Sui Huang received his MD and PhD from the University of Zurich and is an assistant professor at Harvard Medical

School. His research focuses on complexity in multicellularity and cancer.

Back to the biology in systems biology: What can we learn from biomolecular networks?
Sui Huang
Date received (in revised form): 1st December 2003

Abstract
Genome-scale molecular networks, including protein interaction and gene regulatory networks, have taken centre stage in the investigation of the burgeoning disciplines of systems biology and biocomplexity. What do networks tell us? Some see in networks simply the comprehensive, detailed description of all cellular pathways, others seek in networks simple, higher-order qualities that emerge from the collective action of the individual pathways. This paper discusses networks from an encompassing category of thinking that will hopefully help readers to bridge the gap between these polarised viewpoints. Systems biology so far has emphasised the characterisation of large pathway maps. Now one has to ask: where is the actual biology in systems biology? As structures midway between genome and phenome, and by serving as an extended genotype or an elementary phenotype, molecular networks open a new window to the study of evolution and gene function in complex living systems. For the study of evolution, features in network topology offer a novel starting point for addressing the old debate on the relative contributions of natural selection versus intrinsic constraints to a particular trait. To study the function of genes, it is necessary not only to see them in the context of gene networks, but also to reach beyond describing network topology and to embrace the global dynamics of networks that will reveal higher-order, collective behaviour of the interacting genes. This will pave the way to understanding how the complexity of genomewide molecular networks collapses to produce a robust whole-cell behaviour that manifests as tightly-regulated switching between distinct cell fates the basis for multicellular life.

Keywords: systems biology, biocomplexity, genetic networks, intrinsic constraints, adaptation, attractors

INTRODUCTION: MIND THE GAP OF MINDS

Ever since molecular biology came into existence as a discipline devoted to studying the function and functioning of genes some 40 years ago, scientists have entertained the idea of gene regulatory circuits,1,2 and even of genome-scale genetic networks,3 for understanding how the activity of genes by inuencing each other ultimately translates into the phenotypic behaviour of an organism. In the absence of gene-specic data, the concepts of genetic networks were explored mostly at the level of abstract, generic models, without moleculespecic information.4 Despite intriguing insights into the fundamental properties of the collective behaviour of a large number of interacting components,3 such work

Sui Huang, Department of Surgery, Childrens Hospital, Harvard Medical School, 300 Longwood Avenue, Boston, MA 02115 USA Tel: +1 617 919 2393 e-mail: sui.huang@tch.harvard.edu

remained largely unnoticed by practitioners of molecular biology, who emphasised the cloning, characterising and categorising of individual genes. Molecular biologists both experimental and computational used to studying specic, measurable biochemical details, were not ready for the abstraction required to handle networks as an object of study. It was only the recent availability of gene-specic data for genome-scale networks of molecular interactions, made possible by high-throughput genomic technologies, that revived the interest in the idea of large gene networks. The arrival of genome-scale information on networks, ranging from networks of metabolic reactions5 to gene regulatory6,7 and protein interaction

& HENRY STEWART PUBLICATIONS 1473-9550. B R I E F I N G S I N F U N C T I O N A L G E N O M I C S A N D P R O T E O M I C S . VOL 2. NO 4. 279297. FEBRUARY 2004 2 7 9

Huang

A similar polarisation between localists and globalists existed in brain research

networks,810 has led to a recent spate of publications in which the topology (wiring diagram) of molecular networks was analysed.1118 When reading the literature, however, it is important to be aware of two camps of researchers with philosophically opposite mindsets and, hence, disparate motivation. These two camps can be labelled the globalists (or, generalists) and the localists (or, particularists) (Table 1). Historically, a similar polarisation existed in brain research, with globalists emphasising the idea that the entire network of neurons in the brain collectively determine behaviour by distributed information processing, while the localists believed more in localising a brain function to a particular anatomical region.19 It is now known that both were right. The rst wave of work on genomewide networks was conducted mostly by physicists in the globalists perspective, so that networks were analysed in a more

general sense, ie as an entity in their own right.11,12,14,15,20 In this globalists abstraction the identication of individual genes and their relationships and functions were not the primary objectives of analysis. Thus, this early work on networks failed to attract the attention of scientists in the localists camp, which consists largely of mainstream molecular biologists habituated to characterising specic genetic pathways one at a time. There is a sharp, natural, but unarticulated intellectual disconnect between the two camps because of the fundamentally different motivation of the two groups of scientists. The globalists are typically interested in deeper principles of complex systems,2124 they are attracted to biology as a new source of complexity, which led to the notion of biocomplexity.2528 They now seek molecule-specic data, made possible by technological advances, to validate their ideas.

Table 1: Polarity of two points of view in network biology. The table represents a caricature of extreme positions for illustration purposes.
The localist (particularist) view (Those who see the trees rst) Level of original focus New eld created Use of hypothesis Gene- and pathway-centric Systems biology Hypothesis at level of individual pathways. No systemslevel hypothesis: research becomes discovery-driven. Example hypotheses: Gene A inhibits Gene B, is required for function X, etc. System is complicated Properties of systems lie in the property of the components Comprehensiveness. The whole equals the sum of the parts To characterise exhaustively the biochemistry of (all) individual pathways and their functions To describe idiosyncrasy Of primary interest. Specic models with nominal genes and their idiosyncratic properties Precise biochemical characterisation and categorising of physical and regulatory interactions in specic pathways The globalist (generalist) view (Those who see the forest rst) Network-centric Biocomplexity Hypothesis at systems level concerning generic design of network and network position of genes Example hypotheses: Hub proteins are important Power-law architecture favours ordered dynamics System is complex Higher-order system properties emerge from collective behaviour of components Holism The whole is different from sum of parts To understand generic aspects of genome-scale networks as an entity with its higher-order properties To understand universality Of secondary interest. Models with anonymous genes as generic entities may sometime sufce Analysis of large scale features, based on global statistics of local network features (degree distribution, modularity, clustering) Global dynamics of network maps into whole cell behaviour (cell fates) Emphasise emergent whole-cell behaviour, such as switch between discrete cell phenotypes (cell fates)

Philosophy

Practical aims of study

Gene identity in models Network topology

Network dynamics Function

Typical non-biologist partners

Detailed modelling of individual small circuits (modules) in separation as a low-dimensional dynamic system Focus on local cellular functions (eg protein synthesis, vesicle transport, lopodia extension, DNA repair, etc) associated with a specic pathway Computer scientists, engineers Physicists

2 8 0 & HENRY STEWART PUBLICATIONS 1473-9550. B R I E F I N G S I N F U N C T I O N A L G E N O M I C S A N D P R O T E O M I C S . VOL 2. NO 4. 279297. FEBRUARY 2004

Back to the biology in systems biology

Understanding living systems requires both detailed description and abstraction

By contrast, the localists view is rooted in classical molecular biology, hence is shaped by decades of devotion to the study of individual cellular pathways that represent to them linear causal relationships. But the prevalence of pleiotropy and convergence in cell signalling, and of crosstalk between pathways, has led to the increasing awareness that understanding gene function requires that one reaches beyond the narrow focus on individual pathways. The advent of massively-parallel analysis and high-throughput genomics and proteomics technologies29,30 nally led to the acceptance of the idea that molecular pathways are just parts embedded in one genome-scale system,31 leading to the explicit inception of systems biology.32 Nevertheless, despite considering networks and quantitative modelling, the pathway-centric ideology shines through in this new initiative towards integration.3337 To the localists, integration essentially means the combination of different kinds of large datasets in order to nd correlations to or at least to facilitate the generation of hypotheses on relationships between individual genes, proteins and biological functions.3740 Thus, the systems biology approach seeks comprehensive, qualitative and quantitative descriptions of all constituent parts, rather than an understanding of a higher-order quality that may arise from the collective action of the individual parts. The implicit notion is that extension of the existing ways of analysing individual pathways to the exhaustive, simultaneous characterisation of all the molecular interactions in the genome will sufce for understanding how living organisms work.27,35,41 Concretely, the emphasis on the analysis of vast amounts of biochemical detail has led to the involvement of computer scientists for developing computational tools that have become indispensable for representing, mining and modelling the bewildering amount and heterogeneity of collected information.4246 Another natural

extension of pathway-centricism to the entire system, which has become a major research focus of systems biology, is to reconstruct the topology of a network by systematic experimental identication of interactions69,33,47 or to reverse engineer it by inference from measurement of network dynamics.4851 In summary, whereas globalists are attracted by the complex, seeking to understand general principles giving rise to the whole, and are ready to abstract away specic details, localists-turnedsystems biologists are more inclined to attack the complicated, and seek comprehensiveness of detailed description. Because of the unique blend in biology of universality and idiosyncrasy, both approaches are complementary to each other and will be necessary in spite of a persisting, unarticulated mental divide (Table 1). This paper discusses molecular networks in a broadly encompassing category of view, going beyond the characterisation of network maps, and addresses the fundamental questions that investigators of both ideologies studying living systems will ultimately have to address: what do networks tell us about the biology of an organism?

