Structural Learning for Template-free Protein Folding (1311.1422v2)

Abstract: The thesis is aimed to solve the template-free protein folding problem by tackling two important components: efficient sampling in vast conformation space, and design of knowledge-based potentials with high accuracy. We have proposed the first-order and second-order CRF-Sampler to sample structures from the continuous local dihedral angles space by modeling the lower and higher order conditional dependency between neighboring dihedral angles given the primary sequence information. A framework combining the Conditional Random Fields and the energy function is introduced to guide the local conformation sampling using long range constraints with the energy function. The relationship between the sequence profile and the local dihedral angle distribution is nonlinear. Hence we proposed the CNF-Folder to model this complex relationship by applying a novel machine learning model Conditional Neural Fields which utilizes the structural graphical model with the neural network. CRF-Samplers and CNF-Folder perform very well in CASP8 and CASP9. Further, a novel pairwise distance statistical potential (EPAD) is designed to capture the dependency of the energy profile on the positions of the interacting amino acids as well as the types of those amino acids, opposing the common assumption that this energy profile depends only on the types of amino acids. EPAD has also been successfully applied in the CASP 10 Free Modeling experiment with CNF-Folder, especially outstanding on some uncommon structured targets.