From Protein Structure to Function with Bioinformatics.pdf

296 I.A. Cymerman et al.and improvement of computational methods accuracy results in the possibility ofproducing an increasing number of proteins models. According to recent estimates,the novel structures solved by the PSI initiative allow the creation of about 40,000homology models that could not otherwise be made (Service 2008b). To allow fullexploitation of this opportunity, however, these models should be freely available tobiologists, along with information about their reliability.12.2.1 Model QualityThe usefulness of the protein structure for function prediction is dictated by itsquality. Most of the experimentally solved structures represent the data with thesufficient resolution to infer their chemical mode of action but the same can notnecessarily be said of protein models. As for structure, the quality of the modeldetermines the functional information that can be deduced from it. For example,models of moderate quality are of limited use for applications such as computationaldocking of drug-like ligands. As mentioned earlier, the accuracy of homologymodels drops with the decrease of sequence similarity between the templateand the target (see also Chapter 3), so one could use a rule of the thumb that modelsbased on closely related templates are usually “good”, while those based on remotetemplates are usually “bad”. This observation, however, assumes that homologymodels are based on perfect alignments, and that no optimization is carried out tomodify the starting conformation in the direction of the true structure. Besides, thedegree of relationship to a template cannot be exploited to assess models built bytemplate-free (ab initio or de novo) modelling methods (Chapter 1), because theydo not use the templates at all.Recently, a number of model quality assessment programs (MQAPs) have beendeveloped to predict the quality of individual theoretical models, without knowledgeof the native structures. Some of these methods (e.g. PROQ (Wallner andElofsson 2003) ) specialize in discrimination between native, near-native and nonnativestructures, while others (e.g. PROQres (Wallner and Elofsson 2006),ModFold (McGuffin 2008), or MetaMQAP (Pawlowski et al. 2008)) focus on predictinghow far individual parts of the model may deviate from their counterpartsin the native structure. At present there is no universal score that can confidentlyassess global or local accuracy of structural models. However, analyses of proteinfunction based on a theoretical model should take into account the predicted globalquality, if global features are considered (such as charge, surface shape etc.), andthe predicted local quality for consideration of active sites and other particularregions. Here, particularly useful are these methods that directly predict the deviationof different parts of a given model from the unknown native structure.Consideration of their predictions may at least help to avoid over-interpretation ofthe models. Obviously there is no point in analyzing the geometry of interactionsbetween potential catalytic side chains, if the backbone of the corresponding residuesis predicted to within only 5 Å.