From Protein Structure to Function with Bioinformatics.pdf

2 Fold Recognition 51methods (e.g. Jones 1999b; Zhang 2007). The initial dominance of threadingapproaches and their subsequent fall from favour poses interesting questions. Along-running debate in structural biology has concerned homology and analogy.Clearly many diverse sequences can adopt similar folds. Many researchers haveattributed this to the process of divergent evolution of a common ancestral sequenceunder the selection pressure for a particular structure. However, some researchershave suggested that in cases where one observes extreme differences in sequencesadopting the same fold, that these may be cases of convergent evolution – i.e. separateevolution towards the same fold in the absence of a common ancestor. This isakin to the convergent evolution of the wing of a bat and that of a bird.There are clear examples of convergent evolution in proteins where similar localstructures have evolved multiple times. Probably the best-known example is that ofthe Ser/His/Asp catalytic triad (Dodson and Wlodawer 1998) which is found in atleast five different protein folds that cannot easily be considered to be homologous.Evidence such as this in favour of convergent evolution suggests that threadingapproaches may succeed where sequence-profile-based approaches would fail. Andyet the use of threading appears to be steadily diminishing; instead being supplantedby sequence/profile methods.There may be several reasons for this. Firstly, it remains an open questionwhether whole protein folds, rather than localized substructures, have arisen multipletimes during evolution. Nature may have stumbled upon simple folds such asfour helical bundles multiple times, but for more complicated structures this is difficultto demonstrate definitively. Several folds, such as TIM-barrels and beta trefoilswhich were previously thought to be examples of convergence are increasinglybeing revealed as homologues due to enhanced sensitivity in sequence comparisontechniques (Copley and Bork 2000; Ponting and Russell 2000).Secondly, even if true analogues exist it may be that sequence-based methodscan detect them because of common biophysical preferences required for a givenfold, which in turn will be reflected in a well constructed profile from many distanthomologous sequences. Thirdly, the widely accepted measure of success in proteinstructure prediction is CASP. An unfortunate side-effect of the pre-eminence of theCASP competition is that methods that can detect analogous protein fold relationshipswill tend to be obscured by those methods that accurately align proteinsequences to closer homologues in the ever-growing protein structure database.If a homologous relationship can be detected in the structure databases, it willinvariably provide a superior model than an analogous relationship. That is, as thesequence and structure databases grow, the requirement for analogy detectiondecreases as (a) ones ability to search further in sequence space increases, and(b) there are more and closer structural templates to choose from.This leads towards the conclusion that simply solving the structures of a smallnumber of well-chosen proteins (Marsden et al. 2007) would enable the relativelyaccurate modelling of a large proportion of genomic sequences. From the perspectiveof designing novel folds or ab initio fold prediction this is unsatisfactory. Butfor an effective technology for assigning structure to genomes this is fine, as longas we have a sufficient number of well-chosen structures.