A hybrid approach mixing local search and constraint programming applied to the protein structure prediction problem

We present a hybrid system that combines local search techniques and constraint solving. We apply it to the ab-initio protein structure prediction problem, modeled in the Face Centered Cubic Lattice with a pairwise contact energy function. In the literature, the problem is successfully solved using constraint programming for proteins with length up to 160 using the HP energy model. In the case of more complex models, w.r.t. energy and structure, current techniques can not be easily extended and the constraint approach is not applicable to proteins of length over 100. The idea described in this paper is based on the alternation of CSP solving phases and local search phases that modify the predicted spatial conformation. The approach is implemented and tested in Gecode and EasyLocal with encouraging results.