FROM NETWORKS IN BIOLOGY TO THE BIOLOGY OF NETWORKS

As with the information in DNA and amino acid sequences, what networks can tell us about the biology of an organism can be divided into two broad domains: evolution and function (Table 2). While comparative sequence analysis proved to be a powerful tool for determining taxonomic relationships,52 and can shed light on the history and mechanism of genome evolution,5355 it remains of limited use for predicting gene function. Although conservation can point to functional importance, and sequence homology (eg catalytic motifs) allows the prediction of the biochemical function of a protein, the road from sequence to gene function and,

& HENRY STEWART PUBLICATIONS 1473-9550. B R I E F I N G S I N F U N C T I O N A L G E N O M I C S A N D P R O T E O M I C S . VOL 2. NO 4. 279297. FEBRUARY 2004 2 8 1

Huang

Table 2: Molecular networks as intermediate level of description complementing information from easily formalisable genotype (gene sequence) and phenotype (organism).
Gene sequence Formalisation Clues on evolution Linear array easy to compare Simple: relationship from sequence comparison Molecular network Graph Organism Morphometry, etc difcult to compare Difcult: common ancestry or convergence?

Clues on function

Difcult: heuristics of protein domain prediction, comparison

New opportunity: intermediate between genotype and phenotype

Simple: evident from observation or experiment

Context of gene expression is part of gene function

subsequently, to drug development turned out to be bumpier than was anticipated in the dawn of the genomic revolution. It is often overseen that the regulation and the context of expression of a gene is part of what denes its biological function. For example, based on sequence analysis, one readily nds that the protein cyclo-oxygenase-2 (COX-2) has the same enzymatic activity as COX-1, namely, a cyclo-oxygenase in the synthesis of prostaglandins.56,57 This information in itself was of little use until it was established that COX-2 is regulated differently than COX-1, and thus has a different biological function.56 Being induced in inammatory tissues, as opposed to the constitutively active COX-1, COX-2 obviously serves a different function and thus provides a more specic drug target than its nonregulated paralogue. Thus, what determines the distinctive biological functions of COX-1 and COX-2 is encoded in their regulation, not in their enzymatic activity.57,58 Furthermore, a regulatory protein, such as Ras, myc or NFkB, can have multiple, disparate, if not opposite, functions (eg either cell proliferation or differentiation/quiescence or apoptosis), depending on the cellular context, ie the presence of other proteins, and the cell state and type.27,59 Hence, despite great efforts spent in functional annotation,60,61 it is only in rather exceptional cases that a specic physiological function can unconditionally and unambiguously be assigned to a given

protein as a separate entity, dened just by the coding region of its gene. It is likely that most proteins, notably regulatory proteins, exert their biological function as a member of a molecular network. What is less obvious is what networks can tell us about biological evolution (Table 2). Representing an intermediate level of information, between the level of DNA sequence and that of organismal phenotype, network structures are more complex than DNA sequences, yet simple enough for a mathematical representation using the formalism from graph theory. Thus, by serving, to all intents and purposes, as an extended genotype, networks open an entirely new window on the workings of evolution. Before discussing what networks can teach about the evolution of the organism, the author will rstly summarise a few recent, salient ndings on network topology obtained from the globalist point of view.

GLOBAL TOPOLOGY OF NETWORKS

The wiring diagram of the entire network can readily be drawn from data on relationships between biomolecules, including metabolic reactions, physical protein interactions (heterodimeric complexes) and transcriptional gene regulation. Network topology information is available for metabolic networks in various microorganisms5 and, more recently, for a genome-wide proteinprotein interaction network of the yeast Saccharomyces cerevisiae.810,62,63

2 8 2 & HENRY STEWART PUBLICATIONS 1473-9550. B R I E F I N G S I N F U N C T I O N A L G E N O M I C S A N D P R O T E O M I C S . VOL 2. NO 4. 279297. FEBRUARY 2004

Back to the biology in systems biology

Genome-wide gene regulatory network data are currently available for Escherichia coli.7 In the terminology of graph theory, a network consists of nodes (vertices) that can be connected by links (edges) (Figure 1). Genome-wide networks consist of thousands of nodes, and thus are referred to as complex networks,64 by contrast with the networks with low node numbers studied in traditional graph theory. Moreover, an important aspect of biological networks is that they grow over time, by contrast with the networks with constant node numbers seen in traditional graph theory.65 Thus, biomolecular networks have a history of experiencing an increase in the number of nodes.66 What catches the investigators interest is every topological feature that deviates from that of a random network, in which N nodes are randomly connected by L links.65 The non-random structures found in complex networks, then, provide the starting point for further research into their origin and functional signicance.

To illustrate this, focus on the results obtained for the most complete eukaryotic protein interaction network available, the protein interaction network of S. cerevisiae, which is based on genomewide data of pair-wise proteinprotein interactions (Figure 1).810 It is important to note that the incompleteness and inaccuracies in the yeast protein interaction dataset, mostly due to potentially false-positive links,62,63 do not essentially affect the global statistics, as comparisons of various datasets by the authors in the cited works have shown. Many questions posed by the globalist can be addressed with incomplete and not perfectly accurate datasets, while the localist agenda requires precise and detailed information. One elementary global feature that was striking for biologists especially those who think in categories of isolated pathway modules named after their putative cellular functions is that, despite the overall sparseness of the network connectivity

Figure 1: Global picture of the topology of large (complex) networks. Upper panels: graph of the network, displayed using Pajek software (http://vlado.fmf.uni-lj.si/pub/networks/pajek/). A. Computer generated, not growing random network,65 N 500 nodes, mean connectivity degree ,k. $ 4. Note the dominating giant component (containing 487 connected nodes). B. Computer generated, growing network using the generative algorithm with preferential attachment of new nodes,20 N 483, ,k. $ 4. (The algorithm necessarily leads to one single component). C. Yeast protein interaction network consisting of N 2,165 proteins, with ,k. $ 4, based on the CORE data set.63 Lower panels: probability density distribution P(k) in loglog plots. The density distribution exhibits a straight line for B, C indicating a powerlaw (scale-free) distribution

& HENRY STEWART PUBLICATIONS 1473-9550. B R I E F I N G S I N F U N C T I O N A L G E N O M I C S A N D P R O T E O M I C S . VOL 2. NO 4. 279297. FEBRUARY 2004 2 8 3

Huang

Topology of biological networks deviate from that of random networks

(three to four connections per protein, on average), the vast majority (.75 per cent) of the proteins in yeast participate in a single, large sub-network of connected nodes, a so-called giant component12 (Figure 1). Although counterintuitive at rst sight, this nding is not unexpected from a graph theory point of view. Given some minimal connectivity (for a growing, random graph, at least k 1/8 links per node on average66 ), such a giant component is practically unavoidable in protein interaction networks (where average k % 4), and is thus robust in a fundamental way. Hence, this feature does not even qualify as a non-random feature that would point to a particular biological purposefulness. By contrast, Barabasi and coworkers describe an interesting global topology feature that indeed deviates from that of a random graph.12,20 If all the nodes i of the network are treated as one population, and k i of each node i (ie the number k of interaction partners that node i has) is measured and displayed as a distribution, then the occurrence probability P(k) of each value of k exhibits, approximately, a power-law distribution, P(k) $ kg , rather than as would be the case in random networks a Poisson distribution. The exponent g is a characteristic constant (Figure 1). A power-law distribution implies that the more nodes that are sampled, the higher is the mean value of k for the sampled population. Thus, there is no stable mean that is characteristic for an ensemble of proteins (nodes), and such protein interaction networks are called scale-free. Although it remains debated as to whether the degree of distribution precisely follows a power-law, its heavytailed nature implies that there are more proteins with a high-degree k than one would expect in a random network: the protein interaction network is biased to contain more hubs than one would expect. Other non-random topological features that deviate from those of a random graph have been identied in the yeast interaction network. For example, hub

