acceptodds
Under review as a conference paper at ICLR 2027

MOLT: A Self-Evolving Agentic System for Constrained Multi-Property Lead Optimization

Abstract

Lead optimization is inherently iterative, yet current agentic molecular optimizers do not systematically improve from the experience they accumulate. Optimization history is often retained as tool calls or trajectories, leaving the chemical interventions associated with their outcomes difficult to reuse. We introduce MOLT, a self-evolving agentic system for constrained multi-property lead optimization. MOLT consolidates evaluated molecular transitions into executable chemical transformations with outcome evidence; a context-conditioned action policy selects among these transformations, and an LLM controller plans across retained optimization branches. Exploration expands the available chemical knowledge, while optimization decisions provide supervision for updating the action policy. Across five constrained lead-optimization tasks, MOLT achieves the highest success rate on every evaluated task, improving over prior lead-optimization agents on the three shared tasks and over generative baselines on both dual-target tasks.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Lead Optimization, Molecular Optimization, AI Agents, Drug Discovery

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_DPQfOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes MOLT, a self-evolving agentic system for lead optimization. The proposed method converts evaluated molecular transitions into reusable chemical transformations with property evidence. The context-conditioned action policy ranks applicable transformations, and the agent's decisions are used to further update the action policy. Based on the top candidates, an LLM controller plans across retained search branches. The method is evaluated on five constrained lead-optimization tasks, with planning ablations, case studies, and an out-of-distribution ChEMBL generalization experiment.

Critical strengths
  1. The paper is easy to follow. The challenges in learning from molecular optimization experience are clearly illustrated and motivate the design of MOLT.

  2. The proposed method achieves substantial improvements over the agentic baselines on the primary five constrained lead-optimization tasks. The planning ablation also supports the benefit of retaining alternative search branches over other settings. The evaluation on ChEMBL leads shows the approach's ability to transfer beyond the main task.

Critical weaknesses
  1. The contribution of the context-conditioned action policy and its self-evolving updates is not fully ablated. It is unclear whether the self-evolving updates provide substantial improvements over a fixed chemical knowledge base and a frozen action policy, or whether most of the gains come from the learned transformation ranking and retained-branch planning.

  2. Some important implementation details are missing or unclear, like the exact LLM for MOLT planning, the setting of feedforward network for action policy, the prompts, and search budget... The main performance tables do not report variability across independent runs, which makes difficult to assess the robustness of the reported improvements. Since the framework relies on an LLM controller, it is also unclear whether the observed performance generalizes across different LLM backbones or depends on the specific controller used in the experiments.

Questions for authors
  1. Have the authors considered an ablation to compare the action-policy-selected Top-5 transformation against randomly selecting five applicable transformations, or selecting them using historical rule statistics, or selecting by LLM controller?

  2. According to the second weakness, can the authors provide the exact LLM model and its settings, action policy neural network settings, prompts, and optimization budget? If the main results were obtained from multiple runs, reporting mean and standard deviation across seeds would help establish their stability.

  3. Case study 1 shows how accumulated experience introduces and strengthens a useful but late transformation. However, can the same mechanism reinforce suboptimal actions, like the model strengthening its belief in local-optima evidence? It would be interesting to examine failure cases where the agent favors certain transformations while overlooking potentially better frontier branches, and whether the retained-branch mechanism successfully redirects the search toward other feasible paths.

  4. A minor presentation comment, the caption of Table 2 should be placed above the table.

1 comment

W1. Do the self-evolving updates improve over a fixed knowledge bank and frozen action policy, or do most gains come from transformation ranking and retained-branch planning?

Response:

Thank you for your question. The results show that transformation ranking, retained-branch planning, and accumulated knowledge each contribute at a different stage of MOLT.

With the same knowledge bank, the context-conditioned policy improves Hit@5 from 53.6% with historical rule statistics to 70.7% (Fig. 3c). With the same bank and action policy, retained-branch planning improves SR from 58.9% for a single trajectory to 69.6%, while later success increases from 32.4% to 48.9% (Table 2). We further evaluated how performance changes as experience accumulates. With the planner and search budget fixed, increasing exploratory experience from 256 to 32K transitions raises end-to-end SR from 19.5% to 54.4%:

Exploratory transitions2562K8K16K32K
Mean SR (%)19.538.642.643.754.4

The experience-growth curve is reported in Appendix A.6 (Fig. 5) of the revised manuscript.

