From Protein Structure to Function with Bioinformatics.pdf

274 J.D. Watson and J.M. ThorntonThe experimental elucidation of function is a highly resource intensiveprocess so, faced with a large number of structures of unknown function, oneof the major goals of modern bioinformatics is the accurate and automaticprediction of protein function. A variety of computational methods now existwhich aim to predict protein function, many of which have been discussed indetail in the previous chapters, but they effectively fall into two major categories:those which are predominantly sequence-based and those which arestructure-based.Sequence analysis is usually the first step in predicting a protein’s functionas significant sequence similarity is still the most reliable way to infer function.A number of studies have investigated this and have shown that homologousproteins sharing over 40% sequence identity are likely to have conservedfunction (Todd et al. 2001). However, care must be taken when inferring functionas there are a number of exceptions where almost identical proteins havedifferent functions, as well as those where proteins with almost undetectablesequence similarity have evolved the same function (Whisstock and Lesk2003). The development of powerful and sensitive profile- and pattern-basedmethods has increased our ability to infer functional similarities through thedetection of increasingly distant sequence relationships. Other methods developedto help gain functional clues involve looking at residue conservation,phylogenetic profiles, gene location and large scale genome organisation.When the sequence provides few clues to function or there are no detectablehomologues in the databases, a protein’s structure can be used to gain furtherinsight. As elements of a protein’s structure are often conserved for functionalreasons, structure-based approaches can identify more distant relationshipsthan sequence. The methods which have been developed range from large scalefold (Krissinel and Henrick 2004; Holm and Sander 1995) and biologicalassembly comparisons (Krissinel and Henrick 2007) (see also Chapter 6),down through localised pockets and clefts (Laskowski 1995; Glaser et al.2006; Binkowski et al. 2004) (also discussed in Chapter 7), to highly specificthree-dimensional clusters of functional residues (Laskowski et al. 2005a;Stark and Russell 2003; Kristensen et al. 2008) (see Chapter 8).No one method is 100% successful and therefore a more prudent approach isto use as many methods as possible to try to gain functional clues: the moreindependent methods that agree on the same putative function, the more likely itis to be a correct prediction. As a result a number of servers have been developedthat utilise a range of methods to try to predict function. Some of these resources,such as the ProKnow server (Pal and Eisenberg 2005), try to make an overallconsensus prediction, whereas others like the ProFunc server (Laskowski et al.2005b) present the results of a variety of methods for the user to interpret withtheir expert insight (see Chapter 10). The question that arises, however, is howsuccessful have all of the attempts to predict function from structure actuallybeen? In this chapter we shall review the various attempts to answer this question,and the difficulties encountered, with reference to case studies from structuralgenomics projects.