acceptodds
Withdrawn from ICLR 2027

PharmAgent: Constraint-Aware Search with Frozen Language Models for Molecular Optimization

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Molecular optimization must improve target activity and satisfy developability constraints within limited evaluation budgets. Classical methods require tailored rules or training to incorporate chemical instructions and property feedback. Frozen language models can condition edits on this information, but need explicit constraint control and relevant experience. We therefore present PharmAgent, a constraint-aware molecular search method driven by adaptive external state. Its Lagrangian controller translates violations in accepted states into accumulated constraint pressure, keeping this history separate from current property measurements. Structure-indexed replay complements this feedback with relevant evaluated transitions that guide subsequent proposals. As a curriculum progressively activates constraints, candidates and the incumbent are compared under the same current objective, and the accepted state determines the next multiplier update. We derive an exact identity that characterizes how accepted-state violations accumulate in the controller's multipliers. Across five tasks with five independent runs, PharmAgent achieves a summed area under the target-score curves (AUC) of 3.9208 in target-only search, improving over MOLLEO by 37.3%. With online constraints, it achieves a property-adjusted AUC of 0.7076, improving over the strongest online baseline, ExLLM, by 53.8%. These results rank first among all evaluated methods in both target-only and constraint-aware search. The online comparison covers all five baseline frameworks. The full system leads every ablation variant in target quality, property-adjusted performance, and Pareto hypervolume. All five molecular cases reach feasible final states, documenting target gains and trade-offs. These results establish PharmAgent as a powerful framework for budget-limited molecular optimization, achieving leading aggregate performance through adaptive external search state and frozen language models.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Molecular Optimization, Large Language Models, Constrained Optimization, Black-Box Optimization, AI for Science

Reviews

Among:
Shape:
1234this paper, Bottom 25%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.75 over 4 reviews, above 25% of the others.
Reviewer_KEEzOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

PharmAgent is a Generator/Reviewer loop built around a frozen LLM for molecular optimization under a fixed oracle budget. Three external-state mechanisms are claimed: a Lagrangian controller with a quadratic penalty that updates multipliers from the accepted state; Chemical-CER, which retrieves past transitions by Morgan fingerprint proximity; and a curriculum that activates constraints at fixed round boundaries. Evaluation covers PMO (E1) and ADMET-PMO (E2) across five tasks and five runs at 1,000 unique oracle evaluations. The headline claims are a 37.3% gain in summed target AUC over MOLLEO and a 53.8% gain in AA-AUC over ExLLM.

Critical strengths

The separation of current measurements from accumulated multiplier pressure in the prompt—together with the observation that a feasible molecule can still carry substantial pressure, thereby raising the cost of reintroducing a violation—constitutes a substantive conceptual contribution. It is a genuine contribution that merits defense under a parity comparison.

Critical weaknesses
  1. Six checks are enabled by default and applied conjunctively—Lipinski, Veber, PAINS, Brenk, SA ≤ 6.0, and QED ≥ 0.3—thereby discarding candidates before any oracle call. Four of these thresholds match the “easy” constraints in Table A4 exactly, yet the paper describes this procedure as validity checking, and Appendix F states that PAINS, Brenk, and Veber are never used as optimization guidance. This claim is empirically testable: if the gate is active, the 70.87% infeasible fraction reported in Appendix H should fall entirely within hERG, CYP2D6, and Caco-2. If that is the case, E1 is not target-only, the “0 → 4 → 7” schedule does not describe the implementation, and the E2−E1 contrast no longer constitutes a contrast in online constraint control.

  2. With non-decreasing multipliers, boundedness and summability become equivalent assertions; the observed maximum of 1815.53 demonstrates that the premise fails substantively. The abstract selects the most favorable comparator for each metric, and the claim to “rank first among all evaluated methods” is false at the task level. SEISMO, evaluated with only 6–47 evaluations versus 1,000, should not share a ranking column with best/second-best marks. Initialization is unmatched—a curated 26-molecule pool versus ZINC—under a metric in which early feasibility supplies most of the area. On Albuterol, the target equals 1.000 at both endpoints although neither state is albuterol by mass, a saturation effect that merits discussion. The 1,000-call regime structurally disadvantages population methods and should be acknowledged in the limitations. Oracle calls are matched, but LLM computation is not: 420M tokens per E2 arm versus unreported baselines. A single closed-provider backend makes the reported numbers unreproducible in principle.

Questions for authors
  1. Does the MedChem-RAG corpus overlap with the training datasets of the GSK3β, JNK3, and DRD2 oracles? If a deduplication protocol exists, please provide it, together with the E1 results obtained using a corpus cleaned of these target genes.

  2. Please report confidence intervals for each row of Table 4, as well as the results of reversing the task execution order when CER is active (reversing the task order should suffice).

