A tale of human-competitiveness in bioinformatics

Abstract

A key open problem, which has defied scientists for decades is the problem of predicting the 3D structure of proteins (Protein Structure Prediction - PSP) based on its primary sequence: the amino acids that compose a protein chain. Full atomistic molecular dynamics simulations are, for all intents and purposes, impractical as current empirical models may require massive computational resources. One of the possible ways of alleviating this cost and making the problem easier is to simplify the protein representation based on which the native 3D state is searched for. We have proposed a protocol based on evolutionary algorithms to perform this simplification of the protein representation. Our protocol does not use any domain knowledge. Instead it uses a well known information theory metric, Mutual Information, to generate a reduced representation that is able to maintain the crucial information needed for PSP. The evaluation process of our method has shown that it generates alphabets that have competent performance against the original, non-simplified, representation. Moreover, these reduced alphabets obtain better-than-human performance when compared to some classic reduced alphabets.