RAMCF: Rank-Aware Multimodal Contrastive Framework for drug side-effect frequency prediction

Accurate prediction of adverse drug reaction (ADR) frequencies is important for drug safety evaluation and pharmacovigilance. Unlike conventional ADR prediction tasks that focus on binary associations, frequency prediction requires modeling ordinal relationships among multiple frequency levels while integrating heterogeneous biomedical information. However, learning structured representations from diverse modalities while preserving the ordinal structure of ADR frequencies remains challenging. In this study, we propose RAMCF (Rank-Aware Multimodal Contrastive Framework), a multimodal representation learning framework for drug-side effect frequency prediction. RAMCF integrates diverse biomedical modalities, including SMILES sequences, molecular fingerprints, molecular structure images, protein-protein interaction features, and semantic side-effect representations. An adaptive multimodal fusion module dynamically aggregates heterogeneous drug features, while a bidirectional cross-modal interaction module captures dependencies between drug and side-effect representations. To enhance representation learning, RAMCF introduces a rank-aware contrastive learning objective that brings samples with similar ADR frequencies closer while separating dissimilar ones. An ordinal-aware dual-branch prediction module jointly models regression signals and ordinal frequency structures to improve prediction consistency. Experiments on a curated SIDER-based dataset containing 638 drugs, 994 side effects, and 33,905 drug-side effect pairs show that RAMCF outperforms representative baseline methods, achieving an RMSE of 0.6197 and a Spearman correlation of 0.7615. These results demonstrate the potential of multimodal contrastive representation learning for advancing computational pharmacovigilance and improving ADR frequency prediction. The source code and data of RAMCF are available at: https://github.com/Zswsw/RAMCF.