not much enlightenment on the main theme of the book.
Subsequent chapters get down to brass tacks, and fall roughly into two groups. First come those directed towards biological questions – attempts to use computer modelling to understand specific developmental systems, as well as general essays on the application of mathematics or computer modelling to developmental biology. Second, there are accounts of computer constructs inspired by ideas from biology, but studied by computer scientists for their own sake. A good representative of the first category is Meinhardt's contribution, reviewing the work for which he is famous on the role of short-range activation and long-range inhibition in the creation of spatial patterns of differentiation. Even though it has been difficult in most cases to substantiate Meinhardt's models with detailed quantitative data, and some of Meinhardt's specific applications are highly speculative, his basic ideas about how symmetry is broken and signalling centres are set up are simple and illuminating, and are backed up by serious consideration of specific biological examples. They have become a significant part of the developmental biologist's conceptual toolkit, helping us to think about how real systems work. The breaking of symmetry is certainly fundamental to pattern formation, and this is emphasized in a nice (reprinted) essay on the subject by Ian Stewart, although his final disquisition on the symmetry classification of the gaits of quadrupeds strays rather far from the theme of this book.
The book contains relatively few contributions that focus on one particular developmental system and show how computer modelling can really help us understand it. This is a pity, and perhaps a reflection of the editors' backgound as computer scientists rather than developmental biologists. The book does, however, include an essay on one of the most interesting systems from this point of view, the shoot apical meristem, with its regular sequential emergence of leaf primordia, each one positioned precisely in relation to those already present. Work in the last few years has identified many of the key molecules governing this process, to the point where a close interaction between experimentalists and modellers can be very fruitful. The chapter by Jönsson, Shapiro, Meyerowitz and Mjolsness describes one such collaboration, although it is too brief and condensed to do the subject justice.
A central task for computer modellers, addressed by several contributors (e.g. De Jong, Geiselmann and Thieffry, and Reil), is to describe and analyse the behaviour of gene regulatory networks. The standard approach is through partial differential equations. The difficulty is that in most biological systems we have only an incomplete knowledge of network topology, and, worse still, virtually no information about the quantitative parameters of the regulatory mechanisms: the experiments may tell us that X goes up when Y goes down, but they rarely tell us how steeply, or within what range of concentrations, or with what sort of non-linearity. One response is to is to make models that avoid quantitative description and seek to represent the available qualitative information in terms of logical (yes-no) variables instead of continuous variables. The fallacy is that such idealizations, at least for networks with feedback, sadly fail to capture correctly even the qualitative behaviour of the underlying continuum system. Even if we only want to make qualitative predictions, we have to have quantitative information. The lack of quantitation is, I think, the fundamental reason why developmental biology has yet to find its Newton, or its Hodgkin and Huxley.
These problems become more severe, and the computer models markedly more complex and difficult to construct, when cell movement and cell rearrangements also have to be taken into account. Several chapters (e.g. Miodownik, Fleischer, Eggenberger Hotz) discuss this enterprise and convey a flavour of the problem, though with an emphasis on general principles and idealized systems, rather than on analysis of specific real cases.
A still more ambitious undertaking is to model the evolution of developmental programs. This involves setting up a description of the rules of gene regulation, cell proliferation and cell movement, computing the resultant phenotype, and then allowing the rules to change through successive rounds of mutation and reproduction, with selection applied according to some measure of fitness of the computed phenotype. Hogeweg's studies of her `critters' are an example. Here, as in most undertakings of this sort, the assumptions needed to construct a tractable model become so artificial that it is hard to identify any correspondence with a specific real biological system. Whether useful insights are gained into general biological principles is a matter of debate.
The chapters in the final section of the book for the most part abandon any claim to answer specific concrete biological problems. They ask not what computer science can do for biology, but what biology can do for computer science. They are addressed primarily to computer scientists, and make rather hard, though interesting, reading for outsiders. So far, the main achievement of biologically inspired computing has been to cause everyone a lot of trouble, in the form of computer viruses. The book passes over these in silence; the emphasis instead is on more benign forms of `artificial life'– that is, on non-malignant computer programs that evolve by random mutation and selection according to the output that they produce. Ray and Hart (in a reprinted paper) do, however, explore the evolution of software constructs that can migrate from computer to computer to computer, sense their environment, and compete with one another. `Artificial life' is a rich and intriguing topic, with practical applications to such things as the development of control systems for robots (discussed by Jakobi). It raises many general questions. Will such computer constructs tend to evolve in the direction of increasing size and complexity? If the selected phenotype is behavioural, reflecting the connectivity of a control network, how will the structure and performance of that network evolve (Rust, Adams, Schilstra and Bolouri; Dellaert and Beer)? What course will evolution take if the programs operate in a variable environment and are capable of `learning' from `experiences' (discussed by Cangelosii, Nolfi and Parisi)? Will the programs evolve to outstrip the ingenuity of their human creators? The last few chapters of the book do not give straightforward answers to questions such as these – none, at least, that I as an outsider to the field could easily grasp; but they provide much food for thought.
In short, the book is a mixed bag. Reading it makes one all too well aware of the difficulties of the relationship between developmental biology and computer modelling. There are some nicely written essays that make good points of principle, but few shining examples of models that succeed in adding to our understanding of specific events in the development of real organisms.
- © 2004.