  3. How sensitive are PHV and the resulting ranking method to the scales ri and steepness si reported in Table A5?

Reviewer_69gSOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

PharmAgent focuses on multi-constrained molecular optimization under a limited evaluation budget. Rather than training new generative models, it maintains an adaptive external search state around frozen language models. The framework combines a Lagrangian controller, chemical-aware experience replay (Chemical-CER), and curriculum-based constraint activation to guide molecular optimization.

Critical strengths

The paper does not merely package an LLM as an agent that calls chemical tools. Instead, it explicitly investigates how to maintain and update external decision states during the search process, allowing historical experience and constraint violations to guide subsequent optimization. I find this direction valuable, as effective experience accumulation and reuse are important capabilities for agentic AIDD

The authors also ensure that candidate molecules and the current incumbent are compared under the same objective function, which is a reasonable design choice.

Critical weaknesses

The paper addresses an important problem in agentic molecular optimization, and its emphasis on external search states, experience reuse, and adaptive constraint handling is promising. However, the current experimental evaluation does not sufficiently establish computational efficiency or generalizability. In particular, the lack of cost-matched comparisons and the limited coverage of PMO objectives weaken the claimed advantages. I believe the paper has potential, but more comprehensive and fair experimental comparisons are needed to support its conclusions.

  1. The evaluation budget is not sufficiently fair across different methods.

The paper claims to address molecular optimization under limited evaluation budgets. However, the experiments mainly standardize the number of target oracle calls without adequately controlling the actual computational costs of different methods. For LLM-based agents, the number of oracle evaluations alone cannot fully reflect computational efficiency, particularly when methods differ substantially in LLM inference, reasoning, and tool-calling overhead.

  1. The limited selection of PMO tasks raises concerns about generalizability.

The original Practical Molecular Optimization (PMO) benchmark contains 23 oracle functions, including QED, DRD2, GSK3β, JNK3, and 19 objectives from GuacaMol. However, this paper evaluates only five tasks: GSK3β, JNK3, DRD2, Sitagliptin MPO, and Albuterol similarity. What is the rationale for selecting these particular tasks? GSK3β, JNK3, and DRD2 are relatively common molecular optimization benchmarks, but evaluating only these five objectives is insufficient to support strong claims about general molecular optimization capabilities.

This concern is particularly relevant because PharmAgent relies heavily on chemical experience retrieval, structural modification, and constraint control. The effectiveness of these mechanisms may vary substantially across different optimization landscapes. For example, it remains unclear whether Chemical-CER and the Lagrangian controller would remain effective for objectives with numerous local optima, highly discontinuous structure–activity relationships, or strong scaffold dependence.

  1. The independent theoretical contribution is relatively limited. The exact identity derived in the theoretical section mainly follows from directly summing the classical Lagrangian multiplier update equations. While mathematically correct, it does not establish new convergence guarantees or provide substantial theoretical insight into the proposed molecular search framework. In my view, this derivation contributes relatively little to the central methodological claims and could be substantially shortened or omitted. The paper would benefit more from a clearer analysis of how adaptive constraint feedback improves the actual search behavior compared with conventional penalty-based or feasibility-first optimization methods.
Questions for authors

The paper addresses an important problem in agentic molecular optimization, and its emphasis on external search states, experience reuse, and adaptive constraint handling is promising. However, the current experimental evaluation does not sufficiently establish computational efficiency or generalizability. In particular, the lack of cost-matched comparisons and the limited coverage of PMO objectives weaken the claimed advantages. I believe the paper has potential, but more comprehensive and fair experimental comparisons are needed to support its conclusions.

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

This paper proposes PharmAgent, a constraint-aware search framework built around the frozen LLM, aimed at addressing molecular optimization problems under multiple developability constraints. The framework primarily relies on three mechanisms to adjust the search state: (1) a Lagrangian controller that accumulates constraint violations from accepted states and incorporates historical constraint pressure into candidate evaluation and subsequent proposal guidance; (2) Chemical-CER, which retrieves structurally similar evaluated molecular transitions using Morgan fingerprints; and (3) a target-first constraint curriculum that progressively activates ADMET constraints. Experiments were evaluated across five molecular optimization tasks, demonstrating significant improvements in target performance compared to current state-of-the-art baseline methods.

Critical strengths

Strengths 1: Clear and well-justified motivation. The paper identifies an important limitation in molecular optimization: improving target activity does not necessarily yield molecules that satisfy developability constraints. Maintaining the current property measurements and the accumulated constraint pressure as separate representations is a sound design choice in the proposed framework, enabling the search process to account for persistent constraint violations. Strengths 2: Comprehensive empirical evaluation and useful component analysis. The experiments cover five molecular optimization tasks, a variety of baseline methods with two settings, target-only and online-constrained. PharmAgent achieves strong aggregate experimental results, particularly in feasible-solution optimization, where its performance improves by 53.8% over the strongest baseline. The ablation results also provide evidence for the contribution of each component to overall performance.

