TripleVar: A Multimodal Deep Learning Framework for High-Fidelity Clinical Variant Effect Prediction

This poster was part of the ASCPT conference in Denver, Colorado, 2026. It presents TripleVar, a multimodal deep learning framework designed to improve the functional interpretation of genetic variants in the thiamine transporter genes SLC19A2 and SLC19A3, which are associated with severe neurological disorders that can be treated when diagnosed early. A major challenge in precision medicine is that the majority of variants identified in these genes remain unclassified or are labeled as variants of uncertain significance (VUS). TripleVar integrates three complementary data modalities: genomic embeddings derived from DNABERT-2, protein embeddings generated using ESM-2 on simulated mutant protein sequences, and allele-frequency features derived from population databases. The model reconstructs the expected biological protein sequence for each variant using in-silico exon splicing and open reading frame translation before generating protein representations. These modalities are fused into a unified feature representation and processed through a deep neural network classifier. The model was trained on 864 ClinVar variants together with 962 synthetic benign variants for contrastive learning and evaluated on both internal and external datasets. On the ClinVar test set, TripleVar achieved 93.3% classification accuracy with an AUC of 1.0 for binary pathogenic versus benign classification. Evaluation on an independent dataset of experimentally validated variants demonstrated strong generalization, achieving 80% accuracy and 90% precision even without allele frequency information. TripleVar Abstract, minimzed si… Applying the trained model to 28,925 previously unclassified variants, TripleVar prioritized 111 high-confidence pathogenic candidates and 303 likely pathogenic variants for further investigation. These results demonstrate how biologically informed data engineering combined with multimodal machine learning can significantly improve variant interpretation. TripleVar provides a scalable computational pipeline that can be extended to other clinically relevant genes, potentially reducing the burden of variants of uncertain significance and accelerating diagnosis and treatment decisions for rare genetic diseases.