- Rating
- 3(+1)
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper proposes ReMedI for lead molecule optimization with requirements for multi-property improvement, property retention, and structural similarity. The method uses a chemical language model to propose molecular edits, organizes an iterative search with a Metropolis–Hastings (MH) acceptance rule using an objective function that combines property improvement, protected properties, and lead similarity, and adjusts the optimization focus through an outer dynamic focusing mechanism. The authors then construct supervision data from search endpoints, intermediate states, and local transitions to perform LoRA distillation of the language model, and subsequently evaluate direct generation and further search performance. The paper reports gains on C-MuMOInstruct, an adapted SMDD-Bench evaluation, and three language model backbones. The research problem is practically meaningful, but the current evidence is insufficient to fully establish the novelty of the method, the necessity and implementation correctness of the core MH design, and the performance and cost advantages over existing molecular editing methods.
Critical strengths
- The problem setting is practically meaningful. Simultaneously improving target properties, preserving other properties, and limiting structural changes is closer to the actual needs of lead optimization than simply maximizing a single molecular score.
- The framework has a certain flexibility. The language model is responsible for proposing edits, and an explicit objective function is responsible for evaluation, making it convenient to adjust optimization preferences at inference time.
- Search results can be converted into generative capability. For example, in Table 1, GeLLMO-Mistral's IND direct generation SOR increases from 0.324 to 0.524, indicating that search trajectories have exploitable supervision value.
- A variety of experiments are provided. In addition to the main results, the paper includes analyses of search strategies, search depth, distillation rounds, and training data sources, and validates the applicability of the framework on multiple backbone models.
Critical weaknesses
W1. The added contribution relative to mature molecular search–learning frameworks has not been sufficiently established. The overall pipeline of this paper can be summarized as "a generative model proposes candidates—search obtains better molecules—the model is trained using search results—generation or search is performed again." This paradigm already has directly related precedents in molecular optimization. MARS has already combined molecular editing, MCMC with annealing, adaptive proposals, and training of proposal models using search samples; GEGL uses genetic search to generate high-quality molecules to guide neural generative models. Therefore, the search-and-retraining loop itself cannot constitute the main basis for the novelty of this paper. The possible added contributions of this paper lie in instruction-conditioned language model proposals, lead-relative objectives, dynamic focusing, and a specific trajectory distillation design. However, it is currently unclear what capabilities these designs bring that are difficult for existing frameworks to achieve, and whether the gains exceed the natural improvement from switching to a large pretrained editor. Using classical algorithms itself is not a problem, but when the overall paradigm already has precedents, the contribution needs to be established through clear methodological distinctions, mechanistic analysis, or convincing empirical advantages. The current manuscript remains insufficient in these respects. W2. External baseline coverage is insufficient, and the value of large pretrained editors relative to classical molecular editing methods has not been established. The results on three language model backbones mainly show that ReMedI can enhance different language models, but they cannot replace comparison with other molecular editing and optimization paradigms. Although the paper compares with C-MORAL, coverage of non-LLM graph editing, reinforcement learning editing, and search–learning methods remains insufficient. The following representative methods are recommended as priorities: (1) Modof / Modof-pipe: to test the value of large language model editors relative to local graph editing and multi-step fragment modification. (2) Graph-to-Graph Translation: to compare with classical conditional molecular graph transformation methods. (3) MolDQN: to compare with reinforcement learning methods based on valid molecular editing actions. (4) MARS: to compare with the closest molecular editing, MCMC, and proposal learning framework. (5) GEGL: to compare with the closed loop of genetic search and neural model retraining. (6) Graph GA: to provide a search reference that does not rely on large pretrained language models. Among these, Modof explicitly targets local fragment modification of a given molecule and controls the optimization process and molecular similarity through multi-step modification, making it a meaningful editing baseline. The authors need not run all methods, but should at least cover representative non-LLM editing methods and the search–learning