Probing Structure Disrupting Mutations in Helices through Reinforcement Learning

Abstract

Helices are important secondary structural motifs within proteins, formed via hydrogen bonds between amino acid main chain atoms, playing pivotal roles in numerous physiological processes. Variations in the properties of amino acid side chains confer distinct abilities to stabilise or destabilise these helical structures. While amino acids such as Alanine and Leucine are known to promote helix formation, others such as Proline and Glycine possess side chain attributes that can disrupt helical structures. The tendency of amino acids to form or break a helix can be measured through helical propensity. Despite extensive investigations into helical propensity using model peptides, a general model for predicting which amino acids can disrupt or break a helix has yet to be developed. In tackling this challenge, we employ structural biochemistry techniques alongside reinforcement learning methodologies to develop a predictive model for helix-disrupting mutations. We start with a toy model consisting of helices with only 30 amino acids and train different models. Our results underscore the effectiveness of our approach in disrupting helical structures, highlighting the promising potential of reinforcement learning in addressing problems associated with protein structure disruption. Furthermore, this work acts as a proof of concept for similar models to be built and extended to more complex systems, such as protein mutants that cause disease.