Lattice protein design using Bayesian learning

Protein design is the inverse approach of three-dimensional (3D) structure prediction for elucidating relationship between 3D structures and amino acid sequences. In general, computation protein involves a double loop: A loop sequence changes an exhaustive conformational search each sequence. Herein, we propose novel statistical mechanical method using Bayesian learning, which can lattice proteins without search. We consider thermodynamic hypothesis evolution apply it to prior distribution Furthermore, take water effect into account in view grand canonical picture. As result, on applying 2D hydrophobic-polar (HP) model, our successfully finds target conformation has unique ground state. However, performance was not as good HP models compared models. The model improves 20-letter proteins. find strong linearity chemical potential number surface residues, thereby revealing molecules. advantage that greatly reduces time, because does require long calculations partition function corresponding uses general form learning mechanics limited proteins, results presented here elucidate some heuristics used previous methods.