W2/Q2. What are the implementation settings, and how stable are the results across seeds and controller backbones?

Response:

Thank you for asking for these implementation details and robustness checks. The main experiments use Qwen2.5-7B-Instruct without fine-tuning. At each step, the prompt shows the current frontier molecules with their depth, SMILES, similarity to the starting lead, and RDKit descriptors, and asks the controller which branch to expand. The search uses the top five transformations per expanded node, a maximum depth of 10, and a budget of 10 committed outputs. It stops when a feasible molecule is found or the output budget is exhausted. Unselected branches remain available for later backtracking. BBBP, HIA, and Mutagenicity are evaluated only for committed outputs, averaging 4.6 ADMET evaluations per query; unselected branches do not incur these evaluations. The complete search takes 3.4 seconds per query on a single NVIDIA A5000 GPU, including an average of 4.6 controller calls. The controller settings, search budget, computational profile, and representative prompt are reported in Appendix A.13 and Table 12.

To directly assess variability and controller dependence, we additionally ran a robustness study on the same fixed set of 500 queries, with 100 queries from each task, using three independent seeds. The knowledge bank, action policy, prompt, search budget, and evaluation protocol were held fixed across all runs. The SR has a small standard deviation for all three controllers.

ControllerCNSOralSafetyCNS+SafetyOral+Safety
Qwen2.5-7B90.7 ± 0.573.3 ± 0.584.7 ± 0.953.3 ± 0.970.3 ± 0.5
Llama-3.1-8B89.7 ± 1.274.0 ± 2.284.0 ± 0.052.3 ± 0.570.7 ± 1.2
Qwen2.5-0.5B81.7 ± 1.766.0 ± 2.281.0 ± 1.646.0 ± 2.966.7 ± 1.2

Q1. How does the learned Top-5 transformation selector compare with random selection, historical rule statistics, and LLM-based selection?

Response:

Thank you for the question. Appendix Fig. 7 provides this ablation for the transformation selector. Using the same knowledge bank, the context-conditioned policy achieves 70.7% Hit@5 at K=5K=5, compared with 53.6% for historical rule-statistics ranking, while random ranking is substantially lower. The LLM serves a different role: it does not select the Top-5 transformations, but chooses which retained frontier branch to expand after the shortlist has been produced. This separation between transformation ranking and frontier-branch selection is described in Secs. 5.2–5.3 of the revised manuscript.

Q3. Can accumulated evidence reinforce locally attractive but suboptimal actions, and can retained branches recover by returning to a better frontier path?

Response:

Thank you for raising this point. This can happen: a transformation that is useful on average may still be suboptimal for a particular molecule. MOLT re-scores transformations for the current molecule and task, while retaining alternative frontier branches rather than committing permanently to the first choice. Fig. 2b shows a case where the initially attractive branch violates the similarity constraint, while another retained branch later reaches a feasible solution. Consistently, retained branches increase later success from 32.4% to 48.9% (Table 2).

Q4. Please move the caption of Table 2 above the table.

Response:

Thank you for noting this. We have moved the caption above Table 2 in the revised manuscript.

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

MOLT builds a library of reusable molecular edits from exploratory optimization and the associated property evaluations. A policy uses the molecule and task context to shortlist edits, while an LLM controller chooses which branch of the search to continue. The system is designed to update the library and policy through further exploration and planner decisions. The evaluation covers five constrained ADMET tasks and a ChEMBL transfer set, with success requiring the property targets and guardrails, chemical validity, and similarity to the starting molecule.

Critical strengths
  1. Storing experience as executable edits, together with their observed property changes, gives the system a useful way to reuse previous exploration. Evaluating target improvements alongside constraints on other properties also makes the task more informative than optimizing a single score. The paper is careful about what these results mean, distinguishing predicted properties from experimental validation and fingerprint similarity from exact scaffold retention.

  2. The results are encouraging: MOLT reaches 81.8% CNS success versus 51.2% for TRACE-Parallel, and 81% versus 45% on 100 ChEMBL leads. The planning ablation is also useful. Retaining alternative branches raises aggregate SR from 58.9% for a single trajectory to 69.6% in Table 2, providing evidence for this design choice beyond the individual case studies.