proteins tend not to interact with each other.14 Moreover, the proteins or metabolites in real networks have a higher tendency than the nodes in random networks to be organised into clusters (cliques) in which the partners of a given protein are also likely to interact with each other.13,17 Together with the scalefree property, these features give rise to hierarchical modularity, as has been described for metabolic networks.15 The yeast protein network also contains a signicant number of proteins that have low connectivity degrees, yet are central in that they lie on many of the shortest paths that connect any pair of proteins.67 Of interest also is the recent study of network motifs across the network. Network motifs are subgraphs, ie locally dened patterns of interaction between a small number of nodes, such as triangles, branching structures, etc.6,16,18,68 In gene regulatory (transcription) networks, links represent transactivation and are, therefore, arrows rather than undirected connections; thus, transcription networks form directed graphs. Such digraphs exhibit a higher diversity of network motifs than the networks from protein interaction databases, which yield undirected connections. Specically, Shen and coworkers found that in bacterial gene regulatory networks, certain motifs of directed links were signicantly overrepresented, such as the feed-forward loop (Figure 2).18 In view of all of these non-random features of network topology, two interdependent questions immediately come to mind: how do such structures originate, and how do they affect biological function?

NETWORKS AND EVOLUTION: ADAPTATION VERSUS INTRINSIC CONSTRAINTS

A fundamental question in biology is why organisms are the way they are: why does a given trait exist, how does it arise? It is obvious that the non-random network features discussed above can now be

2 8 4 & HENRY STEWART PUBLICATIONS 1473-9550. B R I E F I N G S I N F U N C T I O N A L G E N O M I C S A N D P R O T E O M I C S . VOL 2. NO 4. 279297. FEBRUARY 2004

Back to the biology in systems biology

1
XRE Gene X

2
Gene A ARE

INSERTION

XRE

Gene X

3
Gene A ARE XRE Gene X

Duplicated segment

GENE DUPLICATION DIVERGING MUTATIONS

Gene A

XRE

Gene X

Gene B

ARE

XRE

Duplicated gene A, mutated to gene B

Figure 2: Hypothetical mechanism for the generation of a feed-forward loop in the absence of selection for functionality. 1. Primordial, autoregulative gene X encodes a transcription factor that binds to its own cisregulatory element, X-responsive element (XRE). 2. A second autoregulative gene A, together with its cis-regulatory region, ARE, (eg, a regulon) is inserted into the 59 region of X (eg mediated by transposable elements) in the reverse direction. If XRE is a bidirectional promoter element, XRE can contribute to transactivation of A. (This is the case in the arabinose operon, which is part of a feedforward-loop.18 ) 3. Duplication of the DNA segment Gene A-ARE-XRE creates a new gene that is responsive to both A and X. Due to subsequent diversifying mutations, Gene A becomes Gene B; ARE is partially deleted because it is redundant. Now factor X regulates itself and Gene A via XRE, while both products, A and R, regulate gene B. On top of this structural and procedural scheme, which facilitates the spontaneous creation of a feedforward loop, natural selection may then do the ne tuning, eg change strength and modality of regulation by point mutations in the respective cis or trans regulatory regions

treated as elementary phenotypic traits. In an encompassing view of evolution, one can distinguish between two fundamentally distinct, but not mutually exclusive, groups of mechanisms that can contribute to the very existence and nature of a given trait in living systems. Because authors use a large variety of terms with essentially identical or overlapping core meaning, here they will simply be called Type I and Type II mechanisms, to avoid semantic ambiguity or bias (Table 3). The Type I mechanism concerns the (neo-)Darwinian principles of selection and adaptation. Traits can (smoothly) vary in a random fashion due to mutations; and selection (in a given environment) of the ttest (of the inheritable) variants drives the features towards a(n) (local) optimum in terms of functionality. Because of the underlying connotation of directedness or nalism (even if the mutations themselves are not), the adaptation mechanism is in essence equivalent to the engineers notion of optimisation of a task by deliberate design, or the tinkerers trial and error approach. The vast majority of biologists think in categories of this mechanism to explain almost every feature. The Type II mechanism comprises a variety of alternative, ie nonadaptionist mechanisms, summarised here under the general term intrinsic constraints. The mechanisms have in common the fact that they embody a set of rules that impose constraints so as to make particular features unavoidable, ie intrinsically robust, independent of a notion of usefulness or self-maintenance in the face of external perturbation.69,70 This implies that no external force, like adaptation or any other notion of functionality or purposeful design, is required. Constraint is a creative rather than inhibitory force, for it is that which prevents (constrains) a system from settling down in a random, undifferentiated, pattern less state. The intrinsic constraints can be imposed by elementary physical laws (eg spherical shape), allometric relationships,71,72

& HENRY STEWART PUBLICATIONS 1473-9550. B R I E F I N G S I N F U N C T I O N A L G E N O M I C S A N D P R O T E O M I C S . VOL 2. NO 4. 279297. FEBRUARY 2004 2 8 5

Huang

Table 3: Major points of the two mechanisms for explaining the existence of distinct features in biology. Although the two mechanisms are not mutually exclusive, but complementary and perhaps synergistic, they have become polarised means of explanation among biologists, with the majority today leaning towards Type I
Type I mechanism: Adaptation (natural selection) (Neo)-Darwinian evolution: random mutation generates phenotypic variation, selection of a distinct feature based on tness in a given environment Type II mechanism: Intrinsic constraints Intrinsic inevitability (robustness) of a feature due to constraints in the mechanism of its genesis or design: architecture constraints physical (mechanical, energetic, allometric) constraints self-organisation, pattern formation developmental constraints historical constraints (phylogenetic inertia, genetic drift, etc) Biology can be historical or ahistorical No optimisation necessary, but can secondarily shape trait No notion of purposefulness Self-organisation used to explain unlikely trait Adaptation just maintains favourable traits that are inevitable due to constraints Natural selection is conservative Adaptation is consequence of a trait

Biology as a historical science (Local) optimisation of a trait for functionality Notion of purposefulness is central Gradual, cumulative evolution used to explain unlikely complex trait Adaptation is a powerful moulding force that generates traits from scratch Natural selection is creative Adaptation is cause for a trait

architectural69,7375 (eg symmetry, space lling) and mechanical principles (eg forcebalance rules in tensegrity76 ) and, importantly, self-organisation processes, such as symmetry-breaking patterns24 (see below). Other non-adaptive mechanisms include historical (phylogenic inertia, genetic drift, frozen accidents, etc) and developmental constraints (restrictions in the manufacturing path).69,70,75 A distinct school of thinking explicitly opposite to adaptionism is structuralism, which stresses the study of structures independent of selection for functionality.70,77 Self-organisation is part of complexity theory, which emerged as a eld of study in its own right in the 1980s24,78,79 but which still lacks a rigorous denition, its novelty and necessity is still being questioned by some scientists. (This paper will not refer to theories that address the question of macroscopic evolution and the origin of the complexity of systems itself, which is a different matter.80 ) Nevertheless, by uniting various theoretical concepts aimed at understanding the origin of ordered, ie non-random, patterns through selforganisation, the approach of the complex system theories has added substantial momentum to the community of

researchers that emphasise Type II explanations for particular features. By providing astonishingly simple explanations for the inherent robustness of many counterintuitive patterns, such as stripes, spirals, fractals, power-law relationships, etc, which are found in living as well as non-living systems,24 complexity theory may provide some of the deeper principles of organisation which embody the design constraints and which structuralists were seeking. It is important to note that the Type I and Type II mechanisms are not mutually exclusive, but are consistent with each other and can cooperate in some rather convoluted way. For example, a particular intrinsically robust structure, such as the fractal branching patterns of pulmonary bronchi81 or a network motif in molecular networks (see below), may simply have been exploited because they happen to be advantageous rather than created by natural selection. Variability and selection pressure may not have been strong enough to mould such structures from scratch by stepwise exploration of many alternative, similar but less functional forms in the phenotype space. In other words, functional adaptation may be a byproduct of an inherently robust process

2 8 6 & HENRY STEWART PUBLICATIONS 1473-9550. B R I E F I N G S I N F U N C T I O N A L G E N O M I C S A N D P R O T E O M I C S . VOL 2. NO 4. 279297. FEBRUARY 2004

