Drug-Drug Interaction
Prediction System

Purpose: Polypharmacy safety demands accurate drug–drug interaction (DDI) prediction, yet balancing model complexity with reliability remains challenging. We address the need for enhanced predictive trust and handling of data imbalance by re-evaluating feature representation strategies in neural architectures.

Methods: We introduce tDDI, an uncertainty-aware tabular transformer framework. Departing from purely learned embeddings, our approach leverages explicit physicochemical descriptors for robust signaling and integrates an uncertainty estimator to mitigate severe long-tail class imbalance in pharmacological data.

Results: We demonstrate that explicit descriptors outperform complex categorical embeddings in this context. tDDI significantly improves precision for rare adverse events and, in a prospective evaluation of five novel drugs approved by the FDA in late 2025, successfully identified clinically critical risks.

Conclusion: By synergizing quantitative explanations with natural language reasoning, tDDI advances clinical utility. Our findings suggest that integrating domain-specific descriptors with uncertainty estimation establishes a new standard for trustworthy DDI prediction, offering a robust complement to existing methodologies.