Research map / Label-efficient and robust learning / Recognition and long-tail
Knowledge Enhanced Representation Learning for Drug Discovery
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.
Builds multimodal knowledge graphs from seven public drug/protein sources and pretrains inductive R-GCN embeddings on them (Otter-Knowledge), ensembling them to improve drug-target binding affinity prediction.
Problem
Drug-target binding affinity prediction relies on unimodal pretrained representations (SMILES or protein sequences) that ignore relations among molecules, proteins, and other entities, and multimodal knowledge graphs for this purpose are not publicly available.
Approach
Pretrain an inductive R-GCN over multimodal knowledge graphs whose node features come from modality-specific encoders (ESM-1b, MolFormer or Morgan fingerprints, text encoders), then transfer or fine-tune the embeddings for binding affinity, combining separately trained models via linear ensembling so unaligned KGs need not be merged.
Key results
- Otter Ensemble reaches 0.588 Pearson on DTI DG temporal split vs 0.569 baseline and 0.538 leaderboard, first place on the TDC DTI DG leaderboard.
- With fine-tuning of projection layers, Otter ensemble reaches 0.617 on DTI DG temporal split (baseline 0.557).
- Otter Ensemble achieves 0.839 on DAVIS random split and 0.712 on KIBA drug split in the fine-tuning setting.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace vanilla sequence/SMILES representations with the Otter knowledge-graph ensemble | Drug-target binding affinity prediction · DTI DG Temporal | Pearson correlation: 0.569 -> 0.588 | Helped |
| Replace vanilla sequence/SMILES representations with the Otter knowledge-graph ensemble | Drug-target binding affinity prediction · DAVIS Random | Pearson correlation: 0.805 -> 0.839 | Helped |
| Replace vanilla sequence/SMILES representations with the Otter knowledge-graph ensemble | Drug-target binding affinity prediction · DAVIS Target | Pearson correlation: 0.554 -> 0.578 | Helped |
| Replace vanilla sequence/SMILES representations with the Otter knowledge-graph ensemble | Drug-target binding affinity prediction · DAVIS Drug | Pearson correlation: 0.264 -> 0.168 | Hurt |
| Replace vanilla sequence/SMILES representations with the Otter knowledge-graph ensemble | Drug-target binding affinity prediction · KIBA Random | Pearson correlation: 0.852 -> 0.886 | Helped |
| Replace vanilla sequence/SMILES representations with the Otter knowledge-graph ensemble | Drug-target binding affinity prediction · KIBA Target | Pearson correlation: 0.630 -> 0.678 | Helped |
| Replace vanilla sequence/SMILES representations with the Otter knowledge-graph ensemble | Drug-target binding affinity prediction · KIBA Drug | Pearson correlation: 0.576 -> 0.638 | Helped |
| Fine-tune pretrained GNN projection layers and ensemble models rather than using the baseline | Drug-target binding affinity prediction · DTI DG Temporal | Pearson correlation: 0.557 -> 0.617 | Helped |
| Fine-tune pretrained GNN projection layers and ensemble models rather than using the baseline | Drug-target binding affinity prediction · DAVIS Random | Pearson correlation: 0.840 -> 0.867 | Helped |
| Fine-tune pretrained GNN projection layers and ensemble models rather than using the baseline | Drug-target binding affinity prediction · DAVIS Target | Pearson correlation: 0.559 -> 0.632 | Helped |
| Fine-tune pretrained GNN projection layers and ensemble models rather than using the baseline | Drug-target binding affinity prediction · DAVIS Drug | Pearson correlation: 0.215 -> 0.290 | Helped |
| Fine-tune pretrained GNN projection layers and ensemble models rather than using the baseline | Drug-target binding affinity prediction · KIBA Random | Pearson correlation: 0.857 -> 0.891 | Helped |
Surprising findings
- Noisy, unfiltered links and removing information-flow control gave results comparable to carefully curated links, suggesting resilience to irrelevant triples.
- Fine-tuning the GNN projection layers improved consistently over representation transfer, but the DUDe-only model performed worse than the baseline on several splits.
Limitations
- Enhancements were not significant for the KIBA drug split, which has only 68 drugs.
- Merging heterogeneous KGs into one graph is hard due to schema alignment, motivating the ensemble workaround rather than a unified model.
- Excludes 3D structural modalities.
Experimental setup
Datasets: UniProt BindingDB ChEMBL DrugBank DUDe PrimeKG STITCH TDC DTI DG DAVIS KIBA
Benchmarks: TDC DTI DG DAVIS KIBA
Metrics: Pearson correlation
Baselines: DeepDTA (TDC baseline) Vanilla representation baseline Conplex OntoProtein KeAP
Builds on: R-GCN ESM-1b MolFormer DeepDTA GAS (graph auto-scaling)
Closest papers by meaning
- Learning Molecular Representation in a Cell (ICLR 2025 · similarity 0.87)
- Biomedical Knowledge Graph Embedding with Householder Projection (Student Abstract) (AAAI 2024 · similarity 0.87)
- BounDr.E: Predicting Drug-likeness via Biomedical Knowledge Alignment and EM-like One-Class Boundary Optimization (ICML 2025 · similarity 0.87)
- KnowMol: Advancing Molecular Large Language Models with Multi-Level Chemical Knowledge (NeurIPS 2025 · similarity 0.87)
- Association Pattern-enhanced Molecular Representation Learning (AAAI 2025 · similarity 0.87)
- Dual-Channel Learning Framework for Drug-Drug Interaction Prediction via Relation-Aware Heterogeneous Graph Transformer (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.