Investigation of non-pairwise protein structure score functions using sets of decoy structures

An essential requirement for computational protein structure prediction is a score function that discriminates native-like from non-native-like conformations. Contemporary score functions are either physics-based energy functions or knowledge-based potentials. Physics-based energy functions attempt to model the physical interactions responsible for protein stability. Knowledge-based approaches derive amino acid interaction statistics from experimentally determined protein structures and cast them into “pseudo-energies”. These latter approaches are typefied by united atom representations of residues and the assumption that the energy of a protein conformation can be computed as the summation of pair potentials. This neglect of higher order and detailed atomic interactions hinder their effectiveness. The objective of this thesis was to investigate score functions that addressed these two issues. There are five main results from this work. The first is that the geometrically regular way in which residues pack in protein structures, based on the dense packing of spheres, does not appear to be a useful score function basis. The second result is the development of a novel residue-level burial term that is sensitive to fine-grained packing density. Third is the introduction of a score function that evaluates amino acid combination preferences in residue neighborhoods and offers an alternative to pairwise methods. Fourth is the introduction of high resolution score functions for post processing the results of low resolution score functions. The final result is the observation that the widely used decoy sets are problematic and that any conclusions based on them must be treated with caution.