Protein Disorder Prediction using Jumping Motifs from Torsion Angles Dynamics in Ramachandran Plots

Disordered proteins are functional proteins that do not fold in a fixed 3D structure. The order/disorder prediction in protein sequences is an important task given the biological roles of disordered proteins. In the last decade many computational based methods have been proposed for the disorder identification but currently the most accurate strategies depend on the sequence alignment of large databases of proteins, making the methods slow and hard to apply on proteome-wide analysis. In this paper is proposed an innovative approach for linking the amino acid sequences with transition tendencies in their dihedral torsion angles. The aim is to characterize the dynamical angle variations along the protein chain, as a way of measuring the flexibility of the amino acids and its connection with the disorder state. The features are estimated from empirical propensities computed from Ramachandran Plots. The classification is performed using structural learning in the form of CRF (Conditional Random Fields). The performance is evaluated in terms of AUC (Area Under the ROC Curve), and three suitable performance metrics for unbalanced classification problems. The results show that the proposed method outperforms the most referenced alignment-free predictors and its performance is also competitive with the slower and mature alignment-based methods.