Semantic embedding of variant effect annotations enables rapid and accurate pathogenicity prediction with VUS.Life

Area of Science:

• Genomic Medicine
• Computational Biology
• Artificial Intelligence in Healthcare

Background:

• Interpreting genetic variant pathogenicity is a major obstacle in genomic medicine.
• Millions of variants of uncertain significance (VUS) impede clinical application of genetic data.
• Existing computational methods struggle with complex genomic annotations due to reliance on hand-engineered features.

Purpose of the Study:

• To develop a novel multi-modal framework, VUS.Life, for accurate genetic variant pathogenicity interpretation.
• To leverage semantic text embeddings and protein language modeling to overcome limitations of traditional approaches.
• To enhance the clinical utility of genomic findings by automating the classification of VUS.

Main Methods:

• VUS.Life transforms variant annotations into natural language descriptions, then converts them into vector embeddings using Large Language Models (LLMs).
• Pathogenicity is predicted based on the proximity of variant embeddings to known pathogenic variants.
• The framework incorporates residue-level delta embeddings from ESM-600M for enhanced biophysical and clinical context capture.

Main Results:

• VUS.Life achieved high performance (MCC of 0.895-0.989, F1 ≥ 0.94) on over 10,000 variants across eight ACMG Tier 1 disease genes.
• Unsupervised analysis revealed that delta embeddings can distinguish distinct pathogenic mechanisms, correlating with evolutionary fitness.
• Ablation studies confirmed that the MPNet embeddings alone possess full discriminative power, independent of pre-computed scores.

Conclusions:

• The VUS.Life semantic embedding framework accurately classifies variant pathogenicity from complex annotations, addressing the VUS interpretation bottleneck.
• The approach demonstrates scalability, interpretability, and generalization beyond well-curated genes.
• VUS.Life offers a promising solution for robust automated variant classification in clinical genomics.