Research map / Graph learning / Graph self-supervised learning
3D Infomax improves GNNs for Molecular Property Prediction
Hannes Stärk, Dominique Beaini, Gabriele Corso, Prudencio Tossou, Christian Dallago, Stephan Günnemann, Pietro Lió
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| 3D Infomax pre-training (contrastive 2D-3D MI maximization) | PNA pre-trained on QM9 subset · QM9 quantum properties | MAE (average ~22% reduction): ~22% average MAE reduction | Improved |
| 3D Infomax pre-training vs GraphCL | PNA · QM9 quantum properties (homo) | MAE: 68.96 vs 79.57 (GraphCL) | Improved |
| 3D Infomax pre-training on GEOM-Drugs vs GraphCL | PNA · GEOM-Drugs quantum properties (Gibbs, <E>) | MAE: .1811 vs .1941 (Gibbs) | Improved |
| Pre-training with 2D-only self-supervision (GraphCL node dropping) or property prediction (ProPred) | PNA · QM9 homo | MAE: ProPred 93.11 vs Rand Init 82.10 | Hurt |
| Distance prediction pre-training (DisPred) on GEOM-Drugs | PNA · QM9 quantum properties | MAE: e.g. mu 0.4626 vs 0.4133 worse; homo 80.58 vs 82.10 better | Mixed |
| Using multiple conformers (sum over conformer similarities in contrastive loss) (ablation) | 3D Infomax with QMugs pre-training · QM9 homo | MAE: 82.57 with single conformer vs 70.77 with three | Improved |
| Increasing number of GEOM-Drugs conformers beyond ~3 during pre-training (ablation) | 3D Infomax on GEOM-Drugs · GEOM-Drugs Gibbs and QM9 homo | MAE | Hurt |
| Distance encoding gamma (sin/cos high-frequency features) in 3D network (ablation) | Net3D · QM9 homo pre-training | MAE: 68.96 (Net3D) vs 70.34 (w/o gamma) | Improved |
| Net3D 3D network instead of SMP or EGNN (ablation) | 3D Infomax 3D network · QM9 homo pre-training | MAE: 68.96 vs SMP 72.37 and EGNN 70.46 | Improved |
| 3D Infomax on non-quantum OGB/MoleculeNet datasets (GEOM-Drugs pre-training) | PNA · 10 non-quantum datasets | RMSE / ROC-AUC: improved 4 of 10 datasets; never decreased performance | Mixed |
| 3D Infomax pre-training with RDKit-generated conformers in SMP baseline | SMP (3D GNN, fine-tuning with explicit RDKit 3D) · QM9 homo | MAE: 82.51 vs Rand Init 82.10 | Hurt |
Closest papers by meaning
- Neural P$^3$M: A Long-Range Interaction Modeling Enhancer for Geometric GNNs (NeurIPS 2024 · similarity 0.89)
- Empowering Active Learning for 3D Molecular Graphs with Geometric Graph Isomorphism (NeurIPS 2024 · similarity 0.89)
- RISE: Radius of Influence based Subgraph Extraction for 3D Molecular Graph Explanation (ICML 2025 · similarity 0.88)
- Enhancing Molecular Representation Learning Through the Combination of 3D and 2D Graph Machine Learning (Student Abstract) (AAAI 2025 · similarity 0.88)
- HAGO-Net: Hierarchical Geometric Message Passing for Molecular Representation Learning (AAAI 2024 · similarity 0.88)
- Pushing the Limits of All-Atom Geometric Graph Neural Networks: Pre-Training, Scaling, and Zero-Shot Transfer (ICLR 2025 · 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.
