Full-atom ab initio protein structure prediction with a Genetic Algorithm using a similarity-based surrogate model

The protein structure prediction problem is one of the most interesting challenges of computational biology. One of its critical facets is the optimization method employed. This is often carried out by metaheuristics, such as Genetic Algorithms (GA). The prediction involves optimization of a complex and computationally expensive energy function. Thus, the usual GA requirements of a large number of function evaluations can ultimately result in prohibitive computational costs. We applied a k-nearest neighbors surrogate modeling strategy, with two different similarity criteria, to improve the quality of proteins structures predicted by a crowding-based steady-state GA, without increasing the number of exact fitness evaluations. Additional protein conformations can be investigated using the surrogate model, potentially increasing the exploratory capability of the algorithm. The results obtained from six test proteins suggest that the surrogate model approach has the potential to improve the performance of the described protein structure prediction method.