Research map / Probabilistic and scientific modeling / Molecules and proteins
Order Matters in Retrosynthesis: Structure-aware Generation via Reaction-Center-Guided Discrete Flow Matching
Chenguang Wang, Zihan Zhou, LEI BAI, Tianshu Yu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add positional embedding (absolute PE) to RC-rooted ordering (ablation) | RetroDiT-8M · USPTO-50k | Top-1 (%): 50.4 -> 68.5 | Improved |
| use RoPE instead of absolute PE (ablation) | RetroDiT-8M · USPTO-50k | Top-1 (%): 68.5 -> 71.0 | Improved |
| use RoPE instead of absolute PE (ablation) | RetroDiT-8M · USPTO-50k | Top-10 (%): 94.8 -> 96.5 | Improved |
| use oracle reaction centers instead of predicted | RetroDiT · USPTO-50k | Top-1 (%): 61.2 -> 71.1 | Improved |
| use oracle reaction centers instead of predicted | RetroDiT · USPTO-Full | Top-1 (%): 51.3 -> 63.4 | Improved |
| add RC-rooted ordering instead of canonical ordering (ablation) | RetroDiT 65M · USPTO-Full | Top-1 (%): 50.4 -> 51.2 | Improved |
| add oracle RC ordering to a small model instead of scaling to 65M canonical (ablation) | RetroDiT 280K · USPTO-Full | Top-1 (%): 50.4 -> ~51 | Improved |
| use RC-rooted ordering at low RC accuracy (below 35-40%) (ablation) | RetroDiT-8M · USPTO-50k and USPTO-Full | Top-1 (%): 35--40% | Hurt |
| use RC-rooted ordering at 70% RC accuracy (ablation) | RetroDiT-8M · USPTO-50k and USPTO-Full | Top-1 (%): over 10 points | Improved |
| use structure-aware flow model instead of RetroBridge diffusion | Ours (Pred. RC) · USPTO-50k | Top-1 (%): 50.8 -> 61.2 | Improved |
| use structure-aware flow model instead of RetroSynFlow | Ours (Pred. RC) · USPTO-50k | Top-1 (%): 60.0 -> 61.2 | Improved |
| use DiT backbone instead of Graph Transformer (ablation) | RetroDiT · USPTO-Full | accuracy | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- RetroSynFlow — RETRO SYNFLOW: Discrete Flow-Matching for Accurate and Diverse Single-Step Retrosynthesis (NeurIPS 2025) · better on 1 of 1 shared result
- RetroBridge — RetroBridge: Modeling Retrosynthesis with Markov Bridges (ICLR 2024) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Matching-while-Decoding: Enhancing Template-Free Retrosynthesis via Explicit Structural Alignment (NeurIPS 2026 · similarity 0.93)
- RETRO SYNFLOW: Discrete Flow-Matching for Accurate and Diverse Single-Step Retrosynthesis (NeurIPS 2025 · similarity 0.91)
- RetroBridge: Modeling Retrosynthesis with Markov Bridges (ICLR 2024 · similarity 0.90)
- Advancing Retrosynthesis with Retrieval-Augmented Graph Generation (AAAI 2025 · similarity 0.90)
- LLM-Augmented Chemical Synthesis and Design Decision Programs (ICML 2025 · similarity 0.89)
- GDiffRetro: Retrosynthesis Prediction with Dual Graph Enhanced Molecular Representation and Diffusion Generation (AAAI 2025 · similarity 0.89)
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.
