Research map / Graph learning / Graph self-supervised learning
Does GNN Pretraining Help Molecular Representation?
Ruoxi Sun, Hanjun Dai, Adams Wei Yu
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 |
|---|---|---|---|
| add self-supervised pretraining (node, context, motif, contrastive) | GIN with rich features · MoleculeNet 5 datasets, balanced scaffold | average gain over no pretrain: 0.122 / 0.126 / 0.128 / -0.578 | No clear effect |
| add supervised ChEMBL pretraining after self-supervised | GIN with rich features · MoleculeNet, balanced scaffold | average gain over no pretrain: 0 -> 1.638 | Improved |
| switch from balanced scaffold to scaffold split (self-supervised pretraining) (ablation) | GIN with rich features · MoleculeNet | average gain over no pretrain: Node Prediction 0.122 -> 0.990 | Mixed |
| use basic features instead of rich features (ablation) | no-pretrain GIN · MoleculeNet, balanced scaffold | average ROC-AUC: 78.0 -> 76.7 | Hurt |
| use basic features instead of rich features (ablation) | no-pretrain GIN · MoleculeNet, scaffold split | average ROC-AUC: 71.8 -> 68.7 | Hurt |
| add context-prediction pretraining when only basic features are used (scaffold) (ablation) | GIN basic features · MoleculeNet, scaffold | average gain over no pretrain: 1.792 | Improved |
| pretrain on SAVI (about 500x larger than ZINC15) (ablation) | GIN pretraining · MoleculeNet, balanced scaffold and scaffold | average gain over no pretrain: Context Prediction 0.126 -> 0.546 | No clear effect |
| tune learning rate per setup and average over three splits instead of defaults (ablation) | pretraining evaluation protocol · MoleculeNet | pretraining benefit | Hurt |
| GraphSAGE backbone instead of GIN (ablation) | graph pretraining · MoleculeNet | average gain over no pretrain | No clear effect |
| 7 GNN layers instead of 5 (ablation) | node-prediction pretraining · MoleculeNet | ROC-AUC | No clear effect |
| use released GROVER pretrained weights instead of training from scratch | GROVER base · BACE | ROC-AUC: 86.2 -> 85.1 | No clear effect |
Closest papers by meaning
- Pre-Training Graph Neural Networks on Molecules by Using Subgraph-Conditioned Graph Information Bottleneck (AAAI 2025 · similarity 0.91)
- On the Scalability of GNNs for Molecular Graphs (NeurIPS 2024 · similarity 0.90)
- Exploring Molecular Pretraining Model at Scale (NeurIPS 2024 · similarity 0.88)
- Extracting Training Data from Molecular Pre-trained Models (NeurIPS 2024 · similarity 0.88)
- MORE: Molecule Pretraining with Multi-Level Pretext Task (AAAI 2025 · similarity 0.88)
- Graph Generative Pre-trained Transformer (ICML 2025 · 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.