Critical weaknesses
  1. The main question for me is how much the decision-guided updates contribute to the reported performance. Figure 3c compares the bootstrap policy with rule-statistics ranking, while Case 3 shows a useful multi-step path; neither gives a before/after comparison of the updated policy. The library versions are also a little unclear: the initial 468 rules in Section 6.1 match the size of Table 6's final 128K-transition snapshot, but the version used for each end-to-end result is not identified. Section 5.3 labels selected actions positive without explaining how unsuccessful continuations are handled. Comparing the bootstrap and updated policies with the library fixed would give more direct evidence for this part of the self-evolution claim.

  2. More detail about the search cost would make the efficiency comparison easier to assess and reproduce. Each query allows ten selected outputs, but Section 5.2 retains all valid products at matching sites and supplies auxiliary property estimates when choosing a branch. It is therefore unclear how many candidates receive property evaluations, or how those estimates are obtained. Table 9 reports no neural molecular generation, while describing runtime as "seconds" and the controller as a "local 7B LLM." The controller configuration, evaluation counts, and precise latency would help readers understand the cost of the full search.

Questions for authors
  1. Which library and policy versions were used for Tables 1, 2, and 8, and were they frozen during test evaluation? A comparison of the bootstrap and updated policies on one representative held-out task, with the same library and search budget and with variability reported, would help separate the benefit of decision learning from library growth.

  2. Are all selected actions used as positive labels, or are they filtered by eventual trajectory success? Could the authors describe the update loss and schedule, including how these labels are mixed with the bootstrap data?

  3. Which controller checkpoint and prompt were used, and where do the auxiliary property estimates come from? It would also be useful to report candidate and property-evaluation counts, including unselected branches, alongside precise latency under the stated stopping rule.

  4. A short discussion of RePO, which is already cited, could help explain the choice of selected-action supervision here, compared with reference guidance and verifiable molecular rewards. For the broader discussion of self-evolution, AlphaApollo may also be useful for explaining how MOLT verifies and retains useful chemical edits over successive rounds.

2 comments

W1. How much do decision-guided updates contribute beyond the bootstrap policy when the knowledge bank and search budget are fixed?

Response:

Thank you for the question. With the transformation bank fixed, successful ydecy^{\mathrm{dec}} labeled sequences achieve roughly three times the success rate of random successful sequences. Sec. 5.3 uses ydecy^{\mathrm{dec}} to record transformations selected on successful multi-step trajectories. To test whether these decisions contain reusable information independently of knowledge-bank growth, we replay the recorded sequences on held-out queries from the same task and related molecular contexts. Replaying a random successful sequence serves as the control.

Starting-molecule similarityydecy^{\mathrm{dec}} labeled sequence (%)Random successful sequence (%)
≥0.6\geq 0.641.513.8
≥0.7\geq 0.746.816.8
≥0.8\geq 0.862.218.9

This fixed-bank analysis shows that ydecy^{\mathrm{dec}} captures reusable multi-step utility that is absent from the one-step bootstrap labels: transformations that do not immediately satisfy the task can still provide useful supervision when they belong to a successful trajectory. We have added this analysis to Appendix A.7 (Table 7) of the revised manuscript.

W2/Q3. What controller, prompt, auxiliary property estimates, candidate-evaluation budget, stopping rule, and latency were used?

Response:

Thank you for asking for these details. The complete search uses Qwen2.5-7B-Instruct without fine-tuning. At each step, the prompt shows the current frontier molecules together with their depth, SMILES, similarity to the starting lead, and RDKit descriptors, and asks the controller which branch to expand next. The search uses the top five transformations at each expanded node, a maximum depth of 10, and a budget of 10 committed outputs. All candidates receive inexpensive RDKit descriptors and seed-relative similarity estimates. BBBP, HIA, and Mutagenicity are evaluated only for committed outputs, averaging 4.6 ADMET evaluations per query; unselected branches do not incur these ADMET evaluations. The complete search takes 3.4 seconds per query on a single NVIDIA A5000 GPU, including an average of 4.6 local-controller calls. The complete configuration, computational profile, and representative controller prompt are now reported in Appendix A.13 and Table 12 of the revised manuscript.

