Research map / Label-efficient and robust learning / Recognition and long-tail
Blend the Separated: Mixture of Synergistic Experts for Data-Scarcity Drug-Target 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 MoseDTI, a mixture of two synergistic experts (extrinsic KG-based and intrinsic sequence/graph-based) with a gating model that mutually supervise each other via pseudo labels to handle drug-target interaction prediction under input-data and label scarcity.
Problem
Drug-target interaction prediction when intrinsic or extrinsic input data is scarce for some drugs/targets and when labels for a specific interaction type are very limited (e.g., 10-shot).
Approach
Two heterogeneous experts (an extrinsic KG-embedding MLP and an intrinsic GNN/CNN encoder) are fused by a sample-wise gating model, and they generate pseudo labels for each other from unlabeled drug-target pairs so each expert can work alone when one data perspective is missing.
Key results
- On 10-shot specific-interaction datasets, MoseDTI reaches 67.27% ACC / 75.82% AUC on DGIDB::AGONIST and 92.85% ACC / 98.76% AUC on DGIDB::BLOCKER, outperforming MDTips (64.23 / 73.51 and 91.25 / 97.27).
- On GNBR::E- (10-shot), MoseDTI achieves 78.17% ACC and 88.71% AUC versus MDTips 69.61% and 82.35%.
- Without data scarcity on DrugCentral and DrugBank, MoseDTI achieves 88.40% and 88.15% ACC, slightly above MDTips (88.10 and 87.75).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use MoseDTI instead of MDTips | DTI prediction · DGIDB::AGONIST (10-shot) | ACC: 64.23±4.13 -> 67.27±1.90 | Helped |
| Use MoseDTI instead of MDTips | DTI prediction · DGIDB::BLOCKER (10-shot) | AUC: 97.27±0.86 -> 98.76±0.82 | Helped |
| Use MoseDTI instead of MDTips | DTI prediction · GNBR::E- (10-shot) | ACC: 69.61±3.43 -> 78.17±4.86 | Helped |
| Use MoseDTI instead of MDTips | DTI prediction · DrugCentral | AUPR: 95.32 ± 0.21 -> 95.32 ± 0.22 | No clear effect |
| Use MoseDTI instead of MDTips | DTI prediction · DrugBank | AUPR: 94.02 ± 0.23 -> 95.23 ± 0.28 | Helped |
| Add mutually generated pseudo-label training to the intrinsic expert (ablation) | Mose-intr · Few-shot specific DTI | performance | Helped |
| Add mutually generated pseudo-label training to the extrinsic expert (ablation) | Mose-extr · DGIDB::BLOCKER | performance | Helped |
| Add mutually generated pseudo-label training to the extrinsic expert (ablation) | Mose-extr · DGIDB::AGONIST | performance | No clear effect |
| Add mutually generated pseudo-label training to the full model (ablation) | MoseDTI · Few-shot specific DTI | performance | Helped |
| Add the mixture-of-experts architecture rather than using either expert alone (ablation) | MoseDTI · Few-shot specific DTI | performance | Helped |
| Train the whole model with only ground-truth labels rather than using the synergizing mechanism (ablation) | MoseDTI · GNBR::E- and DGIDB::AGONIST | performance | Mixed |
Surprising findings
- Direct training of the whole model with few-shot true labels (True-all) does not reliably beat the individual true-label experts, e.g., on e- and agonist datasets.
- Pretraining on general DTI labels does not transfer well to specific interaction types, according to the authors' Appendix B experiment.
Limitations
- Only SMILES, amino acid sequences and KG relations are used; other modalities such as text describing entities are not incorporated.
- Performance on general DTI datasets without scarcity is only marginally above MDTips on some metrics.
Experimental setup
Datasets: DGIDB::AGONIST DGIDB::BLOCKER GNBR::E- DrugBank DrugCentral DRKG UniProt
Benchmarks: DGIDB::AGONIST DGIDB::BLOCKER GNBR::E- DrugBank DrugCentral
Metrics: ACC AUC AUPR
Baselines: GNNCPI TransformerCPI TransformerCPI2.0 TransE RotatE TriModel AMGDTI KG-MTL MDTips
Builds on: TransE RotatE ESM-MSA-1b MDTips AMGDTI
Closest papers by meaning
- MOTIVE: A Drug-Target Interaction Graph For Inductive Link Prediction (NeurIPS 2024 · similarity 0.88)
- ExDDI: Explaining Drug-Drug Interaction Predictions with Natural Language (AAAI 2025 · similarity 0.87)
- R-DTI: Drug Target Interaction Prediction Based on Second-Order Relevance Exploration (AAAI 2025 · similarity 0.87)
- Customized Subgraph Selection and Encoding for Drug-drug Interaction Prediction (NeurIPS 2024 · similarity 0.86)
- Multilevel Attention Network with Semi-supervised Domain Adaptation for Drug-Target Prediction (AAAI 2024 · similarity 0.86)
- Rethinking the generalization of drug target affinity prediction algorithms via similarity aware evaluation (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.