Research map / Reinforcement learning / MDPs and planning
Retrosynthesis Planning via Worst-path Policy Optimisation in Tree-structured MDPs
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.
Reframes multi-step retrosynthesis as worst-path optimisation in tree-structured MDPs and trains a Graph2Edits single-step policy via advantage-weighted self-imitation, achieving search-free route generation.
Problem
Multi-step retrosynthesis planning is invalid if any leaf of the synthetic tree is not a purchasable building block, yet existing methods optimise average or cumulative returns and rely on costly search at inference time.
Approach
Model retrosynthesis as a tree-structured MDP with a worst-path (min over root-to-leaf paths) objective, derive Bellman optimality results, and fine-tune a pre-trained single-step policy by advantage-weighted self-imitation of successful subtrees so it can generate routes without search.
Key results
- Achieves 100% success on Retro*-190 with 500 model calls (Retro* search) and 95.78% direct generation without search.
- Reaches 98.2% on ChEMBL-1000 and 99.5% on GDB17-1000 with 500 model calls.
- Average route length 3.92 on Retro*-190 vs 4.38 for Graph2Edits and 4.12 for MEGAN; about 92% of full performance with 10% of training data.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use InterRetro with Retro* search instead of PDVN with Retro* | Multi-step retrosynthesis planning · Retro*-190, 500 calls | success rate: 98.95 -> 100.00 | Helped |
| Use InterRetro with Retro* search instead of PDVN with Retro* | Multi-step retrosynthesis planning · ChEMBL-1000, 500 calls | success rate: 83.50 -> 98.20 | Helped |
| Use InterRetro with MCTS instead of RetroCaptioner with Retro* | Multi-step retrosynthesis planning · GDB17-1000, 500 calls | success rate: 75.20 -> 99.50 | Helped |
| Use InterRetro for direct generation rather than search-based planning | InterRetro · Retro*-190, ChEMBL-1000, and GDB17-1000 | success rate: 95.78%, 93.10%, and 89.00% | Helped |
| Use InterRetro instead of PDVN | Route generation on the shared Retro*-190 targets · Retro*-190 test set (138 shared targets) | average route length: 4.83 -> 3.92 | Helped |
| Train with 10% of the training set instead of the full set (ablation) | InterRetro · Retro*-190 | direct-generation success rate: 92% | Mixed |
| Add advantage weighting to imitation of successful past decisions (ablation) | InterRetro · Retro*-190 | direct-generation success rate: 16% to more than 90%; more than 95% with β ≥ 10 | Helped |
Surprising findings
- Search-free direct generation (95.78% on Retro*-190) exceeds the search-based success of most competing methods.
Limitations
- Assumes all proposed reactions are chemically feasible; no experimental validation.
- Objective ignores building-block cost, reaction conditions and total route size beyond depth.
- Route length is only evaluated on the 138 molecules solved by all methods.
Experimental setup
Datasets: USPTO-50k eMolecules
Benchmarks: Retro*-190 ChEMBL-1000 GDB17-1000
Metrics: success rate average route length sample efficiency
Baselines: MCTS Retro* Self-improve PDVN GraphRetro MEGAN Graph2Edits DreamRetroer RetroCaptioner LocalRetro
Builds on: Graph2Edits Retro* Monte Carlo Tree Search advantage-weighted imitation
Closest papers by meaning
- RetroBridge: Modeling Retrosynthesis with Markov Bridges (ICLR 2024 · similarity 0.90)
- Retro-fallback: retrosynthetic planning in an uncertain world (ICLR 2024 · similarity 0.90)
- Double-Ended Synthesis Planning with Goal-Constrained Bidirectional Search (NeurIPS 2024 · similarity 0.89)
- LLM-Augmented Chemical Synthesis and Design Decision Programs (ICML 2025 · similarity 0.89)
- RETRO SYNFLOW: Discrete Flow-Matching for Accurate and Diverse Single-Step Retrosynthesis (NeurIPS 2025 · similarity 0.88)
- GDiffRetro: Retrosynthesis Prediction with Dual Graph Enhanced Molecular Representation and Diffusion Generation (AAAI 2025 · similarity 0.87)
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.