Research map / Graph learning / Graph self-supervised learning
Fragment-based Pretraining and Finetuning on Molecular Graphs
Kha-Dinh Luong, Ambuj Singh
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 |
|---|---|---|---|
| Add fragment encoder in downstream finetuning (C -> CF) (ablation) | GraphFP contrastive · MoleculeNet 8 tasks | average ROC-AUC: 72.81 -> 73.73 | Improved |
| Add predictive pretraining to contrastive and fragment encoder (CF -> CPF) (ablation) | GraphFP · MoleculeNet 8 tasks | average ROC-AUC: 73.73 -> 74.01 | Improved |
| Add predictive pretraining to contrastive (C -> CP) (ablation) | GraphFP · MoleculeNet 8 tasks | average rank: 5.50 -> 5.19 | Improved |
| Add predictive pretraining to contrastive (C -> CP) (ablation) | GraphFP · MoleculeNet 8 tasks | average ROC-AUC: 72.81 -> 72.59 | No clear effect |
| Use predictive pretraining alone instead of contrastive (ablation) | GraphFP · MoleculeNet 8 tasks | average ROC-AUC: 72.81 -> 69.05 | Hurt |
| Use JT-style smaller fragmentation and add fragment encoder (JT-C -> JT-CF) (ablation) | GraphFP-JT · MoleculeNet 8 tasks | average ROC-AUC: 72.66 -> 70.21 | Hurt |
| Increase fragment vocabulary size from 800 to 1600 (ablation) | GraphFP-C · ClinTox | ROC-AUC: 78.6 -> 76.4 | Hurt |
| Increase fragment vocabulary size from 800 to 3200 (ablation) | GraphFP-C · ClinTox | ROC-AUC: 78.6 -> 75.5 | Hurt |
| Increase fragment vocabulary size from 800 to 3200 (ablation) | GraphFP-C · HIV | ROC-AUC: 76.3 -> 75.1 | Hurt |
| Increase fragment vocabulary size from 800 to 1600 (ablation) | GraphFP-C · SIDER | ROC-AUC: 61.4 -> 60.4 | Hurt |
| Use alpha=0.3 instead of 0.1 in CP loss weighting (ablation) | GraphFP-CP · MoleculeNet 8 tasks | average rank: 2.88 -> 2.13 | Improved |
| Use alpha=0.1 instead of 0.5 in CPF loss weighting (ablation) | GraphFP-CPF · MoleculeNet 8 tasks | average rank: 3.00 -> 1.63 | Improved |
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)
- Expressivity and Generalization: Fragment-Biases for Molecular GNNs (ICML 2024 · similarity 0.89)
- Towards Multiscale Graph-based Protein Learning with Geometric Secondary Structural Motifs (NeurIPS 2025 · similarity 0.88)
- Robust Heterogeneous Graph Classification for Molecular Property Prediction with Information Bottleneck (AAAI 2025 · similarity 0.88)
- Evaluating Representation Learning on the Protein Structure Universe (ICLR 2024 · similarity 0.88)
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.
