(67) Leveraging Vector Embeddings for Rapid and Accurate Pathogenicity Prediction of Genetic Variants

Background: Interpreting the pathogenicity of genetic variants, especially in clinically significant genes like breast cancer susceptibility gene 1 (BRCA1) and 2 (BRCA2), remains a key challenge in genomic medicine. The success of large language models was made possible by vector embeddings technology, which allows semantic representation and contextual understanding of complex data across modalities.

Objectives: To develop and evaluate a novel framework using semantic vector representations of genetic variants for automated pathogenicity classification without relying on existing clinical assertions.

Design/Methods: We transformed Variant Effect Predictor annotations into natural language descriptions while removing confounding fields to prevent data leakage. These descriptions were processed through four embedding models: MPNet, Jina, MedEmbed, and Google. We evaluated performance using curated variants from BRCA Exchange (BRCA1 n=3,311; BRCA2 n=4,100) as ground truth. Evaluation methods included k-nearest neighbors classification (considering the top 20 nearest neighbors) and dimensionality reduction tools (PCA, t-SNE, UMAP) for visual analysis of clustering patterns.

Results: K-nearest neighbors analysis demonstrated exceptional performance across both BRCA genes and all models. For BRCA1, accuracy ranged from 84.6-96.1% for benign and 93.8-98.4% for pathogenic variants. For BRCA2, accuracy ranged from 84.7-98.5% for benign and 92.8-99.3% for pathogenic variants. The MPNet model achieved the highest overall accuracy (BRCA1: 97.6%; BRCA2: 98.9%), followed by Jina (BRCA1: 97.5%; BRCA2: 98.6%), MedEmbed (BRCA1: 96.6%; BRCA2: 97.3%), and Google (BRCA1: 90.8%; BRCA2: 89.9%). In detail, over 92% of benign BRCA1 variants and 97% of pathogenic variants had 80-100% (mean: 99.1% and 99.9%, respectively) of nearest neighbors sharing the same classification with the MPNet model. For BRCA2, similar results were observed with 96% of benign variants and over 99% of pathogenic variants having 80-100% (mean: 99.5% and 99.9%, respectively) of nearest neighbors with matching classifications with the MPNet model. Visual analysis showed distinct clustering by pathogenicity across all models.

Conclusions: Semantic embeddings effectively capture features from variant annotations that correlate with pathogenicity. This approach enables automated classification with high accuracy (>97% with top models). The methodology provides a promising tool for prioritizing VUS for clinical review and supporting evidence-based reclassification.