Co-evolution can be defined as the joint evolution of ecologically interacting species 22 and it implies the evolution of a species in response to selection imposed by another. Co-evolution thus requires the existence of mutual selective pressure on two or more species 23 . Computational methods to predict PPIs through the characteristics of co-evolution have been developed by extrapolating concepts developed for the study of species co-evolution to the molecular level 23 24 . An in silico Two-hybrid (i2h) method has been proposed, based on the study of correlated mutations in multiple sequence alignments 25 26 . The premise is that co-adaptation of interacting proteins can be detected by the presence of a distinctive number of compensatory mutations in corresponding proteins of different species. An interaction index, defined based on the distribution of correlation values is calculated. Correlated mutations can also been used to identify specific residues involved at the interaction sites 26 . Fig. 1D illustrates how correlated mutations can be used to identify functional linkages between proteins.

Protein interactions have also been predicted on the basis of the comparison of evolutionary histories, or phylogenetic trees, under the premise that interacting proteins are subject to similar evolutionary pressures resulting in similar topologies for the corresponding trees 27 28 29 . A more recent method 30 uses the complete network of phylogenetic tree similarities between all protein pairs in the genome to reassess pairwise similarity between the phylogenetic trees of any two proteins, thereby accounting for the co-evolutionary context of the proteins more effectively.

Although homology-based methods are often quite useful for inferring PPIs, there are occasions where homology-based methods may not be effective. For example, Mika and Rost have illustrated earlier that homology-based inference of physical PPIs are accurate only at high levels of sequence identity 31 . Further, homology-based inference of PPIs work better within species than across species, for low and high levels of sequence similarity 31 .

Functional linkages may also be derived by the analysis of correlated mRNA expression levels, or protein co-expression. These techniques do not require any homology information 17 , as they rely on the measurement of additional expression data. These techniques can, therefore, find unique relationships among proteins. The premise of all expression clustering methods is that proteins do not work in isolation and are often co-expressed with functionally related proteins. By altering the conditions for performing the experiments, enough variation in gene expression can be observed to identify co-expressing genes. Protein co-expression analysis is preferable since mRNA levels and protein levels have often been found to be poorly correlated.

Gene expression data has also been shown to be useful in understanding the dynamics of PPI networks 32 33 34 . Lu and collaborators 33 integrated gene expression profiles (from a mice model of asthma) into a network of mouse PPIs derived from the BIND database. They found that highly connected proteins, or hub proteins in the network have less variable gene expression profiles compared to proteins at the network periphery. Mande and collaborators have described the construction of 'conditional networks' by integrating gene expression data under different conditions into protein functional linkage networks 34 . These networks present a picture of the dynamics of the functional linkages between proteins; a comparative analysis of four different conditional networks illustrates important responses in wild-type and mutant Escherichia coli cells treated with ultra-violet rays.

Efforts to mine experimental protein–protein association information from literature have also been made. For example, Hogue and co-workers have described an support vector machine (SVM)-based approach to mine the biomedical literature for PPIs 35 . Databases such as the STRING include such computationally mined interactions 36 . Eisenberg and co-workers have described an approach to identify abstracts that discuss PPIs from literature, which may then be manually scanned to identify PPIs 37 . This approach forms the basis for the rapid expansion of the database of interacting proteins (DIP) 37 . Zaki and collaborators have described a method based on pairwise similarity of protein sub-sequences, to predict PPIs 38 .

Although this review primarily deals with computational methods for predicting PPIs, I here briefly outline some experimental methods for assessing PPIs, for the sake of completeness. There are a number of experimental techniques such as yeast-two hybrid 39 , affinity purification/mass spectrometry 4 5 9 11 40 and protein microarrays 41 42 43 , which are reviewed in detail elsewhere 44 45 . These form the basis of several large-scale datasets on PPIs.

In the yeast-two hybrid assay, two fusion proteins are created: the 'bait' (a protein of interest with a DNA-binding domain attached to its N-terminus) and the 'prey' (its potential interaction partner, fused to an activation domain). If the 'bait' and the 'prey' interact, their binding forms a functional transcriptional activator, which in turn activates reporter genes or selectable markers 39 . This assay has been adapted for high-throughput analyses of PPIs 46 47 .

Gavin and collaborators have described the purification of complexes of 1739 proteins from S. cerevisiae (including the complete set of 1143 human orthologues) using tandem affinity purification coupled to mass spectrometry, illustrating the complexity of connectivity between protein complexes 4 . Mass spectrometry has also been used to construct a large-scale map of human protein interactions 11 .

Protein microarrays aid in the detection of in vitro binary interactions of various types — protein–protein, protein–lipid or antigen–antibody interactions. Proteins covalently attached to a solid support are screened with fluorescently labelled probes (proteins or lipids), to identify interactions 41 . A high density yeast protein microarray comprising 5800 yeast proteins was developed and used to identify novel calmodulin and phospholipid binding proteins 41 .

Although many of these assays can identify PPIs with high confidence, they still have their share of false positives and can suffer from a limited reproducibility. Nevertheless, high-throughput experimental analyses of PPIs are quite important in obtaining the protein interaction map of a cell. Further, combining results from multiple experiments as well as computational methods for predicting functional linkages (as is done in databases such as the STRING) is likely to further improve our understanding of the complex web of interactions within a cell.

