Jacques Fellay

Fine-tuning the ESM2 protein language model to understand the functional impact of missense variants • The study fine-tuned the ESM2 protein language model to classify amino acid-level features, enhancing understanding of missense variants’ functional impacts. • 20 protein features were targeted, including active sites, transmembrane regions, disulfide bonds, and more, to provide detailed insights into pathogenic or benign variant effects. • The researchers demonstrated that some variants of unknown significance (VUS) can be reclassified as pathogenic by identifying enrichment of missense variants in critical protein regions. • The fine-tuned models were used to successfully reclassify 110,304 VUS (6.5%) from the gnomAD database. • This approach enables a comparison of proteins with reference and alternate alleles, helping identify functional gains or losses due to mutations. • The model’s predictions align well with known pathogenicity and conservation scores, showcasing high accuracy and reliability. • Practical applications include evaluating curated missense variants from ClinGen to identify the precise molecular impacts, such as the loss of functional domains. • Notably, mutations like GAA:p.Cys103Gly and HNF4A:p.Arg63Gln were linked to the loss of key protein domains, illustrating the model’s ability to identify disease-relevant changes. • These models are lightweight and accessible, requiring only protein sequences as input, making them user-friendly for both clinicians and researchers. • The study highlights the power of machine learning in variant classification, contributing to advancements in personalized medicine and genetic diagnostics. @jacquesfellay @AliSaadatV 💻Code: github.com/AliSaadatV/ESM2-M… 📜Paper: arxiv.org/abs/2410.10919