framework closest to the mechanism of this paper. Comparisons must use the same lead molecules, property oracles, improvement and retention requirements, similarity constraints, and budget; methods with inconsistent original tasks need to be adapted and cannot simply cite their original paper values. The current results support "ReMedI can enhance existing molecular language models," but are not yet sufficient to establish "this approach, which relies on large pretrained editors, has a competitive advantage over existing molecular editing methods." W3. The necessity of MH relative to other mature search strategies has not been sufficiently validated. The paper compares independent sampling, continuous editing, greedy feedback, and MH. These are useful basic ablations but are insufficient to prove the unique value of choosing MH. In Figure 4, at budget K=20, the SOR of continuous editing, greedy feedback, and MH is approximately 0.39, 0.38, and 0.41, respectively. This shows that the gain from continuous editing is relatively clear, while the additional improvement of MH over simple continuous editing is about two percentage points. The lack of sufficient repeated experiments and paired uncertainty analysis makes the stability of this difference unclear. More importantly, continuous editing is a basic reference and cannot replace comparison with mature alternative search strategies. It is recommended to fix the language model, objective function, prompts, and budget, and compare the following strategies: (1) Full MH versus Metropolis-style acceptance without the Hastings correction. (2) Beam search or search that retains multiple high-quality states. (3) Monte Carlo tree search (MCTS) using the same language model to propose successor edits. (4) Evolutionary search using the same language model as a mutation proposer. It should also compare the effects of data generated by different search strategies when used for distillation, to determine whether the gains depend on MH-specific trajectories or whether generally high-quality search results can produce similar gains. Simulated annealing and MH are not mutually exclusive concepts, so the core here is not a vague comparison of "MH and simulated annealing," but rather distinguishing the contributions of acceptance rules, annealing schedules, state retention, and exploration mechanisms. W4. The actual effective proposal distribution and the computation of the Hastings ratio are not sufficiently clear, affecting judgment of MH correctness. The paper states that generated outputs are truncated, parsed, and resampled for invalid molecules. Therefore, the actual candidate distribution is generally not equal to the original language model's sequence distribution. For example, under the simplified setting of "repeatedly resampling until a valid output is produced": q_valid(y | x, P) = q(y | x, P) / Z_valid(x, P) where Z_valid(x, P) represents the total probability of generating valid outputs from the current input. The corresponding Hastings ratio is: q_valid(x | y, P) / q_valid(y | x, P) = [q(x | y, P) / q(y | x, P)] × [Zvalid(x,P)/Zvalid(y,P)] The validity normalization factor usually depends on the current molecule and cannot be assumed to cancel out. The paper uses the notation of an effective proposal distribution but does not sufficiently explain how this probability is calculated or estimated. In addition, it needs to clarify: (1) How molecular-level proposal probabilities are defined when multiple generated strings map to the same molecule after parsing or normalization. (2) Whether decoding settings such as temperature, top-k, and top-p are included in the probability calculation. (3) Whether the reverse proposal probability is consistent with the actual reverse generation process. (4) How failures and self-transitions are handled when there is an upper limit on invalid retries. These questions do not prove that the implementation is necessarily incorrect, but at present it cannot be confirmed whether the algorithm implements strict MH or uses an acceptance strategy constructed with language model likelihood approximations. The authors should provide complete definitions, derivations, and implementation descriptions. Outer dynamic focusing, stage restarts, changing weights, and best-candidate selection also make the overall process different from a homogeneous MH chain under a fixed target distribution. The paper needs to distinguish between "locally using the MH acceptance rule" and "the entire algorithm having sampling guarantees under a fixed target distribution." W5. Budget and cost accounting are incomplete, and claims about data efficiency and computational efficiency need to be narrowed. The same number of valid candidates does not represent the same computational cost. MH requires additional computation of forward and reverse proposal probabilities, invalid retries consume generation resources, and independent sampling is easier to parallelize in batches. Therefore, best-of-20 or the same number of candidates cannot alone constitute a fair computational budget control. For the version that searches again after distillation, the following should be accounted for separately: Total cost = training trajectory search cost + distillation training cost + test-time search cost This should report total generation count and token volume, number of valid candidates, number of unique oracle queries, probability evaluation overhead, GPU time, memory, and actual runtime. Distillation training itself may not re-call the oracle; statistics should be based on actual calls, and handling of caching and repeated queries should be explained. The comparison of "500 versus 3,883 instances" mainly reflects the number of starting contexts and cannot directly prove that the total amount of supervision data, oracle usage, or overall training cost is lower. It is necessary to also report the number of unique lead molecules, lead–task pairs, trajectories, and final training pairs. Offline trajectory generation and distillation costs can be reused across multiple deployments, while test-time search costs grow with the number of instances. Therefore, offline and online costs should be reported separately, and amortization under different deployment scales should be explained. In addition, the paper should disclose the foundational pretraining, chemical domain adaptation, and existing instruction training resources on which the large editor depends. When sharing the same initialization model, the common pretraining cost can be offset in incremental cost comparisons; but when comparing with non-LLM methods, resource dependency differences must be made explicit. The current evidence is more suitable to support "incremental data savings after reusing a pretrained model" rather than an overall data efficiency advantage. Finally, there is a lack of a combined baseline of C-MORAL + the same refinement. Since the distilled model still benefits from search, it is necessary to test the performance of the RL model connected to the same search in order to distinguish "search can supplement post-training" from "this method is superior to or replaces RL post-training." W6. Property and task specifications are insufficiently disclosed, and the SMDD-Bench adaptation protocol is not transparent enough. The paper mentions BBBP, QED, logP, and properties such as AMES, DILI, and PGP, so the issue is not a complete absence of property names, but a lack of systematic task specifications. It is recommended to list completely for each task: (1) Properties to improve and properties to protect. (2) Optimization direction, improvement thresholds, and retention tolerances. (3) Oracle name, version, output definition, and units. (4) Property scales, objective function parameters, and strict success conditions. For SMDD-Bench, it should be clarified how the original objectives, retention requirements, and hard constraints are handled among the 314 instances, which are retained, filtered, replaced, or not evaluated. Adopting a non-agentic ADMET-AI/RDKit protocol can be valuable, but its difference from the original benchmark should be made explicit to avoid interpreting adapted results as performance under the complete original protocol.
Questions for authors
- What substantive new mechanisms does this paper have relative to MARS and GEGL? Which experiments can isolate the contributions of these mechanisms?
- Compared with representative non-LLM molecular editing methods, does ReMedI still have advantages under the same tasks, constraints, and budget?
- With the language model proposer fixed, what are the advantages of MH relative to MCTS, evolutionary search, and beam search?
- How is the effective proposal distribution calculated? How are invalid resampling, parsing mappings, and the actual decoding distribution handled?
- What is the complete cost of training trajectory search, distillation, and test-time search? What do 500 and 3,883 each count?
- How does C-MORAL perform when connected to the same refinement?
- How is the "approximately 2.8× coverage" calculated? Does the advantage still hold under explicit structural constraints?
- What are the properties and thresholds for each task? Which original constraints of SMDD-Bench are not included in the current evaluation?
5 comments
Rebuttal (Part 2)
Q3. Comparison with non-LLM molecular optimization methods
To address the request for broader comparisons, we evaluated adapted implementations of MARS, GEGL, and GP-MoLFormer-base+GA on 150 held-out lead–task instances. All methods use identical property oracles, optimization objectives, candidate-selection criteria, and evaluation metrics, with a matched budget of 320 molecular proposals per lead.
We compare against ReMedI-MH using the frozen GeLLMO-Mistral proposal, without additional distillation. For MARS and GEGL, we additionally match mean structural similarity to the original lead.
| Method | SOR ↑ | SSOR ↑ | Sim ↑ |
|---|---|---|---|
| ReMedI-MH (Ours) | 0.6533 | 0.3800 | 0.5155 |
| MARS (adapted) | 0.5414 | 0.2567 | 0.5121 |
| GEGL (adapted) | 0.5333 | 0.2733 | 0.5142 |
| GP-MoLFormer-base+GA | 0.5333 | 0.2533 | 0.1868 |
ReMedI achieves the highest SOR and SSOR across all evaluated methods. At comparable structural similarity, it improves SOR by 11.2 and 12.0 percentage points over MARS and GEGL, respectively, and SSOR by 12.3 and 10.7 percentage points. GP-MoLFormer-base+GA also achieves lower success rates, despite substantially lower structural similarity.