STRING (Search Tool for the Retrieval of Interacting Genes/Proteins; http://string.embl.de/) 36 48 is a pre-computed database for the exploration and analysis of protein–protein associations. The associations are derived from high-throughput experimental data, mining of databases and literature, analyses of co-expressed genes and also from computational predictions, including those based on genomic context analysis. STRING employs a unique scoring framework based on benchmarks of the different types of associations against a common reference set, to produce a single confidence score per prediction. The graphical user interface is appealing and user-friendly, backed by an excellent visualisation engine. Medusa http://coot.embl.de/medusa/, a general graph visualisation tool, is a front end (interface) to the STRING protein interaction database 49 .

Human Protein Reference Database (HPRD; http://www.hprd.org/) 50 integrates information relevant to the function of human proteins in health and disease. The database is almost completely manually curated by biologists who have read and interpreted over 300,000 published articles during the annotation process. Data pertaining to thousands of PPIs, post-translational modifications, enzyme/substrate relationships, disease associations, tissue expression and sub-cellular localisation have been extracted from literature into the database.

The DIP (Database of Interacting Proteins; http://dip.doe-mbi.ucla.edu/) database 51 catalogues experimentally derived PPIs. Due to the variety of experiments and their corresponding reliabilities, DIP applies some quality assessment methods to pick out subsets of most reliable interactions. The DIP is generally considered as a valuable benchmark or verify the performance of any new method for prediction of PPIs.

The Predictome 52 database houses links between the proteins of 44 genomes based on the implementation of gene context functional linkage methods, viz. chromosomal proximity, phylogenetic profiling and domain fusion. It also contains information on large-scale experimental screenings of PPI data, from experiments such as yeast two-hybrid, immuno-co-precipitation and correlated expression. The Predictome database is presently accessible through the visual front-end provided by VisANT 53 , which is a versatile tool for visualisation and analysis of interaction data. Website http://visant.bu.edu/.

Network theory provides a quantifiable description of networks; there are several network measures that enable the comparison and characterisation of complex networks:

The understanding of the topology or the architectural principles of a biological network can directly give an insight into various network characteristics. There are several known topologies of networks, characterised by their distinctive network parameters. The following are some network models that are relevant to the understanding of biological networks.

Networks can be perturbed through the removal of nodes and edges. A typical analysis would be to probe the effect of disrupting a node and its corresponding edges. Networks of different topologies vary in their resilience to various types of perturbations. A number of studies have been carried out to analyse the response of networks to the deletion of their nodes and edges. A review of how nodes in a network can be prioritised based on network analysis has been presented elsewhere 68 .

Barabási and co-workers have analysed the response of scale-free and random networks to various types of 'attacks' 69 . In particular, they have analysed the networks representing the topologies of the Internet and the World-Wide Web. The common observation is that scale-free networks are quite insensitive to random node removals; they are highly robust in the face of random node failures and the characteristic path length was found to be almost unaffected. This is intuitively reasonable, since most of the vertices in these networks have low degree and therefore lie on few paths between others; thus their removal rarely affects communications substantially. On the other hand, directed attacks targeting the highly connected hubs led to a rapid disruption of the communication through the network. The characteristic path length was found to increase very sharply with the fraction of hubs removed and typically only a small fraction of the hubs needed to be 'knocked out' before essentially all communication through the network was destroyed 67 69 .

Jeong and co-workers have analysed the effect of node deletions on S. cerevisiae PPI network 70 . They report that although proteins with five or fewer links constituted about 93% of the total number of proteins, only about 21% of them were essential. On the other hand, only 0.7% of the proteins had more than 15 links, but single deletion of 62% of these proved lethal. This implies that highly connected proteins with a central role in the architecture of the network are three times more likely to be essential than proteins with only a small number of links to other proteins.

Another comprehensive analysis of vulnerability of complex networks to various types of attacks has been discussed in 71 . In addition to node deletions studied earlier 69 , they have also studied the effects of edge removals. Further, for each case of attacks on vertices and edges, four different attacking strategies were employed: removals by the descending order of the degree and the betweenness centrality, calculated for either the initial network or the modified network during the iterative removal procedure. They report that the removals based on the re-calculated degrees and betweenness centralities are often more harmful than the attack strategies based on the initial network's parameters, underlining the importance of the changes in network structure following the removal of important edges or nodes.

Wingender and co-workers have proposed a measure, known as pairwise disconnectivity index 72 , which quantifies how crucial a node or an edge (or a group of nodes/edges) is, for sustaining the communication between connected pairs of vertices in a directed network. This is one metric that explicitly considers paths between the various nodes in a network; it is thus quite useful in analysing how node deletions in a network can disrupt the flow of information.

We have earlier reported an analysis of the number of disrupted shortest paths in the network, to identify nodes that may be critical to a network 73 . Network analysis has also been used for identifying pathways to drug resistance 74 . Ge and collaborators have developed an 'information flow analysis', to identify proteins central for information transmission in interactome networks of S. cerevisiae and C. elegans 75 ; the proteins so identified were also likely to be essential for survival. The method employs confidence scores for PPIs and also considers multiple paths in a network while evaluating the importance of each protein 75 . The analysis of node deletions from PPI networks has been used for the identification of potential drug targets 73 76 .