DESCRIPTION (provided by applicant): We propose to develop a more accurate refinement algorithm that addresses the major goals from the RFA on High Accuracy Protein Structure Modeling. Better refinement of the starting template toward the native structure is a primary step in improving the predictions from close as well as remote homologs. At this level of structure prediction, where the conformational space is limited to a single fold family, sequence-specific differences in tertiary structure determine the perturbations necessary to refine a template toward its native structure. However, current refinement methods are dominated by random searches of local backbone conformations and only consider tertiary structure (such as side-chain packing) indirectly as an outcome to these main-chain movements. As supported by the data in the Preliminary Results section, this proposal is based on the hypothesis that a more accurate refinement method needs to be driven by tertiary structure. Therefore, the major goal of this proposal is to statistically model and apply more exact descriptions of the variation in tertiary structure to improve protein structure refinement in comparative modeling. In particular, our analysis will more clearly define the contributions to protein conformation from multi-bodied, tertiary interactions versus those determined by the linear protein backbone. As a new investigator, this proposal continues my group's long-term objective of discovering the determinants of protein structure and function, and we have assembled a collaborative, multi-disciplinary team of computational biochemists and statisticians with expertise in development of algorithms modeling protein structure, Bayesian non-parametric techniques, and high performance computing. With our computational resources and environment, we will complete the following objectives framed in the three stages of our refinement algorithm. First, we will create conformationally "relaxed" starting structures that will have a higher likelihood of reaching the native state. Secondly, we will use our relative packing group construct to develop a side-chain centric, refinement move set. This move set will be incorporated into a structure build-up routine based on distance geometry. Lastly, we will derive selection algorithms that will identify near native models. By emphasizing that sequence specific variation in tertiary structure determines a protein's backbone, the proposed research represents a subtle but innovative shift in perspective to protein refinement of comparative models.