Artificial neural networks for dihedral angles prediction in enzyme loops: a novel approach.

Abstract

Structure prediction of proteins is considered a limiting step and determining factor in drug development and in the introduction of new therapies. Since the 3D structures of proteins determine their functionalities, prediction of dihedral angles remains an open and important problem in bioinformatics, as well as a major step in discovering tertiary structures. This work presents a method that predicts values of the dihedral angles φ and ψ for enzyme loops based on data derived from amino acid sequences. The prediction of dihedral angles is implemented through a neural network based mining mechanism. The amino acid sequence data represents 6342 enzyme loop chains with 18,882 residues. The initial neural network input was a selection of 115 features and the outputs were the predicted dihedral angles φ and ψ. The simulation results yielded a 0.64 Pearson's correlation coefficient. After feature selection through determining insignificant features, the input feature vector size was reduced to 45, while maintaining close to identical performance.