Evolutionary insights into protein structure.

Abstract

Proteins perform most of the specialized functions we require at a molecular level--they serve in catalysis, signal transduction, and even structural roles. Yet even after decades of study, we still do not understand many of the forces at work determining protein structure. In an attempt to answer these questions, we have focused on computational methods that make use of evolutionary information in the form of phylogenetic trees. We used phylogenetic trees (as well as multiple sequence alignments) to derive optimal structure-dependent mutation matrices, 20 x 20 matrices describing the probability of one amino acid mutating to another in a given period of evolutionary time. Mutation matrices are necessary tools in tasks such as sequence alignment, identification of homologous protein, and in another topic we investigated, the reconstruction of ancestral protein sequences by maximum-likelihood methods. With our optimal mutation matrices it was possible to reconstruct ancestral proteins, and to include structural information in the reconstruction through the use of structure-dependent matrices. The numerous examples of a protein retaining structure and function over geologic time scales motivated us to analyze the correlations of our optimal matrices with physical-chemical characteristics of the amino acids. Those characteristics most important in determining protein structure are likely to be the most conserved during evolution, as reflected in our matrices. This analysis was done for general structure-dependent matrices, as well as for matrices derived specifically for 2 other classes of proteins: antibody molecules and thermophilic proteins. Finally, we incorporated site-heterogeneity into our analysis of protein evolution using simple mathematical models of amino acid substitution. These mathematical models allow the derivation of a mutation matrix with only 10 free variables, as opposed to the hundreds normally needed to set a mutation matrix. Given the much smaller number of parameters, it was possible to derive multiple mutation matrices without an a priori division of the data set. Our simple models of amino acid substitution were validated against a data set of HIV-1 and HIV-2 proteins and also used to investigate the issue of HIV-1 subtype evolution.