Protein Structure Prediction in the 3D HP Model Using Deep Reinforcement Learning

Simulating protein folding is a fundamental challenge in biophysics and computational biology, yet it is crucial for understanding protein structure, function, and dynamics, with significant implications for drug discovery and disease diagnosis. The Hydrophobic-Polar (HP) model serves as a simplified yet powerful framework for studying protein folding, classifying amino acids as either hydrophobic (H) or polar (P) on a lattice structure. Despite its apparent simplicity, finding optimal conformations in the HP model remains NP-complete, making it particularly challenging for larger proteins. Early approaches to this problem employed various computational methods, including genetic algorithms [Unger and Moult, 1993], Monte Carlo simulations with evolutionary algorithms [Patton et al., 1995], and memetic algorithms with self-adaptive local search [Krasnogor, 2010]. Additional methodologies encompassed ant colony optimization [Shmygelska and Hoos, 2005], core-directed chain growth [Beutler and Dill, 1996], and the pruned-enriched Rosenbluth method (PERM) [Grassberger, 1997].