Accurate prediction of docked protein structure similarity using neural networks and restricted Boltzmann machines

One of the major challenges for protein-protein docking is to accurately discriminate native-like structures from false-positives. While there is an agreement on the existence of a relationship between various favorable intermolecular interactions (e.g., Van der Waals, electrostatic, desolvation forces, etc.) and the similarity of a conformation to its native structure, the exact nature of this relationship is not clear. Different docking algorithms often formulate this relationship as a weighted sum of selected terms and calibrate their weights against a training set to evaluate and rank candidate complexes. Despite improvement in the predictive abilities of recent docking methods, even state-of-the-art methods often fail to predict the binding of many complexes and still output a large number of false positive complexes. We propose a novel machine learning approach that not only ranks candidate structures relative to each other, but also predicts how similar each candidate is to the native conformation. We trained a two-layer neural network, a deep neural network and a network of Restricted Boltzmann Machines against extensive datasets of unbound complexes. We tested these methods with a set of candidate structures. Our method is able to predict the RMSDs of unbound docked complexes with a very small, often <; 1.5Å error margin.