MMI-DDI: Multimodal Mutual Influence Framework for Drug-Drug Interaction Prediction under Drug Cold Start

Accurate prediction of drug-drug interactions (DDIs) is essential for safe polypharmacy and pharmacovigilance, yet traditional experimental approaches remain resource-intensive. While deep learning has advanced DDI prediction, existing frameworks face two fundamental limitations: (1) they predominantly focus on molecular structures while treating other critical pharmacological modalities (e.g., targets, enzymes) as independent features, neglecting the asymmetric mutual influences both within a single drug and between drugs; and (2) prevalent dataset partitioning strategies introduce evaluation bias by failing to distinguish drug cold start from label cold start, thereby obscuring true model generalizability. To address these limitations, we propose the Multimodal Mutual Influence Drug-Drug Interaction (MMI-DDI) framework. MMI-DDI explicitly mines and fuses asymmetric multimodal mutual influences at two levels: within a single drug (capturing intra-modal dependencies and inter-modal interactions within one drug) and between drugs (modeling bidirectional asymmetric influence propagation across modalities between drug pairs). These comprehensively derived representations are then fed into a custom-designed multi-type DDI predictor. In addition, we construct a new benchmark dataset and develop a label-aware and drug-aware partitioning algorithm to enable rigorous evaluation under both standard and cold-start scenarios. We conduct extensive experiments on three tasks: standard prediction, single-drug cold start, and double-drug cold start. Evaluations are performed on four datasets, including randomly split datasets and a curated dataset constructed using our automated partitioning algorithm. Across all settings, MMI-DDI consistently outperforms representative baselines in terms of accuracy, macro precision, and macro recall. Furthermore, case study provides interpretable pharmacological insights, while error analysis identifies directions for future improvement.