Cross-protein transfer learning substantially improves disease variant prediction
Problem
Most human missense genetic variants have unknown disease effects, leaving clinicians unable to interpret them.
Why it matters
Unknown variant effects block genetic diagnosis and disease risk assessment for patients.
What the AI did
Trained cross-protein transfer deep learning models on deep mutational scanning data from five proteins to predict missense variant pathogenicity across the human proteome.
Results
At 95% sensitivity on ClinVar disease variants, specificity improved to 68% vs 27% (ESM-1v) and 55% (EVE); state-of-the-art on unseen proteins.
Impact
Improves clinical interpretation of missense variants beyond existing tools.
Record
| Venue | Genome biology |
|---|---|
| Published | 2023 |
| Work type | article |
| Authors | Milind Jagota, Chengzhong Ye, Carlos Albors, Ruchir Rastogi, Antoine Koehl, Nilah M. Ioannidis, Yun S. Song |
| Citations | 67 (OpenAlex, at import time) |
| References | 79 |
| Model or method | CPT-1 (cross-protein transfer model) |
| Organization | not stated in source |
| Primary source | https://doi.org/10.1186/s13059-023-03024-6 |
| Open access | gold |
| Keywords | missense variants, pathogenicity prediction, transfer learning, deep mutational scanning, ClinVar |
Importance rating
significant — thousands to millions affected
Abstract states genetic variation is a major determinant of disease risk and the vast majority of missense variants have unknown effects; proteome-wide clinical interpretation is a broad, significant capability.