Q1. Which knowledge-bank and policy snapshots were used for Tables 1, 2, and 8, and were they frozen during test evaluation?

Response:

Thank you for your question. The results referred to as Tables 1 and 8 in the submitted manuscript, which are Tables 1 and 11 in the revised manuscript, use the same final knowledge bank snapshot built from 128K exploratory transitions, containing 468 deployed transformations, together with the corresponding frozen action policy. Table 2 fixes the same transformation bank while varying the planning strategies. No bank or policy update is performed during test evaluation. We have clarified these settings in Sec. 6.1 of the revised manuscript.

Q2. How are ydecy^{\mathrm{dec}} labels filtered, combined with bootstrap labels, and used to update the policy?

Response:

Thank you for pointing this out. We retain ydecy^{\mathrm{dec}} labels only for selected actions on trajectories that eventually reach a feasible molecule. Decisions from unsuccessful continuations are discarded. The policy is therefore not trained to imitate every LLM decision regardless of outcome. This filtering and the update procedure are now explicit in Sec. 5.3 of the revised manuscript.

Q4. How does MOLT differ from RePO in its supervision and from AlphaApollo in the experience it retains?

Response:

We would like to clarify the differences. RePO trains a molecular generator using final-reference guidance and verifiable property and similarity rewards. MOLT instead ranks a discrete set of executable transformations, so the selected transformation provides direct action-level supervision in the current molecule and task context. AlphaApollo retains and revises complete candidate solutions through verifier feedback, whereas MOLT retains reusable local chemical edits extracted from evaluated transitions, together with the contexts and property effects observed for each edit. Across rounds, MOLT therefore updates its transformation knowledge and supporting evidence rather than a memory of complete solutions.

Thank you for the detailed response and the additional analysis. The clarifications on evaluation snapshots, label filtering, and computational cost address several of my concerns. The sequence-replay results are encouraging, although they do not directly establish the benefit of updating the action policy. A comparison of the bootstrap and updated policies with the same bank and search budget would make this contribution clearer. I will retain my current assessment while encouraging the authors to further develop this promising direction.

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

This paper introduces MOLT, a self-evolving agentic system that learns from past experiences to conduct rigorous and exploratory multi-objective lead optimization. The authors demonstrate state-of-the-art performance on several constrained lead-optimization tasks, beating several baselines.

Critical strengths

-The core methodology is very interesting and well-designed. It is an agentic system in which the LLM agent only acts as a decision-maker among the current choices within the molecular tree. This avoids the syntactic and semantic headaches that often come with having LLMs output raw SMILES.

  • I also like how the agent’s decisions impact the future policy for candidate selection, making the system self-evolving and utilizing every bit of past experience.

  • I think the design choices for the method are solid and logical.

  • The ablations are also good and demonstrate the isolated impact of the decision-making agent against random/greedy selection.

Critical weaknesses
  • Given that this paper advertises the method as self-evolving, it would be nice to see a graph/learning curve of SR as experience accumulates.

  • MolSearch as a baseline is missing; the authors mention it in the related work, and given its similar methodology to MOLT (lead optimization with MCTS), it might provide a stronger argument for agentic impact to see MOLT outperform classical MCTS methods.

  • I have a few concerns about some parts of the methodology and evaluation. To begin, while it’s good that the agent’s decision feeds into future policy updates, it seems like the policy is fully following the LLM’s decisions, whether or not that decision leads to a positive change in future trajectories. Is there any current way for the policy to only follow “successful” LLM choices, or is the policy eventually trained to fully imitate LLM decision-making regardless of outcome? similar methodology to MOLT (lead optimization with MCTS), it might provide a stronger argument of agentic impact to see MOLT outperform classical MCTS methods.

-My biggest concern is that there doesn’t seem to be any evaluation or control of important lead optimization properties such as synthesizability or binding affinity. One of the main goals of optimization is to preserve or improve the target affinity of the starting molecule, but the properties being filtered for do not include binding affinity or direct SA. Therefore, while the SR result for MOLT is indeed very strong, it doesn’t include evaluations of actual target activity, and there’s no way to tell whether or not the resulting molecules are actually strong/successful candidates.