Overall, these results support both the methodological contributions and empirical effectiveness of ReMedI. Our instruction-conditioned, likelihood-aware refinement provides effective multi-property optimization under structural preservation constraints, while trajectory distillation further improves the proposal and enables transfer to unseen tasks.
Rebuttal (Part 3)
Q4. Effective Proposal Distribution and Implementation
We clarify how invalid-generation retries, molecular parsing, and decoding determine the effective proposal distribution. We show that validity-conditioned sampling admits an exact MH equilibrium interpretation, while molecular normalization introduces a tractable approximation whose practical benefits are supported by ablation experiments.
1. Effective proposal under validity resampling.
Let be the raw LM probability of generating string from molecule under focused prompt , and let denote valid, oracle-scoreable outputs.
Because invalid generations are retried without advancing the refinement chain, the effective proposal, considering validity filtering alone, is
The state-dependent normalization represents the probability of generating a valid edit. Thus, retries condition the proposal on chemical validity rather than introducing an independent search heuristic.
2. Relationship to the implemented Hastings correction.
Although Eq. (6) uses , our implementation computes acceptance using raw LM probabilities:
\alpha_n(x,x')=\min\left\\{1,e^{U_n(x')-U_n(x)}\frac{q_\theta(x\mid x',P)}{q_\theta(x'\mid x,P)}\right\\}.
Proposition. For a fixed prompt and utility , under validity-only conditioning, this acceptance rule satisfies the exact MH equilibrium condition for
Proof. Substituting the validity-conditioned proposal and target into the probability flow gives
&\widetilde{\pi}_{n,P}(x)\\,q_\theta^{\mathrm{valid}}(x'\mid x,P)\\,\alpha_n(x,x') \\\\ &\quad\propto \min\left\\{e^{U_n(x)}q_\theta(x'\mid x,P),e^{U_n(x')}q_\theta(x\mid x',P)\right\\} \\\\ &\quad=\widetilde{\pi}_{n,P}(x')\\,q_\theta^{\mathrm{valid}}(x\mid x',P)\\,\alpha_n(x',x). \end{aligned}$$ The expression is symmetric in $x$ and $x'$, establishing the equilibrium condition. **Therefore, validity retries combined with raw-probability acceptance define an exact MH procedure for a feasibility-aware target**, favoring both molecular utility and the ability to propose valid edits. **3. Molecular parsing and decoding.** Generated strings undergo deterministic parsing, canonicalization, and salt stripping through a mapping $h(z)$. Since multiple strings may represent the same molecular state, the effective proposal over molecules is $$q_H(x'\mid x,P)=\frac{\displaystyle\sum_{\substack{z\in\mathcal V\\\h(z)=x'}}q_\theta(z\mid x,P)}{Z_\theta(x,P)}.$$ Rather than explicitly marginalizing over all equivalent representations, **we evaluate the LM probabilities of canonical representations**, yielding a computationally tractable approximation. Accordingly, the exact equilibrium result applies to validity-only conditioning with fixed prompts and utilities; the full procedure additionally incorporates molecular normalization and adaptive refinement. Importantly, **generation and likelihood evaluation use identical decoding settings** (temperature $=1$, top-$p=1$), so the probabilities used for acceptance correspond to the actual LM sampling distribution. **4. Empirical validation of preprocessing.** We ablate validity retries and salt stripping across three backbones using the frozen $q_0$ proposal. **Validity retries:** *On* redraws invalid generations without advancing the chain, whereas *Off* consumes the refinement step. | Backbone | SOR (On / Off) | SSOR (On / Off) | | :------------- | :---------------- | :---------------- | | GeLLMO-Mistral | **0.448** / 0.412 | **0.172** / 0.152 | | GeLLMO-Llama | **0.500** / 0.380 | **0.172** / 0.132 | | mCLM-4B | 0.295 / 0.295 | 0.071 / 0.081 | **Salt stripping:** | Backbone | SOR (On / Off) | | :------------- | :---------------- | | GeLLMO-Mistral | 0.448 / 0.448 | | GeLLMO-Llama | 0.500 / 0.500 | | mCLM-4B | **0.292** / 0.000 | Validity retries improve success for both GeLLMO backbones. **Salt stripping is particularly important for mCLM-4B**, whose molecule-block outputs can contain salts: disabling normalization reduces SOR from **0.292 to 0.000**. In contrast, the SMILES-based GeLLMO backbones are unaffected, consistent with their more directly usable molecular representations. **Together, the theoretical analysis and ablations explain the roles of validity conditioning, the Hastings correction, and representation-dependent molecular preprocessing in our implementation.**
Rebuttal (Part 1)
We thank the reviewer for their thoughtful evaluation and suggestions, which helped clarify our methodology and strengthen our experiments. We address their concerns below.
