[Paper Review] Retro-fallback: retrosynthetic planning in an uncertain world

Austin Tripp, Krzysztof Maziarz+3 more|arXiv (Cornell University)|Oct 13, 2023

Machine Learning in Materials Science4 citations

TL;DR

This paper proposes retro-fallback, a novel retrosynthetic planning algorithm that explicitly accounts for uncertainty in reaction feasibility and buyability by modeling synthesis plans as stochastic processes. It maximizes the probability of at least one successful lab execution through greedy optimization of successful synthesis probability (SSP), outperforming MCTS and retro* in in-silico benchmarks by generating more robust, diverse backup plans.

ABSTRACT

Retrosynthesis is the task of planning a series of chemical reactions to create a desired molecule from simpler, buyable molecules. While previous works have proposed algorithms to find optimal solutions for a range of metrics (e.g. shortest, lowest-cost), these works generally overlook the fact that we have imperfect knowledge of the space of possible reactions, meaning plans created by algorithms may not work in a laboratory. In this paper we propose a novel formulation of retrosynthesis in terms of stochastic processes to account for this uncertainty. We then propose a novel greedy algorithm called retro-fallback which maximizes the probability that at least one synthesis plan can be executed in the lab. Using in-silico benchmarks we demonstrate that retro-fallback generally produces better sets of synthesis plans than the popular MCTS and retro* algorithms.

Motivation & Objective

• To address the critical gap in retrosynthesis planning where algorithms produce plans that fail in real labs due to uncertain reaction feasibility.
• To formalize uncertainty in reaction outcomes and molecule buyability as a stochastic process, enabling probabilistic evaluation of synthesis success.
• To develop a search algorithm that prioritizes plans with high probability of at least one working execution, improving reliability in real-world applications.
• To demonstrate that generating diverse, backup-ready plans is more effective than optimizing for a single optimal path under uncertainty.
• To provide a new evaluation metric—successful synthesis probability (SSP)—that captures the practical likelihood of lab execution.

Proposed method

• Formalize retrosynthesis as a stochastic process where each reaction's feasibility and each molecule's buyability are modeled as independent Bernoulli random variables.
• Define a new evaluation metric, successful synthesis probability (SSP), as the probability that at least one synthesis plan in a set will succeed in the lab.
• Propose retro-fallback, a greedy algorithm that selects actions to maximize SSP at each step, favoring nodes that increase the likelihood of at least one viable path.
• Use Monte Carlo sampling to estimate SSP for candidate reactions, enabling efficient approximation of the true success probability.
• Integrate SSP estimation into a tree search framework, where nodes are expanded based on their contribution to overall SSP rather than deterministic cost or quality.
• Maintain a set of candidate plans and dynamically re-evaluate SSP as new reactions are added, ensuring robustness against failure of individual paths.

Figure 2: Results with optimistic heuristic on “non-trivial” molecules. C and R refer to constant and rank marginal probabilities, while I and G refer to independent and GP-induced correlations. Solid lines are sample means (averaged across molecules), with error bars representing standard errors.

Experimental results

Research questions

• RQ1 Can retrosynthesis planning be improved by explicitly modeling uncertainty in reaction feasibility and molecule buyability?
• RQ2 Does optimizing for the probability that at least one plan succeeds (SSP) lead to more reliable lab outcomes than optimizing for a single optimal path?
• RQ3 How does retro-fallback compare to established algorithms like MCTS and retro* in terms of SSP and plan diversity under uncertainty?
• RQ4 Can a greedy algorithm based on SSP estimation outperform heuristic-based or cost-optimized approaches in real-world synthesis reliability?
• RQ5 What is the impact of plan diversity and backup planning on the overall success rate of retrosynthetic planning in uncertain environments?

Key findings

• Retro-fallback significantly outperforms MCTS and retro* in terms of successful synthesis probability (SSP) on in-silico benchmarks, demonstrating higher reliability under uncertainty.
• The algorithm generates more diverse and robust sets of synthesis plans, reducing the risk of total failure when primary plans fail.
• By prioritizing SSP over individual plan quality or cost, retro-fallback produces plans that are more likely to succeed in real lab settings.
• The method effectively mitigates the risk of 'hallucinated' reactions by favoring reactions with higher estimated feasibility, even when predictions are uncertain.
• Empirical results show that retro-fallback maintains high SSP across multiple target molecules, indicating consistent performance under varying uncertainty levels.
• The approach demonstrates that backup planning is not merely a side effect but a core design principle that enhances real-world applicability of retrosynthesis algorithms.

Figure F.1: Sample pairs of reactions where $K_{\text{total}}>0.8$ . Top: both reactions join a COCl group to an NH group in a ring to form molecules which differ only by the location of the Cl atom on the right side ring (far away from the reaction site). Middle: two reactions transforming a tert-b