Back to the biology in systems biology

Not all traits are evolved by natural selection alone

Structural constraints may have helped shape network topology features

that creates structure based on internal design principles. Selection then does the ne-tuning. Unfortunately, the above dichotomy a matter of two non-competing and overlapping explanations becomes a matter of competing, polarised points of view. The majority of modern molecular biologists gravitate towards the adaptionist perspective and put all faith in the extraordinary power of natural selection in the creation and moulding of traits. This unquestioned belief in natural selection as the core, irreducible principle of explanation in biology hinders the pluralistic endeavour to disentangle and assess the relative contribution to a trait of either mechanism. Even Darwin himself cautioned us in the introduction to The Origin of Species: I am convinced that Natural Selection has been the main, but not the exclusive, means of modication.82

EVOLUTION OF NETWORK STRUCTURES: INTRINSIC CONSTRAINTS

The dualistic scheme of explanation can now be applied to the non-random architectural features of networks. In the spirit of the prevailing bias towards adaptionism, the identication of particular network structures was almost always automatically followed by the interpretation that they must reect some optimisation of functionality by evolution.18,83,84 Such statements, although particularly attractive to the engineers mind accustomed to purposeful, optimal design, are dangerous. When invoking the survival of the ttest in the absence of rigorous analysis of the actual mechanism of the tness associated with a trait and of its history, one can easily succumb to the inherently selfreferential logic of this statement (if what is t survives, then what survives must be t). Although pure Darwinian adaptionism applies in many specic scenarios, eg in one of its purest forms the generation of antibiotic resistance, in most cases both mechanisms appear to

contribute, in a synergistic manner, to the prevalence of a given trait. Topological features of networks are ideal objects of study for determining the relative contribution of either type of mechanism to a distinct trait. The deviation from randomness of a network structure can be precisely determined and a mechanism for the structural and historical constraint formalised. A network feature represents an elementary phenotype, but unlike classical phenotypic traits it does not undergo ontogenesis, which complicates debates on evolution. Thus, developmental constraints do not have to be considered. From the discussion presented above, it becomes obvious that before attributing a feature to adaptation, one needs to exclude any constraint accounting for the inherent inevitability of that structure. The demonstration that a network feature is signicantly more abundant than one would expect if the network were to have been generated randomly, by comparing it with a random graph, does not sufce to support natural selection. Furthermore, the demonstration of convergence, ie similarity of a trait in the absence of genetic evidence for common ancestry, is often used as an argument in favour of adaptation.85 Other than dispelling the idea of a common ancestry, however, convergence does not exclude the role of various kinds of intrinsic constraints in enforcing a particular feature, but may instead point precisely to some universality of the constraining design rules. The origin of a network motif needs to be examined in more detail. The feedforward loop has been shown, based on a comparison with the random graphs, to be a prominent, hence adaptive feature in the E. coli gene regulatory network (Figure 2).18 This conclusion is problematic, since real gene regulatory networks cannot be randomly redrawn like the nodes and links in a graph. On the contrary, evolution of the genome and, hence, of the network that it encodes, is subject to a series of

& HENRY STEWART PUBLICATIONS 1473-9550. B R I E F I N G S I N F U N C T I O N A L G E N O M I C S A N D P R O T E O M I C S . VOL 2. NO 4. 279297. FEBRUARY 2004 2 8 7

Huang

Natural selection and historical/structural constraints cooperate in network evolution

mechanistic constraints that may signicantly reduce and bias the space of the possible random network architectures represented by the immaterial, random graphs. Genes are embodied by a linear physical structure, the DNA, which rearranges in particular ways. Genome evolution occurs via a nite set of physicochemical mechanisms, including chromosome and tandem gene duplication, followed by diversication of the duplicated gene pair due to mutations of the redundant genes.5355 Moreover, rearrangements of DNA are caused by DNA segment inversion or deletion, or insertion of DNA fragments via transposable elements, crossovers, etc.54,55,86 These processes, as well as promoter-generating mutations,87 drive the rewiring of gene regulatory networks by shufing cis- and trans-regulatory regions. For example, the duplication of a gene, with subsequent mutations and acquisition of new properties in the coding region but not in the promoter region,53 would create a branching network motif (one gene activates two genes). These mechanisms could facilitate the genesis of particular network motifs over others, independent of functionality. Figure 2 shows a hypothetical mechanism that could explain the genesis of a feedforward loop18 via only a gene insertion and a gene duplication event, plus a few minor mutations. Mechanistic and historical considerations help to place the potential effect of adaptation by selection in the appropriate light. Unlike the ad hoc assumption of an effective selective advantage, the assumption of mechanistic and historical constraints can be rigorously tested.75 Analysis of the DNA sequence to trace back the history of the cis- and transregulatory elements within a specic network motif, combined with the study of generative algorithms constrained by mechanistic considerations of DNA rearrangement, could reveal the contribution by Type II mechanisms to a network structure. In the case of global structures, like the

power-law degree distribution of k discussed above, Barabasi and coworkers proposed a generative algorithm for network growth (increase in node number) that inevitably results in scalefree networks.20 An essential ingredient of their algorithm is that genes newly added to the existing, growing network preferentially establish a connection to an already highly connected gene. Although the model of genesis is cast in generic terms, one can now examine whether simulations of the actual physical mechanism (increase in genome size by gene duplication, rewiring by mutations, particular exon shufing, etc) are compatible with the generic model.88 In fact, protein sequence comparisons that allowed the determination of the evolutionary age of a protein appear to suggest that the hubs are older and that preferential attachment might have taken place during evolution.89 A power-law architecture has also been found in a large number of non-biological networks, such as social and computer networks,90,91 suggesting that adaptive evolution is not likely to be the creator of this architecture. This section on the evolution of network features concludes by emphasising again that each of the types of explanation for a particular structure adaptation and intrinsic constraints does not exclude the other. Evolution will secondarily, by chance, coopt an intrinsically robust feature that happens to have desirable properties and maintain it. For example, a giant component of the network which is an unavoidable structure may coincide with functional advantages: in the case of the metabolic network, it might facilitate global coordination of energy utilisation upon change of nutritional source; and in the case of the regulatory network, global connectivity allows regulation of wholecell behaviour, such as cell fate switches, as discussed below. Similarly, structural modularity of networks, for which several advantages have been attributed in terms of functionality and evolvability,92 may

2 8 8 & HENRY STEWART PUBLICATIONS 1473-9550. B R I E F I N G S I N F U N C T I O N A L G E N O M I C S A N D P R O T E O M I C S . VOL 2. NO 4. 279297. FEBRUARY 2004

Back to the biology in systems biology

simply have been a fortuitous accident of horizontal (vector-mediated) transfer of entire clusters of genes. Such groups of genes could collectively confer a biochemical function to the host, allowing it to invade a new metabolic niche, and hence be maintained as a module by selection.87 Since the expression of these genes is coordinated inter se such as to perform a metabolic task, they necessarily form a network module. Finally, as proposed below, the power-law feature of regulatory networks may have the functional benet of generating ordered global network dynamics. Thus, adaptation is part of the game, but it may be conservative, or just doing the ne tuning, rather than creative.

therefore, necessary to move from the domain of topology to that of dynamics. Network dynamics come into play when considering two essential properties of molecular regulation networks: (i) A node i in the network (gene or protein) takes a value, xi (t), which represents its expression or activation level, which changes with time in response to incoming inuences from other nodes j, k, l, . . . , represented by the links. (ii) The links of the network represent inuences (rather than ows) and have a modality (eg inhibition or stimulation). Together with interaction rules (eg additive effects, or some Boolean functions), they determine how the level xi (t) of a given protein i (output) is jointly affected by the values x j,k,l (t) of its interaction partners j, k, l, etc (inputs). The dynamics of the system is the concerted change in the levels of xi (t) for all of the nodes i of the network, and can be represented by the N-dimensional state vector S(t) [x1 (t), x2 (t), . . . x N (t)] for a network with N nodes. While the topology is represented by a graph, the dynamics can be imagined as a more abstract N-dimensional state space in which S(t) moves its position with time t as dened by its components [x1 (t), x2 (t), . . . x N (t)] (see Figure 3 for details). The interactions of the network now impose restrictions on the movement of the state vector S(t): a given state can only move in certain directions in the state space because its components xi (t) cannot change independently but inuence each other. This frustration of the freedom of movement is where the unfathomable complexity of large networks collapses. If the wiring diagram is right, this restriction will give rise to simple, robust, higher-order behaviour, as will be seen below. Mathematically, the dynamics of the network can be described and modelled

