Research map / Image generation / Autoregressive and latent generation
Graph Diffusion Transformers for Multi-Conditional Molecular 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 Graph DiT, a discrete graph diffusion transformer with a clustering-based condition encoder, AdaLN conditioning, and a graph-dependent noise model for predictor-free multi-conditional molecular and polymer generation.
Problem
Inverse molecular design under multiple simultaneous numerical and categorical property conditions (e.g., synthesizability, gas permeability, bioactivity) is underexplored for graph diffusion models, which struggle with heterogeneous property scales and mixed condition types and treat atom and bond
Approach
A Transformer graph denoiser conditioned via AdaLN on a condition encoder (clustering-based soft-assignment encoding for numerical properties, one-hot for categorical), trained with a graph-dependent transition matrix that couples node and edge noise, and sampled with predictor-free guidance.
Key results
- On 10K polymer generation, Graph DiT reports average MAE 0.9205 versus 1.1206 for the best baseline JTVAE-BO, a reduction of about 17.9%.
- On small molecule tasks, Graph DiT achieves property accuracy of 0.9135 (BACE), 0.9417 (BBBP), and 0.9777 (HIV), ranking first in average rank across the three tasks.
- In expert evaluation of O2/N2 polymer inverse design, the top-3 polymers were all generated by Graph DiT; median rank from multi-conditional generation was 4, 9, and 11 for Synth., O2Perm, and N2Perm versus over 30 for single-conditional sets.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use Graph DiT instead of the strongest baseline | multi-conditional polymer generation · Gas Perm | Avg. MAE: 1.1206 -> 0.9205 | Helped |
| use Graph DiT instead of Graph DiT-LCC | multi-conditional molecular generation · Synth. & BACE | Property accuracy: 0.9050 -> 0.9135 | Helped |
| use Graph DiT instead of Graph DiT-LCC | multi-conditional molecular generation · Synth. & BBBP | Property accuracy: 0.9390 -> 0.9417 | Helped |
| use Graph DiT instead of Graph DiT-LCC | multi-conditional molecular generation · Synth. & HIV | Property accuracy: 0.9766 -> 0.9777 | Helped |
| use Graph DiT instead of the strongest baseline | polymer generation · Synth. & O2 Perm | Avg. MAE: 3.2971 -> 2.9442 | Helped |
| use Graph DiT instead of the strongest baseline | polymer generation · Synth. & N2 Perm | Avg. MAE: 3.2911 -> 3.1124 | Helped |
| use Graph DiT instead of the strongest baseline | polymer generation · Synth. & CO2 Perm | Avg. MAE: 3.2622 -> 3.1155 | Helped |
| use Graph DiT instead of DiGress v2 | polymer inverse design · O2/N2 gas separation polymer design | Avg. MAE: 8.4707 -> 2.9770 | Helped |
| connect all generated graph components instead of keeping only the largest connected component (ablation) | Graph DiT · multi-conditional molecular generation | condition controllability | Helped |
| use clustering-based numerical condition encoding instead of direct or interval-based encoding (ablation) | Graph DiT · gas permeability condition control | condition controllability | Helped |
| use clustering-based numerical condition encoding instead of direct or interval-based encoding (ablation) | Graph DiT · generation diversity | diversity | Hurt |
| use AdaLN instead of In-Context conditioning or Cross-Attention (ablation) | Graph DiT · multi-conditional molecular generation | condition controllability | Helped |
Surprising findings
- Without rule-based checking, validity of DiGress, GDSS, and MOOD drops from around 0.99 to below 0.6, while Graph DiT keeps above 0.8.
- GDSS and MOOD achieve lower polymer MAE but poor distribution learning, which the authors attribute to carbon-element collapse acting as a confounder.
Limitations
- Trade-off between distribution fitting and internal diversity observed during training.
- Validity without rule checking is lower (0.82 vs 0.97 for LCC variant), so connecting components introduces some structural alteration.
- Expert evaluation involves only four polymer scientists and 25 polymers per case study.
Experimental setup
Datasets: polymer gas permeability dataset MoleculeNet HIV MoleculeNet BBBP MoleculeNet BACE
Benchmarks: polymer gas permeability (Synth., O2Perm, N2Perm, CO2Perm) MoleculeNet HIV, BBBP, BACE with synthetic score conditions O2/N2 gas separation polymer inverse design
Metrics: Validity Coverage Diversity Similarity Fréchet ChemNet Distance MAE Accuracy Avg. Rank Utility Score Agreement Score
Baselines: Graph GA MARS LSTM-HC JTVAE-BO GDSS DiGress DiGress v2 MOOD
Builds on: DiGress GDSS MOOD Diffusion Transformer (DiT) classifier-free guidance
Closest papers by meaning
- Graph Diffusion that can Insert and Delete (NeurIPS 2025 · similarity 0.91)
- Geometric-Facilitated Denoising Diffusion Model for 3D Molecule Generation (AAAI 2024 · similarity 0.89)
- Sampling 3D Molecular Conformers with Diffusion Transformers (NeurIPS 2025 · similarity 0.89)
- SubgDiff: A Subgraph Diffusion Model to Improve Molecular Representation Learning (NeurIPS 2024 · similarity 0.88)
- Graph Generation with Diffusion Mixture (ICML 2024 · similarity 0.88)
- Chain-of-Generation: Progressive Latent Diffusion for Text-Guided Molecular Design (NeurIPS 2026 · 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.