Leveraging protein structural information to improve variant effect prediction.

19^th June 2025, 14:00
Joe Marsh
University of Edinburgh, UK

Abstract

Interpreting the effects of genetic variation remains a major challenge, even as genome sequencing becomes routine in research and medicine. Most human genomes carry thousands of protein-coding changes, but we often cannot tell which ones alter function in a meaningful way. Computational variant effect predictors (VEPs) aim to address this by estimating the impact of individual mutations, although their accuracy varies widely. Early approaches focused on evolutionary conservation and sequence patterns. More recently, advances in protein structure prediction, particularly through AlphaFold, have enabled the integration of three-dimensional structural information into these models. I will describe our work benchmarking the current generation of VEPs and show that the best-performing methods now all incorporate structural information. I will also explain how protein structures can be used not only to predict whether a mutation has a clinical impact, but also to infer the mechanism by which it acts. In particular, I will show how structural features can help distinguish between mutations that cause a disruptive loss of function and those that act through more specific mechanisms, such as dominant-negative or gain-of-function effects.