Research map / Learning theory and algorithms / Combinatorial and search optimization
Retro-fallback: retrosynthetic planning in an uncertain world
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.
Reformulates retrosynthesis with stochastic feasibility and buyability models, defines successful synthesis probability (SSP), and proposes the greedy retro-fallback search that maximizes SSP for sets of backup plans.
Problem
Retrosynthesis search algorithms assume perfect knowledge of feasible reactions and buyable molecules, so plans may fail in the lab; existing methods do not produce sets of plans that serve as backups.
Approach
Model reaction feasibility and molecule buyability as binary stochastic processes, define SSP as the probability that at least one plan in a set succeeds, and greedily expand frontier molecules by expected SSP improvement estimated from samples, using recursive psi/rho value functions.
Key results
- Retro-fallback achieves higher mean SSP than MCTS, retro*, and breadth-first search across feasibility models with SA score and optimistic heuristics on 190 hard molecules.
- The gap is largest for independent feasibility models, where similar reactions provide backup plans; it shrinks under GP-induced correlations.
- Best individual plan success probability is similar across algorithms, so the advantage comes from finding effective backup plans.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use retro-fallback instead of breadth-first search, retro*, or MCTS | Retrosynthesis search on 190 hard USPTO molecules · USPTO hard test set | SSP | Helped |
| Use retro-fallback instead of baseline algorithms | Retrosynthesis search · USPTO hard test set | success probability of the best individual plan | No clear effect |
| Use retro-fallback instead of baseline algorithms | Retrosynthesis search · USPTO hard test set | fraction solved | Helped |
| Use retro-fallback instead of retro* | Retrosynthesis search · USPTO hard test set | shortest plan length | Mixed |
| Reduce the number of feasibility and buyability samples (ablation) | Retro-fallback · 25-molecule subsample | SSP | Hurt |
| Reduce the number of feasibility and buyability samples (ablation) | Retro-fallback · 25-molecule subsample | SSP variance | Hurt |
| Use more than 100 feasibility and buyability samples (ablation) | Retro-fallback · 25-molecule subsample | SSP | No clear effect |
| Use a GP-correlated feasibility model instead of independent feasibility outcomes (ablation) | Retrosynthesis search algorithms · USPTO hard test set | SSP | Hurt |
| Evaluate with the optimistic heuristic instead of the SA score heuristic (ablation) | Retro-fallback and baseline search algorithms · USPTO hard test set | SSP | No clear effect |
| Evaluate on easier molecules instead of hard molecules (ablation) | Retro-fallback and baseline search algorithms · Easier-molecule test set | SSP | No clear effect |
Surprising findings
- The SSP of the single best plan is similar across algorithms; the gain comes from backup plans rather than better primary plans.
- Performance gap is driven by a large difference on a small number of molecules rather than a consistent small improvement.
Limitations
- Feasibility models are synthetic and not validated against real chemistry, so SSP estimates are only as good as the assumed models.
- Retro-fallback is slower and may scale worse than other search algorithms.
- Does not account for plan length/cost or limits on number of backup plans.
Experimental setup
Datasets: USPTO eMolecules
Benchmarks: 190 hard molecules from USPTO test set
Metrics: successful synthesis probability fraction solved shortest plan length best individual plan success probability
Baselines: MCTS retro* breadth-first search proof number search
Builds on: retro* Monte Carlo Tree Search AO* syntheseus
Closest papers by meaning
- Retrosynthesis Planning via Worst-path Policy Optimisation in Tree-structured MDPs (NeurIPS 2025 · similarity 0.90)
- RetroBridge: Modeling Retrosynthesis with Markov Bridges (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)
- Trustworthy Retrosynthesis: Eliminating Hallucinations with a Diverse Ensemble of Reaction Scorers (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.