Multidimensional mutual information methods for the analysis of covariation in multiple sequence alignments

TL;DR

This study evaluates multidimensional mutual information methods for analyzing covariation in multiple sequence alignments, comparing their effectiveness to other statistical models and exploring their ability to identify residue contacts and structural features.

Contribution

The paper introduces and tests multidimensional mutual information methods as tools for covariation analysis, demonstrating their comparable performance to existing models and clarifying their ability to distinguish direct from indirect residue correlations.

Findings

Multidimensional MI methods perform comparably to maximum entropy models.

Different covariation detection methods have less than 65% overlap in top pairs.

Partial correlation methods better identify close contacts by filtering out fitness-related correlations.

Abstract

Several methods are available for the detection of covarying positions from a multiple sequence alignment (MSA). If the MSA contains a large number of sequences, information about the proximities between residues derived from covariation maps can be sufficient to predict a protein fold. If the structure is already known, information on the covarying positions can be valuable to understand the protein mechanism. In this study we have sought to determine whether a multivariate extension of traditional mutual information (MI) can be an additional tool to study covariation. The performance of two multidimensional MI (mdMI) methods, designed to remove the effect of ternary/quaternary interdependencies, was tested with a set of 9 MSAs each containing <400 sequences, and was shown to be comparable to that of methods based on maximum entropy/pseudolikelyhood statistical models of protein sequences. However, while all the methods tested detected a similar number of covarying pairs among the residues separated by < 8 {\AA} in the reference X-ray structures, there was on average less than 65% overlap between the top scoring pairs detected by methods that are based on different principles. We have also attempted to identify whether the difference in performance among methods is due to different efficiency in removing covariation originating from chains of structural contacts. We found that the reason why methods that derive partial correlation between the columns of a MSA provide a better recognition of close contacts is not because they remove chaining effects, but because they filter out the correlation between distant residues that originates from general fitness constraints. In contrast we found that true chaining effects are expression of real physical perturbations that propagate inside proteins, and therefore are not removed by the derivation of partial correlation between variables.