Protein Structure Parameterization via Mobius Distributions on the Torus

TL;DR

This paper introduces new statistical models for protein backbone torsion angles using Mobius transformations on the torus, enabling better analysis and prediction of protein structures.

Contribution

It proposes novel Mobius distribution models for toroidal data, including marginal, conditional, and sine-skewed variants, applied to protein structure data.

Findings

Models effectively fit large protein datasets

Simulation confirms accurate maximum likelihood estimates

Proposed methods improve protein 3D structure prediction

Abstract

Proteins constitute a large group of macromolecules with a multitude of functions for all living organisms. Proteins achieve this by adopting distinct three-dimensional structures encoded by the sequence of their constituent amino acids in one or more polypeptides. In this paper, the statistical modelling of the protein backbone torsion angles is considered. Two new distributions are proposed for toroidal data by applying the M\"obius transformation to the bivariate von Mises distribution. Marginal and conditional distributions in addition to sine-skewed versions of the proposed models are also developed. Three big data sets consisting of bivariate information about protein domains are analysed to illustrate the strength of the flexible proposed models. Finally, a simulation study is done to evaluate the obtained maximum likelihood estimates and also to find the best method of generating samples from the proposed models to use as the proposal distributions in the Markov Chain Monte Carlo sampling method for predicting the 3D structure of proteins.