From Protein Structure to Function with Bioinformatics.pdf

2 Fold Recognition 41The idea of combining sequence and secondary structure when searching a databaseis schematically represented in Fig. 2.4b. This general idea demonstrated significantlysuperior performance over standard sequence searching and was ordersof magnitude faster to compute than most threading algorithms – a feature which isparticularly important when searching large databases of templates.In the early days of the international CASP competition, threading approacheswere generally the top performers with the hybrid sequence-structure approachesjust described following closely behind. However, two factors were to push threadingoff the top spot in the structure prediction game: (1) the explosion in the size ofthe sequence database and (2) the development of PSI-BLAST.2.3.2 Sequence Profiles and Hidden Markov ModelsAs the sequence databases were rapidly growing in size due to worldwide efforts atgenome sequencing, technological developments geared towards using this informationefficiently were underway. A simple approach by Park et al. (1997) illustratedhow two homologous sequences, which have diverged beyond the pointwhere their homology can be recognised by a simple direct comparison, can berelated through a third sequence that is suitably intermediate between the two.Known as ‘intermediate sequence search’, this ‘hopping’ through sequence spacewas clearly going to be powerful, and a more refined approach was developed inPSI-BLAST (Altschul et al. 1997). Instead of using a fixed 20 × 20 scoring matrixfor every protein, and for every position in a protein, one could construct an n × 20scoring matrix, or profile that captures the specific mutational propensities of eachposition in a specific protein sequence. For this reason such a profile is often calleda position specific scoring matrix or PSSM.After an initial standard BLAST scan to collect relatively close homologues, the(pseudo) multiple sequence alignment of these homologues to the query sequencepermits one to calculate statistics based on the observed mutations at each positionin the query sequence. These statistics form the basis of a new scoring matrix whichcan be used for a subsequent round of searching. This process of collecting homologues,building a new scoring function and searching again with this new scoringfunction can be iterated many (usually between 5 and 10) times and is calledPosition Specific Iterated BLAST (PSI-BLAST). Coupling this powerful iterativeapproach with the growing sequence database permitted a substantial improvementin the detection of extremely remote homology, and this was reflected in theFig. 2.4 (continued) a profile used for secondary structure, i.e. Each position in each sequencehas a probability for each of the three types of secondary structure. Note, in this case one wouldprobably use predicted secondary structure for the templates even though one knows the truesecondary structure. This has been shown to perform well (e.g. Bennett-Lovsey et al. 2008)