Research map / Image generation / Autoregressive and latent generation
SubgDiff: A Subgraph Diffusion Model to Improve Molecular Representation Learning
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.
Proposes SubGDiff, a diffusion model that adds noise to sampled subgraphs (via a mask) at each step, with subgraph prediction, expectation-state diffusion and k-step same-subgraph diffusion, used both for molecular pretraining and conformation generation.
Problem
Existing molecular diffusion models add independent Gaussian noise to each atom, ignoring substructure dependencies, which limits their usefulness for molecular representation learning and generation.
Approach
Introduce a subgraph mask sampled at each diffusion step so that only a connected torsion-based subgraph is noised, add a subgraph-prediction BCE loss on the shared denoiser, and use an expectation-state forward approximation plus k-step same-subgraph diffusion to make training and sampling tractable.
Key results
- On MD17 force prediction, SubGDiff reports the best MAE on all eight molecules (e.g., Aspirin 0.880 vs. 1.107 for the next best GeoSSL).
- On MoleculeNet scaffold split (GIN backbone), SubGDiff achieves average ROC-AUC 74.85 vs. 73.73 for MoleculeSDE (VP).
- On GEOM-QM9 conformation generation with 500 steps, SubGDiff gets COV-R mean 89.78% vs. 80.20% for GeoDiff at the same step count.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SubGDiff pretraining instead of MoleculeSDE (VP) | GIN molecular property prediction · MoleculeNet (2D topology only), Avg | ROC-AUC: 73.73 -> 74.85 | Helped |
| Use SubGDiff pretraining instead of the strongest listed baseline | MD17 force prediction · MD17 force, eight molecules | MAE | Helped |
| Replace GeoDiff with SubGDiff | GFN conformation generation · GEOM-QM9, 5000 timesteps | COV-R mean (%): 80.36 -> 90.91 | Helped |
| Replace GeoDiff with SubGDiff | GFN conformation generation · GEOM-QM9, 5000 timesteps | COV-P mean (%): 53.66 -> 50.16 | Hurt |
| Replace GeoDiff with SubGDiff | GFN conformation generation · GEOM-QM9, 5000 timesteps | MAT-P mean (Å): 0.6673 -> 0.6114 | Helped |
| Replace GeoDiff with SubGDiff | GFN conformation generation · GEOM-QM9, 500 timesteps | COV-R mean (%): 80.20 -> 89.78 | Helped |
| Replace GeoDiff with SubGDiff | GFN conformation generation · GEOM-QM9, 500 timesteps | MAT-R mean (Å): 0.3617 -> 0.2417 | Helped |
| Use SubGDiff instead of GeoDiff | Conformation generation · GEOM-QM9, 500 timesteps | COV-R mean (%) | Helped |
Surprising findings
- On GEOM-QM9 with 5000 steps, SubGDiff's COV-P mean is lower than GeoDiff's (50.16 vs 53.66), so it does not win on all metrics.
Limitations
- Subgraph distribution is fixed to a torsion-based decomposition, not flexible.
- Gains on COV-P at 5000 steps are negative relative to GeoDiff.
Experimental setup
Datasets: PCQM4Mv2 MD17 MoleculeNet GEOM-QM9 GEOM-Drugs
Benchmarks: MD17 MoleculeNet GEOM-QM9 GEOM-Drugs
Metrics: mean absolute error (force) ROC-AUC COV-R MAT-R COV-P MAT-P
Baselines: GeoDiff MoleculeSDE GraphMVP GeoSSL InfoNCE EBM-NCE 3D InfoMax AttrMask
Builds on: DDPM GeoDiff MoleculeSDE Torsional diffusion MDSM
Closest papers by meaning
- Geometric-Facilitated Denoising Diffusion Model for 3D Molecule Generation (AAAI 2024 · similarity 0.91)
- 3D Denoisers Are Good 2D Teachers: Molecular Pretraining via Denoising and Cross-Modal Distillation (AAAI 2025 · similarity 0.90)
- Graph Diffusion that can Insert and Delete (NeurIPS 2025 · similarity 0.89)
- Equivariant Blurring Diffusion for Hierarchical Molecular Conformer Generation (NeurIPS 2024 · similarity 0.88)
- Graph Diffusion Transformers for Multi-Conditional Molecular Generation (NeurIPS 2024 · similarity 0.88)
- DiffMS: Diffusion Generation of Molecules Conditioned on Mass Spectra (ICML 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.