Q1–Q2. Methodological Contributions and Comparison with Non-LLM Methods
We explain ReMedI's distinctions from MARS and GEGL and provide additional comparisons under matched conditions.
1. Methodological differences from MARS and GEGL
While building on established techniques, ReMedI combines instruction-conditioned editing, likelihood-aware refinement, and trajectory-based self-distillation for multi-objective lead optimization. Three mechanisms distinguish our approach:
(a) Instruction-conditioned proposals with MH refinement. Unlike MARS's fragment-level graph edits, ReMedI uses a pretrained molecular LM conditioned on the current molecule and requested properties. Unlike MARS's approximate acceptance rule, ours incorporates forward and reverse proposal likelihoods:
α(x,x′)=min{1,exp(U(x′)−U(x))⋅qθ(x′∣x,P)qθ(x∣x′,P)}
Adaptive prompting directs edits toward the least-satisfied property or structural similarity constraint.
(b) Explicit multi-objective lead optimization. Unlike MARS and GEGL's original de novo settings, ReMedI jointly optimizes requested properties, preserves protected properties, and maintains lead similarity through
U(x)=βSlogS(x)+βGG(x)+βLL(x),
where S(x) measures aggregate property satisfaction, G(x) penalizes protected-property violations, and L(x) encourages lead similarity. The satisfaction term prevents large gains in one property from compensating for unmet requirements.
(c) Context-conditioned trajectory distillation. Unlike GEGL's reward-prioritized imitation, ReMedI distills successful endpoints conditioned on original leads and intermediate states, alongside accepted transitions. Rejected proposals also provide information about unsuccessful edits, helping improve subsequent MH refinement. This trains a shared proposal that improves direct generation and refinement across property combinations.
2. Experimental evidence isolating these contributions
(a) Search strategy ablation. We compare independent best-of-K sampling, unguided sequential editing, greedy feedback, and MH under matched candidate budgets (SOR):
MH leads at every candidate budget. At K=20, it achieves 0.4631, versus 0.3929 for sequential, 0.3790 for greedy, and 0.2923 for independent sampling, demonstrating gains beyond iterative or greedy editing.
(b) Trajectory depth ablation. With 80 proposals, increasing depth from 80 independent one-step proposals to ten eight-step trajectories improves SSOR from 0.070 to 0.140, demonstrating the benefit of composing edits.
(c) Proposal distillation. On GeLLMO-Mistral, direct-generation SOR improves from 0.324 to 0.524 on IND and 0.268 to 0.372 on OOD tasks. Subsequent MH refinement improves from 0.528 to 0.692 on IND and 0.400 to 0.532 on OOD tasks, showing gains in both generation and refinement, including unseen property combinations.
3. Comparison with non-LLM molecular optimization methods
We evaluate adapted MARS, GEGL, and GP-MoLFormer-base+GA on 150 held-out lead–task instances (15 per task across ten tasks), using identical property oracles, objectives, candidate-selection criteria, and metrics, with 320 proposals per lead. We use ReMedI-MH with a frozen GeLLMO-Mistral proposal, without distillation, and match mean structural similarity to the original lead for MARS and GEGL.
ReMedI achieves the highest SOR and SSOR. At comparable similarity, SOR improves by 11.2 and 12.0 percentage points over MARS and GEGL, respectively, and SSOR by 12.3 and 10.7 percentage points. GP-MoLFormer-base+GA also has lower success despite substantially lower similarity.
Overall, these results support ReMedI's methodological contribution: instruction-conditioned, likelihood-aware refinement enables multi-property lead optimization with structural preservation, while trajectory distillation strengthens the proposal and transfers to unseen tasks.