REVIEWS Structural systems biology: modelling protein interactions Patrick Aloy*‡ and Robert B. Russell‡ Abstract | Much of systems biology aims to predict the behaviour of biological systems on the basis of the set of molecules involved. Understanding the interactions between these molecules is therefore crucial to such efforts. Although many thousands of interactions are known, precise molecular details are available for only a tiny fraction of them. The difficulties that are involved in experimentally determining atomic structures for interacting proteins make predictive methods essential for progress. Structural details can ultimately turn abstract system representations into models that more accurately reflect biological reality. Structural-genomics Structural genomics Systems biology means different things to different initiatives and the advancing 1 Initiatives to solve X-ray or people . There are those who see it as a logical con- pace of structural biology mean that it is increasingly NMR structures in a high- tinuation of functional genomics — that is, carrying rare to find a single protein for which no structural throughput manner. They are out experiments on the genome scale with the aim of information is available or for which structural infor- usually focused on a single understanding how the whole is greater than the sum mation is not readily accessible by straightforward organism, pathway or disease, 6 or are aimed at providing a of its parts (see, for example, REFS 2,3). Others see it homology modelling . It is probable that a near-com- complete set of protein folds as a branch of mathematical biology (see, for example, plete structural picture will be available for most of (by solving representative REFS 4,5), which consists of the study of small systems the proteins in any given organism soon. However, structures, on the basis of for which sufficient parameters have been measured structural biology remains limited in terms of what which all other structures can be modelled). to allow simulations of how the molecules function it can deliver, and still struggles with the structures together to achieve a particular outcome. In our view, that are of the most relevance to systems biology — that Homology modelling it is both of these things. Molecular biology is no longer is, those in which two or more macromolecules are A method of protein-structure dominated by studies of single macromolecules — study- in contact. Large protein complexes or whole sys- prediction that uses a known ing pathways, complexes or even entire organisms is now tems require years of study for a detailed structural structure as a modelling template for a homologue that the norm. understanding to be reached. To address this prob- has been identified on the Genome-sequencing projects have provided a near lem, several new techniques have emerged to predict basis of sequence similarity. complete list of the components that are present in an and model the structures of interacting proteins. We organism, and post-genomic projects have aimed to review these here, and discuss how they are already catalogue the relationships between them. Systems having an impact on our understanding of complex biology is mainly about making sense of these relation- biological systems. *Institució Catalana de ships when they are considered together. For example, Recerca i Estudis Avançats understanding metabolic and signalling pathways What makes structural biology so hard? (ICREA) and Institute for or gene-regulatory networks relies on a detailed Determining the 3D structures of proteins has been Research in Biomedicine (IRB), knowledge of protein–metabolite, protein–protein and hard work since the beginning. The first X-ray struc- Parc Científic de Barcelona, protein–nucleic-acid interactions. tures took decades to solve (see, for example, REF. 7). Josep Samitier 1–5, 08028 Barcelona, Spain. A full understanding of how molecules interact However, the situation has markedly progressed, ‡European Molecular Biology comes only from three-dimensional (3D) structures, and now individual protein structures can be deter- Laboratory (EMBL), as they provide crucial atomic details about binding. mined in a matter of days when sufficient material Meyerhofstrasse 1, Knowledge of these details allows the more rational is available. Modern overexpression and purification 69117 Heidelberg, Germany. Correspondence to R.B.R. design of experiments to disrupt an interaction and procedures can usually supply sufficient material for e-mail: [email protected] therefore to perturb any system in which the inter action structural studies on a single protein, but obtain- doi:10.1038/nrm1859 is involved. ing sufficient material can be an enormous problem 188 | MARCH 2006 | VOLUME 7 www.nature.com/reviews/molcellbio © 2006 Nature Publishing Group FOCUS ON MODELLING CELLULAR SYSTEMSREVIEWS Interactome for large complexes. The reasons are fairly simple: Reassuringly, several advances are beginning to address The protein-interaction complex assembly, although not well understood, is these problems. For example, attempts to express the equivalent of the genome. something that requires precise control and timing in subunits of a complex together in various organisms have It denotes the set of the cell, and this is not easy to reproduce in a labora- shown promise10–12, and improvements in both crystalliza- interactions that occur in tory setting. Working with large complexes therefore tion techniques and synchrotron radiation facilities mean an organism. usually involves years of tinkering with systems to that smaller amounts of material can be used to solve large obtain material from natural sources and growing structures. Elsewhere, relatively new techniques such as many hundreds or thousands of litres of culture (see, cryo-electron microscopy (for example, REF. 13), which for example, REFS 8,9). This is necessary because mil- can reconstruct structures from samples at very low tem- ligram concentrations of material are often needed to peratures, can provide lower-resolution structures for large attempt the difficult task of growing crystals that will complexes using much smaller amounts of material. diffract to a high resolution — a task that is also more However, there is still a large gap between the number difficult for complexes than for individual proteins. of complexes that are thought to exist on the basis of data from, for example, two-hybrid14–18 or affinity- purification19–21 techniques and the number for which experimental 3D structures are available. And this gap is Box 1 | Uncovering protein interactions growing with the arrival of the first drafts of the human Experimental methods interactome22,23. This has essentially defined the next a b generation of structure prediction — rather than focus- Y X ing on single proteins, prediction techniques must now Z AD X Bait tackle whole complexes or systems to have the most Y 24,25 DBD impact in biology . Reporter gene W Affinity column Yeast two-hybrid Affinity purifications What and how to predict There are many prediction challenges in the protein- Computational methods interaction world. Perhaps the most obvious is simply to predict ‘who interacts with whom’ (BOX 1). The first cdGene fusion X ? Y drafts of whole-organism interactomes from high- Sp1 throughput protein-interaction approaches are still Sp1 26,27 Sp2 far from complete , and can therefore be comple- Sp2 mented by computational predictions. The late 1990s Sp3 Sp3 saw the emergence of several methods that have this Sp4 Sp4 Sp5 aim. Perhaps the best known are those that are based Sp5 Sp6 around ‘genomic context’. Here, the unifying theme is to Genomic context Co-evolution propose interactions between proteins for which there Many efforts have been undertaken to provide comprehensive lists of protein–protein is evidence of an association, because of similarities interactions and uncover the secrets behind cell networks. Experimental methods either in how they are placed relative to each other in include chemical crosslinking, chemical footprinting, protein arrays, fluorescence resonance the hundreds of known genome sequences or in their energy transfer and, more recently, fluorescence cross-correlation spectroscopy95. expression profile28–30 (BOX 1). For example, genes that lie However, the most widely used systems remain the yeast two-hybrid system and affinity in the same bacterial operon often encode proteins that purifications. The idea behind the yeast two-hybrid system is simple (see figure, part a). In are functionally associated. the most common variant, a bifunctional transcription factor (usually GAL4) is split into Functional associations that are derived from a its DNA-binding domain (DBD) and its activation domain (AD). Each segment is then genomic context do not necessarily imply a direct fused to a protein of interest (X and Y) and if these two proteins interact, the activity of physical interaction between two molecules. Proteins the transcription factor is reconstituted. The system has been scaled up and applied in at opposite ends of a single pathway or complex can genome-scale screens14–18,22,23. For affinity purification (see figure, part b), a protein of interest (bait) is tagged with a molecular label (dark purple in the figure) to allow easy give the same signal as those in tight, direct, physical purification. The tagged protein is then co-purified together with its interacting partners contact. Moreover, errors in the underlying genome or (W–Z), which