Protein co-evolution, co-adaptation and interactions.

Co-evolution has an important function in the evolution of species and it is clearly manifested in certain scenarios such as host-parasite and predator-prey interactions, symbiosis and mutualism. The extrapolation of the concepts and methodologies developed for the study of species co-evolution at the molecular level has prompted the development of a variety of computational methods able to predict protein interactions through the characteristics of co-evolution. Particularly successful have been those methods that predict interactions at the genomic level based on the detection of pairs of protein families with similar evolutionary histories (similarity of phylogenetic trees: mirrortree). Future advances in this field will require a better understanding of the molecular basis of the co-evolution of protein families. Thus, it will be important to decipher the molecular mechanisms underlying the similarity observed in phylogenetic trees of interacting proteins, distinguishing direct specific molecular interactions from other general functional constraints. In particular, it will be important to separate the effects of physical interactions within protein complexes ('co-adaptation') from other forces that, in a less specific way, can also create general patterns of co-evolution.

Introduction

Co-evolution is a well-documented phenomenon that is both an important force in the organization of biological communities and a key component of current evolutionary theory. The knowledge we have accumulated regarding the co-evolution of species is particularly relevant in the context of this review as the relationship between pairs of genes and proteins can be described in terms of co-evolution, extrapolating the concepts and methodologies developed for the study of species co-evolution to the molecular level.

The original formulation of the term co-evolution is usually attributed to Ehrlich and Raven (1964), even if the initial ideas on mutual influence between species can be traced back to Darwin's (1862) work on orchids and pollinators. Strictly defined, co-evolution is the joint evolution of ecologically interacting species (Thompson, 1994) and it implies the evolution of a species in response to selection imposed by another. In this definition, co-evolution requires the existence of mutual selective pressure on two or more species.

Ecologists have described a number of examples of co-evolution from paired species, including inter-specific competition for resources, the interaction between parasites and their hosts, the relationship between predator and prey, as well as symbiotic relationships (see for example, Moya ). In some cases, it has been possible to pinpoint morphological traits developed as a consequence of co-evolution, including direct or inverse concordances between characters (Thompson, 1994). In general, some similarity of the corresponding phylogenetic trees would be expected in these cases, for example the taxonomy of parasites and their hosts tend to be topologically similar (see for example, Stone and Hawksworth, 1985; Hafner and Nadler, 1988). However, it is important to note here that although the congruence of such trees reflects a similarity between the evolutionary processes (co-evolution) this is not sufficient evidence to demonstrate the existence of mutual influence (co-adaptation). Indeed, this resemblance does not necessarily imply that one has influenced the shape and structure of the other, or vice versa.

In general, a species evolves in response to a complex interaction with many other species. In extreme cases, when the process of co-evolution involves whole groups of species and specific examples of co-evolution between pairs of species cannot be identified, the process is called ‘diffuse co-evolution' or ‘guild co-evolution' (Thompson, 1994; Futuyma, 1997). This general ‘diffuse co-evolution' is the background process behind the constant improvement in the fitness of species (often referred to as the ‘arms race between competing species' and formulated as the ‘Red Queen Hypothesis'; Van Valen, 1973, 1977).

To study the molecular mechanisms that cause co-evolution, it is important to bring back the concept of ‘co-adaptation', which was first introduced by Dobzhansky (1950, 1970); see also Wallace (1953, 1991). The concept of co-adaptation was coined to refer to the coordination of specific changes in functional features (initially to the selective superiority of inversion heterozygotes; Ridley, 2003). In a number of cases, it has been possible to detect the adaptation of a set of genes to optimize physiological performance and reproductive success (see for example, Burton ).

Taking this concept to the molecular level, co-adaptation can be applied to direct mutual interaction between proteins, for example physical contact as part of protein complexes, that are ‘complex and that require mutually adjusted changes' following the definition of co-adaptation by Ridley mentioned above.

