An ensemble machine learning approach to chemical shift prediction from protein structure.

摘要

This thesis presents a novel ensemble machine learning approach to the problem of predicting nuclear magnetic resonance (NMR) chemical shifts from three-dimensional protein structures (METASHIFT). Chemical shift prediction is a computational technique with important applications in NMR assignment, chemical shift reference checking, structure refinement and structural model evaluation amongst other domains.;A comprehensive study of the impact of the use of state-of-the-art Random Forests machine learning technique, expanded predictor variable sets and increased training and test data set sizes on structure-based chemical shift prediction accuracy is carried out. The benefit of the observed improved chemical shift prediction accuracy on a simulated NMR assignment problem is also assessed. Additionally, a framework for developing partitioned secondary structure specific shift prediction regressors is presented and justified. Potential sources of error are analyzed and the details of the practical implementation of the METASHIFT algorithm are discussed.