MolBridge: Atom-Level Joint Graph Refinement for Robust Drug-Drug Interaction Event Prediction

TL;DR

MolBridge tackles the challenge of predicting DDI events by explicitly modeling atom-level cross-molecular interactions in a joint graph of drug pairs and mitigating over-smoothing with a Structure Consistency Module. By combining explicit intra-molecular bonds with learned cross-molecular associations through adaptive balancing $A = (1 - \alpha)A' + \alpha A_r$, MolBridge captures both local and long-range dependencies, achieving state-of-the-art results on Deng and Ryu datasets and strong inductive generalization to DrugBank. The method demonstrates robustness on rare DDI types and offers mechanistic interpretability through case studies that identify known interaction motifs, suggesting practical value for safer drug combination strategies. Overall, MolBridge advances DDI prediction by fusing fine-grained structural modeling with hierarchical refinement, enabling more accurate, robust, and interpretable reasoning about drug interactions.

Abstract

Drug combinations offer therapeutic benefits but also carry the risk of adverse drug-drug interactions (DDIs), especially under complex molecular structures. Accurate DDI event prediction requires capturing fine-grained inter-drug relationships, which are critical for modeling metabolic mechanisms such as enzyme-mediated competition. However, existing approaches typically rely on isolated drug representations and fail to explicitly model atom-level cross-molecular interactions, limiting their effectiveness across diverse molecular complexities and DDI type distributions. To address these limitations, we propose MolBridge, a novel atom-level joint graph refinement framework for robust DDI event prediction. MolBridge constructs a joint graph that integrates atomic structures of drug pairs, enabling direct modeling of inter-drug associations. A central challenge in such joint graph settings is the potential loss of information caused by over-smoothing when modeling long-range atomic dependencies. To overcome this, we introduce a structure consistency module that iteratively refines node features while preserving the global structural context. This joint design allows MolBridge to effectively learn both local and global interaction outperforms state-of-the-art baselines, achieving superior performance across long-tail and inductive scenarios. patterns, yielding robust representations across both frequent and rare DDI types. Extensive experiments on two benchmark datasets show that MolBridge consistently. These results demonstrate the advantages of fine-grained graph refinement in improving the accuracy, robustness, and mechanistic interpretability of DDI event prediction.This work contributes to Web Mining and Content Analysis by developing graph-based methods for mining and analyzing drug-drug interaction networks.

Figures (9)

• Figure 1: (a) Metabolism-related interactions account for 73.4% of DDI types tian2025ddinter, yet existing GNN methods (e.g., GCN kipf2017semisupervised) rely on coarse drug-level representations that cannot capture metabolic DDI complexity, making fine-grained atomic-level modeling crucial. (b) Compared to GCN that treats a drug as an isolated graph, MolBridge achieves better performance by explicitly modeling cross-molecular atomic associations. (c) MolBridge comprehensively outperforms GCN on metabolism-related samples across all benchmark metrics.
• Figure 2: Framework of MolBridge. Given a pair of molecular SMILES as input, the model performs: (i) drug initialization to obtain node features and adjacency relationships for each molecule, (ii) joint molecular graph construction to capture inter-molecular connections and cross-drug interactions, (iii) structure consistency module to learn atomic dependencies within the joint graph, and (iv) DDI event prediction via a final classifier that outputs interaction probabilities.
• Figure 3: Ablation study results of MolBridge and its variants on Deng's dataset. The Structure Consistency Module (SCM) is the most critical component, while both $A'$ (explicit chemical bonds) and $A_r$ (attention-based cross-molecular interactions) contribute to optimal performance.
• Figure 4: Visualization and cluster analysis of DDI representations. (a) t-SNE comparison of embedding quality across methods for high-frequency (upper) and low-frequency (lower) DDI event. (b) DBSCAN clustering identifies seven distinct sub-clusters within one interaction type, with drugs labeled by DrugBank IDs and ATC codes.
• Figure 5: Representative substructures identified by MolBridge. The model highlights key atomic interactions (red box) linked to known reactive groups (e.g., N-nitrosourea), suggesting potential metabolic competition in DDI cases.