acceptodds
Under review as a conference paper at ICLR 2027

ReMedI: Refinement-based Molecular Editing via Inference-time Sampling

Abstract

In drug development, lead optimization must improve competing properties, such as increasing permeability and reducing toxicity, while limiting scaffold changes that could compromise the lead’s potency. RL-based post-training, such as C-MORAL, addresses this problem by fine-tuning molecular language models to directly optimize molecular leads for objectives of interest, which relies on expensive task-specific datasets and faces challenges in achieving multi-objective optimization (MOO). We argue a pretrained model may already contain useful molecular edits that zero-shot generation cannot efficiently reach. Motivated by the limited search capabilities of zero-shot generation, we develop an efficient Metropolis–Hastings search strategy to better explore the proposal distribution. We propose ReMedI, which uses the molecular language model as a base proposer for inference-time alignment. Given a lead and a set of multi-objective tasks, the model proposes edits conditioned on the current molecule, while Metropolis–Hastings refinement guides sequential edits under an objective for property improvement, preservation, and lead similarity. On both C-MuMOInstruct and SMDD benchmarks, refinement improves frozen base models across three backbones, reaching significantly higher multi-objective optimization success rates compared to baselines. These search traces can then distilled into the base model which improves zero-shot generation and subsequent search using less training molecular leads than that of C-MORAL. ReMedI achieves the strict-success coverage of RL baseline C-MORAL, including on out-of-distribution tasks, with matched lead-relative similarity.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Drug Discovery, AI for Science, Chemical Language Models

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_qffiOpenReview ↗
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
  1. 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.
  2. 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.
  3. 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.
  4. 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
  1. What substantive new mechanisms does this paper have relative to MARS and GEGL? Which experiments can isolate the contributions of these mechanisms?
  2. Compared with representative non-LLM molecular editing methods, does ReMedI still have advantages under the same tasks, constraints, and budget?
  3. With the language model proposer fixed, what are the advantages of MH relative to MCTS, evolutionary search, and beam search?
  4. How is the effective proposal distribution calculated? How are invalid resampling, parsing mappings, and the actual decoding distribution handled?
  5. What is the complete cost of training trajectory search, distillation, and test-time search? What do 500 and 3,883 each count?
  6. How does C-MORAL perform when connected to the same refinement?
  7. How is the "approximately 2.8× coverage" calculated? Does the advantage still hold under explicit structural constraints?
  8. What are the properties and thresholds for each task? Which original constraints of SMDD-Bench are not included in the current evaluation?
5 comments
AuthorsOpenReview ↗

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)}\alpha(x,x') = \min\{1, \exp(U(x')-U(x)) \cdot \frac{q_\theta(x\mid x',P)}{q_\theta(x'\mid 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)=βSlog⁡S(x)+βGG(x)+βLL(x),U(x)=\beta_S\log S(x)+\beta_G G(x)+\beta_L L(x),

where S(x)S(x) measures aggregate property satisfaction, G(x)G(x) penalizes protected-property violations, and L(x)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-KK sampling, unguided sequential editing, greedy feedback, and MH under matched candidate budgets (SOR):

KIndependentSequentialGreedyMH (Ours)
10.05090.06980.06870.0815
40.14470.19330.18850.2478
80.20710.27480.26870.3214
120.24430.32740.31230.3924
160.27430.36360.34930.4254
200.29230.39290.37900.4631

MH leads at every candidate budget. At K=20K=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.

MethodSOR ↑SSOR ↑Sim ↑
ReMedI-MH (Ours)0.65330.38000.5155
MARS (adapted)0.54140.25670.5121
GEGL (adapted)0.53330.27330.5142
GP-MoLFormer-base+GA0.53330.25330.1868

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.

AuthorsOpenReview ↗

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.

MethodSOR ↑SSOR ↑Sim ↑
ReMedI-MH (Ours)0.65330.38000.5155
MARS (adapted)0.54140.25670.5121
GEGL (adapted)0.53330.27330.5142
GP-MoLFormer-base+GA0.53330.25330.1868

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.

AuthorsOpenReview ↗

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 qθ(z∣x,P)q_\theta(z\mid x,P) be the raw LM probability of generating string zz from molecule xx under focused prompt P=Pϕ(t)P=P_\phi(t), and let V\mathcal V denote valid, oracle-scoreable outputs.

