Retrorlff improves chemical retrosynthesis predictions using forward feedback
Single-step Retrosynthesis via Reinforcement Learning from Forward Feedback
Artificial Intelligence
Summary
Chemists try to figure out what starting materials create a desired product, but often multiple answers exist and models only learn from one example. The authors developed RetroRLFF, a method that uses feedback from predicting the forward reaction to better guess all possible starting materials. This approach improves the accuracy of identifying reactants and is preferred by expert chemists compared to previous methods. However, the method’s effectiveness in real lab experiments still needs separate confirmation.
What this means in practice
- •For chemical synthesis teams: Generate more accurate and chemically consistent reactant candidates for planning organic synthesis steps using improved retrosynthesis prediction.
- •For drug discovery teams: Assist in designing synthetic routes by providing better predictions of viable reactants from target drug molecules.
Authors
Zixun Lan, Zhenfu Liu, Yi Li, Xinheng Wang, Mingmin Chi, Qibiao Wang, Jie Zhu, Xiaohui Tian, Fei Ma
Abstract
Single-step retrosynthesis is a one-to-many inverse problem, yet most models are trained and evaluated against a single recorded reactant set. Reference-based training gives no explicit positive signal for unrecorded alternatives, and exact-match evaluation rejects non-matching predictions regardless of chemical plausibility. We propose Retrosynthesis via Reinforcement Learning from Forward Feedback (RetroRLFF), which optimizes reactant generation using both reference agreement and forward consistency with the target product. Starting from a supervised decoder-only Transformer, RetroRLFF generates complete reactant candidates and uses a frozen forward reaction model to predict products. A sequence-level reward combines reference agreement and forward consistency while penalizing invalid outputs. Direct Advantage-Based Policy Optimization updates the policy without human preference annotations or a preference-trained reward model. On USPTO-50K, RetroRLFF outperforms all 14 baselines in candidate-level round-trip accuracy at k = 1, 3, 5, and 10. With approximately 50 million parameters, it achieves 90.1% top-1 round-trip accuracy, exceeding R-SMILES, the strongest competing baseline on this metric, by 8.5 percentage points. Compared with its supervised policy, round-trip accuracy increases from 83.1% to 90.1%, while exact-match accuracy is maintained at 57.9% versus 57.5%. In an anonymized expert assessment of 131 randomly selected test cases, organic synthesis experts prefer RetroRLFF over EditRetro in 67.9% of comparisons. These results support forward feedback as a training signal for improving forward consistency while retaining top-1 reference recovery. Expert assessment provides complementary evidence of synthetic feasibility on the selected cases, although neither forward consistency nor expert preference establishes experimental feasibility.