A multitask neural network trained on embeddings from ESMFold can accurately rank order clinical outcomes for different cystic fibrosis mutations

Abstract

Advancements in AI-based protein structure prediction have opened new research avenues in computational biology. Previous work has shown mixed results when the predicted 3D structure or model embeddings have been used as features for predicting phenotypes. Many existing tools which predict the pathogenicity of variants (e.g. SIFT 4G, Polyphen-2, REVEL, LYRUS) have incorporated structural protein features, but have shown mixed success when applied to mutations on the CFTR gene. This paper explores whether embeddings from the ESM-Fold model, which predicts the 3D structure of a protein from an amino acid sequence, can be used to predict clinical outcomes for different Cystic Fibrosis (CF) genotypes found in the CFTR2 database. A neural network model trained on these embeddings is able to obtain both non-trivial and statistically significant correlations on CF-related phenotypes (sweat chloride, pancreatic insufficiency, and infection rates). Overall, AI-based protein structure prediction models show a promising ability to assess the relative severity of CFTR-based mutations on key phenotypic outcomes associated with CF. The entire processing and analysis pipeline for this work can be found at this repo.