Research map / Graph learning / Knowledge and temporal graphs
MKG-FENN: A Multimodal Knowledge Graph Fused End-to-End Neural Network for Accurate Drug–Drug Interaction Prediction
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 MKG-FENN, which builds four drug knowledge graphs and trains a four-channel GNN with an end-to-end MLP fusion head to predict multi-class drug-drug interaction events.
Problem
Predicting drug-drug interaction (DDI) event types accurately, since existing methods model drug relationships only partially and use non-end-to-end feature fusion.
Approach
Constructs four drug knowledge graphs (drug-target/transporter entities from DrugBank, SMILES, DDI-matrix neighbors, and MACCS substructures) and learns a four-channel GNN whose outputs are fused by an MLP trained end-to-end for DDI event classification.
Key results
- On Task 1 (known drugs), MKG-FENN reaches ACC 0.9409 and AUPR 0.9786, beating MDDI-SCL (ACC 0.9378, AUPR 0.9782) with 64 wins, 2 ties, 0 losses across metrics.
- On Task 2 (known vs new drugs), ACC 0.6805 and AUPR 0.7049 versus best baseline MDDI-SCL at ACC 0.6767 and AUPR 0.6947.
- Ablation shows the full four-channel model (ACC 0.9409) outperforms any single channel (ACC about 0.916-0.917).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use P1 channel only instead of all four channels (ablation) | MKG-FENN · DDI Task 1 (known drugs) | ACC: 0.9409 -> 0.9172 | Hurt |
| Use P2 channel only instead of all four channels (ablation) | MKG-FENN · DDI Task 1 (known drugs) | ACC: 0.9409 -> 0.9169 | Hurt |
| Use P3 channel only instead of all four channels (ablation) | MKG-FENN · DDI Task 1 (known drugs) | ACC: 0.9409 -> 0.9169 | Hurt |
| Use P4 channel only instead of all four channels (ablation) | MKG-FENN · DDI Task 1 (known drugs) | ACC: 0.9409 -> 0.9157 | Hurt |
| Use P1 and P2 channels instead of all four channels (ablation) | MKG-FENN · DDI Task 1 (known drugs) | ACC: 0.9409 -> 0.9341 | Hurt |
| Use P1 and P3 channels instead of all four channels (ablation) | MKG-FENN · DDI Task 1 (known drugs) | ACC: 0.9409 -> 0.9307 | Hurt |
| Use P1 and P4 channels instead of all four channels (ablation) | MKG-FENN · DDI Task 1 (known drugs) | ACC: 0.9409 -> 0.9322 | Hurt |
| Use P2 and P3 channels instead of all four channels (ablation) | MKG-FENN · DDI Task 1 (known drugs) | ACC: 0.9409 -> 0.9318 | Hurt |
| Use P2 and P4 channels instead of all four channels (ablation) | MKG-FENN · DDI Task 1 (known drugs) | ACC: 0.9409 -> 0.9303 | Hurt |
| Use P3 and P4 channels instead of all four channels (ablation) | MKG-FENN · DDI Task 1 (known drugs) | ACC: 0.9409 -> 0.9297 | Hurt |
| Use P1, P2, and P3 channels instead of all four channels (ablation) | MKG-FENN · DDI Task 1 (known drugs) | ACC: 0.9409 -> 0.9344 | Hurt |
| Use P1, P2, and P4 channels instead of all four channels (ablation) | MKG-FENN · DDI Task 1 (known drugs) | ACC: 0.9409 -> 0.9355 | Hurt |
Surprising findings
- Only a single GNN layer worked best, which the authors attribute to over-smoothing in deeper GNNs.
- On the new-drug tasks, absolute performance is much lower (Task 3 ACC about 0.46 for MKG-FENN), and AUC is not the best metric for it (MDF-SA-DDI and DeepDDI have higher AUC in some tasks).
Limitations
- Small dataset of 572 drugs limits generalization.
- Performance drops substantially on new drugs (Tasks 2 and 3), with Task 3 ACC about 0.46.
- Supplementary details (notations, hyperparameter analysis) are deferred to an external file.
Experimental setup
Datasets: DrugBank 5.1.7 DDIMDL dataset MACCS keys features
Benchmarks: DDIMDL drug-drug interaction dataset (572 drugs, 65 event classes)
Metrics: ACC AUPR AUC F1 precision recall Wilcoxon signed-rank test Friedman test
Baselines: MDDI-SCL MDF-SA-DDI DDIMDL MDNN DeepDDI GNN DNN RF KNN LR
Builds on: KGNN MUFFIN DDIMDL
Closest papers by meaning
- Dual-Channel Learning Framework for Drug-Drug Interaction Prediction via Relation-Aware Heterogeneous Graph Transformer (AAAI 2024 · similarity 0.89)
- ExDDI: Explaining Drug-Drug Interaction Predictions with Natural Language (AAAI 2025 · similarity 0.87)
- Customized Subgraph Selection and Encoding for Drug-drug Interaction Prediction (NeurIPS 2024 · similarity 0.85)
- MOTIVE: A Drug-Target Interaction Graph For Inductive Link Prediction (NeurIPS 2024 · similarity 0.85)
- Biomedical Knowledge Graph Embedding with Householder Projection (Student Abstract) (AAAI 2024 · similarity 0.84)
- Motif-Oriented Representation Learning with Topology Refinement for Drug-Drug Interaction Prediction (AAAI 2025 · similarity 0.84)
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.