Abstract

Successful application of physics-based protein-structure prediction methods depends on sophisticated computational approaches to global optimization of the conformational energy of a polypeptide chain. One of the most effective procedures for the global optimization of protein structures appears to be the Conformational Space Annealing (CSA) method. CSA is a hybrid method which combines genetic algorithms, essential aspects of the build-up method and a local gradient-based minimization. CSA evolves the population of conformations through genetic operators (mutations, i.e. perturbations of selected geometric parameters, and crossovers, i.e. exchange of selected subsets of geometric parameters between conformations) to a final population optimizing their conformational energy. Implementation of the CSA method with the united-residue force field (UNRES, in which each amino-acid residue is represented by two interaction sites, namely the united peptide group and the united side-chain) was enhanced by introducing new crossover operations consisting of (i) copying -hairpins, (ii) copying remote strand pairs forming non-local -sheets, and (iii) copying -helical segments. A mutation operation, which shifts the position of a -turn, was also introduced. The new operations promote -structure, and are essential for searching the conformational space of proteins containing both - and -structure; without these operations, excessive preference of -helical structures is obtained, even though these structures are high in energy. Parallelization of the CSA method has also been enhanced by removing most of the synchronization steps; the improved algorithm scales almost linearly up to 1,000 processors with over 75% average performance.

Author Keywords: Protein structure prediction; Global optimization; Conformational space annealing