Research map / Image generation / Fast diffusion sampling
Conditional Synthesis of 3D Molecules with Time Correction Sampler
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 TACS, which combines training-free online guidance with a Time Correction Sampler that uses an EGNN time predictor to correct the effective diffusion timestep, improving property targeting on 3D molecules while keeping stability and validity.
Problem
Conditional 3D molecule generation with diffusion models suffers a trade-off: training-free online guidance hits target quantum properties but drives samples off the data manifold, lowering molecular stability and validity.
Approach
A time predictor (EGNN trained with cross-entropy to classify the diffusion timestep of noisy molecules) estimates the effective timestep after guidance; Tweedie's formula is then evaluated at the corrected timestep (with a time window clip), so online guidance is applied while samples stay on the correct marginal at each reverse step.
Key results
- On QM9, TACS achieves MAE 0.387 for mu (vs EDM 1.118, EEGSDE 0.854) with MS 83.3%, and MAE 332 meV for Delta epsilon (vs EEGSDE 539) with MS 88.8%.
- Online guidance alone (OG) reaches lowest MAE (e.g., 0.184 for Cv) but MS drops to 30.2%, showing the stability cost TACS aims to fix.
- TCS alone reaches 90.3% MS and 94.9% validity on Cv, surpassing EDM and EEGSDE in stability.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| combine online guidance with time correction | QM9 conditional molecular generation · QM9, ε_LUMO | MS: 21.1$\pm$0.5 -> 82.7$\pm$0.7 | Helped |
| combine online guidance with time correction | QM9 conditional molecular generation · QM9, ε_LUMO | Valid: 42.3$\pm$0.9 -> 91.3$\pm$0.2 | Helped |
| remove online guidance (z=0) (ablation) | TACS · QM9, ε_LUMO | MAE: 236 -> 493 | Hurt |
| remove online guidance (z=0) (ablation) | TACS · QM9, ε_LUMO | MS: 74.9 -> 91.7 | Helped |
| use expectation instead of argmax for timestep prediction (ablation) | TACS · QM9, α | MAE: 0.659 -> 0.703 | Hurt |
| use expectation instead of argmax for timestep prediction (ablation) | TACS · QM9, α | MS: 83.3 -> 84.6 | Helped |
| increase time window to Δ=16 (ablation) | TACS · QM9, α | MS: 86.08 -> 79.10 | Hurt |
| start TCS and OG at timestep 400 rather than 600 (ablation) | TACS · QM9, ε_LUMO | MS: 74.9 -> 86.6 | Helped |
| start TCS and OG at timestep 400 rather than 600 (ablation) | TACS · QM9, ε_LUMO | MAE: 236 -> 360 | Hurt |
| increase Monte Carlo samples from 1 to 5 (ablation) | TACS · QM9, α | MAE: 1.440 -> 1.395 | No clear effect |
| increase Monte Carlo noise scale to σ=0.3 (ablation) | TACS · QM9, α | MS: 86.35 -> 75.67 | Hurt |
| increase Monte Carlo noise scale to σ=0.3 (ablation) | TACS · QM9, α | MAE: 1.395 -> 2.936 | Hurt |
Surprising findings
- Expectation-based time prediction gave higher stability but higher MAE than argmax, and the novelty of TACS samples dropped to 71.6% versus 84.5% for EDM.
- Applying TCS and OG only from later steps (t=200) reduced performance again, which the authors leave unexplained.
Limitations
- Relies on a trained neural property predictor rather than exact computational chemistry for main QM9 experiments.
- Scaling exact quantum-chemistry guidance to QM9 is infeasible on current noisy quantum hardware.
- Sensitive to time window size; performance drops for large windows, and the method is evaluated mainly on QM9.
Experimental setup
Datasets: QM9 Geom-Drug
Benchmarks: QM9 Geom-Drug H3+ synthetic
Metrics: MAE molecule stability validity atom stability Tanimoto similarity novelty uniqueness
Baselines: EDM EEGSDE OG U-bound L-bound CFG+OG GDM G-SchNet EDM-Bridge
Builds on: EDM EEGSDE DPS Loss Guided Diffusion classifier-free guidance
Closest papers by meaning
- Graph Diffusion Transformers for Multi-Conditional Molecular Generation (NeurIPS 2024 · similarity 0.88)
- Navigating the Design Space of Equivariant Diffusion-Based Generative Models for De Novo 3D Molecule Generation (ICLR 2024 · similarity 0.87)
- Chemistry-Inspired Diffusion with Non-Differentiable Guidance (ICLR 2025 · similarity 0.87)
- Sampling 3D Molecular Conformers with Diffusion Transformers (NeurIPS 2025 · similarity 0.87)
- Straight-Line Diffusion Model for Efficient 3D Molecular Generation (NeurIPS 2025 · similarity 0.87)
- Equivariant Blurring Diffusion for Hierarchical Molecular Conformer Generation (NeurIPS 2024 · 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.