Here, we will use the term ‘co-evolution' to refer to the similarity of evolutionary histories, which can be quantified through the similarity of the corresponding phylogenetic trees. By way of contrast, we will use ‘co-adaptation' to refer to the molecular mechanism that would explain co-evolutionary changes by the specific influence of protein families on each other's evolutionary histories. According to our definition, co-adaptation will imply that changes in one family will influence those in the other, and vice versa, in a way that will be mostly specific for those proteins. With this definition, co-adaptation will be a mechanism that requires a direct relationship between the corresponding families (e.g. physical interaction), but it will not necessarily the only cause of co-evolution. Other factors that would have a common general influence on two proteins, without requiring interaction between them, could also influence their evolution and cause them to present co-evolutionary characteristics.

Here, we shall review the evidence for co-evolution and co-adaptation at the molecular level, indicating the practical consequences of their study on our understanding of the organization of protein interactions, and how they are exploited to predict protein interactions.

Co-evolution at the residue level

It is tempting to think that mutations at a given position in a protein are not completely independent of mutations at other positions within the same protein. The most widely studied characteristic related to concerted mutational behaviour is the so-called ‘correlated mutations' within multiple sequence alignments (MSAs).

In MSAs, homologous proteins are represented in such a way that equivalent residues are placed in the same column. Hence, a column in an MSA contains a representation of amino-acid changes permitted during evolution at that position. As functional and structural requirements impose constraints on these changes, MSAs are a rich source of structure–function information. In some cases, it is possible to observe concerted mutations at two positions (columns) in MSAs, the amino-acid changes in one position being related to those in the other. Some time ago, a weak but consistent relationship was found between this correlated mutational behaviour and spatial proximity (Göbel ; Olmea and Valencia, 1997). One hypothesis to explain such relationships states that destabilizing changes in one position can be evolutionarily fixed if they are ‘accommodated' or ‘compensated' by a modification nearby. Co-evolution between residues in the same proximity seems to have an important function in protein structure and function (Shim Choi ; Socolich ). Nevertheless, the relationship between correlated mutations (evident in MSAs) and compensatory changes (a possible explanation for these observations) has remained largely elusive. In practice, a number of variations in the specific methods to predict physical proximity based on the detection of correlations have been implemented with moderate success (see Fodor and Aldrich, 2004; Shackelford and Karplus, 2007 for systematic comparisons of methods).

Various arguments can be used to explain the difficulties in detecting signs of compensatory mutations in MSAs. For example, the presence of binding sites and active sites imposes a strong constraint on the variability of sequences that is difficult to separate from the purely structural one. The conservation of apolar residues in the protein core and the constraints imposed during folding tend to occlude possible signs of correlated changes. Furthermore, the relationship between correlated changes and physical proximity is complicated by the dependence between distant residues that cooperate in signal transmission processes (e.g. induced fit movements).

In any case, it is important to take into account that compensation can be achieved by cooperation between relatively close sets of residues organized into local structures without the need of direct physical contact between all the participating residues. This type of local compensation fits well with the co-variation model (Fitch, 1971; Shindyalov ; Susko ; Wang ). In this model, the induction of mutations can be explained in terms of the increased local capacity to accept mutations after the introduction of an initial isolated change, with no need of direct interactions between the mutated residues.

The relationship between correlated mutations and spatial proximity (not always direct contact) has not only been found between residues in the same protein but also between residues in different proteins (Pazos ; Yeang and Haussler, 2007; Burger and van Nimwegen, 2008). The hypothesis invoked to explain these inter-protein correlations is the same as that for the intra-protein ones, and involves co-adaptation between interacting proteins at the residue level. In some cases, it has been shown experimentally that compensatory changes at interfaces can indeed recover the stability of complexes lost due to an earlier mutation (Mateu and Fersht, 1999; del Alamo and Mateu, 2005). Correlated inter-protein changes seem to be more evident in obligate complexes, in which the two proteins must constantly interact to perform their biological function (Mintseris and Weng, 2005). Signs of inter-protein correlation can be used in some cases as constraints to select the arrangement of two interacting proteins or protein domains (Pazos ), or to guide protein docking experiments (Tress ), even though the corresponding residues might not enter in direct physical contact (Halperin ). Moreover, inter-protein correlated pairs can also be used to look for interaction partners (Pazos and Valencia, 2002).

