Research map / Graph learning / Graph self-supervised learning
Enhancing Molecular Representation Learning Through the Combination of 3D and 2D Graph Machine Learning (Student Abstract)
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 a dual-GNN molecular representation learner that feeds SchNet 3D atomic embeddings into a GINE-based 2D GNN, pre-trained contrastively on conformer pairs, and evaluates frozen embeddings with random forests on MoleculeNet.
Problem
Self-supervised molecular representation learning typically relies on 2D graph information and unreliable positive pairs from graph augmentations, limiting the capture of 3D spatial information that affects properties like binding and solubility.
Approach
Use conformers of the same molecule (lowest-energy pairs) as contrastive positives, and combine an E(3)-invariant SchNet 3D encoder with a 2D GINEConv GNN (via a global dummy node) so the representation encodes both spatial arrangement and topology.
Key results
- Achieves best AUROC on BACE (0.828) and BBBP (0.731) among compared embeddings, with random forests on frozen embeddings.
- Achieves best RMSE on ESOL (1.597) and Lipophilicity (1.151).
- Does not beat baselines on Clintox (0.604 vs 0.882 for MACCS) or Freesolv (3.738 vs 3.676 for MACCS).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Combine 3D and 2D molecular information in the representation | Our Method · BBBP | AUROC: 0.717 (0.0349) -> 0.731 (0.0333) | Helped |
| Combine 3D and 2D molecular information in the representation | Our Method · ESOL | RMSE: 2.169 (0.1538) -> 1.597 (0.1123) | Helped |
| Combine 3D and 2D molecular information in the representation | Our Method · Lipophilicity | RMSE: 1.243 (0.0487) -> 1.151 (0.0410) | Helped |
| Combine 3D and 2D molecular information in the representation | Our Method · BACE | AUROC: 0.847 (0.0317) -> 0.828 (0.0325) | Hurt |
| Combine 3D and 2D molecular information in the representation | Our Method · Clintox | AUROC: 0.882 (0.0433) -> 0.604 (0.0891) | Hurt |
| Combine 3D and 2D molecular information in the representation | Our Method · Freesolv | RMSE: 3.676 (0.4578) -> 3.738 (0.4328) | Hurt |
Surprising findings
- Simple MACCS fingerprints outperform the proposed method on Clintox and Freesolv, which the authors attribute to small dataset size and overfitting risk.
Limitations
- Evaluation limited to a few MoleculeNet benchmarks and random forest probes; no comparison to stronger 3D or pretrained models.
- Underperforms simple fingerprints on Clintox and Freesolv.
- Freesolv dataset is small (~650 molecules), making embeddings prone to overfitting.
Experimental setup
Datasets: QM9 GEOM-Drugs MoleculeNet BACE MoleculeNet BBBP MoleculeNet Clintox MoleculeNet ESOL MoleculeNet Freesolv MoleculeNet Lipophilicity
Benchmarks: MoleculeNet BACE MoleculeNet BBBP MoleculeNet Clintox MoleculeNet ESOL MoleculeNet Freesolv MoleculeNet Lipophilicity
Metrics: AUROC RMSE
Baselines: MACCS ECFP RDKit HiMol MolCLR
Builds on: SchNet GINE MolCLR NT-Xent
Closest papers by meaning
- 3D Denoisers Are Good 2D Teachers: Molecular Pretraining via Denoising and Cross-Modal Distillation (AAAI 2025 · similarity 0.89)
- Empowering Active Learning for 3D Molecular Graphs with Geometric Graph Isomorphism (NeurIPS 2024 · similarity 0.89)
- Learning Over Molecular Conformer Ensembles: Datasets and Benchmarks (ICLR 2024 · similarity 0.88)
- HAGO-Net: Hierarchical Geometric Message Passing for Molecular Representation Learning (AAAI 2024 · similarity 0.87)
- MOL-Mamba: Enhancing Molecular Representation with Structural & Electronic Insights (AAAI 2025 · similarity 0.87)
- A probability contrastive learning framework for 3D molecular representation learning (NeurIPS 2024 · similarity 0.86)
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.