Critical weaknesses

Weakness 1: The methodological novelty is relatively limited, as the proposed mechanisms are mainly adaptations and combinations of existing techniques.

The three main components are based on existing ideas: Lagrange multipliers and quadratic penalty methods for constrained optimization, fingerprint-based nearest-neighbor retrieval for experience replay, and curriculum learning. It would strengthen the contribution if the authors more clearly distinguished the proposed mechanisms from direct applications of existing methods, and demonstrated the necessity of the specific design choices.

Weakness 2: The experiments and theoretical claims have limitations that restrict the generality of the conclusions.

Some baseline comparisons are not fully equivalent in terms of optimization protocols. Specifically, SEISMO follows its original short-run protocol, performing only 27–48 unique target evaluations and carrying its final score over the common 1,000-call horizon to compute AUC. While this follows the original reporting convention, it makes comparisons with methods that use the full evaluation budget less direct. In addition, MOLLEO and ExLLM were independently reimplemented using MiniMax-M3, and their online-adapted versions were designed after the original PharmAgent runs. These choices make it particularly important to assess how sensitive the reported performance gains are to baseline implementation and protocol differences.

The scope of the theoretical analysis is also relatively limited. The result that constraint violations tend to zero relies on the uniform boundedness of the multipliers, but the paper explicitly acknowledges that this assumption is not established for the proposed search process. Consequently, the analysis does not guarantee the feasibility or convergence of the actual algorithm under a finite budget.

Questions for authors

Q1: Is the "Historical Pressure" component of the Lagrangian controller necessary? The current ablation only removes the entire Lagrangian module, failing to isolate the specific contribution of historical multiplier accumulation versus immediate quadratic penalties.

Q2: Can the contributions of Chemical-CER’s structure-based retrieval and cross-task experience reuse be decoupled? Authors can provide ablations for within-task vs. cross-task retrieval and report the frequency and effectiveness of cross-task queries.

Q3: Is the constraint curriculum sufficiently adaptive? The current implementation uses fixed activation rounds (20 and 60) rather than dynamically triggering constraints based on search feedback, which may limit generalization of proposed method.

Reviewer_1Z6mOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes PharmAgent, a molecular optimization framework that combines frozen LLMs with adaptive Lagrangian constraint control, structure-aware experience replay, and curriculum-based constraint activation. Experiments on five molecular optimization tasks show improved aggregate performance over several classical and LLM-based baselines under a fixed target-oracle evaluation budget. The authors also conduct ablation studies to evaluate the contributions of individual components.

Critical strengths

The paper demonstrates that the proposed framework can be used effectively for molecular optimization. The authors provide comparisons with several existing methods and conduct ablation studies to evaluate the contributions of individual components. The reported results show strong performance under a fixed target-oracle evaluation budget.

Critical weaknesses
  • The experiments limit target-oracle evaluations but do not account for other computational costs. According to Table A6, PharmAgent E2 consumes approximately 420 million tokens across 25 runs. If these resources were allocated to traditional baselines for additional oracle evaluations, they might achieve comparable or better performance. Without comparisons under equivalent total computational budgets, the reported efficiency advantages remain unclear. Furthermore, PharmAgent is essentially an LLM harness, yet all experiments use only MiniMax-M3. Evaluations across different LLM families and model scales are needed to establish the generalizability of the proposed harness.

  • The paper relies heavily on terminology and complicated descriptions, making the central ideas difficult to follow. Although the proposed mechanisms are connected to the identified limitations of existing methods, the paper does not sufficiently justify why these particular designs are preferable to simpler alternatives. Moreover, the main theoretical identity follows directly from the multiplier update rule, while the conditional feasibility result relies on assumptions not established for the actual search procedure. Thus, the theoretical analysis offers limited insight into practical search efficiency or finite-budget performance.

Questions for authors

Question 1: Is PharmAgent's additional computational cost justified?

The experiments limit the number of target-oracle evaluations but ignore the substantial computational cost of LLM inference. PharmAgent consumes approximately 420 million tokens across 25 runs.

What if the computational resources spent on these LLM calls were instead allocated to traditional baselines, allowing them to perform more oracle evaluations? Could those baselines achieve equal or better performance?

Restricting only oracle evaluations while ignoring LLM inference costs does not ensure a fair comparison. Could the authors compare the methods under the same total computational or monetary budget, rather than just the same number of oracle evaluations?

Question 2: Does the proposed harness generalize across different LLMs?

PharmAgent is essentially an LLM harness, yet all experiments use only MiniMax-M3. Since the effectiveness of a harness may depend heavily on the underlying LLM, evaluating it on a single model is insufficient to demonstrate its generalizability.

Could the authors evaluate the proposed harness across different LLM families and model scales to verify whether its improvements consistently hold?

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

Related papers

Discussion (0)

Sign in to comment.