Co-evolution at the protein level, similarity of phylogenetic trees

As mentioned in the Introduction, the protein feature most intuitively related to co-evolution is the similarity of the phylogenetic trees of interacting protein families. Qualitative similarities between phylogenetic trees have been observed in a number of interacting families (e.g. insulins and their receptors (Fryxell, 1996), dockerins/cohexins (Pages ) and vasopressins/vasopresin receptors (van Kesteren )). Recent studies that have quantified the relationship between tree similarities and protein interactions in large data sets (Goh ; Pazos and Valencia, 2001) have demonstrated that such similarity is not anecdotal. For example, the phylogenetic trees of the NuoE and NuoF subunits of the Escherichia coli NADH dehydrogenase complex display a clear similarity (0.86 in a 0–1 scale; in this methodology, the similarity between the phylogenetic trees is quantified indirectly as the Pearson's correlation coefficient between the sequence similarity matrices of the two families). These two subunits interact tightly as reflected in the 3D structure of the complex (PDB: 2fug; Sazanov and Hinchliffe, 2006; Figure 1). Many pairs of proteins in the flagellar machinery also co-evolve, as reflected by the similarity of their corresponding trees (Juan ). Additionally, the similarity of phylogenetic histories can be used to predict the function of hypothetical proteins. For example, the hypothetical E. coli protein YecS strongly co-evolves with the flagellar protein FliY, suggesting that it is potentially acting in the flagella machinery.

Figure 1

Co-evolution of interacting proteins. Example of two E. coli proteins that are tightly interacting (nuoE––blue and nuoF––red) and co-evolving (as reflected in the similarity of their phylogenetic trees, below). The observed co-evolution between these proteins is affected by many factors besides the co-adaptation of the two proteins, such as the interactions with other proteins in the cell (grey).

It has recently been shown that the tree similarity of interacting proteins is more evident when it is calculated for the residues that make up the actual interaction surfaces (Mintseris and Weng, 2005) or when relatively conserved regions are used rather than the full protein sequences (Kann ). Co-evolution is also evident between interacting domains to such an extent that it is possible to pinpoint the domains responsible for the interaction using domain-restricted calculations of tree similarity (Jothi ).

Obviously, the trees of any pair of protein families share a certain degree of similarity due to the underlying speciation of the host organisms. This similarity is in part related to the archetypal ‘tree of life' that represents the global evolutionary relationship of the species. Indeed correcting for this ‘background similarity' improves the performance of these methods (Pazos ; Sato ). Such background similarity can be extracted from an accepted ‘tree of life', for example that obtained from a molecular marker such as 16SrRNA or from a set of conserved genes (Pazos ; Sato ), or it can be directly inferred from the main tendencies in the data (Sato ). An additional advantage of incorporating this information about the species tree is that non-standard evolutionary events (such as horizontal gene transfer) can be detected along with the predictions of interactions. This can be achieved by looking for incongruences between the species phylogeny represented in the ‘tree of life' and that of a given protein family (Pazos ).

Because of its simple and intuitive nature, this ‘mirrortree' method (Pazos and Valencia, 2001) has been applied to many protein families (i.e., Devoto ; Labedan ; Dou ) and different variations have been developed for a variety of applications (i.e., Goh and Cohen, 2002; Gertz ; Ramani and Marcotte, 2003; Kim ; Tan ; Jothi ; Pazos ; Sato , 2006; Izarzugaza ; Tillier ; Waddell ). For example, the concept of tree similarity was used to look for the correct mapping between two families of interacting proteins, that is, to choose which ligand within a family interacts with which receptor within another family. The idea is that the correct mapping (set of relationships between the leaves of both trees) will be that which maximizes the similarity between the trees (Gertz ; Ramani and Marcotte, 2003; Jothi ; Izarzugaza ; Tillier ). Mirrortree can also be used in a ‘supervised' way by training machine learning systems with examples of interacting and non-interacting pairs, using descriptors based on the phylogenetic trees of the proteins and the species involved (Craig and Liao, 2007).

