Machine learning of three-dimensional protein structures to predict the functional impacts of genome variation

Abstract

Research in the human genome sciences generates a substantial amount of genetic data for hundreds of thousands of individuals, which concomitantly increases the number of variants with unknown significance (VUS). Bioinformatic analyses can successfully reveal rare variants and variants with clear associations to disease-related phenotypes. These studies have made a significant impact on how clinical genetic screens are interpreted and how patients are stratified for treatment. There are few, if any, comparable computational methods for variants to biological activity predictions. To address this gap, we developed a machine learning method that uses protein three-dimensional structures from AlphaFold to predict how a variant will influence changes to a genes downstream biological pathways. We trained state-of-the-art machine learning classifiers to predict which protein regions will most likely impact transcriptional activities of two proto-oncogenes, nuclear factor erythroid 2 (NFE2)-related factor 2 (Nrf2) and c-MYC. We have identified classifiers that attain accuracies higher than 80%, which have allowed us to identify a set of key protein regions that lead to significant perturbations in c-MYC or Nrf2 transcriptional pathway activities. SignificanceThe vast majority of mutations are either unspecified and/or their downstream biological implications are poorly understood. We have created a method that utilizes protein structure to cluster mutations from population-scale repositories to predict downstream functional impacts. The broader impacts of this approach include advanced filtering of mutations that are likely to impact genome function.