acceptodds
Under review as a conference paper at ICLR 2027

LLM-Guided Dynamic Action Spaces for Synthesizable Molecular Optimization

Abstract

Synthesizable molecular optimization seeks to improve target properties while ensuring that molecular modifications follow feasible synthetic pathways. Existing synthesis-aware methods typically rely on exploring a large space of candidate transformations defined by reaction templates and purchasable building blocks. This search becomes even more challenging when property improvement requires multiple reaction steps, as the space expands further along the pathway. To address this challenge, we introduce **MolReAct**, which reformulates molecular optimization as search over compact reaction spaces proposed by a tool-augmented large language model (LLM). At each step, the LLM combines its prior chemical knowledge with cheminformatics tools to identify a molecule-specific set of compatible reactions, preserving synthesizability while making multi-step optimization feasible. Given this compact action space, we further leverage Group Relative Policy Optimization (GRPO) with the terminal oracle reward to improve long-term decision-making over multiple reaction steps. Across diverse molecular optimization tasks, **MolReAct** achieves the highest Top-10 score on 11 of 14 tasks and the best sample efficiency on 12 of 14 tasks, outperforming existing baselines under limited oracle budgets. Beyond these gains, **MolReAct** also provides each optimized molecule with a template-grounded synthetic pathway.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Synthesizable Molecular Optimization, Tool-Augmented Large Language Models, Multi-step Molecular Optimization

Reviews

Among:
Shape:
1234this paper, Bottom 9%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.50 over 2 reviews, above 9% of the others.
Reviewer_E9juOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes MolReAct, a framework for molecule-conditioned optimization. It reformulates molecular optimization as a search over compact reaction spaces proposed by a tool-augmented large language model. It proposes potential molecules based on chemical knowledge with cheminformatics tools, and also applies GRPO to optimize the generation process.

Critical strengths
  1. Comprehensive Empirical Evaluation: The results span 14 objectives and outperform several relevant baselines. Furthermore, the paper includes comparisons against random, greedy, and MCTS selection, alongside model-scale ablations. This provides strong evidence isolating the contributions of the proposer and policy.

  2. Tool-based Integration. Incorporating external chemical tools in LLMs is a good idea to improve the proposal of better potential molecules.

Critical weaknesses
  1. Novelty. The paper mainly integrates a standard LLM proposer and evaluator to establish the initial framework and then uses GROP to perform optimization. But all the methods have appeared in the existing literature, and I do not see much novelty here.

  2. Overstated Practical Synthesizability: The evidence supports template-grounded accessibility rather than true practical synthesizability. Compatible templates, purchasable reagents, and valid products are useful constraints. However, they do not guarantee reaction selectivity or experimental yields.

Questions for authors
  1. What is the novelty here? Could you please list some methodological novelties compared with existing similar work?

  2. This paper [1] also do molecule optimization; I am wondering about the performance comparison with it. [1] POLO: Preference-Guided Multi-Turn Reinforcement Learning for Lead Optimization, KDD 2026

  3. What fraction of successful trajectories require an intermediate decrease in the task-oracle score, and how does performance scale with maximum reaction depth?

Reviewer_JNTMOpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes MolReAct, which aims to improve molecular properties starting from given initial molecules under synthesis constraints. Its core design enables a tool-augmented large language model to construct a compact dynamic reaction action space: first matching 115 reaction templates, then using the frozen Llama-3.3-70B model combined with RDKit tools to propose reactions and additional reactants. After exact matching against the Enamine catalog and filtering via template execution, candidate reactions are retained. The Qwen3-4B policy model learns to select candidate reactions or stop via GRPO, using the final oracle model score as the reward, for up to 5 steps; meanwhile, a reaction cache is used to reduce repeated LLM inference.

Critical strengths

The methodological design in this paper addresses a practical search bottleneck issue and provides a clear constrained interface.

Under a budget that explicitly accounts for strategy training queries, the experimental results demonstrate broad coverage and do not rely on success in a single task.

Critical weaknesses

The core output of the synthetic pathway still lacks sufficient reliable and verifiable evidence, and the presented cases exhibit inconsistencies in specific structures. My main concern is that the existing pathway records themselves cannot be adequately validated, rather than viewing wet experiments as a necessary acceptance criterion for all such methods. This directly affects whether the central contribution of "synthesizable molecule optimization" holds true.

The existing ablation studies are insufficient to determine whether the gains stem from the emphasized multi-step long-term decision-making. The comparison between random templates and Tool-LLM primarily focuses on the quality of single-step proposals, without adequately testing a non-LLM candidate space that is equally compact and trained with the same trainer. Additionally, Tables 6 and 7 use final checkpoints, whereas Tables 1 and 4b employ candidates accumulated across checkpoints; thus, differences among these tables cannot be directly attributed to gains from any specific component.

Questions for authors

Q1. In Figure 8D, why does the urea formation step simultaneously convert the oxygen on the heterocycle to nitrogen? Which reactant provides the newly added urea carbonyl group? Please explain based on the actual reaction template and SMILES whether this is a drawing error or an execution result.

Q2. It is hoped that the author can conduct additional comparative experiments with maximum reaction depths of 1, 2, and 5 under the same oracle budget, in order to verify the independent benefits of multi-step planning. In Figures 8C and 8D, why did this strategy not stop at the intermediate structures with higher scores?

Q3. How were the 2500 endpoints in Appendix A.5 de-duplicated? If it was a resampling process, were the additional Oracle calls included in the budget? Did the main results, ablation studies, and baselines use the Top-10 consistent candidate set definition and aggregation rules?

Q4. Figure 4 specifies whether the candidate molecules of MolReAct are restricted to those generated solely from each starting point itself, while allowing the unconditional method to use the entire generation pool? Please clarify the source and sample size of each method, and explain whether the comparison of success rates is fair.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.