The main problem of mirrortree-like approaches is the need to construct good phylogenetic trees on a genomic scale. These are necessary to assess the similarity of all possible pairs in the search for those that are correlated. The automatic generation of reliable phylogenetic trees, with all the steps involved (orthologue detection, distance estimation, methods to generate the tree and so on) is not trivial. Thus, advances in generating reliable phylogenetic trees on a genomic scale will greatly improve this approach (Huerta-Cepas ).

Similarity of phylogenetic profiles as another case of co-evolution at the protein level

An extreme case of co-evolution involves the simultaneous loss of two interdependent proteins. One hypothesis seeks to explain this concerted disappearance of interacting proteins as ‘reductive evolution'. If the two proteins are needed to perform a given function and one of them is lost for any reason, the evolutionary pressure to maintain the other disappears as it cannot work alone. In a similar way, if one of the two proteins is ‘acquired' (i.e., horizontal gene transfer), the required partner must also be acquired. This is related to the concept of ‘selfish operons' (Lawrence, 1997), groups of related genes that are subject to concerted horizontal transfer. In practical terms, all this means that two related proteins will tend to be present in the same subset of organisms and absent in the rest. The pattern of presence/absence of a given protein (gene) in a set of genomes was termed ‘phylogenetic profile'. The similarity of phylogenetic profiles has been used extensively to detect protein functional relationships from genomic information (Gaasterland and Ragan, 1998; Marcotte ; Pellegrini ). These profiles constitute the simplest way of looking for protein co-evolution.

Initial attempts to detect protein interactions and functional relationships using ‘phylogenetic profiles' represented the distribution of proteins qualitatively, as binary vectors where ‘1' coded for the presence of the protein in an organism and ‘0' for its absence. Later, quantitative information was added by incorporating the similarity of a protein in an organism with respect to a reference organism (Date and Marcotte, 2003) into the vector positions. Once these vectors of species distributions (phylogenetic profiles) are defined, different measures of similarity can be used. It has repeatedly been shown that similar vectors are related to interactions or functional relationships between the corresponding proteins. For example, the phylogenetic profile of E. coli ribosomal protein RL7 reflects that this protein is present in most eubacterial genomes but not in archaea. Indeed, if we look for families with similar distributions, many ribosomal proteins functionally related to RL7 show up (Pellegrini ). Other examples of functionally related families with similar species distributions include flagellar proteins (which also display similar phylogenetic trees as mentioned earlier) and proteins involved in amino-acid metabolism (Pellegrini ). As with the similarity of phylogenetic trees or any other method for predicting functional relationships, the similarity of phylogenetic profiles can also be used for the ‘context-based' prediction of cellular activity. For example, the hypothetical E. coli protein YBGR has a species distribution similar to many proteins involved in amino-acid biosynthesis, supporting its function in this activity (Pellegrini ).

Gene copy number appears to be another protein feature that co-evolves, in the sense that gene expansion in one family could be ‘accommodated' by corresponding expansions in a related family, and vice versa. In this sense, ‘quantitative' phylogenetic profiles coding the number of copies of a given protein family in a set of organisms can also be used to detect functionally related families (Ranea ).

The selection of the set of organisms used to build such profiles has been shown to affect the performance of the method in predicting interactions (Sun ). Indeed, the optimal set of organisms depends on the type of functional relationship we are trying to detect, a given set of organisms being better for detecting relationships between proteins of a specific functional class (Jothi ). Incorporating evolutionary models into the methodology to differentially weight the gain/loss of genes depending on the phylogenetic context also improves performance (Zhou ; Barker ; Cokus ). As with mirrortree, phylogenetic profiles encoding the presence/absence of protein domains rather than entire proteins can also be used to detect functional relationships (Pagel ).

