Research map / Image generation / Autoregressive and latent generation
Geometric-Facilitated Denoising Diffusion Model for 3D Molecule Generation
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 GFMDiff, a diffusion model for de novo 3D molecule generation with a triplet-aware Dual-Track Transformer denoiser and a Geometric-Facilitated Loss that penalizes valence mismatch to guide bond formation.
Problem
De novo 3D molecule generation with diffusion models struggles to capture multi-body interatomic geometry beyond pair-wise distances and generates bonds indirectly through predefined rules, hurting stability and validity.
Approach
A Dual-Track Transformer Network that jointly updates atom-pair and pair-triplet features (using distances and bond angles) as the E(n)-equivariant denoiser, plus a GFLoss that derives predicted valencies from distance-based bond predictions and penalizes mismatch with ground-truth valencies during training.
Key results
- On GEOM-QM9, GFMDiff reaches 87.7% molecule stability vs. 89.4% for GeoLDM in the bolded best? no: GeoLDM is best at 89.4%, GFMDiff 87.7% (second), and 96.3% validity vs. 93.8% for GeoLDM.
- On GEOM-Drugs, GFMDiff reports 86.5% atom stability and 3.9% molecule stability, outperforming GCDM's 3.7%.
- On conditional QM9 generation, GFMDiff achieves the lowest MAE on all six properties, e.g., 1.74 Bohr^3 for alpha vs. 1.97 for GCDM.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add triplet-wise geometric information through the pair-triplet attention track (ablation) | GFMDiff · GEOM-QM9 | Mol Stability (%): 85.9$\pm$0.2 -> 87.7$\pm$0.2 | Helped |
| Add GFLoss to guide bond formation (ablation) | GFMDiff · GEOM-QM9 | Validity (%): 95.2$\pm$0.0 -> 96.3$\pm$0.3 | Helped |
| Use GFMDiff instead of GCDM | QM9 conditional generation · QM9 conditional generation | α MAE: 1.97 -> 1.74 | Helped |
| Use GFMDiff instead of GCDM | QM9 conditional generation · QM9 conditional generation | Δε MAE: 602 -> 558 | Helped |
| Use GFMDiff instead of GCDM | QM9 conditional generation · QM9 conditional generation | εHOMO MAE: 344 -> 321 | Helped |
| Use GFMDiff instead of GCDM | QM9 conditional generation · QM9 conditional generation | εLUMO MAE: 479 -> 430 | Helped |
| Use GFMDiff instead of GCDM | QM9 conditional generation · QM9 conditional generation | μ MAE: 0.844 -> 0.728 | Helped |
| Use GFMDiff instead of GCDM | QM9 conditional generation · QM9 conditional generation | Cv MAE: 0.689 -> 0.593 | Helped |
| Use GFMDiff instead of GeoLDM | GFMDiff · GEOM-QM9 | NLL: -171.0$\pm$0.2 -> -128.0$\pm$0.2 | Hurt |
| Use GFMDiff instead of GeoLDM | GFMDiff · GEOM-QM9 | Mol Stability (%): 89.4$\pm$0.5 -> 87.7$\pm$0.2 | Hurt |
| Use GFMDiff instead of GCDM | GFMDiff · GEOM-Drugs | Mol Stable (%): 3.7$\pm$0.3 -> 3.9$\pm$0.2 | Helped |
Surprising findings
- GFMDiff has worse NLL than GCDM on QM9 (-128.0 vs. -171.0), though the authors still report it beats the other baselines.
- Removing the triplet track (w/o tri) hurts results more than removing GFLoss, according to the ablations.
Limitations
- Lower NLL than GCDM on QM9
- Drugs results only report stability; novelty not reported
- Molecule stability on QM9 is below GeoLDM
Experimental setup
Datasets: GEOM-QM9 GEOM-Drugs
Benchmarks: GEOM-QM9 GEOM-Drugs
Metrics: atom stability molecule stability validity uniqueness NLL MAE on conditional property prediction
Baselines: E-NF G-SchNet EDM GCDM GeoLDM Bridge+Force
Builds on: EDM GeoLDM GCDM
Closest papers by meaning
- Enhancing Generalizability in Molecular Conformation Generation with METRIZATION-Informed Geometric Diffusion Pretraining (AAAI 2025 · similarity 0.91)
- SubgDiff: A Subgraph Diffusion Model to Improve Molecular Representation Learning (NeurIPS 2024 · similarity 0.91)
- Equivariant Blurring Diffusion for Hierarchical Molecular Conformer Generation (NeurIPS 2024 · similarity 0.90)
- Towards Unified and Lossless Latent Space for 3D Molecular Latent Diffusion Modeling (NeurIPS 2025 · similarity 0.90)
- Graph Diffusion that can Insert and Delete (NeurIPS 2025 · similarity 0.90)
- All-atom Diffusion Transformers: Unified generative modelling of molecules and materials (ICML 2025 · similarity 0.90)
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.