Applications of statistical geometry to the functional analysis of protein mutants.

摘要

With the advent of recombinant DNA and PCR techniques in the 1970s and 1980s, numerous strategies have emerged for performing site-directed and random mutagenesis. Indeed, it is now possible for laboratories to undertake experiments in which every amino acid in a protein is replaced with any of the 19 alternatives, with the goal being to characterize the role of each residue in the protein by measuring how the various amino acid replacements affect protein function. Since such experiments are expensive and time-consuming, there is growing interest in computational approaches for studying protein mutagenesis. We have developed a computational mutagenesis methodology whose underpinnings are based on the application of a Delaunay tessellation-derived four-body statistical potential function. Since the potential is derived via an ab initio approach that utilizes the atomic coordinates of non-homologous, high-resolution protein structures, the computational mutagenesis incorporates information about both sequence and structure. Using our methodology, every single or multiple mutant of a protein can be characterized by a scalar residual score, which measures the relative change in overall sequence-structure compatibility from wild-type, as well as a vector residual profile, which quantifies environmental perturbations from wild-type at every amino acid position. With a focus on proteins for which the relative activities of ample numbers of single point mutants have been experimentally determined, we illustrate how the residual scores can be used to group the amino acids of a protein into structural or functional classes, as well as to elucidate the structure-function relationship inherent in a protein. Additionally, the residual profiles of the functionally annotated mutants of a protein are used as a training set for supervised machine learning algorithms, which yield accurate inferential models of mutant activity. Finally, we successfully apply supervised learning to a training set of residual profiles associated with single and multiple mutants of HIV-1 protease, isolated and sequenced from patients enrolled in clinical trials, and for which fold-levels of resistance to inhibitor drugs are available from phenotypic assays.