A Peaceman-Rachford Splitting Method for the Protein Side-Chain Positioning Problem

This paper considers the NP-hard protein side-chain positioning (SCP) problem, an important final task of protein structure prediction. We formulate the SCP as an integer quadratic program and derive its doubly nonnegative (DNN) (convex) relaxation. Strict feasibility fails for this DNN relaxation. We apply facial reduction to regularize the problem. This gives rise to a natural splitting of the variables. We then use a variation of the Peaceman-Rachford splitting method to solve the DNN relaxation. The resulting relaxation and rounding procedures provide strong approximate solutions. Empirical evidence shows that almost all our instances of this NP-hard SCP problem, taken from the Protein Data Bank, are solved to provable optimality. Our large problems correspond to solving a DNN relaxation with 2,883,601 binary variables to provable optimality.