NETWORK FUNCTION: FROM TOPOLOGY TO DYNAMICS

Most functional interpretation of networks has been based on their topology alone. For example, it has been suggested that highly connected hubs in networks are more likely to be essential genes because they interact with many proteins. In fact, for the yeast proteome there is a weak association between degree k of a protein and the probability that it is essential.12 The power-law network has been shown to be generically more tolerant to random failure of individual nodes, but vulnerable to targeted attacks on the hubs.93 This would be in line with the suggestion that hub proteins have a slower rate of evolution.94 Much of the topology-based reasoning about function rests on the unarticulated premise that the molecular network acts like a communication network in which some information ows in the links from node to node. Although this may be appropriate for metabolic reaction networks, it certainly does not apply to networks of regulation, like the protein or transcription networks, where a link represents an inuence rather than a ow. To understand functionality, it is,

& HENRY STEWART PUBLICATIONS 1473-9550. B R I E F I N G S I N F U N C T I O N A L G E N O M I C S A N D P R O T E O M I C S . VOL 2. NO 4. 279297. FEBRUARY 2004 2 8 9

Huang

TOPOLOGY N=2 B
p
1

DYNAMICS
State X

PHENOTYPIC BEHAVIOUR
State X State Y

SMALL CIRCUITS

A
State Y

low A high B

high A low B

2D phase plane

Projection as potential wells attractors = cell fates

LARGE NETWORKS

N = 24

( N-dimensional state space )

Schematic projection as 3D epigenetic landscape

Figure 3: From topology to dynamics; illustrated for a small circuit and a large network. Top panel: Small circuit (signalling module) consisting of two proteins A and B, which are mutually inhibitory, and promote their own decay. For a wide range of kinetic parameters, this topology gives rise to bistability. The middle diagram shows a phase plane (two-dimensional state space) that represents the dynamics. Each dot represents a state S(t) [x A (t), x B (t)] of the circuit at time t, dened by the expression or activity values x(t) of the two nodes A and B (see text). The ticks emanating from each dot point to the direction in state space in which that respective state will move in the next time unit to comply with the interactions (ie, systems equations) of the circuit. There are two stable states, X and Y, to which the ticks point. The right-hand diagram shows an intuitive projection of the state space (along the line p), emphasising the bistability of the system, with states X and Y as potential wells (attractors), and a hill in between representing an unstable region. Bottom panel: Analogous illustration of topology and dynamics, but for a larger network (N 24). Now the state space is N-dimensional; it contains all the network states, dened by S(t) [x1 (t), x2 (t), . . . x N (t)], and cannot be drawn, however, it can be schematically considered as a landscape in which the pits represent (high-dimensional) attractor states (right-hand diagram). In this projection, every point in the xy plane represents a network state S(t), arranged so as to generate the landscape, where the vertical axis would represent some state change potential value reecting the dissimilarity of the state Sxy at position (x,y) to that of the attractor state that it will eventually reach. Each attractor state corresponds to a distinct phenotypic cell state, the cell fate. The landscape can be thought of as the formal and molecular equivalent of Waddingtons epigenetic landscape (see Figure 4)

using a large variety of formalisms that differ in the level of coarse-graining. For example, systems of ordinary differential equations, with continuous values of concentrations for xi (t) that are subjected to mass action kinetics (like chemical reactions), are often used for smaller circuits consisting of a handful of proteins

or genes.9598 This is already an abstraction, in particular when used to model gene regulatory or signal transduction, since xi (t) represents a hypothetical average of the stochastically uctuating activity of node i, taken over time and over an ensemble of heterogeneous cells. At the even more

2 9 0 & HENRY STEWART PUBLICATIONS 1473-9550. B R I E F I N G S I N F U N C T I O N A L G E N O M I C S A N D P R O T E O M I C S . VOL 2. NO 4. 279297. FEBRUARY 2004

Back to the biology in systems biology

abstract end of the spectrum of models are the discrete networks in which genes are represented by binary, ON-OFF switches, with Boolean functions determining the interaction modalities. Because of their computational simplicity, such Boolean networks have been used as a convenient, tractable caricature for studying fundamental aspects of the global behaviour of large networks using statistical ensembles of networks of varying topologies.4,22 In fact, Boolean functions (eg AND, OR, . . .) capture the link modalities and their cooperation rather well/although these are, in reality, embodied by the molecular interactions at promoters and in proteinprotein complexes.47,99,100 Quantitative modelling of local modules of signalling and gene regulatory pathways that consider non-linear aspects (such as positive and negative feedback loops) can predict a variety of interesting behaviours, such as robustness, multistability, oscillation and switch-like behaviour.101 These emergent qualities are not evident from study of the topology, nor can they intuitively be predicted by the linear upstream downstream schemes of the experimentalist (Figure 3). This pathwaycentred modelling, however, predicts the dynamics only of the components of a particular signalling module, not of phenotypic cell behaviour.

CELL FATES AS ATTRACTORS IN GENOME-SCALE NETWORKS

The dynamics of large networks is highly constrained

Cells undergo almost discrete transitions between a few distinct cell phenotypes

A globalist view of the dynamics of the network may also, therefore, be necessary. Two ndings suggest that global network dynamics may be relevant. First, the existence of a giant component in the protein interaction network12 (and probably also in the gene regulatory network6 ) raises the question of whether there is a collective, higher-order behaviour that arises from the large number (N 1,00010,000) of interdependent genes and proteins, or

whether their activity changes simply average out and get lost over the large population of the nodes, so that no interesting global pattern emerges. In other words, is the genome-wide network (or large portions of it) wired in such a way that it acts coherently as one entity with a distinct macroscopic dynamics and, thus, contains additional information which would not be accessible in the study of pathway modules in separation? Secondly, cells in multicellular organisms exhibit a simple, coherent whole-cell behaviour which may precisely reect a higher-order dynamics of the global network: the switching between cell fates. This strictly regulated, rule-based systems behaviour is robust and remarkably simple compared with the complexity of the underlying molecular network.100 Cells in multicellular organisms are able to rapidly and reliably integrate a multitude of simultaneous and often conicting signals from their tissue environment, and respond by selecting just one of a few discrete cell fates, such as quiescence, proliferation, migration or differentiation into distinct cell types.100 Tight regulation of the balance between cell fates is required for the development and homeostasis of the tissue. Cell fates are discrete, mutually exclusive behavioural types, as Waddington described back in the 1940s.102 What is particularly counterintuitive from the localist, pathway-centred perspective, is that the same cell fate switch can be triggered by a broad variety of unrelated biochemical and physical stimuli, while the same biochemical stimulus can elicit entirely different cell fate switches depending on the context.59,100 This dees the localist notion of modular pathways that serve a specic cellular function. The mutual exclusivity of cell fates is often taken for granted but is a systems property that requires that the pathways that have been associated with particular cell fates be globally coordinated. This would be consistent with the function of a giant component in

& HENRY STEWART PUBLICATIONS 1473-9550. B R I E F I N G S I N F U N C T I O N A L G E N O M I C S A N D P R O T E O M I C S . VOL 2. NO 4. 279297. FEBRUARY 2004 2 9 1

