Is protein classification necessary? Towards alternative approaches to function annotation

Introduction

Protein classification schemes codify various relationships between groups of proteins that are often based on global sequence or structural similarities. However, the complex nature of evolutionary relationships between proteins raises questions about the possibility of generating a reliable classification [1]. The function of a protein is not easy to define and difficulties in describing it can occur at all levels of a classification hierarchy, even when an unambiguous sequence relationship is evident. The problem is exemplified by the so-called “moonlighting” proteins [2,3], which can play multiple roles in the cell. Such proteins have been able to acquire new functions with only minimal changes in either sequence or structure. Conversely, it has become increasingly apparent that proteins can undergo significant changes in sequence and/or structure while still maintaining the same or a similar function [4●,5●]. The picture that is emerging of protein sequence/structure/function space is one of multiple and complex relationships that in many ways defy classification.

One example of the problem involves the organization of proteins into discrete categories based on their “fold”. Structural alignments have revealed numerous geometric relationships between local fragments of proteins that have been classified, globally, as belonging to different folds. Such observations have led to the view that protein structure space is continuous rather than discrete [6–9●]. The importance of this issue is amplified by observations that proteins can have different global topologies, and hence be assigned different fold classifications, but can still share a biological function [10●]. Indeed that evolutionary relationships might exist between proteins that have been classified differently can be inferred from the very existence of so many different folds and topologies (over 1000 topologies in CATH and folds in SCOP). As has been recently pointed out, it is unlikely that each of these folds appeared independently but, rather they probably evolved from a smaller, less diverse set of ancestral proteins [11●]. This in turn suggests that there are numerous functional relationships between proteins with seemingly unrelated structure. The potential existence of such relationships implies that there is a richness of diverse information in structural databases that has yet to be uncovered. This article provides specific examples and suggests a general strategy for mining this information.

Mechanisms for the evolution of structurally diverse proteins with common functions

A conceptual problem that arises in understanding the origins of structural diversity is that most mutations are neutral in the sense that they do not usually cause structural changes while those that do would generally be expected to result in disordered structures (see [12], for example) and hence would be expected to also abrogate function. However, evidence has recently accumulated implying that the standard mechanisms of evolution, ranging from point mutations, to large deletions, insertions or rearrangements of secondary structure elements (SSEs) can result in structural diversity while maintaining function. That small changes in sequence can lead to large changes in structure has been known for some time and there have been recent reports of this phenomenon using both designed [13] and naturally occurring proteins [14,15]. At the extreme are “chameleon” sequences which can adopt multiple conformations depending on factors such as oligomeric state [16] or temperature [17].

The conservation of the sequence and structure of a functional site combined with structural diversity in other regions provides a simple mechanism to conserve function but not global structure. The α/β and all-α ferrodoxins offer an apparent example of this mechanism. Despite a structural similarity that only involves a pair of helices surrounding the functional Fe-S cluster, an evolutionary relationship between these two groups of proteins [18] is suggested both by their sequence similarity and by the existence of both types of ferredoxins as individual domains in the protein dihydropyrimidine dehydrogenase.

Insertion, deletion and rearrangement of individual domains are other mechanisms for the evolution of novel functions [19,20]. A related mechanism can occur within domains, for example the rearrangement of structural fragments that consist of a number of SSEs that play a functional role [4●,5●,9●,10●] and would also account for proteins with different global topologies and related functions. A possible objection to the idea is that significant changes within a domain would be expected to be highly destabilizing. However, recent experimental evidence suggests that this is not an issue. Graziano et al. [21] created a library of DNA elements representing random SSEs from proteins with known structure. Members of the library were randomly combined, producing several stably folded proteins one of which was highly homologous to the protein aspartate racemase from Polaromonas sp. In another example, individual members of the family of cobaltochelatases still preserve function even after significant deletions, insertions, duplications and substitutions of multiple SSE’s [22]. Finally, Peisajovich et al. [23] proposed and experimentally verified a mechanism involving circular permutation in a DNA methyltransferases which preserves function even after significant rearrangement.

