Research map / Label-efficient and robust learning / Self-supervised and contrastive
DecoyDB: A Dataset for Graph Contrastive Learning in Protein-Ligand Binding Affinity Prediction
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.
Introduces DecoyDB, a large decoy-augmented protein-ligand complex dataset, and a graph contrastive pretraining framework (two-category RMSD-weighted InfoNCE plus denoising score matching) that improves GNN binding affinity prediction on PDBbind.
Problem
Binding affinity prediction is limited by scarce labeled protein-ligand complexes (PDBbind under 20K), and generic graph contrastive learning augmentations break biochemical structure and lack well-defined positive/negative complex pairs.
Approach
Construct a large unlabeled dataset of PDB complexes plus AutoDock Vina decoys annotated with RMSD, and use an InfoNCE loss whose negative weights scale with decoy RMSD, combined with a denoising score matching regularizer on ligand coordinates.
Key results
- On GIGN, pretraining with DecoyDB reduces RMSE from 1.460 to 1.386 on PDBbind core set 2013 and from 1.263 to 1.188 on core set 2016.
- On EGNN, the framework reduces RMSE on PDBbind 2016 from 1.304 to 1.250, versus 1.285 for ConBAP.
- On LP-PDBbind leakage-proof split, GIGN RMSE drops from 1.496 to 1.371 with pretraining.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Pre-train GIGN with OURS on DecoyDB | GIGN · PDBbind core set 2013 | RMSE: 1.460 (0.050) -> 1.386 (0.043) | Helped |
| Pre-train GIGN with OURS on DecoyDB | GIGN · PDBbind core set 2016 | RMSE: 1.263 (0.020) -> 1.188 (0.022) | Helped |
| Pre-train EGNN with OURS on DecoyDB | EGNN · PDBbind core set 2016 | RMSE: 1.304 (0.033) -> 1.250 (0.017) | Helped |
| Use DecoyDB pre-training | GIGN · LP-PDBbind split | RMSE: 1.496 (0.029) -> 1.371 (0.006) | Helped |
| Use DecoyDB pre-training | GIGN · Random split | RMSE: 1.294 (0.017) -> 1.269 (0.015) | Helped |
| Add decoy negative samples to the contrastive loss (ablation) | GIGN pre-training with one-category graph contrastive loss · PDBbind core set 2013 | RMSE: 1.406(0.047) -> 1.394(0.036) | Helped |
| Add decoy negative samples to the contrastive loss (ablation) | GIGN pre-training with one-category graph contrastive loss · PDBbind core set 2016 | RMSE: 1.246(0.014) -> 1.221(0.022) | Helped |
| Remove graph contrastive learning from the full pre-training framework (ablation) | GIGN pre-training with OURS · PDBbind core sets 2013 and 2016 | RMSE | Hurt |
| Remove DSM-based regularization from the full pre-training framework (ablation) | GIGN pre-training with OURS · PDBbind core sets 2013 and 2016 | RMSE | Hurt |
| Use edge-perturbation GCL instead of OURS | GIGN · PDBbind core sets 2013 and 2016 | RMSE | No clear effect |
| Use node-dropping GCL instead of OURS | GIGN · PDBbind core sets 2013 and 2016 | RMSE | No clear effect |
| Increase the number of fine-tuning labels from 2000 to 12000 | GIGN pre-trained with OURS · PDBbind fine-tuning across label sizes | RMSE | Helped |
Surprising findings
- General graph contrastive learning (edge perturbation or node dropping) gives no significant improvement on GIGN, and GCL-ND yields a lower R on core set 2013 (0.700).
Limitations
- Decoys are generated by a single docking program (AutoDock Vina), so decoy quality depends on its scoring and sampling.
- Gains are reported mainly on PDBbind core sets with small absolute differences and reliance on the GIGN base model for ablations.
- Pretraining is restricted to GNN-based models; other model families are trained only on PDBbind without pretraining.
Experimental setup
Datasets: DecoyDB PDBbind2013 PDBbind2016 LP-PDBbind Redocked 2020 PCQM4Mv2
Benchmarks: PDBbind core set 2013 PDBbind core set 2016 LP-PDBbind
Metrics: RMSE Pearson R
Baselines: AutoDock Vina DeepDTA GraphDTA Pafnucy OnionNet SchNet EGNN GIGN TorchMD-Net Frad ConBAP
Builds on: InfoNCE denoising score matching ConBAP Frad GIGN
Closest papers by meaning
- Protein-ligand binding representation learning from fine-grained interactions (ICLR 2024 · similarity 0.87)
- UniSite: The First Cross-Structure Dataset and Learning Framework for End-to-End Ligand Binding Site Detection (NeurIPS 2025 · similarity 0.87)
- Towards precision protein-ligand affinity prediction benchmark: A Complete and Modification-Aware DAVIS Dataset (NeurIPS 2025 · similarity 0.87)
- CBGBench: Fill in the Blank of Protein-Molecule Complex Binding Graph (ICLR 2025 · similarity 0.87)
- M²N: A Progressive Macro-to-Micro 3D Modeling Scheme for Unveiling Drug-Target Affinity (AAAI 2025 · similarity 0.86)
- PPDiff: Diffusing in Hybrid Sequence-Structure Space for Protein-Protein Complex Design (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.