Hierarchical Multi-Relational Graph Representation Learning for Large-Scale Prediction of Drug-Drug Interactions

TL;DR

This work tackles large-scale drug-drug interaction prediction by bridging explicit relational information and implicit correlations between drug pairs. It introduces HMGRL, a hierarchical framework that learns relation-aware graph embeddings (RaGSE) via multi-relational graphs and a differentiable multi-view spectral clustering (MVDSC) module to capture diverse implicit DP correlations. The model integrates multi-source features (RaGSEs, targets, enzymes, substructures, SMILES) and uses MVDSC to produce multiple DP representations before predicting DDI types, achieving superior performance on two DrugBank-derived datasets across three tasks. The results demonstrate HMGRL’s ability to leverage both explicit drug connections and implicit pairwise patterns, enabling accurate large-scale DDI predictions with potential for discovering novel interactions in practice.

Abstract

Most existing methods for predicting drug-drug interactions (DDI) predominantly concentrate on capturing the explicit relationships among drugs, overlooking the valuable implicit correlations present between drug pairs (DPs), which leads to weak predictions. To address this issue, this paper introduces a hierarchical multi-relational graph representation learning (HMGRL) approach. Within the framework of HMGRL, we leverage a wealth of drug-related heterogeneous data sources to construct heterogeneous graphs, where nodes represent drugs and edges denote clear and various associations. The relational graph convolutional network (RGCN) is employed to capture diverse explicit relationships between drugs from these heterogeneous graphs. Additionally, a multi-view differentiable spectral clustering (MVDSC) module is developed to capture multiple valuable implicit correlations between DPs. Within the MVDSC, we utilize multiple DP features to construct graphs, where nodes represent DPs and edges denote different implicit correlations. Subsequently, multiple DP representations are generated through graph cutting, each emphasizing distinct implicit correlations. The graph-cutting strategy enables our HMGRL to identify strongly connected communities of graphs, thereby reducing the fusion of irrelevant features. By combining every representation view of a DP, we create high-level DP representations for predicting DDIs. Two genuine datasets spanning three distinct tasks are adopted to gauge the efficacy of our HMGRL. Experimental outcomes unequivocally indicate that HMGRL surpasses several leading-edge methods in performance.

Figures (10)

• Figure 1: Flowchart of the proposed HMGRL. Firstly, HMGRL utilizes DDI and DDS graphs to produce RaGSEs for all drugs. Secondly, within the multi-source feature learning module, HMGRL integrates RaGSEs with multiple biochemical attributes to generate comprehensive DP features. Thirdly, HMGRL develops an MVDSC module that employs multiple DP features to uncover implicit correlations between DPs. The MVDSC module generates multiple DP representations, each highlighting distinct implicit correlations. Ultimately, by merging all these DP representation views, we obtain a high-level DP representation suitable for DDI prediction.
• Figure 2: The generating process for drug RaGSEs. Initially, obtaining RaGSEs for known drugs by aggregating features of neighbors under different relations in the DDI graph. Following this, RaGSEs for new drugs are obtained by aggregating features of n-hop neighbors under different similarity relations in the DDS graph. Through this operation, we ensure that all drugs, including new drugs, are equipped with powerful RaGSEs.
• Figure 3: The obtaining process for DP representations within DSC 2.
• Figure 4: Six metrics versus the embedding propagation distance $L$ on different tasks of Dataset 1.
• Figure 5: AUPR (%) versus the dimensions $d^{att}$ and $d^{emb}$ on different tasks of Dataset 1.