Teaching computers to fold proteins

A new general algorithm for optimization of potential functions for protein folding is introduced. It is based upon gradient optimization of the thermodynamic stability of native folds of a training set of proteins with known structure. The iterative update rule contains two thermodynamic averages which are estimated by (generalized ensemble) Monte Carlo. We test the learning algorithm on a Lennard-Jones (LJ) force field with a torsional angle degrees-of-freedom and a single-atom side-chain. In a test with 24 peptides of known structure, none folded correctly with the initial potential functions, but two-thirds came within $3\AA$ to their native fold after optimizing the potential functions.

Figure 1

(Color online) Decoy plots–energy vs RMSD $\left(\mathrm{\AA }\right)$ for two peptides at the folding temperature and below. These two peptides “cold denaturate.” Cluster centers are marked with a circle. The largest cluster (in blue) is, in all four cases, the one with the lowest RMSD.

Figure 2

(Color online) Representative decoys from the lowest free energy cluster for four peptides from top left to right (ID, RMSD from native, native structure type): 5CYT 88-101, $1.04\mathrm{\AA }$, $\alpha$-helix; 2MHR 67-78, $4.74\mathrm{\AA }$, $\alpha$-helix; 2I1B 69-82, $4.4\mathrm{\AA }$, $\beta$-turn and 2I1B 103-112, $1.63\mathrm{\AA }$, $\beta$-turn.