ETD

• drugs
• genomic large language models
• interpretability
• machine learning
• pharmacogenomics
• TabPFN
• variants

Pharmacogenomics enables therapies tailored to an individual’s genetic background, but testing every variant-drug interaction experimentally is impossible. To tackle this problem, we developed an interpretable machine-learning framework to predict whether a variant significantly alters drug response. We assembled two multi-source datasets, one restricted to missense variants and one spanning all variant types, integrating variant-, protein- and drug-level information. Genomic language models were also employed to represent non-coding variants, which lack protein-level information. Predictions were made with the tabular foundation model TabPFN and interpreted through SHAP analysis and structural mapping onto drug-protein complexes. The generalizable missense model predicted interaction significance to a modest but meaningful degree, with AlphaMissense pathogenicity as the strongest positive signal. Extending the model to all variant types covered non-missense variation, where genomic language model scores and regulatory annotation were especially valuable. Evo2 agreed most strongly with classical predictions and clinical annotations, and was the only genomic language model contributing meaningfully.
To our knowledge, this is the first framework combining a tabular foundation model with variant, protein, drug and genomic language model features, offering interpretable insights into the determinants of variant-drug interactions.