Research map / Graph learning / Knowledge and temporal graphs
Dual-Channel Learning Framework for Drug-Drug Interaction Prediction via Relation-Aware Heterogeneous Graph Transformer
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes TIGER, a dual-channel relation-aware heterogeneous graph Transformer that fuses graph-level embeddings from drug molecular graphs with node-level embeddings from subgraphs of biomedical knowledge graphs to predict DDIs.
Problem
Drug-drug interaction prediction is limited by GNN local-neighborhood aggregation (over-squashing, weak long-range modeling) and by difficulty integrating heterogeneous information from biomedical knowledge graphs and drug molecular graphs.
Approach
A Transformer with relation-aware self-attention that treats each typed relation between node pairs as a distinct edge and adds shortest-path distance relations for distant nodes, applied to both molecular graphs (graph-level) and sampled knowledge-graph subgraphs (node-level), with a mutual-information regularizer.
Key results
- TIGER beats the best baseline on each dataset with average gains of 3.66% ACC, 4.17% F1, 3.78% AUC and 4.83% AUPR.
- On OGB-biokg, TIGER surpasses 0.90 ACC on a dense dataset.
- Dual-channel (MG + BKG) model outperforms either single channel on all three datasets.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Combine the molecular graph and biomedical knowledge graph channels (ablation) | TIGER · DrugBank | ACC: 0.6347 -> 0.7903; 0.7126 -> 0.7903 | Helped |
| Combine the molecular graph and biomedical knowledge graph channels (ablation) | TIGER · KEGG | ACC: 0.7716 -> 0.8752; 0.8593 -> 0.8752 | Helped |
| Combine the molecular graph and biomedical knowledge graph channels (ablation) | TIGER · OGB-biokg | ACC: 0.7987 -> 0.8548; 0.8476 -> 0.8548 | Helped |
| Use TIGER instead of MDNN | DrugBank DDI prediction · DrugBank | ACC: 0.7394 -> 0.7905 | Helped |
| Use TIGER instead of Bi-GNN | KEGG DDI prediction · KEGG | ACC: 0.8503 -> 0.8850 | Helped |
| Use TIGER instead of MIRACLE | OGB-biokg DDI prediction · OGB-biokg | ACC: 0.8923 -> 0.9162 | Helped |
| Use the probability-based subgraph extractor rather than the DeepWalk-based extractor (ablation) | TIGER · DrugBank | ACC: 0.7849 -> 0.7905 | No clear effect |
| Use the DeepWalk-based subgraph extractor rather than the probability-based extractor (ablation) | TIGER · OGB-biokg | ACC: 0.8791 -> 0.9162 | Helped |
| Use the k-subtree-based extractor rather than the strongest alternate extractor (ablation) | TIGER · DrugBank, KEGG, and OGB-biokg | ACC | Mixed |
| Set k to 2 in the k-subtree extractor (ablation) | TIGER · DrugBank | ACC | Helped |
| Increase k beyond 4 in the k-subtree extractor (ablation) | TIGER · DrugBank | ACC | Hurt |
| Increase the number of transformer layers L while fixing k=2 (ablation) | TIGER · DrugBank | ACC | No clear effect |
Limitations
- Performance degrades and becomes unstable when k grows beyond 4 due to subgraph noise
- Baselines retrained with original hyperparameters rather than tuned on the same setup
Experimental setup
Datasets: DrugBank KEGG OGB-biokg
Benchmarks: DrugBank KEGG OGB-biokg
Metrics: ACC F1 AUC AUPR
Baselines: SSI-DDI Molormer KGNN KG2ECapsule DDKG Bi-GNN MIRACLE MDNN
Builds on: Transformer Graphormer Deep Graph Infomax-style JS mutual information estimator DeepWalk PageRank
Closest papers by meaning
- MKG-FENN: A Multimodal Knowledge Graph Fused End-to-End Neural Network for Accurate Drug–Drug Interaction Prediction (AAAI 2024 · similarity 0.89)
- Biomedical Knowledge Graph Embedding with Householder Projection (Student Abstract) (AAAI 2024 · similarity 0.88)
- Customized Subgraph Selection and Encoding for Drug-drug Interaction Prediction (NeurIPS 2024 · similarity 0.87)
- Motif-Oriented Representation Learning with Topology Refinement for Drug-Drug Interaction Prediction (AAAI 2025 · similarity 0.87)
- MOTIVE: A Drug-Target Interaction Graph For Inductive Link Prediction (NeurIPS 2024 · similarity 0.87)
- Knowledge Enhanced Representation Learning for Drug Discovery (AAAI 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.