From Protein Structure to Function with Bioinformatics.pdf

46 L.A. Kelleyin ever greater detail. Profile-based approaches attempt to generate a single statisticalrepresentation for a set of related proteins – a kind of ‘average’ representative.However, this discards much of the information present in the network ofrelationships. A simple approach to retrieve some of this information was used byBateman and Finn (2007). Their method compares the output of two profilesearches and asks whether there are more sequences found in common betweenthe two outputs than expected by chance. For highly related query profiles, therewill be a large number of sequences in common. For unrelated queries, the outputswill only share sequence regions in common due to chance. This approach is analogousto investigating the first order structure of the homology network, i.e. comparingthe neighbours of one sequence to the neighbours of another. This simpleapproach was found to be highly effective at homology detection (there is noalignment generated by this method) significantly surpassing state-of-the-art profile-profilecomparison methods.Weston et al. (2004) used more of the global structure of the homology networkwith their Rankprop algorithm. The critical innovation that led to the successof the Google search engine is its ability to exploit global structure byinferring it from the local hyperlink structure of the Web. Google’s Pagerankalgorithm models the behaviour of a random web surfer who clicks on successivelinks at random and also periodically jumps to a random page. The webpages are ranked according to the probability distribution of the resulting randomwalk. The Rankprop algorithm begins with a precomputed protein similaritynetwork defined on the entire protein database. Analogous to a diffusion process,a query protein is added to the network and link information from the query toits direct sequence neighbours is propagated through the network to the neighboursof the neighbours, and so on. After propagation, database proteins areranked according to the amount of link information they received from thequery. This approach is shown to outperform standard sequence-profile searchingand is comparable to profile-profile searching, despite using PSI-BLAST togenerate the initial similarity network.Finally, Heger et al. (2008) have developed an algorithm called Maxflow whichis capable of traversing large homology networks at the level of individual residues.It searches across the network of pairwise alignments for consistently alignedpairs of residues. This method stands out from the others because of its focus onalignment generation which is critical for protein modelling.All of these new network-centric approaches are exciting developments inhomology detection. One serious drawback is the enormous computational burdenrequired to generate all versus all similarity networks. It seems clear thattheir performance would increase if it were possible to create truly completenetworks from the current databases of ∼6 million sequences. However, reduceddatabases containing sequences with