Huang

the regulatory networks that spans almost the entire genome (even if this structural feature is statistically almost unavoidable, and thus may not have evolved for this function, as discussed above). Do the dynamics of large networks in fact generate robust, simple, macroscopic behaviour that is manifest as regulated cell fate switching? To overcome the lack of specic information on genes and the precise network topology, yet be able to study the universal behaviour of large networks, more than 30 years ago Kauffman used random Boolean networks as models to show that, given some minimal topological constraints (such as sparse connectivity), the global dynamics of a network with N genes will be far from chaotic (disordered), but that due to the constraints imposed by the particular genegene interactions, it will inevitably exhibit some ordered collective behaviour: most states S(t) are not stable, thus forcing S(t) to move in the state space until it arrives at a stable state to which unstable states will be attracted (see Figure 3 for details).3,22 These attractor states can now be designated as the cell fates or, in Kauffmans more general view, as the various cell types in a multicellular organism.22,100 In fact, the dynamics of a network with attractor states naturally captures the essential properties of cell fate dynamics, including mutual exclusivity, robustness and all-or-none transitions between cell fates in response not to a single specic instructive signal but to a large variety of signals. Although Kauffman used randomly connected networks (with restrictions only with regard to connectivity degree and type of Boolean function), he was able to show some universal dynamic properties, such as the existence of robust attractors.22 Kauffmans pioneering work receives little attention from modern day experimental and computational molecular biologists, however,34,35 who in their localist point of view emphasise the description of specic, idiosyncratic details over abstraction and analysis of higher-order behaviour. Interestingly, it

turned out that the scale-free network topology that is now found in real networks is even more likely than randomly connected networks to generate a reasonable, ordered dynamics, in the sense of Kauffman (ie containing stable, compact attractors, consistent with cell fate behaviour).103 Thus, the selforganised scale-free topology relaxes Kauffmans restrictions on gene network topology for producing ordered behaviour. The idea that attractors correspond to distinct, stable phenotypic states is just one aspect of a more general view. The genome-wide molecular network and, hence, the cell do not randomly and indifferently visit all possible network states S(t). Instead, because of the interactions between the molecular constituents, and the particularity of their wiring diagram, the state space is not isotropic, but highly sub structured, or compartmentalised, into stable attractor states and their basins of attraction. Thus, the interactions between genes and proteins are wired such as to impose restrictions on the dynamics of the network S(t). The complexity of the genomic network collapses into simple, robust behaviour with relatively few stable network states the biological cell fates. The attractors in the state space landscape (Figure 3) may thus represent a molecular and formal explanation for Waddingtons epigenetic landscape (Figure 4), which he proposed as an intuitive metaphor to explain the discreteness of cell fates more than 60 years ago, without the notion of genes and networks.

CONCLUSION
In this paper the author has discussed disparate but complementary viewpoints in biology that the burgeoning notion of system-wide networks has exposed. On the one hand, a localist view that stresses the importance of detailed pathway analysis at the level of nominal genes and proteins, and that considers a network as the sum of all the pathways; and on the

2 9 2 & HENRY STEWART PUBLICATIONS 1473-9550. B R I E F I N G S I N F U N C T I O N A L G E N O M I C S A N D P R O T E O M I C S . VOL 2. NO 4. 279297. FEBRUARY 2004

Back to the biology in systems biology

Figure 4: Epigenetic landscape proposed by Waddington in the 1940s as a metaphor to capture the observable dynamics of cell fates. The marble represents a cell state. Since Waddington was studying development, the overall slope represents the driving force of development. As the marble rolls downwards, the cell is forced to choose between discrete fates represented by the valleys. Reprinted from Waddington (1957), The Strategy of the Genes, Allen and Unwin, London.

other hand, a globalist view that emphasises the additional emergent information harboured by a network as an entity. Since living systems are both complicated and complex,27 both viewpoints are equally important. Network structures provide an ideal handle with which to determine the relative contribution of natural selection versus intrinsic constraints to particular traits. Global network dynamics, however, open a new perspective to cell fate decisions in multicellular organisms which will foster understanding of phenomena such as the multi-potency of stem cells. In conclusion, however, it is important to remember that networks are just a convenient abstraction, which fail to capture some essential properties of living systems. First, they represent a level of description that is devoid of space and physicality, such as the subcellular localisation of the proteins representing

the network nodes, as well as mechanical forces within cells.76 Secondly, in multicellular organisms, every cell contains its own individual network. One forgets that in a theoretical treatment one analyses and models a single proxy network whose behaviour is then mentally multiplied to billions of identical replicate cells in order to represent a tissue with its billions of cells. This subconscious, amplifying projection is problematic, since there is no representative average cell104 because of molecular noise105107 and other, epigenetic variations.108 Instead, a population, even of clonal cells, can dramatically vary in its network state, as evidenced by the dispersion of behaviour of the same genes in individual cells within a population.104,109 Perhaps this population heterogeneity itself is a desired feature, because the dispersion of single cell behaviour establishes robust, characteristic macroscopic response dynamics for the tissue as an entity. In this case, the detailed behaviour of a single cellular network would not matter, while sloppiness within the network of individual cells may even be benecial for the behaviour of tissues. Finally, at the tissue level, there is another network: the cellcell communication network, mediated by secreted factors, extracellular matrix and mechanical forces. Understanding intracellular molecular network is, therefore, all but a rst step towards this network of networks that makes multicellular living systems work.
Acknowledgment This work was funded by the Air Force Ofce of Scientic Research (AFOSR/USAF) and the Patterson Trust. The author also wishes to thank Joy P. Maliackal for helpful discussions on networks and Donald E. Ingber for his continuous support.

References
1. Monod, J. and Jacob, F. (1961), General conclusions: Teleonomic mechanisms in cellular metabolism, growth and differentiation, Cold Spring Harbor Symp. Quant. Biol., Vol. 26, pp. 389401.

& HENRY STEWART PUBLICATIONS 1473-9550. B R I E F I N G S I N F U N C T I O N A L G E N O M I C S A N D P R O T E O M I C S . VOL 2. NO 4. 279297. FEBRUARY 2004 2 9 3

Huang

2.

Delbruck, M. (1949), Discussion, in CNRS, Eds, Unites biologiques douees de continuite genetique, Colloques Internationaux du Centre National de la Recherche Scientique, Paris. Kauffman, S. A. (1969), Metabolic stability and epigenesis in randomly constructed genetic nets, J. Theor. Biol., Vol. 22, pp. 437467. Huang, S. (2001), Genomics, complexity and drug discovery: Insights from Boolean network models of cellular regulation, Pharmacogenomics, Vol. 2, pp. 203222. Overbeek, R., Larsen, N., Pusch, G. D. et al. (2000), WIT: integrated system for highthroughput genome sequence analysis and metabolic reconstruction, Nucleic Acids Res., Vol. 28, pp. 123125. Lee, T. I., Rinaldi, N. J., Robert, F. et al. (2002), Transcriptional regulatory networks in Saccharomyces cerevisiae, Science, Vol. 298, pp. 799804. Salgado, H., Santos-Zavaleta, A., GamaCastro, S. et al. (2001), RegulonDB (version 3.2): transcriptional regulation and operon organization in Escherichia coli K-12, Nucleic Acids Res., Vol. 29, pp. 7274. Mewes, H. W., Frishman, D., Guldener, U. et al. (2002), MIPS: A database for genomes and protein sequences, Nucleic Acids Res., Vol. 30, pp. 3134. Uetz, P., Giot, L., Cagney, G. et al. (2000), A comprehensive analysis of proteinprotein interactions in Saccharomyces cerevisiae, Nature, Vol. 403, pp. 623627. Ito, T., Chiba, T., Ozawa, R. et al. (2001), A comprehensive two-hybrid analysis to explore the yeast protein interactome, Proc. Natl. Acad. Sci. USA, Vol. 98, pp. 4569 4574. Jeong, H., Tombor, B., Albert, R. et al. (2000), The large-scale organization of metabolic networks, Nature, Vol. 407, pp. 651654. Jeong, H., Mason, S. P., Barabasi, A.-L. and Oltvai, Z. N. (2001), Lethality and centrality in protein networks, Nature, Vol. 411, pp. 4142. Wagner, A. and Fell, D. A. (2001), The small world inside large metabolic networks, Proc. R. Soc. Lond., B, Biol. Sci., Vol. 268, pp. 18031810. Maslov, S. and Sneppen, K. (2002), Specicity and stability in topology of protein networks, Science, Vol. 296, pp. 910913. Ravasz, E., Somera, A. L., Mongru, D. A. et al. (2002), Hierarchical organization of modularity in metabolic networks, Science, Vol. 297, pp. 15511555.

16.

