Disease-specific variant pathogenicity prediction using multimodal biomedical language models.

Area of Science:

• Genomics and Bioinformatics
• Computational Biology
• Machine Learning in Healthcare

Background:

• Missense variants are crucial for genetic disorder diagnosis and risk prediction.
• Current methods lack disease-specific context, limiting clinical utility.
• Accurate interpretation of missense variants requires disease-specific pathogenicity prediction.

Purpose of the Study:

• To introduce a novel deep learning framework, DIVA (disease-specific variant pathogenicity prediction).
• To predict both disease type and deleteriousness probability for missense variants.
• To enhance clinical variant interpretation by incorporating disease specificity.

Main Methods:

• DIVA integrates protein sequence and disease-related textual annotations.
• Utilizes two pre-trained language models for information encoding.
• Employs a contrastive learning paradigm to align variants with relevant diseases.

Main Results:

• DIVA outperforms existing baseline methods in disease prediction accuracy.
• Achieves high relevance to clinically curated disease annotations for missense variants.
• Incorporation of AlphaMissense scores enhances variant deleteriousness prediction.

Conclusions:

• DIVA provides accurate, disease-specific variant pathogenicity predictions.
• Addresses the unmet need for context-aware variant interpretation in clinical settings.
• Offers new insights into predicting variant pathogenicity with disease awareness.