Research map / Image generation / Autoregressive and latent generation
Learning Generative Dynamics for 3D Molecule Generation via Sequential Neuro-Symbolic Constraints
Zhuoran Li, Xu Sun, Chang Chen, Wanyu Lin
Read the paperCheck your idea against 137,217 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 sequential neuro-symbolic constraints integrated into diffusion dynamics to make generated 3D molecules chemically valid, not just statistically plausible.
Problem
Diffusion models for 3D molecule generation learn atomic distributions implicitly and can produce statistically plausible but chemically invalid molecules.
Approach
Explicitly encodes physicochemical rules as sequential neuro-symbolic constraints into the generative dynamics of a 3D molecular diffusion model, rather than learning them implicitly from data.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Enforce all four constraints simultaneously instead of sequentially (ablation) | SensDiff · QM9 | molecule stability: 88.8% with one constraint -> 78.8% with all four | Hurt |
| Apply constraints in the coarse-to-fine order (global geometry first, valence last) instead of other permutations (ablation) | SensDiff · QM9 | validity: 74.6% to 90.9% across permutations | Improved |
| Activate fine-grained local rules early (ablation) | SensDiff · QM9 | validity | Hurt |
| Use hierarchical weights (global geometry over local bonds/valence) rather than flat or inverted (ablation) | SensDiff · QM9 | generation quality | Improved |
| Add linear ramp-up of newly introduced constraint weights (ablation) | SensDiff · QM9 | training stability | Improved |
| Add energy-adaptive weighting (ablation) | SensDiff · QM9 | generation quality | Improved |
| Apply constraints during training rather than sampling only (ablation) | SensDiff · QM9 | generation quality | Improved |
| Apply constraints more frequently during sampling (ablation) | SensDiff · QM9 | stability and constraint satisfaction | Improved |
| Add symbolic constraints to GEOLDM diffusion training | GEOLDM · QM9 | molecule stability: 86.2 -> 95.6 | Improved |
| Add symbolic constraints to GEOLDM diffusion training | GEOLDM · GEOM-Drugs | atom stability: 80.7 -> 87.9 | Improved |
| Add symbolic constraints to GEOLDM diffusion training | GEOLDM · QM9 conditional | alpha MAE: 2.46 -> 1.49 | Improved |
Closest papers by meaning
- Equivariant Neural Diffusion for Molecule Generation (NeurIPS 2024 · similarity 0.90)
- Geometric-Facilitated Denoising Diffusion Model for 3D Molecule Generation (AAAI 2024 · similarity 0.90)
- Navigating the Design Space of Equivariant Diffusion-Based Generative Models for De Novo 3D Molecule Generation (ICLR 2024 · similarity 0.89)
- Chain-of-Generation: Progressive Latent Diffusion for Text-Guided Molecular Design (NeurIPS 2026 · similarity 0.89)
- NExT-Mol: 3D Diffusion Meets 1D Language Modeling for 3D Molecule Generation (ICLR 2025 · similarity 0.89)
- All-atom Diffusion Transformers: Unified generative modelling of molecules and materials (ICML 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.