Milo, R., Shen-Orr, S., Itzkovitz, S. et al. (2002), Network motifs: Simple building blocks of complex networks, Science, Vol. 298, pp. 824827. Wuchty, S. (2002), Interaction and domain networks of yeast, Proteomics, Vol. 2, pp. 17151723. Shen-Orr, S. S., Milo, R., Mangan, S. and Alon, U. (2002), Network motifs in the transcriptional regulation network of Escherichia coli, Nat. Genet., Vol. 31, pp. 6468. Sacks, O. (2002), Forgetting and Neglect in Science, in Silver, R. B., Ed., Hidden Histories of Science, New York Review of Books, pp. 141179. Barabasi, A. L. and Albert, R. (1999), Emergence of scaling in random networks, Science, Vol. 286, pp. 509512. Goodwin, B. (1993), How the Leopard Changed Its Spots: The Evolution of Complexity, Princeton University Press, Princeton, NJ. Kauffman, S. A. (1993), The Origins of Order, Oxford University Press, New York, NY. Kauffman, S. A. (2000), Investigations, Oxford University Press, New York, NY. Ball, P. (1998), The Self-made Tapestry: Pattern Formation in Nature, Oxford University Press, New York, NY. Coffey, D. S. (1998), Self-organization, complexity and chaos: The new biology for medicine, Nat. Med., Vol. 4, pp. 882885. Strohman, R. C. (1997), The coming Kuhnian revolution in biology, Nat. Biotechnol., Vol. 15, pp. 194200. Huang, S. (2000), The practical problems of post-genomic biology, Nat. Biotechnol., Vol. 18, pp. 471472. Oltvai, Z. N. and Barabasi, A. L. (2002), Systems biology. Lifes complexity pyramid, Science, Vol. 298, pp. 763764. Murphy, D. (2002), Gene expression studies using microarrays: Principles, problems, and prospects, Adv. Physiol. Educ., Vol. 26, pp. 256270. Mollay, M. P. and Witzmann, F. A. (2002), Proteomics: Technologies and applications, Brief. Funct. Genom. Proteom., Vol. 1, pp. 23 29. Marcotte, E. M. (2001), The path not taken, Nat. Biotechnol., Vol. 19, pp. 626627. Aebersold, R., Hood, L. E. and Watts, J. D. (2000), Equipping scientists for the new biology, Nat. Biotechnol., Vol. 18, p. 359. Ideker, T., Thorsson, V., Ranish, J. A. et al. (2002), Integrated genomic and proteomic

17.

3.

18.

4.

19.

5.

20.

6.

21.

7.

22.

8.

23. 24.

9.

25.

10.

26.

27.

11.

28.

12.

29.

13.

30.

14.

31. 32.

15.

33.

2 9 4 & HENRY STEWART PUBLICATIONS 1473-9550. B R I E F I N G S I N F U N C T I O N A L G E N O M I C S A N D P R O T E O M I C S . VOL 2. NO 4. 279297. FEBRUARY 2004

Back to the biology in systems biology

analyses of a systematically perturbed metabolic network, Science, Vol. 292, pp. 929934. 34. Endy, D. and Brent, R. (2001), Modelling cellular behaviour, Nature, Vol. 409(Suppl.), pp. 391395. Bray, D. (2003), Molecular networks: The top-down view, Science, Vol. 301, pp. 1864 1865. Nurse, P. (2003), Systems biology: Understanding cells, Nature, Vol. 424, p. 883. Stuart, J. M., Segal, E., Koller, D. and Kim. S. K. (2003), A gene-coexpression network for global discovery of conserved genetic modules, Science, Vol. 302, pp. 249255. Tornow, S. and Mewes, H. W. (2003), Functional modules by relating protein interaction networks and gene expression, Nucleic Acids Res., Vol. 31, pp. 62836289. Bader, G. D., Heilbut, A., Andrews, B. et al. (2003), Functional genomics and proteomics: Charting a multidimensional map of the yeast cell, Trends Cell. Biol., Vol. 13, pp. 344356. Ge, H., Walhout, A. J. and Vidal, M. (2003), Integrating omic information: a bridge between genomics and systems biology, Trends Genet., Vol. 19, pp. 551560. Evans, G. (2000), Designer science and the omic revolution, Nat. Biotechnol., Vol. 18, p. 127. Eisen, M. B., Spellman, P. T., Brown, P. O. and Botstein, D. (1998), Cluster analysis and display of genome-wide expression patterns, Proc. Natl. Acad. Sci. USA, Vol. 95, pp. 1486314868. Steffen, M., Petti, A., Aach, J. et al. (2002), Automated modelling of signal transduction networks, BMC Bioinformatics, Vol. 3, pp. 34. Shannon, P., Markiel, A., Ozier, O. et al. (2003), Cytoscape: a software environment for integrated models of biomolecular interaction networks, Genome Res., Vol. 13, pp. 24982504. Bader, G. D. and Hogue, C. W. (2000), BIND-a data specication for storing and describing biomolecular interactions, molecular complexes and pathways, Bioinformatics, Vol. 16, pp. 465477. Chen, Y. and Xu, D. (2003), Computational analyses of high-throughput proteinprotein interaction data, Curr. Protein Pept. Sci., Vol. 4, pp. 159181. Davidson, E. H., Rast, J. P., Oliveri, P. et al. (2002), A genomic regulatory network for development, Science, Vol. 295, pp. 1669 1678. Hasty, J., McMillen, D., Isaacs, F. and 60. 57. 55. 50. 49.

Collins, J. J. (2001), Computational studies of gene regulatory networks: in numero molecular biology, Nat. Rev. Genet., Vol. 2, pp. 268279. Dhaeseleer, P., Liang, S. and Somogyi, R. (2000), Genetic network inference: From co-expression clustering to reverse engineering, Bioinformatics, Vol. 16, pp. 707726. Gardner, T. S., di Bernardo, D., Lorenz, D. and Collins, J. J. (2003), Inferring genetic networks and identifying compound mode of action via expression proling, Science, Vol. 301, pp. 102105. Tamada, Y., Kim, S., Bannai, H. et al. (2003), Estimating gene networks from gene expression data by combining Bayesian network model with promoter element detection, Bioinformatics, Vol. 19 (Suppl. 2), pp. II227II236. Tatusov, R. L., Galperin, M. Y., Natale, D. A. and Koonin, E. V. (2000), The COG database: A tool for genome-scale analysis of protein functions and evolution, Nucleic Acids Res., Vol. 28, pp. 3336. Lynch, M., Conery, J. S. (2000), The evolutionary fate and consequences of duplicate genes, Science, Vol. 290, pp. 1151 1155. Eichler, E. E. and Sankoff, D. (2003), Structural dynamics of eukaryotic chromosome evolution, Science, Vol. 301, pp. 793797. Long, M., Deutsch, M., Wang, W. et al. (2003), Origin of new genes: Evidence from experimental and computational analyses, Genetica, Vol. 118, pp. 171182. Williams, C. S. and DuBois, R. N. (1996), Prostaglandin endoperoxide synthase: why two isoforms?, Am. J. Physiol., Vol. 270, pp. G393G400. Tanabe, T. and Tohnai, N. (2002), Cyclooxygenase isozymes and their gene structures and expression, Prostaglandins Other Lipid Mediat., Vol. 6869, pp. 95114. Kam, P. C. and See, A. U. (2000), Cyclooxygenase isoenzymes: Physiological and pharmacological role, Anaesthesia, Vol. 55, pp. 442449. Huang, S. and Ingber, D. E. (2000), Shapedependent control of cell growth, differentiation and apoptosis: Switching between attractors in cell regulatory networks, Exp. Cell Res., Vol. 261, pp. 91103. Ashburner, M. and Lewis, S. (2002), On ontologies for biologists. The Gene Ontology untangling the web, Novartis Found Symp., Vol. 247, pp. 6680. Wu, C. H., Huang, H., Arminski, L. et al.

35.

36.

37.

51.

38.

52.

39.

53.

40.

54.

41.

42.

56.

43.

44.

58.

45.

59.

46.

47.

48.

61.