‘Anticorrelated' phylogenetic profiles (a protein is present when the other is absent, and vice versa) can also be informative and they have been linked to enzyme ‘displacement' in metabolic pathways (Morett ). Recently, phylogenetic profiling was extended to triplets of proteins, facilitating the search for more complicated distributions (e.g. ‘protein C is present if A is absent and B is also absent'). This allows the detection of interesting cases related to biological phenomena beyond binary functional interactions, such as complementation (Bowers ). Phylogenetic profiling also helps in structure-based functional transfer: similar structures do not ensure similar functions (Devos and Valencia, 2000, 2001). However, if the phylogenetic profiles of two structurally similar proteins are also related, the chances that the two proteins have the same function are much higher (Shakhnovich, 2005).

This powerful and intuitive approach has some disadvantages. The main one is that it can only be applied to complete genomes, as only then is it possible to be sure of the absence of a given gene. In addition, it cannot be used with essential proteins that are present in most organisms as they would produce ‘flat' profiles (‘1' in all the positions) without information to match with other profiles. Moreover, this approach is more suitable for species with a strong tendency towards genomic reduction of unnecessary genes (bacteria and archea).

The idea of functional relationships between proteins has been extended to include other features together with the co-evolution related ones discussed above, leading to the concept of ‘functional neighbourhoods' (Danchin, 2003). Apart from the two methods discussed in detail here, there are many others for the prediction of functional associations between proteins based on co-evolution and other genomic features, which are termed ‘context-based' methods (Valencia and Pazos, 2002; Shoemaker and Panchenko, 2007). At the practical level, many of these methods are available through web resources such as STRING (von Mering ), Prolinks (Bowers ) and ECID (Pazos )). As illustrated above with some examples, these methods can be used for the context-based functional transfer and, in this aspect, are orthogonal and complementary to the traditional homology-based strategy.

Co-evolution at the protein network level

Network concepts are becoming increasingly popular in molecular biology (Barabasi and Oltvai, 2004; Xia ). Some biological phenomena cannot be deduced by simply summing the properties of the molecular components, but are ‘hidden' in the complex networks representing the relationships that exist between them. In the case of protein co-evolution, it is clear that if a given protein interacts with many different partners, the changes in its amino-acid sequence (and therefore in the topology of the tree) will be a complex combination of the effects produced by the interactions with all these partners. In this sense, the full network of molecular interactions in a cell can be seen as a co-evolving system.

Recently, a new method based on the construction of the whole network of inter-protein tree similarities has been proposed (Juan ). In this case, the significance of the similarity of the trees of two protein families is evaluated in the context of the similarities to the trees of the rest of the proteome. Taking the complete co-evolutionary context into account substantially improves the detection of interacting proteins (Juan ). This procedure not only corrects the interdependence between the pairwise co-evolutions discussed above but it also corrects for other factors that influence tree similarity. These factors include the bias introduced by the underlying species phylogeny (discussed above) and methodological errors during the detection of orthologues and the construction of the trees. Additionally, the information contained in this whole network of co-evolutions enables specific co-evolutionary trends to be separated from global trends (affecting many pairs), thereby providing important information on the structure and functioning of molecular complexes. For example, the interactions between members of the flagellar machinery are detected by this method with sensitivity and specificity higher than those obtained using the pairwise similarities alone (Juan ).

The origin of co-evolution between protein families

One important question that arises is to what extent the observed co-evolution is due to direct compensatory changes in the corresponding proteins (co-adaptation) or to indirect factors that affect the sequences of both families in a similar magnitude. These include similar expression patterns, common functions in a given pathway, participation in a metabolic channelling event or collaboration in a specific cellular process.