Analysis of protein structures has also suggested evolutionary mechanisms by which different structures can share a common function. For example, it has been shown that proteins with different topologies can evolve from a conserved structural core [4●,5●,10●,24●]. Moreover, analyses of individual functional families [5●] and of homologous superfamilies in CATH [4●] have shown that structural changes associated with the evolutionary mechanisms summarized above are in fact quite common and often occur while still preserving overall function. An important example is provided by the evolution of heteromeric protein-protein interactions via the duplication of genes that are involved in homodimeric interactions [25]. It was found that in the evolution of heteromeric proteins, the general location of the interface seen in homodimers was conserved. This suggests that the prediction of protein-protein interactions sites may benefit from the exploitation of apparently remote structural relationships.

Classification can obscure functional relationships

The observation of functional relationships between proteins that have been classified as structurally unrelated provides some of the most striking consequences of the evolutionary mechanisms summarized in the previous section. A recent example is the identification of an evolutionary path relating the P22 and phage λ Cro proteins. Both proteins are transcription factors and both contain a DNA-binding helix-turn-helix motif but they have very different global topologies; P22 Cro is an all α protein and λ Cro is an α/β protein. The two proteins have a low level of sequence identity (25%) and could not be unambiguously related based on sequence alone. However, a relationship has been established through the identification of a series of proteins with sequence similarities (>40%) using transitive sequence searches [26●,27●]. Classifications that would place the two Cro proteins in different categories would clearly obscure the relationship between them. Related studies has been reported by Lupas and coworkers [28,29] who demonstrated an evolutionary connection between proteins in different folds. As in the Cro protein example, they identified functional relationships that involve only a few SSEs, in one case a small βαβ fragment containing conserved residues [24●].

We have recently discussed other cases where common structural fragments in proteins with different overall topologies play a similar functional role [9●]. For example, Figure 1 represents the SSE’s of three sugar-binding proteins. Each is classified as a different fold in SCOP and a different architecture in CATH (a “jelly roll”, a “β-prism” and a “β-propeller”). Nevertheless, they all contain a common substructure consisting of a 4-stranded and 3-stranded sheet with identical connectivity. Moreover, the overall locations of sugar binding sites on the surface of this substructure are also conserved [9●]. An evolutionary relationship between these proteins is also suggested by their more general biological function shown in Figure 1B. Each protein plays a similar role in distinct but related pathways: the jelly-roll facilitates viral entry into bacterial cells and the β-propeller facilitates bacterial entry into eukaryotic cells. Moreover, both of these activities are mediated through interactions with sugar-modified proteins on the cell surface. Although the specific function of the β-prism proteins has not been determined, it is known that they are involved in apoptosis [30]. Based on the relationships identified here, it seems likely that they play a role in autophagy, a particular type of apoptotic mechanism where sugar-binding is a component of the fusion of the phagosome with the lysozome.

Observations such as these suggest that the identification of remote similarities between differently classified proteins can have important practical applications, even when the structural similarity between two proteins is highly remote. For example, Xie and Bourne have identified relationships in the ligand binding properties of highly divergent proteins based on similarities in the sequence profiles of residues surrounding the binding site [31] and have used this information to identify off-targets for pharmaceuticals [32●]. What general strategies can be used to uncover the type of apparently remote relationships described in this section?

Enhanced function annotation and prediction

Structure-based function annotation often begins with a protein of unknown function but known structure, either determined experimentally or from a model. A set of structurally similar proteins is then identified and this often constitutes the end-point for protein function annotation servers, i.e. the identification of a set of possible functions derived from structural neighbors. But how are these neighbors found? If we restrict our definition of structural neighbors to proteins that have already been classified as belonging to the same fold or superfamily, we will be missing a large number of relationships that are not covered by such classifications [9●]. An alternate strategy is to identify structural neighbors, independent of classification, for example to accept the alignment of as few as three SSEs as “hits” [9●]. This will clearly significantly increase the number of possible relationships but many of these will constitute false positives so that list has to be paired down with other information.

Figure 2 provides a schematic of how this can be accomplished. It assumes that a researcher interested in a particular protein could, if interactive tools were available, pose a series of detailed queries that would enable the assignment of function based on a structural similarity. Relevant questions would include determining the “important” parts of each protein on the list and their specific functions and, conversely, given a particular function that may be known from experiment, identifying those regions of the sequence and structure that actually participate in this function. Such information can be crucial in the decision of whether homology transfer to the protein of unknown function is appropriate.

