HLN-DDI: hierarchical molecular representation learning with co-attention mechanism for drug-drug interaction prediction

Abstract

Background: Accurate identification of drug-drug interactions (DDIs) is critical in pharmacology, as DDIs can either enhance therapeutic efficacy or trigger adverse reactions when multiple medications are administered concurrently. Traditional methods for identifying DDIs are labor-intensive and time-consuming, prompting the development of computational alternatives. However, existing computational approaches frequently encounter challenges related to interpretability and struggle to effectively capture the complex, multi-level structures inherent in drug molecules. Specifically, they often fail to adequately analyze substructural components and neglect interactions across hierarchical structural levels, resulting in incomplete molecular representations.

Results: In this study, we propose a Hierarchical Learning Network with a co-attention mechanism tailored to molecular structure representation for predicting DDIs, named HLN-DDI. The proposed method advances existing approaches by explicitly encoding motif-level structures and capturing hierarchical molecular representations at atom-level, motif-level, and whole-molecule scales. These hierarchical representations are integrated using a co-attention mechanism and combined with interaction-type information to enhance predictive performance. Comprehensive evaluations demonstrate that HLN-DDI significantly outperforms state-of-the-art methods across multiple benchmark datasets, achieving over 98% accuracy under transductive scenarios and surpassing 99% on various evaluation metrics. Moreover, HLN-DDI achieves a notable accuracy improvement of 2.75% in predicting DDIs involving unseen drugs. Practical assessments with real-world DDI scenarios further validate the efficacy and utility of our proposed model.

Conclusion: By leveraging hierarchical molecular structures and employing a co-attention mechanism to effectively integrate multi-level representations, HLN-DDI generates comprehensive and precise drug representations, leading to substantially improved predictions of potential drug-drug interactions.

Keywords: Co-attention mechanism; Drug-drug interaction; Molecular graph learning.

Fig. 1

Framework of HLN-DDI. HLN-DDI first decomposes the input molecular graphs of two drugs into motifs, subsequently augmenting these graphs by incorporating motif-level and molecular-level nodes and establishing corresponding edges. Graph Neural Networks (GNNs) with shared weights are then employed to hierarchically encode the augmented graphs, generating atom-level, motif-level, and molecular-level representations. These hierarchical representations are integrated via a co-attention mechanism to compute a pairwise interaction importance matrix. Finally, the likelihood of drug-drug interactions is determined by weighting and aggregating the drug representations and relationship information based on the interaction importance matrix

Fig. 2

Summary of Motif Creation Process. The entire construction encompasses three sequential stages: 1 Initially, the provided molecular graph is broken down utilizing the BRICS framework, 2 Subsequently, the molecule undergoes additional decomposition guided by our supplementary rules, and 3 Finally, the resultant motifs are assembled into motif-level nodes

Fig. 3

Visualization results of drug pair features after t-SNE processing at atom-level (A,D), motif-level (B,E), and molecule-level (C,F)

Fig. 4

Performance evaluation of SSI-DDI, GMPNN-CS, DSN-DDI and HLN-DDI for new FDA-approved drugs