From Protein Structure to Function with Bioinformatics.pdf

120 P. Tompabeen developed into several variants, enabling prediction of disorder within theterminal regions of proteins (Li et al. 1999), prediction of regions likely to serveas recognition motifs (VL-XT (Iakoucheva et al. 2002) ) or a combined predictionof short and long regions of disorder (VSL2 (Peng et al. 2006) ). Becauseshort disordered regions are context dependent, i.e. their lack of structuredepends on their structural environment, whereas disorder of long regions standson its own, this combined approach results in one of the most powerful algorithmsof disorder prediction.A computationally different ML approach is the application of support vectormachines (SVMs), as exemplified by DISOPRED2 (Ward et al. 2004). This algorithmsearches for a hyperplane in a feature space that separates ordered and disorderedproteins. The hyperplane may either be linear or non-linear, and unbalancedclass frequencies of data from ordered (e.g. proteins in PDB) and disordered (e.g.proteins in DisProt (Sickmeier et al. 2007) ) proteins are taken into consideration.Sequence profiles generated by PSI-BLAST are also incorporated as an input.5.3.6 Prediction Based on Contact PotentialsSome predictors operate on a completely different principle, based on the idea thatIDPs cannot fold because their amino acids cannot make sufficient inter-residueinteractions to overcome the unfavourable decrease in entropy accompanying folding.There are several predictors based on this principle, which apply simple statisticalprinciples (FoldUnfold (Galzitskaya et al. 2006) ), compare pairwise interactionpotentials (Ucon (Schlessinger et al. 2007) ), or estimate the total inter-residueinteraction energy of a chain (IUPred (Dosztanyi et al. 2005a, b) ). This latter isdescribed in some detail.To estimate the total pair wise interaction energy realized by a polypeptidechain, IUPred uses low-resolution force fields (statistical potentials) derivedfrom globular proteins. The underlying idea is that the contribution of a residuedepends not only on its type, but also on other amino acids, i.e. its potentialpartners, in the sequence. Because a probabilistic treatment of the potentialinteractions of all residues with all others is not tractable, the problem is simplifiedby a quadratic expression in the amino acid composition. The contributionof an amino acid is approximated by an energy predictor matrix, which relatesthe energy contribution of amino acid i to that of amino acid j. The parametersof the matrix are determined by least squares fitting to actual globular proteins.By this approach, the average energy level of disordered proteins (−0.07 arbitraryunits) is significantly more unfavourable than that of globular proteins(−0.81 arbitrary units), which suggests that the approach is informative on thegross structural status of proteins (Fig. 5.3). When only a pre-defined localsequential neighbourhood is considered in the calculations, the approach providessequence-specific information on disorder, forming the basis of IUPred(Dosztanyi et al. 2005a, b).