From Protein Structure to Function with Bioinformatics.pdf

122 P. Tompa5.3.8 Comparison of Disorder Prediction MethodsAs already alluded to, different predictors perform at different levels, as demonstratedin the critical assessment of structure prediction algorithms experimentsCASP 6 (Jin and Dunbrack 2005) and CASP 7 (Bordoli et al. 2007). Of course, theperformance of disorder predictors depends on the criteria of evaluation, and alsoon the datasets used for evaluation. Basically, there are two factors limiting directcomparison of predictors: (1) comparisons cannot be based on simple percent valuesof predicted disorder, because the number of positive hits can be easilyincreased at the expense of false positives, i.e. predicting disorder for orderedregions; and (2) the amount of data on order and disorder differ significantly, whichis difficult to handle when prediction accuracies are simply compared. Accordingly,the performance of methods is compared by a variety of measures, basically bydefining sensitivity and specificity, i.e. the ratio of correctly predicted disorder vs.the ratio of incorrectly predicted ordered regions. As a fair assessment, one mightstate that the predictors mentioned above perform at a level approaching the bestsecondary-structure prediction algorithms. To arrive at a dependable assessment ofdisorder, it is recommended that several predictors based on different principlesshould be used.5.4 Functional Classification of IDPsPredicting function of IDPs is even more challenging than predicting their structurefor several reasons. First, IDPs evolve very fast, and even though their structuralstate as such is often preserved, there is very little information on how much theirfunctions change. Another reason is that the functional classification of proteins/genes is usually done at the level of the whole gene, and it is often very obscure inwhat way and to what extent disorder of a segment (IDR) contributes to these. Inaddition, in many cases the functions of IDPs cannot be incorporated into functionalclassification schemes developed for ordered proteins. The area of functionalclassification of IDPs witnesses immense activity, which has so far resulted in twofundamentally different approaches to classification. Key aspects of these arereviewed next.5.4.1 Gene Ontology-Based Functional Classification of IDPsIn several studies the prevalence of disorder in functional classes of proteins hasbeen addressed (Iakoucheva et al. 2002; Ward et al. 2004; Tompa et al. 2006; Xie et al.2007). These are usually based on the Gene Ontology (GO) scheme (Ashburneret al. 2000), and have addressed the prevalence of disorder in all three ontologies,