Minimizing and learning energy functions for side-chain prediction.

摘要: Side-chain prediction is an important subproblem of the general protein folding problem. Despite much progress in side-chain prediction, performance far from satisfactory. As example, ROSETTA program that uses simulated annealing to select minimum energy conformations, correctly predicts first two angles for approximately 72% buried residues a standard data set. Is further improvement more likely come better search methods, or functions? Given exact minimization NP hard, it difficult get systematic answer this question. In paper, we present novel method and learning functions training are both based on Tree Reweighted Belief Propagation (TRBP). We find TRBP can obtain global optimum function few minutes computation 85% proteins benchmark also effectively bound partition which enables using Conditional Random Fields (CRF) framework learning. Interestingly, finding does not significantly improve ROSETTA's default terms (less than 0:1%), while new weights gives significant boost 78%. Using recently modified with softer Lennard-Jones repulsive term, accuracy 77% Here again, improves modeling even 80%. Finally, highest (82.6%) obtained extended rotamer library CRF learned weights. Our results suggest combining machine approximate inference state-of-the-art prediction.