A Reinforcement-Learning-Based Approach to Enhance Exhaustive Protein Loop Sampling.

Motivation: Loop portions in proteins are involved in many molecular interaction processes. They often exhibit a high degree of flexibility, which can be essential for their function. However, molecular modeling approaches usually represent loops using a single conformation. Although this conformation may correspond to a (meta-)stable state, it does not always provide a realistic representation. Results: In this paper, we propose a method to exhaustively sample the conformational space of protein loops. It exploits structural information encoded in a large library of three-residue fragments, and enforces loop-closure using a closed-form inverse kinematics solver. A novel reinforcement-learning-based approach is applied to accelerate sampling while preserving diversity. The performance of our method is showcased on benchmark datasets involving 9-, 12- and 15-residue loops. In addition, more detailed results presented for streptavidin illustrate the ability of the method to exhaustively sample the conformational space of loops presenting several meta-stable conformations.