From Protein Structure to Function with Bioinformatics.pdf

5 Structure and Function of Intrinsically Disordered Proteins 119Fig. 5.2 Foldindex plot of the disorder of p53. Disorder of the tumor suppressor p53 has beenpredicted by the FoldIndex algorithm (Prilusky et al. 2005). The plot is colour coded, with reddenoting predicted disorder and green denoting order, in agreement with biophysical data thatsuggest disorder within the N-terminal trans-activator domain and the C-terminal tetramerizationand regulatory domains (Bell et al. 2002; Dawson et al. 2003) (Reprinted from Prilusky et al.2005 by permission of Oxford University Press)of predicted regular secondary structure are structurally disordered (Liu and Rost2003). Whereas the predictor, NORSp, performs comparably to other disorder predictors,it should be noted that there are well-ordered proteins composed entirely ofnon-repetitive local structural elements (termed loopy proteins (Liu et al. 2002) ),and IDPs, which contain transient local structural elements (Fuxreiter et al. 2004).The tendency of these latter to form structure is well-predictable, which sets a conceptuallimit to predictions based on the above principle.5.3.5 Machine Learning AlgorithmsArguably the most advanced approaches to disorder prediction are machine learning(ML) algorithms, i.e. predictors trained to distinguished sequences that encodeordered or disordered structures. Compared to the previous simpler approaches,these incorporate non-trivial amino acid features and hidden sequence properties,which probably explains their superior performance. At the same time, their correctprediction often does not rely on known principles, and thus they do not add to ourunderstanding of what defines disorder.The classical ML algorithm is PONDR (predictor of natural disorderedregions), a neural network (NN) algorithm, which is based on local amino acidcomposition, flexibility and other sequence features (Romero et al. 1998). It has