Questions for authors

See weakness

  • Provide some metric of affinity within the results, e.g. an additional check for Vina score within an acceptable range of the starting affinity.

  • I’d be interested in seeing if the accumulation of experience actually improves lead optimization results over time. Or does the agentic system immediately give way to the strong results in Table 1?

Overall, given stronger evaluation and more convincing results on the target activity, I would be highly inclined to raise my score, given that the methodology is very well done and interesting. However, in its current state, I think it’s hard to claim that the method is performing lead optimization successfully when binding affinity is never even considered.

2 comments

W1/Q2. Does MOLT’s end-to-end SR improve as exploratory experience accumulates?

Response: Thank you for asking for this analysis. We find that MOLT’s mean SR increases monotonically from 19.5% to 54.4% as exploratory experience grows from 256 to 32K transitions. We evaluated MOLT at five experience checkpoints. At each checkpoint, the knowledge bank and context-conditioned action policy used only the exploratory transitions collected up to that point. The planner, search budget, evaluation set, and success criterion were kept fixed.

Exploratory transitions2562K8K16K32K
Mean SR (%)19.538.642.643.754.4

MOLT therefore does not immediately attain its later-stage performance; its end-to-end optimization capability improves as experience accumulates. We have added this learning curve to Appendix A.6 (Fig. 5) of the revised manuscript.

W2. How does MOLT compare with MolSearch under the same tasks, constraints, predictors, and evaluation budget?

Response: Thank you for suggesting this baseline. Under the same maximum budget of ten candidate evaluations per query, MOLT substantially outperforms MolSearch on both CNS and Safety. We ran the official MolSearch implementation with its original ChEMBL MMP library and retargeted it to our CNS and Safety tasks. Both methods use the same 500 starting molecules per task, property predictors, target and guardrail requirements, and Sim≥0.50\mathrm{Sim}\geq 0.50.

TaskMethodVal (%)Sim (%)Prop (%)SR (%)
CNSMolSearch99.298.852.051.6
CNSMOLT100.095.286.681.8
SafetyMolSearch96.696.040.240.2
SafetyMOLT100.097.284.281.8

Both methods maintain high validity and structural similarity. The main difference is property satisfaction: MolSearch often remains close to the starting molecule without crossing the target-property threshold, whereas MOLT selects transformations that satisfy the target while retaining the guardrails. We have added this comparison to Appendix A.9 (Table 9) of the revised manuscript.

W3. Are ydecy^{\mathrm{dec}} labels retained only from successful trajectories, or are all LLM-selected actions treated as positive supervision?

Response: Thank you for pointing this out. We retain ydecy^{\mathrm{dec}} labels only for selected actions on trajectories that eventually reach a feasible molecule. Decisions from unsuccessful continuations are discarded. The policy is therefore not trained to imitate every controller decision regardless of its outcome. This filtering and the update procedure are now explicit in Sec. 5.3 of the revised manuscript.

W4/Q1. Can MOLT improve the target liability while preserving target activity and synthetic accessibility?

Response: Thank you for raising this concern. We added a target-specific experiment on 200 held-out GSK3$\beta$-active leads with predicted GSK3$\beta$ activity ≥0.6\geq 0.6 and a Mutagenicity liability. MOLT was asked to reduce Mutagenicity to ≤0.3\leq 0.3, while keeping predicted GSK3$\beta$ activity ≥0.5\geq 0.5, SA≤4\mathrm{SA}\leq 4, and seed-relative similarity ≥0.5\geq 0.5.

SettingActivity retained (%)Mean predicted GSK3$\beta$ activity
With activity and SA guardrails100.00.90 → 0.75
Without activity and SA guardrails71.80.90 → 0.63

With predicted activity and SA enforced as explicit guardrails, mean predicted Mutagenicity decreases from 0.51 to 0.18, and every final output retains predicted GSK3$\beta$ activity above the prescribed threshold. Without these guardrails, activity retention falls to 71.8% and mean predicted activity to 0.63. These results show that MOLT can correct the Mutagenicity liability while explicitly preserving predicted target activity and synthetic accessibility. We have added this experiment to Appendix A.8 (Table 8) of the revised manuscript.

Thanks for the response to my concerns. I will discuss with other reviewers and update my review.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.