& HENRY STEWART PUBLICATIONS 1473-9550. B R I E F I N G S I N F U N C T I O N A L G E N O M I C S A N D P R O T E O M I C S . VOL 2. NO 4. 279297. FEBRUARY 2004 2 9 5

Huang

(2002), The protein information resource: An integrated public resource of functional annotation of proteins, Nucleic Acids Res., Vol. 30, pp. 3537. 62. von Mering, C., Krause, R., Snel, B. et al. (2002), Comparative assessment of largescale data sets of protein-protein interactions, Nature, Vol. 417, pp. 399403. Deane, C. M., Salwinski, L., Xenarios, I. and Eisenberg, D. (2002), Protein interactions: Two methods for assessment of the reliability of throughput observations, Mol. Cell. Proteomics, Vol. 1, pp. 349356. Strogatz, S. H. (2001), Exploring complex networks, Nature, Vol. 410, pp. 268276. Erdos, P. and Renyi, A. (1959), On random graphs, Publ. Math., Vol. 6, p. 290. Callaway, D. S., Hopcroft, J. E., Kleinberg, J. M. et al. (2001), Are randomly grown graphs really random?, Phys. Rev. E Stat. Nonlin. Soft Matter Phys., Vol. 64, 041902. Joy, M. P., Brock, A. and Huang, S., manuscript in preparation. Itzkovitz, S., Milo, R., Kashtan, N. et al. (2003), Subgraphs in random networks, Phys. Rev. E Stat. Nonlin. Soft Matter Phys., Vol. 68, 026127. Gould, S. J. and Lewontin, R. C. (1979), The spandrels of San Marco and the Panglossian paradigm: A critique of the adaptationist programme, Proc. R. Soc. Lond., B, Biol. Sci., Vol. 205, pp. 581598. Webster, G. and Goodwin, B. (1981), History and structure in biology, Perspect. Biol. Med., Vol. 25, pp. 3962. Gould, S. J. (1966), Allometry and size in ontogeny and phylogeny, Biol. Rev. Camb. Philos. Soc., Vol. 41, pp. 587640. West, G. B., Brown, J. H. and Enquist, B. J. (1999), The fourth dimension of life: fractal geometry and allometric scaling of organisms, Science, Vol. 284, pp. 16771679. Thompson, D. W. (1942), Growth and Form, Macmillian, New York, NY. Seilacher, A. (1970), Arbeitskonzept zur Konstruktionsmorphologie, Lethaia, Vol. 3, pp. 393396. Autumn, K., Ryan, M. J. and Wake, D. B. (2002), Integrating historical and mechanistic biology enhances the study of adaptation, Q. Rev. Biol., Vol. 77, pp. 383408. Ingber, D. E. (2003), Tensegrity I. Cell structure and hierarchical systems biology, J. Cell Sci., Vol. 116, pp. 11571173. Hughes, A. J. and Lambert, D. M. (1984), Functionalism and structuralism, and Ways of Seeing , J. Theor. Biol., Vol. 111, pp. 787800.

78.

Mitchell Waldrop, M. (1992), Complexity. The emerging science at the edge of order and chaos, Simon and Schuster, New York, NY. Bar-Yam, Y. (1997), Dynamics in complex systems (studies in nonlinearity), Perseus Publishing New York, NY. Carlson, J. M. and Doyle, J. (2002), Complexity and robustness, Proc. Natl. Acad. Sci. USA, Vol. 99(Suppl. 1), pp. 25382545. Weibel, E. R. (1991), Fractal geometry: A design principle for living organisms, Am. J. Physiol., Vol. 1, pp. L361L369. Darwin, C. (1964), On the origin of species: A facsimile of the rst edition, Harvard University Press, Cambridge, MA. Alon, U. (2003), Biological networks: The tinkerer as an engineer, Science, Vol. 301, pp. 18661867. Wuchty, S., Oltvai, Z. N. and Barabasi, A. L. (2003), Evolutionary conservation of motif constituents in the yeast protein interaction network, Nat. Genet., Vol. 35, pp. 176179. Conant, G. C. and Wagner, A. (2003), Convergent evolution of gene circuits, Nat. Genet. Vol. 34, pp. 264266. Lawrence, J. G. (2002), Shared strategies in gene organization among prokaryotes and eukaryotes, Cell, Vol. 110, pp. 407413. Dabizzi, S., Ammannato, S. and Fani, R. (2001), Expression of horizontally transferred gene clusters: Activation by promotergenerating mutations, Res. Microbiol., Vol. 152, pp. 539549. Berg, J., Lassig, M. and Wagner, A. (2003), Structure and evolution of protein networks (in review). Available at http://arxiv.org/abs/ cond-mat/0207711. Eisenberg, E. and Levanon, E. Y. (2003), Preferential attachment in the protein network evolution, Phys. Rev. Lett., Vol. 91, 138701. Amaral, L. A., Scala, A., Barthelemy, M. and Stanley, H. E. (2000), Classes of small-world networks, Proc. Natl. Acad. Sci. USA, Vol. 97, pp. 1114911152. Barabasi, A.-L. (2002), Linked: How everything is connected to everything and what it means, Perseus Publishing New York, NY. Hartwell, L. H., Hopeld, J. J., Leibler, S. and Murray, A. W. (1999), From molecular to modular cell biology, Nature, Vol. 402(Suppl.), pp. C47C52. Albert, R., Jeong, H. and Barabasi, A.-L. (2000), Error and attack tolerance of complex networks, Nature, Vol. 406, pp. 378382. Fraser, H. B., Wall, D. P. and Hirsh, A. E.

79.

80.

63.

81.

64. 65. 66.

82.

83.

84.

67. 68.

85.

86.

69.

87.

70.

88.

71.

72.

89.

73. 74.

90.

91.

75.

92.

76.

93.

77.

94.

2 9 6 & HENRY STEWART PUBLICATIONS 1473-9550. B R I E F I N G S I N F U N C T I O N A L G E N O M I C S A N D P R O T E O M I C S . VOL 2. NO 4. 279297. FEBRUARY 2004

Back to the biology in systems biology

(2003), A simple dependence between protein evolution rate and the number of proteinprotein interactions, BMC Evol. Biol., Vol. 23, p. 11. 95. Tyson, J. J., Chen, K. and Novak, B. (2001), Network dynamics and cell physiology, Nat. Rev. Mol. Cell. Biol., Vol. 2, pp. 908916. Ferrell Jr, J. E. and Machleder, E. M. (1998), The biochemical basis of an all-or-none cell fate switch in xenopus oocytes, Science, Vol. 280, pp. 895898. Bhalla, U. S., Ram, P. T. and Iyengar, R. (2002), MAP kinase phosphatase as a locus of exibility in a mitogen-activated protein kinase signaling network, Science, Vol. 297, pp. 10181023. Hoffmann, A., Levchenko, A., Scott, M. L. and Baltimore, D. (2002), The IkappaB-NFkappaB signaling module: temporal control and selective gene activation, Science, Vol. 295, pp. 12411245. Dueber, J. E., Yeh, B. J., Chak, K. and Lim, W. A. (2003), Reprogramming control of an allosteric signaling switch through modular recombination, Science, Vol. 301, pp. 1904 1908.

(2003), Sniffers, buzzers, toggles and blinkers: Dynamics of regulatory and signaling pathways in the cell, Curr. Opin. Cell Biol., Vol. 15, pp. 221231. 102. Waddington, C. H. (1956), Principles of Embryology, Allen and Unwin Ltd, London. 103. Fox, J. J. and Hill, C. C. (2001), From topology to dynamics in biochemical networks, Chaos, Vol. 11, pp. 809815. 104. Levsky, J. M. and Singer, R. H. (2003), Gene expression and the myth of the average cell, Trends Cell Biol., Vol. 13, pp. 46. 105. McAdams, H. H. and Arkin, A. (1999), Its a noisy business! Genetic regulation at the nanomolar scale, Trends Genet., Vol. 15, pp. 6569. 106. Elowitz, M. B., Levine, A. J., Siggia, E. D. and Swain, P. S. (2002), Stochastic gene expression in a single cell, Science, Vol. 297, pp. 11831186. 107. Blake, W. J., Kaern, M., Cantor, C. R. and Collins, J. J. (2003), Noise in eukaryotic gene expression, Nature, Vol. 422, pp. 633637. 108. Spudich, J. L. and Koshland Jr, D. E. (1976), Non-genetic individuality: Chance in the single cell, Nature, Vol. 262, pp. 467471. 109. Takasuka, N., White, M. R., Wood, C. D. et al. (1998), Dynamic changes in prolactin promoter activation in individual living lactotrophic cells, Endocrinology, Vol. 139, pp. 13611368.

96.

97.

98.

99.

100. Huang, S (2002), Regulation of cellular states in mammalian cells from a genomewide view, in Collado-Vides, J. and Hofestadt, R., Eds, Gene Regulation and Metabolism: Post-Genomic Computational Approach, MIT Press, Cambridge, MA. 101. Tyson, J. J., Chen, K. C. and Novak, B.

& HENRY STEWART PUBLICATIONS 1473-9550. B R I E F I N G S I N F U N C T I O N A L G E N O M I C S A N D P R O T E O M I C S . VOL 2. NO 4. 279297. FEBRUARY 2004 2 9 7