Research map / Graph learning / Graph self-supervised learning
Bi-level Contrastive Learning for Knowledge-Enhanced Molecule Representations
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.
Gode pre-trains a molecular-graph GNN (M-GNN, initialized from GROVER) and a KG sub-graph GNN (K-GNN over a molecule-centric KG, MolKG built from PubChemRDF and PrimeKG) and aligns them with InfoNCE-style contrastive learning before fine-tuning on MoleculeNet property prediction.
Problem
Molecular property prediction from molecular graphs alone ignores the broader relational context of molecules in biomedical knowledge graphs, limiting representation quality.
Approach
Treat each molecule as a node in a molecule-centric KG, pre-train a GNN on the molecular graph and another GNN on the molecule's KG sub-graph with self-supervised tasks, then align the two views via contrastive learning so the molecular embedding absorbs biochemical knowledge.
Key results
- Gode achieves an average 12.7% ROC-AUC improvement on classification and 34.4% RMSE/MAE improvement on regression over baselines, and 2.2% / 7.2% over KANO.
- Gode reaches 94.8 ROC-AUC on BBBP, 94.7 on ClinTox, 92.0 on BACE, and 84.3 on Tox21, outperforming the best baselines in those tasks.
- Ablation: adding MolKG via a static KG embedding to GROVER gives 14.3% overall enhancement; full Gode gives 23.2% over GROVER.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add M-GNN pre-training (ablation) | Gode · BBBP | ROC-AUC: 92.2 -> 94.8 | Helped |
| add K-GNN pre-training (ablation) | Gode · BBBP | ROC-AUC: 93.2 -> 94.8 | Helped |
| remove M-GNN pre-training (ablation) | Gode · Lipophilicity | RMSE: 0.743 -> 0.708 | Helped |
| remove indication relationship from MolKG (ablation) | Gode · ClinTox | ROC-AUC: 94.7 -> 93.4 | Hurt |
| remove xlogp3 and xlogp3-aa relationships from MolKG (ablation) | Gode · FreeSolv | RMSE: 1.048 -> 1.189 | Hurt |
| remove xlogp3 and xlogp3-aa relationships from MolKG (ablation) | Gode · Lipophilicity | RMSE: 0.743 -> 0.782 | Hurt |
| remove tautomer_cnt and covalent_unit_cnt relationships from MolKG (ablation) | Gode · QM7 | MAE: 57.2 -> 61.7 | Hurt |
| remove nbr_2d, nbr_3d, and has_same_conn relationships from MolKG (ablation) | Gode · BBBP | ROC-AUC: 94.8 -> 95.0 | Helped |
| integrate MolKG static KG embeddings with GROVER features (ablation) | GROVER · MoleculeNet tasks | overall performance: 14.3% | Helped |
| add K-GNN pre-training alone (ablation) | Gode Case 0 · MoleculeNet tasks | overall performance: 4.5% | Helped |
| add contrastive learning and use both molecule and KG embeddings (ablation) | Gode Case 0 · MoleculeNet tasks | overall performance: 13.6% | Helped |
Surprising findings
- Omitting M-GNN pre-training improved performance on the Lipophilicity dataset, which the authors flag for further investigation.
- Removing the structural-similarity (neighbor 2D/3D) relations slightly improved BBBP and Tox21 results.
Limitations
- MolKG coverage is limited to the included PubChemRDF subdomains and PrimeKG drugs; the authors plan to expand it.
- Gain over GROVER is attributed to KG knowledge, but the KG-only embeddings alone (K-GNN without contrastive alignment) give only a modest 4.5% boost.
- Gode is evaluated only on MoleculeNet property prediction tasks; generalization to other downstream tasks is not shown.
Experimental setup
Datasets: ZINC15 ChEMBL PubChemRDF PrimeKG MoleculeNet BBBP SIDER ClinTox BACE Tox21 ToxCast FreeSolv
Benchmarks: BBBP SIDER ClinTox BACE Tox21 ToxCast FreeSolv ESOL Lipophilicity QM7 QM8
Metrics: ROC-AUC RMSE MAE
Baselines: GCN GIN SchNet MPNN DMPNN MGCN N-GRAM GROVER MGSSL KGE_NFM MolCLR KANO
Builds on: GROVER TransE G-Meta KANO MolCLR
Closest papers by meaning
- Enhancing Molecular Representation Learning Through the Combination of 3D and 2D Graph Machine Learning (Student Abstract) (AAAI 2025 · similarity 0.86)
- MOL-Mamba: Enhancing Molecular Representation with Structural & Electronic Insights (AAAI 2025 · similarity 0.86)
- Robust Heterogeneous Graph Classification for Molecular Property Prediction with Information Bottleneck (AAAI 2025 · similarity 0.86)
- A probability contrastive learning framework for 3D molecular representation learning (NeurIPS 2024 · similarity 0.86)
- UniCorn: A Unified Contrastive Learning Approach for Multi-view Molecular Representation Learning (ICML 2024 · similarity 0.85)
- Knowledge Enhanced Representation Learning for Drug Discovery (AAAI 2024 · 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.