Research map / Label-efficient and robust learning / Recognition and long-tail
Motif-Oriented Representation Learning with Topology Refinement for 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 MOTOR, a motif-oriented graph learning model that encodes molecular motif structure, context, and semantics (LDA-style topic codebook) and iteratively refines the DDI network topology for drug-drug interaction link prediction.
Problem
AI-driven drug-drug interaction (DDI) prediction under-explores intra-molecular motifs and assumes the observed DDI network topology is complete and fixed, which limits expressive power.
Approach
Combines three motif-level views (BRICS-decomposed motif structure, motif-graph context via message passing, and LDA/DMM-based motif semantic codebook) with an iterative loop that alternates graph representation learning and learned topology refinement, trained with supervised, inconsistency (KL) and contrastive (JSD) losses.
Key results
- MOTOR achieves AUROC 98.96, AP 97.55, F1 92.42, ACC 96.62 on ZhangDDI, outperforming the best baseline MIRACLE (96.44 AUROC).
- On DeepDDI MOTOR reports AUROC 94.84 and ACC 92.15, versus 93.54 AP-level runner-up MIRACLE per the table.
- Average improvements of 3.80%, 2.25%, and 0.96% over the strongest baselines on the three datasets, with F1 gains up to 4.74%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use MOTOR rather than the strongest baseline | DDI prediction · ZhangDDI | F1: 87.68 -> 92.42 | Helped |
| use MOTOR rather than the strongest baseline | DDI prediction · ChCh-Miner | F1: 94.67 -> 98.30 | Helped |
| use MOTOR rather than the strongest baseline | DDI prediction · DeepDDI | F1: 93.54 -> 94.84 | Helped |
| add motif structure learning (ablation) | MOTOR · ZhangDDI | F1: 90.32 -> 92.42 | Helped |
| add motif context learning (ablation) | MOTOR · ZhangDDI | F1: 91.25 -> 92.42 | Helped |
| add motif semantic learning (ablation) | MOTOR · ZhangDDI | F1: 91.80 -> 92.42 | Helped |
| add graph representation learning (ablation) | MOTOR · ZhangDDI | F1: 90.30 -> 92.42 | Helped |
| add graph topology learning (ablation) | MOTOR · ZhangDDI | F1: 90.62 -> 92.42 | Helped |
| add inconsistency loss LI (ablation) | MOTOR · ZhangDDI | F1: 91.39 -> 92.42 | Helped |
| add contrastive loss LC (ablation) | MOTOR · ZhangDDI | F1: 90.26 -> 92.42 | Helped |
Limitations
- Relies on fixed-size hyperparameters (L1=L2=L3=3, P=3, C=8, T=15) without reported sensitivity analysis in the text.
- Evaluated only on link prediction over three DDI datasets with a stratified random split, not on cold-start or temporal splits.
- Interpretability is shown through qualitative case studies only.
Experimental setup
Datasets: ZhangDDI ChCh-Miner DeepDDI
Benchmarks: ZhangDDI ChCh-Miner DeepDDI
Metrics: AUROC AP F1 ACC
Baselines: GCN GAT MR-GNN EPGNN-DS DeepDrug MIRACLE CSGNN SSI-DDI DeepDDS DSN-DDI
Builds on: BRICS graph attention networks Latent Dirichlet Allocation Dirichlet-Multinomial Mixture
Closest papers by meaning
- Customized Subgraph Selection and Encoding for Drug-drug Interaction Prediction (NeurIPS 2024 · similarity 0.87)
- Dual-Channel Learning Framework for Drug-Drug Interaction Prediction via Relation-Aware Heterogeneous Graph Transformer (AAAI 2024 · similarity 0.87)
- Knowledge Enhanced Representation Learning for Drug Discovery (AAAI 2024 · similarity 0.86)
- MOTIVE: A Drug-Target Interaction Graph For Inductive Link Prediction (NeurIPS 2024 · similarity 0.86)
- ExDDI: Explaining Drug-Drug Interaction Predictions with Natural Language (AAAI 2025 · similarity 0.85)
- MAGE: Model-Level Graph Neural Networks Explanations via Motif-based Graph Generation (ICLR 2025 · similarity 0.85)
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.