Research map / Probabilistic and scientific modeling / Molecules and proteins
GDiffRetro: Retrosynthesis Prediction with Dual Graph Enhanced Molecular Representation and Diffusion 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 GDiffRetro, a semi-template retrosynthesis model that combines dual-graph face features for reaction center identification with a 3D conditional diffusion model (EGNN-based) for completing synthons into reactants, reporting top-1 gains on USPTO-50k.
Problem
Semi-template retrosynthesis methods ignore face (ring/region) information in molecular graphs when identifying reaction centers, and generate reactants with 2D sequence or graph methods that ignore 3D structure.
Approach
Integrate each product's original molecular graph with its planar dual graph (faces as nodes) via RGCN to inject face information into atom embeddings, and generate missing reactant atoms with an E(3)-equivariant conditional diffusion model in 3D conditioned on the synthon's fixed coordinates.
Key results
- On USPTO-50k (unknown class), GDiffRetro reaches 58.9% top-1, 79.1% top-3, 81.9% top-5, above the best semi-template baseline (G2Gs/RetroDiff ~52.6%), a 12.0% relative top-1 gain in that category.
- Dual graph improves reaction center top-1 from 81.4% to 86.2% (ablation).
- With known reaction class, top-1 reaches 67.6%, a 4.3% relative gain over the best semi-template method.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add dual graphs to the reaction center model (ablation) | GDiffRetro · USPTO-50k reaction center prediction, unknown class | top-1 accuracy: 81.4 -> 86.2 | Helped |
| add dual graphs to the reaction center model (ablation) | GDiffRetro · USPTO-50k reaction center prediction, unknown class | top-2 accuracy: 93.4 -> 95.1 | Helped |
| add dual graphs to the reaction center model (ablation) | GDiffRetro · USPTO-50k reaction center prediction, unknown class | top-3 accuracy: 96.5 -> 97.4 | Helped |
| add dual graphs to the reaction center model (ablation) | GDiffRetro · USPTO-50k reaction center prediction, unknown class | top-5 accuracy: 98.6 -> 98.8 | Helped |
| add dual graphs to the reaction center model (ablation) | GDiffRetro · USPTO-50k reaction center prediction, unknown class | top-10 accuracy: 99.6 -> 99.6 | No clear effect |
| add dual graphs to the retrosynthesis model (ablation) | GDiffRetro · USPTO-50k retrosynthesis prediction, unknown class | top-1 accuracy: 53.1 -> 58.9 | Helped |
| add dual graphs to the retrosynthesis model (ablation) | GDiffRetro · USPTO-50k retrosynthesis prediction, unknown class | top-3 accuracy: 76.1 -> 79.1 | Helped |
| add dual graphs to the retrosynthesis model (ablation) | GDiffRetro · USPTO-50k retrosynthesis prediction, unknown class | top-5 accuracy: 79.7 -> 81.9 | Helped |
| add dual graphs to the retrosynthesis model (ablation) | GDiffRetro · USPTO-50k retrosynthesis prediction, unknown class | top-10 accuracy: 81.4 -> 83.8 | Helped |
Surprising findings
- Top-5 accuracy (known class) of MEGAN (87.5%) exceeds GDiffRetro (86.1%), attributed by authors to limited sampling iterations and synthon-reactant similarity reducing diversity.
Limitations
- Top-3 and top-5 accuracy only comparable to, not clearly above, template-free and semi-template baselines.
- Limited sampling (300 draws) may restrict candidate diversity for top-k results.
- Evaluated only on USPTO-50k; template-based methods are not directly comparable.
Experimental setup
Datasets: USPTO-50k
Benchmarks: USPTO-50k
Metrics: top-k exact match accuracy (k=1,3,5)
Baselines: MHNreact GLN LocalRetro GraphRetro RetroComposer Dual-TB Transformer SCROP Retroformer GTA Graph2SMILES Transformer (Aug.)
Builds on: RGCN EGNN G2Gs DDPM-style diffusion (Ho et al.) RetroDiff
Closest papers by meaning
- Advancing Retrosynthesis with Retrieval-Augmented Graph Generation (AAAI 2025 · similarity 0.92)
- RetroOOD: Understanding Out-of-Distribution Generalization in Retrosynthesis Prediction (AAAI 2024 · similarity 0.90)
- RetroBridge: Modeling Retrosynthesis with Markov Bridges (ICLR 2024 · similarity 0.90)
- RETRO SYNFLOW: Discrete Flow-Matching for Accurate and Diverse Single-Step Retrosynthesis (NeurIPS 2025 · similarity 0.89)
- Trustworthy Retrosynthesis: Eliminating Hallucinations with a Diverse Ensemble of Reaction Scorers (NeurIPS 2026 · similarity 0.88)
- LLM-Augmented Chemical Synthesis and Design Decision Programs (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.