Because invalid generations are retried without advancing the refinement chain, the effective proposal, considering validity filtering alone, is

qθvalid(x′∣x,P)=qθ(x′∣x,P)Zθ(x,P),Zθ(x,P)=Pr⁡z∼qθ(⋅∣x,P)(z∈V).q_\theta^{\mathrm{valid}}(x'\mid x,P)=\frac{q_\theta(x'\mid x,P)}{Z_\theta(x,P)},\qquad Z_\theta(x,P)=\Pr_{z\sim q_\theta(\cdot\mid x,P)}(z\in\mathcal V).

The state-dependent normalization Zθ(x,P)Z_\theta(x,P) 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 qθvalidq_\theta^{\mathrm{valid}}, 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 PP and utility UnU_n, under validity-only conditioning, this acceptance rule satisfies the exact MH equilibrium condition for

π~n,P(x)∝eUn(x)Zθ(x,P).\boxed{\widetilde{\pi}_{n,P}(x)\propto e^{U_n(x)}Z_\theta(x,P).}

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.**
Reviewer_MACeOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents REMEDI, which is a framework designed for molecular lead optimization by using the Metropolis-Hasting (MH) search strategy to refine the edits proposed by chemical language models. Using the C-MuMOInstruct and SMDD benchmarks, the authors show that REMEDI can increase the frozen base models success rate in multi-objective optimization. Ablation experiments suggest that while sequential editing generally outperforms independent sampling by a large margin, the proposed MH method can provide additional advantage in the evaluation. Additionally, the search trajectories are used to finetune the base CLM, which also increases model performance in the zero-shot generation. Ablation experiments suggest that one round distillation is the most effective and the lead-conditioned successful endpoints provide the main signal for learning edits satisfying the optimization targets.

Critical strengths
  • Novel idea: this paper introduces MH sampling into the sequential molecular editing proposed by LLM, which is a training-free method and demonstrates better performance compared with the RL baseline (C-MORAL). On a high level, this also agrees with recent work in using MCMC on general LLM reasoning (Karan ICLR 2026), so I think this direction is well-motivated.
  • Great presentation: the main text of this paper is well organized and easy to follow.
Critical weaknesses
  • Choice of baseline: this is my main concern. The authors use C-MORAL as the only baseline for all the evaluation, which is also a very recent work. Using AI/LLM for lead optimization is an active research area with different methodology, so the author should at least compare the proposed method with some well established methods like MARS and GP-MoLFormer-Sim, which the authors have also cited in the manuscript.
  • Customized metric of SSOR: the authors explicitly mention that they ‘emphasize SSOR because it captures simultaneous satisfaction of all property targets at near-optimality’. However, according to equation 9 in Appendix A1, it seems that SSOR depends heavily on the choice of near-optimal thresholds (which seems to be 0.1 according to the example in text below SSOR). The choice of this value seems to be arbitrary, making the near-optimal conclusion less convincing. Moreover, from equation 9, SSOR is balancing the different properties only (without considering similarity). Considering the authors also argued that ReMedI forms an empirical Pareto frontier in SSOR-similarity space, it would be better to make it clear what they refer to for the ‘optimality’ in SSOR.
Questions for authors
  1. Results presented in Figure 2 and Table 2 appear to suggest REMEDI could maximize property increase at the cost of lower similarity (non-local edits). In Figure 2, the main advantage of REMEDI at the Pareto Frontier is in the low similarity regime (below 0.6); In Table 2, REMEDI presents higher likelihood of property increase with lower similarity. How should we interpret these results together with the conclusion that REMEDI can achieve more success under similarity requirements?
  2. As I also mentioned in W2, how would the choice of threshold affect evaluation using SSOR? Will the current gain at high minimum similarity zone (Figure 2) become more significant when we use a lower threshold?
  3. Would the self-distillation affect the foundation model performance in general molecular editing / lead optimization tasks? (Tasks beyond C-MUMOInstruct like SMDD-Bench)

Minor points:

  1. What does REMEDI stand for? The author should add it.

  2. The authors may consider pointing to the Appendix A1 when introducing SOR and SSOR in section 5.1

  3. The authors may consider moving the legend in Figure 2 to the bottom or add a box for it, as I almost miss it during reading.

3 comments
AuthorsOpenReview ↗

Rebuttal (Part 1)

W1. Choice of Baselines

We thank the reviewer for the constructive feedback and agree that comparisons beyond C-MORAL are important for establishing ReMedI's effectiveness against established molecular optimization methods. To address this concern, we conducted additional experiments with MARS [1], GEGL[2], and GP-MoLFormer[3], adapting their search procedures to our lead-optimization setting.

