Research map / Probabilistic and scientific modeling / Molecules and proteins
Pre-training via Denoising for Molecular Property Prediction
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| pre-training via denoising on PCQM4Mv2 added to GNS-TAT + Noisy Nodes (ablation) | GNS-TAT + NN · QM9 | MAE (12 targets): improves on all targets; SOTA on 10 of 12 | Helped |
| pre-training via denoising on PCQM4Mv2 added to TorchMD-NET (ablation) | TorchMD-NET · QM9 HOMO/LUMO | MAE: HOMO 18.1->15.6; LUMO 15.6->13.2 (vs +Noisy Nodes) | Helped |
| adding Noisy Nodes to TorchMD-NET (ablation) | TorchMD-NET · QM9 HOMO/LUMO | MAE: HOMO 22.0->18.1; LUMO 18.7->15.6 | Helped |
| pre-training via denoising on PCQM4Mv2 (ablation) | TorchMD-NET · MD17 aspirin forces | Test MAE: 0.268 -> 0.222 | Helped |
| pre-training via denoising on OC20 itself (denoising target = relaxed structure) (ablation) | GNS + Noisy Nodes · OC20 IS2RE validation | convergence speed / final validation performance: faster convergence, same final performance | Mixed |
| pre-training via denoising on PCQM4Mv2 (upstream from different domain) (ablation) | GNS + Noisy Nodes · OC20 IS2RE validation | validation performance: no benefit | No clear effect |
| pre-training via denoising on PCQM4Mv2 added to GNS-TAT + NN (fine-tuning on DES15K) (ablation) | GNS-TAT + NN · DES15K | test performance: further improvement over Noisy Nodes alone | Helped |
| AttrMask (2D atom-number masking) pre-training instead of denoising (ablation) | GNS-TAT + NN · DES15K | test performance: underperforms denoising | Hurt |
| freezing pre-trained backbone and training only decoder (HOMO) (ablation) | GNS-TAT pre-trained · QM9 HOMO | test MAE: 40 meV vs >100 meV with random backbone; worse than full fine-tuning | Mixed |
| GNS-TAT architecture change (vs GNS) with Noisy Nodes (ablation) | GNS + NN · QM9 | MAE (HOMO, LUMO, U0 etc.): HOMO 20.4->17.3; LUMO 17.5->17.1; U0 7.30->6.39 | Helped |
| pre-training data scaling (PCQM4Mv2 upstream size) (ablation) | GNS-TAT + NN · QM9 HOMO | test MAE: downstream performance generally improves as upstream data increases | Helped |
| pre-training via denoising on PCQM4Mv2 at small downstream dataset sizes (ablation) | GNS-TAT + NN · QM9 HOMO (varying downstream size) | test MAE: gain grows as downstream data shrinks | Helped |
Closest papers by meaning
- Sliced Denoising: A Physics-Informed Molecular Pre-Training Method (ICLR 2024 · similarity 0.91)
- Learning 3D Anisotropic Noise Distributions Improves Molecular Force Fields (NeurIPS 2025 · similarity 0.90)
- 3D Denoisers Are Good 2D Teachers: Molecular Pretraining via Denoising and Cross-Modal Distillation (AAAI 2025 · similarity 0.90)
- Structure-based drug design by denoising voxel grids (ICML 2024 · similarity 0.88)
- SubgDiff: A Subgraph Diffusion Model to Improve Molecular Representation Learning (NeurIPS 2024 · similarity 0.88)
- A Denoising Pre-training Framework for Accelerating Novel Material Discovery (AAAI 2025 · 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.