From Protein Structure to Function with Bioinformatics.pdf

10 Integrated Servers for Structure-Informed Function Prediction 253Plus there are various spanners that gum up the works. Firstly, it is often difficultto solve the whole intact protein. In these cases the best one can get is a structuralmodel for part of the protein – say, just a single domain. On its own this domainmay say little about the protein’s function. Secondly, even if the whole protein issolved, it may be just one component of a multi-protein complex. Again, the structuregives only part of the story. More dastardly still are the so-called moonlightingproteins which can actually have more than one function, depending on their context:cellular location, environment, and so on (Jeffery 1999). And some proteinscan alter their function according to which alternatively spliced variant is expressedat any given time (Stamm et al. 2005).Another problem with function prediction is the difficulty of assessing the successor failure of a given prediction method and, indeed, even defining what ismeant by function. Function can be described at many levels, ranging from biochemicalfunction through biological processes and pathways, all the way up tothe organ or organism level (Shrager 2003). Consequently, a given protein may beannotated at several different levels of functional specificity: for example, ubiquitinlikedomain, signalling protein, predicted serine hydrolase, probable eukaryoticD-amino acid tRNA deacetylase, and so on. Thus it is difficult to judge theaccuracy of any such assignment, especially if the assignment is one of the morevague ones.A common strategy for assessing function prediction methods is to use the GeneOntology (GO) (The Gene Ontology Consortium 2000; Camon et al. 2004). This isan open source scheme for functional annotation of protein sequences. It is amachine-readable ontology based on a controlled vocabulary of functional descriptorsand many function prediction methods couch their results in terms of GOcodes. Although not strictly hierarchical, the GO functional descriptors range fromthe truly unspecific (e.g. enzyme) down to the highly precise (e.g. 1-pyrroline-4-hydroxy-2-carboxylate deaminase).10.1.2 Structure-Function Prediction MethodsAs the previous chapters show, there are very many different methods for predictingfunction from structure. Several reviews have described them and considered theirusefulness (Kim et al. 2003; Watson et al. 2005; Rigden 2006). None of the methodsis perfect and none can hope to be successful in all situations. For example,some methods are only suitable for enzymes – and so cannot help at all if the proteinin question is not an enzyme. Other methods rely strongly on some match –whether of the fold, or a motif, or a binding site, etc. – to a protein of knownstructure. So if no match can be found, or the match is merely to another hypotheticalprotein, such a method effectively returns a blank.Consequently, a sensible approach is to throw a large number of these predictivemethods at the protein structure and see what drops out. The two servers describedin this chapter do just that. They are ProKnow from UCLA at http://proknow.mbi.