We evaluate all methods on 150 held-out lead–task instances using the same property oracles, optimization objectives, candidate-selection criteria, and evaluation metrics, under a matched budget of 320 molecular proposals per lead. We use ReMedI with its frozen GeLLMO-Mistral proposal, without additional distillation. For MARS and GEGL, we additionally match mean structural similarity to ensure a meaningful comparison of molecular preservation.

MethodSOR ↑SSOR ↑Sim ↑
ReMedI-MH (Ours)0.65330.38000.5155
MARS (adapted)0.54140.25670.5121
GEGL (adapted)0.53330.27330.5142
GP-MoLFormer-base+GA0.53330.25330.1868

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.

We adapted MARS, GEGL, and GP-MoLFormer to our lead-optimization setting while preserving their core search mechanisms. Our results demonstrate that ReMedI outperforms these established approaches under matched molecular proposal budgets, particularly in multi-property optimization with structural preservation constraints.

AuthorsOpenReview ↗

Rebuttal (Part 2)

W2. Customized metric of SSOR

Choosing SSOR’s near-optimal thresholds. We thank the reviewer for highlighting this. Both the thresholds and the SSOR metric are defined by the source benchmark C-MuMOInstruct rather than by ReMedI. C-MuMOInstruct sets its property-specific absolute thresholds (Θp\Theta_p) using the 60th percentile of training-molecule property distributions among 256K molecular pairs (§2.3). It motivates strict success as reaching “pharmaceutically relevant” property thresholds while preserving already desirable properties [4]. Accordingly, our use of “near-optimality” follows this benchmark terminology and refers to simultaneous satisfaction of the prescribed property targets. We highlight that the thresholds distinguish properties needing improvement from those already meeting the benchmark targets, and SSOR requires every property to meet its target, so gains in one cannot offset failure in another.

Additionally, we clarify that the cutoffs are property-specific, rather than a universal tolerance of 0.1 in the illustrative example below Eq. 9. The SSOR targets are shown in the table below. Figure 2 separately measures strict-success coverage under the same minimum lead-similarity requirement for our method versus the baseline. Our intended claim is an empirical success–similarity tradeoff that is Pareto-dominant relative to C-MORAL.

PropertyNear-optimal threshold Θp\Theta_p
AMP ↑0.8
BBBP ↑0.8
CARC ↓0.2
DRD2 ↑0.4
hERG ↓0.3
HIA ↑0.4
LIV ↓0.9
MUT ↓0.2
PlogP ↑1.5
QED ↑0.9

Thresholds’ effects on ReMedI performance. To assess whether our performance gains depend on the source benchmark cutoffs, we re-evaluated fixed outputs under stricter and relaxed property targets. We define a relaxation parameter (rr) which shifts each default property cutoff by 100∣r∣%100|r|\%, with negative values imposing stricter targets and positive values relaxing them. We show that for both stricter and less strict thresholds, MH continues to improve SSOR beyond the base model, and distillation + MH sampling improves this further across all three backbones.

1. Stricter property targets. We re-evaluated the same fixed outputs under stricter property thresholds. At 5% stricter targets (r=−0.05r=-0.05), the q0-base → q0-MH → q1-MH SSOR rates are 8.05% → 13.68% → 16.10% for Llama, 7.65% → 11.67% → 15.09% for Mistral, and 3.02% → 8.85% → 11.67% for mCLM-4B. On IND instances, the same trend holds. Thus, MH improves performance under stricter targets, and distillation provides further gains across all three backbones.

2. Relaxed property targets. The benefits also persist under relaxed thresholds. At 5% relaxation (r=+0.05r=+0.05), the q0-base → q0-MH → q1-MH SSOR rates on IND are 8.20% → 22.20% → 30.00% for Llama, 8.00% → 17.60% → 27.20% for Mistral, and 2.60% → 10.20% → 18.80% for mCLM-4B. On OOD, the corresponding rates are 18.51% → 28.57% → 33.00%, 17.10% → 27.57% → 33.20%, and 7.24% → 17.71% → 25.55%, respectively.

We tested ten evaluated settings spanning r=−0.09r=-0.09 to r=+0.20r=+0.20. Across all ten, the curves consistently follow q0-base < q0-MH < q1-MH across all three backbones for both IND and OOD tasks. This demonstrates ReMedI’s performance persists beyond default benchmark targets.