It would make sense, and it is probably the first hypothesis that one might formulate, to think that coordinated changes in protein sequences are mechanistically related to the co-adaptation of the corresponding sequences and structures. The importance of compensatory changes can be justified in terms of maintaining the stability of protein complexes and/or the specificity of their binding to other proteins. As described above (see ‘Co-evolution at the residue level'), inter-protein compensatory changes, whereby a destabilizing mutation at the interface of one interacting partner is compensated for by a mutation in the other partner, have been found experimentally in different systems. Inter-protein compensatory mutations have also been proposed as an explanation for mutations that are pathogenic in one organism and neutral in others (Kondrashov ; Kulathinal ; Ferrer-Costa ), as well as in cases where protein families are evolving very fast while having to maintain highly specific interactions with no cross-talk (Watanabe ; Kachroo ; Liu ; Wang and Kimble, 2001; Haag ). Given that inter-protein co-adaptation at the residue level has been repeatedly observed and it has a plausible physical interpretation, it makes sense to think that the observations of co-evolution at other ‘subprotein' levels (i.e., protein regions or domains) could, to some extent, also be the result of physical compensation. As mentioned above, co-evolution has been detected between entire proteins, protein domains (Jothi ), conserved regions (Kann ) and between protein surfaces in obligate complexes (Mintseris and Weng, 2005).

Alternatively, a number of forces affecting sets of proteins and genes can generate similar evolutionary rates, such as similar expression patterns, common cellular localization and functioning in a given biochemical pathway. These external forces can create in sets of genes under common pressure signatures of co-evolution without the need for specific co-adaptation between the corresponding proteins. Families with similar evolutionary rates in different organisms would ultimately present similar trees, because the changes that occur in both families and that are responsible for shaping their trees will be of a similar magnitude. Indeed, direct (Fraser ; Hakes ) and indirect (Eisen ; Pal ; Fraser ; Subramanian and Kumar, 2004; Chen and Dokholyan, 2006; Drummond ) relationships between similar evolutionary rates and protein interactions have been found.

Therefore, even if there are indications that compensatory co-adaptive changes occur between interacting proteins and they could moderately influence the similarity of the corresponding trees, it is difficult to think that co-adaptation is the only process responsible for the observed co-evolution. It is clear that a large number of accumulated compensatory changes would be needed to affect the inter-sequence distances and hence the phylogenetic trees. In summary, it is possible that a large proportion of the observed tree similarity is due to similarities in evolutionary rates (‘diffuse co-evolution' under general selective pressure) and that specific co-adaptation (directly related to mutual effects) has a function in shaping the details of the regions of interaction.

One factor that could provide some insight into the causes of any observed co-evolution is its specificity. One would intuitively relate specific co-evolution (particular of a given pair of proteins) to co-adaptation between these proteins, whereas broader nonspecific co-evolution (‘diffuse co-evolution') involving many proteins would be more easily related to the similarity in evolutionary rates. It is even possible to think of a gradient of specificity in the factors affecting the evolution of proteins, from highly specific factors affecting only a pair of proteins to highly unspecific factors (i.e., grown temperature, osmolarity, …) which affect the whole proteome (i.e. through the differential use of codons).

Further progress in this area will require a better understanding of the co-adaptation process at the molecular level, identifying the residues/positions in the protein sequences and structures, as well as the chain of events leading to compensation and their consequences for adaptation. It has been shown that inter-protein-correlated residues are closer than the average (Pazos ; Yeang and Haussler, 2007), although co-evolution is not always evident at the protein interface itself (Hakes ). Indeed, compensation could occur even over relatively large distances through chains of interactions (i.e., allosteric effects).

The discussion of co-evolution versus co-adaptation has scientific and practical implications. The first is related to the role of natural selection in the organization of molecular networks (e.g. gene control and protein interaction networks) and to what extent co-adaptation shapes the structure and evolution of protein networks. A practical consequence is the improvement of co-evolutionary-based methods to detect protein interactions, as discussed earlier. Other practical consequences are related to the possibility of modelling protein interactions, engineering specific interactions and designing molecules to interfere with the protein–protein recognition process (i.e., in signalling pathways), which would certainly benefit from a more precise understanding of the potential physical co-adaptation between proteins in protein complexes.