VarDrug: A Machine Learning Approach for Variant-Drug Interaction, Application to Drugs for Psychiatric Disorders

Predicting variant-drug interactions is essential for advancing precision medicine across therapeutic areas. The Pharmacogenomics Knowledge Base (PharmGKB) dataset, with ~11,000 samples, is underutilized in machine learning (ML) due to its limited size. After filtering for variant mappings and excluding metabolizer-related conditions, we obtain ~4,000 samples for a six-class prediction task (increasing or decreasing toxicity, efficacy, and dosage). We introduce VarDrug, the first ML framework for variant-drug interaction prediction using PharmGKB, designed to model interactions between genetic variants and drugs. VarDrug integrates a self-supervised VariantEncoder pre-trained on 100,000 GRCh38 variants, Fingerprint method for drug encoding, and gene co-expression profiles for enhanced variant representation. Using SMOTE for class balancing and 5-fold cross-validation, we evaluate five ML models (CatBoost, RandomForest, ExtraTree, DecisionTree, SVC) against label encoding and rule-based baselines. RandomForest achieves a weighted F1 score of 90%, significantly outperforming baselines (best weighted F1: 68%). Ablation studies confirm the VariantEncoders critical role, while a case study on psychiatric disorders, focusing on borderline personality disorder (BPD), demonstrates biological plausibility with alignment to known pharmacogenetic annotations for genes like ABCB1 and CYP2D6. VarDrugs approach, mapping drug-gene and mechanism-of-action-gene interactions, offers a scalable framework for optimizing treatment strategies and reducing adverse drug reactions across pharmacogenomic applications.