Cross-protein transfer learning substantially improves disease variant prediction.

Area of Science:

• Genomics
• Computational Biology
• Molecular Biology

Background:

• Genetic variations, particularly missense variants, significantly influence individual disease risk.
• The functional impact of most missense variants remains largely uncharacterized.
• Accurate prediction of variant pathogenicity is crucial for clinical genetics and personalized medicine.

Purpose of the Study:

• To develop a robust computational framework for predicting missense variant pathogenicity across the human proteome.
• To leverage deep mutational scanning (DMS) data for training predictive models.
• To enable accurate interpretation of genetic variants for clinical applications.

Main Methods:

• Utilized cross-protein transfer (CPT) models trained on DMS data from five proteins.
• Integrated features from protein sequence models, vertebrate sequence alignments, and AlphaFold structures.
• Evaluated model performance on unseen proteins and compared with existing methods like ESM-1v, EVE, and REVEL.

Main Results:

• Achieved state-of-the-art performance in predicting missense variant pathogenicity for unseen proteins.
• CPT-1 model demonstrated high sensitivity (95%) and improved specificity (68%) for detecting human disease variants compared to other methods.
• Vertebrate sequence alignments provided complementary predictive signals to deep learning models.

Conclusions:

• Deep mutational scanning data is effective for learning transferable variant properties.
• The developed framework accurately predicts variant pathogenicity and aids in clinical variant interpretation.
• Released predictions for missense variants in 90% of human genes to facilitate research and clinical use.