Unveiling rare drug interactions via a BioGPT-enhanced dual graph framework for robust pharmacovigilance

Uncovering rare drug-drug interactions (DDIs) is pivotal for patient safety, yet hampered by severe class imbalance and semantic complexity in biomedical text. To address this, we present BIDGNN, a framework synergizing generative augmentation with dual-graph reasoning. We employ a BioGPT-driven generative adversarial network (BioGPT-GAN) to synthesize realistic training samples for underrepresented interactions, resolving data scarcity. Complementarily, a dual-graph architecture captures both explicit syntactic dependencies and implicit semantic correlations, fused via cross-modal attention. BIDGNN achieves an F1-score of 83.57% on the DDI 2013 benchmark, demonstrating superior sensitivity to minority interaction types. This work offers a robust, interpretable solution for automated pharmacovigilance, effectively translating massive literature into clinical safety insights.