acceptodds
Withdrawn from ICLR 2027

Chain-of-Molecules: Agentic Reasoning in Chemical Space for Constrained Molecular Optimization

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

10 views

Abstract

Chain-of-thought reasoning has enabled language models to solve complex problems through explicit intermediate steps. We ask whether this paradigm can be transferred from language into chemical space. We introduce Chain-of-Molecules (CoM), a molecular-generation paradigm in which intermediate reasoning states are molecules and transitions between them are explicit chemical edits. In Agentic CoM, a symbolic environment executes valid edits, rejects infeasible ones, and returns the resulting molecule, allowing the model to explore alternatives, backtrack, and adapt its subsequent actions. We evaluate CoM in two complementary settings. First, a controlled property-targeting benchmark compares four generation paradigms across in-distribution, interpolation, and extrapolation regimes, disentangling the contributions of intermediate molecular states and interleaved environment execution. Second, we ask which generation paradigm yields the most adaptable base model for molecular optimization: starting directly from the respective pretrained base models, we use reinforcement learning to adapt them to three downstream bioactivity objectives without task-specific supervised finetuning. Agentic CoM performs strongest across both settings. Together, these results support explicit molecular trajectories and environment feedback not only as an effective optimization paradigm, but as a basis for reusable pretrained molecular optimizers. Code, data, and model checkpoints are publicly available.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
molecular generation, molecular optimization, drug discovery, bioactivity optimization, molecular reasoning

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_1DyGOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates the effectiveness of Agentic CoM, which performs iterative molecular editing using external tools for constrained molecular optimization. The authors define and compare four paradigms that take target molecular properties and a reference molecule (SMILES) as input and generate optimized molecules. Experimental results suggest that Agentic CoM achieves the best performance in terms of constraint satisfaction and molecular diversity.

Critical strengths
  1. The authors make an effort to clearly distinguish the four paradigms by providing intuitive figures and detailed explanations.

  2. Although the paper adopts independent experimental settings, the authors make an effort to ensure fair comparisons through consistent data splits and experimental configurations.

  3. The authors openly and honestly acknowledge the limitations of their work.

Critical weaknesses
  1. The motivation for the paradigm categorization and its novelty are unclear. The paper does not sufficiently justify the classification into four paradigms or explain how they relate to existing molecular generation/optimization methods. In particular, given the presence of prior studies on agentic molecular generation, the authors should clarify what distinguishes Agentic CoM from existing approaches and why it should be considered a new paradigm.

  2. The experimental analysis and interpretation of the results are insufficient. The paper adopts experimental settings that differ from prior work, making direct performance comparisons difficult. Moreover, the reported advantages of Agentic CoM and the verifiability of its trajectories appear largely expected, with limited additional insights. Although the paper reports findings on the benefits of intermediate molecular states, the importance of structural control, and the effects of scaling, SFT, and RL, it provides insufficient analysis of why these effects occur. In particular, the scaling experiments, which only consider 1M and 20M settings, are too limited to support general conclusions about scaling trends.

Questions for authors

Major Questions

  1. Is the primary contribution of this paper the proposal of a new methodology or an empirical comparison of four molecular generation paradigms? What is the key message that Agentic CoM aims to convey to the molecular generation community?

  2. What specific hypotheses are the SFT and RL experiments intended to validate? In particular, what is the motivation for including settings without SFT?

Minor Questions and Suggestions

  1. Section 2.5 (Evaluation Metrics) would be more appropriately placed in the experimental section. Additionally, important details for reproducibility, such as dataset construction, training configurations, and tool specifications, should be sufficiently described in the main text rather than being mostly deferred to the appendix.

  2. The authors should clearly distinguish experimental settings adopted from prior work from those newly introduced in this paper. For example, the rationale behind the training data size and the number of samples (128) should be provided.

  3. How is the candidate pool of atoms or functional groups used for addition operations in the molecular editing trajectory defined?

  4. Sections 4 and 5 could be combined into a single section for better organization.

Reviewer_sbUYOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

The paper proposes CoM, a combination of a small transformer that proposes molecular edits, an environment to executes and validate the proposals, and pipeline to generate training data and RL the model to produce molecules that have the desired property and structural constraints. The authors train 1/20M Qwen architecture transformers from scratch and evaluate CoM on molecular property optimization tasks and adaptations to downstream bioactivity objectives.

Critical strengths

Paper is well written and easy to read, good set of experiments covering in-distribution/interpolation/extrapolation settings, and models as small as 1M parameters doing well on this task is surprising.

Critical weaknesses
  1. In sections 4 and 5, the authors only compared 4 variants of their own modelling framework. There are no comparisons against existing molecular optimizers. For example the authors themselves implemented a GA that optimizes the same property to construct training data for SFT and RL, but the performance of this GA, or other standard baselines like Graph GA/REINVENT are never reported. Appendix D.5 does provide external results on TOMG-Bench, but there are no comparisons with matched training/oracle budgets.
  2. Follow up on the missing baselines, I think even adding a simple baseline, such as prompting an off the shelf LLM with the same restricted edit grammar and putting the RDKit computed properties and the edit distance constraint in the context, would be a very informative baseline. This should be a simple addition either using locally hosted models such as Qwen3.5/Gemma4, or frontier models via API.
  3. For each objective specific model, the reported RL training budget is around 2 million rollouts, which is a huge oracle budget for molecular optimization. For example, the practical molecular optimization benchmark which the authors cited, limits oracle budget to 10K because sample efficiency matters for molecular design.
Questions for authors

See weaknesses

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

This paper introduces Chain-of-Molecules (CoM), a molecular-generation framework in which autoregressive models produce sequences of intermediate molecules connected through chemical editing operations. The authors investigate whether explicitly representing molecular optimization as a sequential reasoning process improves generation compared with direct molecular prediction.