AuthorsOpenReview ↗

Rebuttal (Part 3)

Q1. Achieving Performance Improvements at Matched Similarity

Figure 2 and Table 2 should be interpreted as complementary views of the tradeoff between property improvement and structural preservation. Our goal is not to maximize similarity, but to improve molecular properties while maintaining sufficient resemblance to the original lead. Accordingly, ReMedI uses a soft similarity target of 0.7 during optimization, consistent with the similarity constraint of SMDD-Bench.

Figure 2 shows what happens when similarity is treated as a hard requirement during evaluation. At each threshold, a molecule counts as successful only if it satisfies all property targets while meeting that minimum similarity. ReMedI achieves higher strict-success coverage over much of the evaluated range, although its advantage decreases at stringent thresholds. This shows where ReMedI's additional property improvements can be achieved without violating the specified similarity requirement.

Table 2 explains the tradeoff behind these gains. ReMedI achieves substantially higher SOR than C-MORAL on GeLLMO-Llama (0.764 vs. 0.580) and GeLLMO-Mistral (0.771 vs. 0.640), with moderately lower mean similarity (0.633 vs. 0.707 and 0.663 vs. 0.737, respectively). We also observe that C-MORAL sometimes returns the original lead unchanged, preserving similarity without improving its properties. ReMedI can instead select more substantial edits when they offer meaningful property improvements, rather than favoring structural preservation alone.

The key distinction is that Table 2 measures the average extent of editing, whereas Figure 2 measures how often those edits successfully satisfy both property and similarity requirements. ReMedI's advantage is therefore greater optimization success at a given minimum similarity, particularly when the constraint permits meaningful structural changes—not higher similarity for every generated molecule.

Q2. Thresholds’ effects on ReMedI performance.

This question has been answered above in W2.

Q3. Self-distillation’s effects on foundation model performance in general molecular editing / lead optimization tasks (Tasks beyond C-MuMOInstruct like SMDD-Bench).

We additionally evaluate on 314 SMDD-Bench instances, comparing our original (q0q_0) and distilled (q1q_1) proposals against pooled C-MORAL.

Backboneq0q_0-Direct SOR ↑q1q_1-Direct SOR ↑C-MORAL Pooled SOR ↑
GeLLMO-Mistral0.69110.74410.6401
GeLLMO-Llama0.61150.65420.5796
mCLM-4B0.52870.57610.5064

Our distilled proposal achieves the highest SOR across all three backbones, outperforming pooled C-MORAL by 10.4, 7.5, and 7.0 percentage points, respectively. Notably, C-MORAL's C-MuMOInstruct-trained models perform worse than the original q0q_0 models on SMDD-Bench, whereas our distilled q1q_1 improves over q0q_0 by approximately 7–9% across all backbones. This demonstrates that search-derived supervision transfers beyond C-MuMOInstruct and improves general molecular editing performance.

For minor points: ReMedI stands for Refinement-based Molecular Editing via Inference-time Sampling. We will add a pointer to appendix A1 for the SOR and SSOR sections and add a box for the legend for Figure 2. We thank the reviewer for the careful review and constructive feedback!

[1] Xie et. al., MARS: Markov Molecular Sampling for Multi-objective Drug Discovery

[2] Ahn et. al., Guiding Deep Molecular Optimization with Genetic Exploration

[3] Navratil et. al., GP-MoLFormer-Sim: Test Time Molecular Optimization through Contextual Similarity Guidance

[4] Dey et. al., Large language models for controllable multi-property multi-objective molecule optimization

Reviewer_pu8DOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

ReMedI combines a molecular language model with Metropolis–Hastings (MH) refinement for multi-objective lead optimization. The model proposes molecular edits, while an acceptance rule balances property improvement, preservation, and similarity to the original lead. Search-generated examples are subsequently used for supervised distillation. Although the reported improvements are encouraging, the methodological novelty and experimental controls are insufficiently established.

Critical strengths
  1. The method addresses a practically relevant problem: improving multiple molecular properties while limiting deviation from the original lead.
  2. Experiments cover three model backbones, two benchmarks, and held-out property combinations.
  3. Learning from search-generated examples provides a useful way to improve both direct generation and subsequent refinement.
Critical weaknesses

Limited methodological novelty. The framework largely combines established components: instruction-conditioned molecular editing, MH acceptance, adaptive prompting, and weighted supervised distillation. The paper does not sufficiently establish what constitutes a substantive advance over existing iterative molecular search and search-based self-training methods.

