Protein folding classes -- High-dimensional geometry of amino acid composition space revisited
Abstract
In this study, the distributions of protein folding classes of experimentally determined structures from a legacy dataset and a comprehensive database SCOPe are modeled with precise geometric objects as convex polytopes in the high-dimensional amino acid composition space. This is a follow-up of a previous non-statistical, geometry-motivated modeling of protein classes with ellipsoidal models, which are superseded presently in three important respects: (1) as a paradigm shift descriptive 'distribution models' of experimental data are de-coupled from, and serves as the basis for, future potential predictive 'domain models' generalizable to proteins in the same structure class for which the 3-dimensional structures have yet to be determined experimentally, (2) the geometric and analytic characteristics of class distributions are obtained via exact computational geometry calculations, and (3) the full data from a comprehensive database are included in such calculations, eschewing training set selection and biases. In contrast to statistical inference and machine-learning approaches, the analytical, non-statistical geometry models of protein class distributions demonstrated in this study, complete with precise information on the size of class distributions and their relative disposition in the high-dimensional space, and intended not principally for prediction but as accurate description of the complex relationships between amino acid composition and protein classes, suggest that they may ultimately be useful adjuncts for validating sequence-based methods for predicting protein structures and contribute to the mechanistic understanding of secondary structure formation and the folding of polypeptide chains into 3-dimensional conformation of functional protein molecules.