As indicated schematically in Figure 2, biophysical information combined with bioinformatics analyses of an entire set of structurally related proteins can address such questions, especially when facilitated by the availability of interactive query and visualization tools directed at widely used databases such as UniProt, the Gene Ontology (GO) and the PDB. For example in cases where there are no original functional hypotheses, sequence conservation patterns and the identification of cavities can be used in combination with UniProt sequence “features” to locate functionally important regions on a protein surface. The presence of such regions in a subset of structural neighbors should then enable a plausible function annotation. In cases where a researcher already has a functional hypothesis, the initial set of structural neighbors can be filtered based on the GO term representing that function. The reduced set of proteins generated in this way can then be further analyzed to identify the sequence and structural determinants of that function. We have implemented a server, called MarkUs (http://luna.bioc.columbia.edu/honiglab/mark-us/cgi-bin/submit.pl), that enables this type of analysis.

Although we have focused here on structure, categorizing proteins into families based on global sequence similarity is also problematic, as highlighted by the example shown in Figure 3. The figure shows a sequence alignment of a functionally important region for three members of the thioredoxin superfamily. Sequence differences unambiguously place thioredoxin and glutaredoxin-4 in different families. However, although the protein Q8L5D4 is included in the thioredoxin family (see the Pfam classification, for example), it has a CXFS motif that is characteristic of the glutaredoxin and not the thioredoxin family. In fact, it has been suggested that thioredoxin family members with the CXFS motif be reclassified as glutaredoxins [33]). This then is a case where global and local sequence similarities lead to different classifications.

The examples presented above highlight the difficulties associated with describing protein function based on the classification of proteins into specific groups. The alternative that we propose begins with the identification of possible evolutionary relationships by using sequence and/or structural alignment tools to identify related proteins independent of their classification, but then filtering the list based on more local criteria. The successful implementation of such a strategy, which should be maximally flexible and adaptable to the needs of a particular problem, will require the development of tools necessary both to generate hypotheses and to eliminate obvious false positive that are generated in the first stage.

An important step in this direction would be a more fine-grained description of function. For example, significant improvements in the accuracy of annotation transfer have been observed when GO terms are associated with specific residues [34,35] or specific structural features such as similar loop conformations [36,37], and where the validity of annotation transfer was based on the conservation of such specific features. While, as in these studies, the association between GO term and structure/sequence features was done automatically, a combination of expert-identified associations and computational tools would probably be optimal. For the proteins shown in Figure 3, a GO/residue association would both prevent the misclassification of the protein Q8L5D4 as a thioredoxin and facilitate an alternate functional hypothesis based on its local similarity to the glutaredoxins.

Summary

The difficulties associated with classification based on structural similarity exist in other areas of biology as well. For example, ambiguities in Linnaean classification of biological organisms equivalent to the type of classification used in SCOP and CATH with “fold”, “superfamily”, “family” replaced by “kingdom”, “phylum”, etc. continue to come to light as more complete genomes become available. Indeed, whole kingdoms have been removed or replaced [38]. The possibility of actually generating an unambiguous classification that represents the “tree of life” has been questioned [39] as alternatives to the traditional descent with modification for the transfer of genes between organisms have been discovered. Such complexities have led to the proposition that evolutionary relationships are more appropriately represented as a “web of life” [40].

In full analogy, we have argued here that the complex web of relationships that characterize protein structure/function space cannot be properly represented by placing proteins into discrete categories, especially when based on global structural similarity. On this basis, we are suggesting a number of changes in how function annotation is addressed. Specifically, we propose a more dynamical approach to annotation which involves the use of tools based on sequence and structure alignments that identify many possible relationships that can then be used as a starting point for more detailed analysis. This analysis should be based on local, residue-level information offered by as the integration of biophysical tools, evolutionary conservation patterns, and expanded versions of databases such as GO and UniProt. It is our opinion that the development of computational technologies that enable this overall approach will lead to a far more productive exploitation of the information contained in sequence/structure/function databases than is currently possible. The implications of such an approach, for example on the impact of structural genomics initiatives, could be enormous. Every new structure that is solved and every new homology model that is constructed would then have the potential of uncovering multiple unanticipated relationships thus playing an important role in the development of a new description of the web of relationships that characterizes “protein space”.