Model Comparison in Structural Biology

点击翻译按钮获取中文摘要

DESCRIPTION (provided by applicant): X-ray crystallography has traditionally been used to generate three-dimensional structural models of biological molecules, which provide fundamental insights into biological mechanisms. The progress of refining a structural model is monitored using a powerful cross-validation statistic, R-free. However, recent advances in refinement techniques have created new classes of models that model conformational heterogeneity using ensembles or multiple conformations. There is currently a critical need to create new model selection criteria to evaluate different classes of models, as vastly different interpretations of biologically important motions can be drawn from these datasets. Bayesian model selection presents disciplined methods to determine the level of modeling detail appropriate for a given dataset. We will develop comparison techniques to rigorously trade off the quality of fit and parsimony of distinct model types. First, we will create synthetic X-ray diffraction datasets to be processed using standard data integration pipelines. Synthetic datasets afford us knowledge of the "correct" answer and allow us to vary the input conformational heterogeneity and noise. After model refinement, we will use information criteria to evaluate the tradeoffs between model complexity and parsimony. Next, we will evaluate real datasets, focusing on the refinement of high-resolution enzyme and low-resolution membrane protein data sets. We will rigorously explore the effect of global parameter grid searches on the resulting models. Finally, we will implement and distribute software that automates model comparisons. This software will be integrated into leading structure refinement and integrative modeling suites. These statistical methods will provide a general and significant improvement to the inference of protein ensembles from diverse structural data. With our research program, we will provide the structural biology community with statistically rigorous, computationally tractabl model comparison techniques integrated into existing popular software suites, and evidence for their utility. These advances will enable the exploitation of conformational heterogeneity to identify new inhibitors using in silico docking and to guide engineering of new protein functions, while avoiding futile explorations of imprecise models caused by poor data quality.