Development and validation of statistical potential functions for the prediction of protein/nucleic-acid interactions from structure

Abstract

This work outlines the development, validation and application of a series of novel statistical (knowledge-based) potential functions to the prediction of protein/nucleic-acid interactions from structure. Three methods are described: a statistical potential for the evaluation of inter-molecular hydrogen bonds at protein/nucleic-acid interfaces; an all-atom, distance-dependent statistical potential for protein-DNA interactions, based upon the naive Bayes classifier formalism; and a similar approach, specific to the structural properties of protein-RNA interactions. These three methods are shown to be able to reliably discriminate non-native and near-native structures from native protein/nucleic-acid complexes, and are successfully demonstrated in applications to computational molecule/molecule docking (the prediction of molecular interactions from structure), rational (structure-based) protein design, and the recapitulation of experimentally determined binding energies for mutations to protein/nucleic-acid complexes. Despite their simplicity, these statistical techniques are found to be sensitive to subtle structural and chemical changes at protein/nucleic-acid interfaces, and in several cases, are demonstrated to possess performance characteristics on par with significantly more complicated, physics-based methods. These results suggest that simple, statistical potential functions can serve as a generally useful tool for the computational prediction, design and simulation of protein interactions with nucleic-acid molecules.