Predicting Genetic Variant Pathogenicity Using Vector Embeddings

Background

Interpreting the pathogenicity of genetic variants remains a critical bottleneck in genomic medicine. Millions of variants of uncertain significance (VUS) hinder the clinical application of genetic findings. Traditional computational approaches often rely on hand-engineered features and fail to fully capture the complexity of multidimensional genomic annotations.

Methods

We developed a novel semantic embedding framework, VUS.Life, that transforms variant annotations into natural language descriptions and leverages pre-trained language models to capture pathogenicity-relevant relationships in high-dimensional vector space. This approach enables direct pathogenicity prediction through representation learning rather than traditional feature-based methods.

Results

We evaluated the framework using curated variants from three disease-associated genes: BRCA1 (n=3,311) and BRCA2 (n=4,074) from BRCA Exchange, and FBN1 (n=1,532) from ClinVar. Variant annotations from the Variant Effect Predictor (VEP) were converted into natural language while excluding fields linked to known pathogenicity assertions to prevent data leakage. We then embedded these descriptions using three models: MPNet (all-mpnet-base-v2), Google’s text-embedding-004, and MedEmbed-large-v0.1. A k-nearest neighbor (k-NN) approach (up to 20 neighbors) was used to predict pathogenicity for new or unreviewed variants. Dimensionality reduction techniques (PCA, t-SNE, UMAP) enabled visualization of the embedding spaces.

k-NN classification showed exceptional performance across all genes and embedding models. For BRCA1 , overall accuracy ranged from 97.3% (Google) to 97.9% (MPNet), with benign/likely benign accuracy of 95.1–97.2% and pathogenic/likely pathogenic accuracy of 97.9–98.4%. For BRCA2 , overall accuracy ranged from 97.9% (Google) to 99.1% (MedEmbed), with benign/likely benign accuracy of 96.6–99.2% and pathogenic/likely pathogenic accuracy of 98.6–99.4%. FBN1 validation confirmed the method’s generalizability, with accuracy exceeding 96% across all embeddings. Application to not-yet-reviewed BRCA1/2 variants demonstrated the framework’s practicality and scalability, with unknown variants aligning closely to known benign or pathogenic clusters.

Conclusions

This semantic embedding framework, VUS.Life, accurately captures pathogenicity-relevant features from complex variant annotations, enabling high-accuracy (>96%) automated classification across multiple genes and models. The approach generalizes beyond well-curated genes and supports scalable, interpretable, and representation-based classification of VUS. It holds significant promise for alleviating the variant interpretation bottleneck in clinical genomics.