The proposed Agentic CoM variant additionally incorporates an external chemical environment that executes edits and provides feedback, enabling iterative refinement and correction. The framework is evaluated on constrained molecular property optimization across several generalization regimes and on downstream bioactivity objectives using reinforcement learning.

The results indicate that explicit molecular trajectories, particularly when combined with environment feedback, improve optimization performance over the alternative generation paradigms considered.

Critical strengths
  1. Meaningful formulation of sequential molecular reasoning.

The central question is relevant: whether intermediate molecular states can provide a useful computational structure for molecular generation, analogous to intermediate reasoning steps in language models. The distinction between self-generated trajectories and trajectories executed through an external chemical environment is conceptually interesting and provides a reasonable experimental framework for studying the contribution of environment feedback.

  1. Consistent improvements within the proposed experimental framework.

The experiments compare several generation paradigms and suggest that intermediate molecular trajectories and environment execution contribute positively to constrained optimization. The additional reinforcement-learning experiments provide evidence that the approach can be adapted beyond the original supervised objectives.

Critical weaknesses
  1. Insufficient evidence of competitiveness against established molecular optimization approaches.

The principal comparisons are between variants of the proposed autoregressive framework. Although these comparisons are useful for isolating architectural choices, they do not establish whether Agentic CoM constitutes a meaningful advance over established molecular optimization techniques.

In particular, sequential molecular editing, reinforcement-learning-based optimization, and iterative search through chemical space are well-established approaches. Comparisons against competitive molecular editing and optimization baselines, under comparable evaluation budgets, are necessary to establish the practical significance of the contribution.

The authors discuss existing molecular optimization benchmarks and explain why some are not directly applicable to their reference-conditioned setting. Nevertheless, this does not entirely resolve the absence of competitive external comparisons.

  1. Computational efficiency and practical relevance are insufficiently established.

The proposed approach requires generating and executing potentially long molecular trajectories rather than directly producing candidate molecules. This introduces additional inference-time computation and potentially substantial training costs associated with generating supervised trajectories.

The paper does not provide a sufficiently systematic analysis of the resulting computational trade-offs. In particular, it remains unclear whether the improvements persist when competing methods are allocated equivalent computational or molecular-evaluation budgets.

Furthermore, the controlled property-targeting tasks primarily involve relatively inexpensive molecular descriptors. Although downstream bioactivity experiments broaden the evaluation, the evidence remains limited regarding the benefits of CoM for more demanding structure-based molecular optimization involving docking or other computationally expensive objectives.

These limitations are important because the practical value of iterative molecular reasoning depends on whether its additional computational cost translates into meaningful optimization gains.

Questions for authors
  1. Computational efficiency: What are the complete training and inference costs of the different generation paradigms, including supervised trajectory construction, average trajectory length, environment calls, and wall-clock time per successful molecule? How do the methods compare under matched computational budgets?
  2. External baselines: How does Agentic CoM compare with competitive sequential molecular editing, reinforcement-learning, or search-based optimization methods under the same structural constraints and evaluation budget?
  3. Task complexity: Do the improvements persist for more demanding molecular optimization objectives, particularly structure-based tasks involving docking scores or expensive bioactivity oracles?

Comments on the paper

1 comment
AuthorsOpenReview ↗

Withdrawal statement

Dear Reviewers,

Thank you for evaluating our work. While your score-based assessments hurt a bit, we value your suggestions for improvement and believe we learned valuable lessons from your feedback.

We understand from all three reviews - and reviewer 1DyG raises exactly the right question (Question 1) - that we apparently failed to make our research question sufficiently explicit. Our goal is not to propose a new SOTA molecular optimization method. Rather, we present an empirical study of whether explicit reasoning in chemical space can be genuinely helpful for molecular optimization. While we understand reviewer 1DyG's perspective that some of the results may seem expected, we found that making a reasoning model work from scratch in a domain different from natural language is far from straightforward, and we believe that demonstrating its feasibility is already a meaningful contribution. Perhaps including some of our unsuccessful experiments would have helped to illustrate this challenge.

With this research question in mind, we believe it becomes clearer why direct molecular generation is the baseline we primarily care about. In fact, many established molecular optimization approaches, such as REINVENT, rely on direct SMILES generation models under the hood. We therefore see our work as largely orthogonal to approaches investigating which optimization algorithms perform best. Instead, we ask: Which type of generative model is a suitable backbone for molecular optimization? For example, REINVENT primarily provides an RL-based optimization framework using direct feedback from a scoring function. Our results suggest that using a CoM-style model instead of a direct generation model could be beneficial. In principle, these two directions could be combined.

Nevertheless, we understand the reviewers' rationale that comparisons against established external baselines are important to demonstrate the practical relevance of our approach.

Lastly, including information about computational costs is certainly feasible, and we will take this suggestion on board as well.

Taken together, we believe that clarifying our research question, restructuring the narrative accordingly, and including additional baselines and analyses would require substantial changes to the manuscript. As we want to respect best practices during the rebuttal period, we do not think it would be appropriate to ask the reviewers to evaluate a substantially revised version of the paper at this stage.

We have therefore decided to withdraw our submission, take the lessons learned from your feedback, and use them to improve our manuscript.

Thank you again for your time, thoughtful reviews, and constructive feedback.

Best regards,
The Authors

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

Related papers

Discussion (0)

Sign in to comment.