The computational fairness of the comparison is unclear. Selecting the best of 20 final candidates does not equalize the resources used to produce them. MH refinement requires additional proposals, oracle evaluations, and reverse-probability computations. Comparisons with larger-budget direct sampling and greedy iterative editing under matched query and runtime budgets are necessary to isolate the benefit of the proposed search mechanism.

The implementation of the Hastings correction requires clarification. The method retries invalid generations and parses model outputs into molecules, which can change the effective proposal distribution. The paper should explain how qvalidq^{\mathrm{valid}} is computed, including state-dependent normalization and the mapping between generated strings and molecular states. Without this clarification, it is difficult to assess whether the implemented acceptance rule corresponds to the stated MH formulation.

The data-efficiency claim does not establish overall resource efficiency. Using fewer initial lead–task instances does not account for the cost of generating search trajectories or the resulting number of training examples. Total oracle calls, generated tokens, and data-collection and training costs should be reported alongside the comparison with C-MORAL.

Independent validation of molecular improvements is insufficiently established. Optimizing and evaluating against the same property predictors can favor candidates that exploit predictor errors. Independent rescoring or higher-fidelity validation would strengthen the conclusions. Fingerprint similarity alone also does not establish preservation of the original scaffold or biological potency.

Questions for authors

None

4 comments
AuthorsOpenReview ↗

Rebuttal (Part 1)

W1. Limited Methodological Novelty

We thank the reviewer for the constructive feedback and agree that ReMedI builds on MCMC and search-guided learning. Our novelty is a constraint-aware search-and-distillation pipeline for multi-objective molecular editing that outperforms existing approaches. MARS [1] and GEGL [2] also combine established techniques but primarily address de novo molecular optimization. ReMedI instead edits existing leads, improving multiple properties while limiting structural drift. Unlike MARS's graph-based proposals, ReMedI uses instruction-conditioned molecular language models and incorporates state-dependent proposal likelihoods into MH acceptance. Adaptive prompting targets unsatisfied constraints; the acceptance objective balances property improvements, protected properties, and lead similarity. Unlike GEGL's imitation of high-reward molecules, our distillation learns from successful endpoints, intermediate edits, and unsuccessful proposals, improving both direct generation and subsequent MH refinement across unseen objective combinations.

Three experimental findings support these contributions:

1. Distillation improves generation and subsequent refinement. For GeLLMO-Mistral, direct-generation SOR improves from 0.324 to 0.524 on IND tasks and 0.268 to 0.372 on OOD tasks. With MH refinement, SOR improves from 0.528 to 0.692 on IND and 0.400 to 0.532 on OOD tasks. These gains come from one shared adapter trained on five objective combinations, demonstrating transfer to unseen tasks.

2. ReMedI outperforms established molecular optimization methods. Using the same 150 held-out lead–task instances, property oracles, optimization objectives, and 320-proposal budget per lead, ReMedI achieves 0.653 SOR and 0.380 SSOR, versus 0.541/0.257 for MARS [1] and 0.533/0.273 for GEGL [2], respectively, at comparable mean structural similarity.

3. MH refinement improves over conventional search. Under matched generation budgets, MH achieves 0.463 SOR, compared with 0.393 for unguided sequential editing, 0.379 for greedy refinement, and 0.292 for independent sampling.

Together, these results show that our combination of refinement and distillation yields improvements beyond those achieved by the individual search strategies and existing molecular optimization baselines. We will revise the manuscript to clarify our contribution and benefits.

W2. Computational Fairness of the Comparison

We address the computational fairness concern through two controlled comparisons: matching the total generation budget and the number of candidates used for final selection. This tests whether the improvements come from the search procedure rather than additional sampling.

1. Comparison under a fixed generation budget.

We fix the budget to 320 attempted generations per lead, retaining all unique valid molecules for final selection. This ensures identical generation and retention protocols for MH refinement, direct sampling, and C-MORAL.

BackboneMetric ↑C-MORALq0-baseq0-MHq1-baseq1-MH
GeLLMO-LlamaSOR0.3070.4540.6980.6030.757
SSOR0.1210.1510.4720.3830.571
GeLLMO-MistralSOR0.4600.4510.6650.6280.751
SSOR0.1660.1720.3910.3020.524
mCLM-4BSOR0.3720.3930.4770.4860.562
SSOR0.0900.1330.1820.2050.279

Under the same generation budget, MH achieves the highest SOR and SSOR across all three backbones, outperforming direct sampling and C-MORAL. These improvements cannot be attributed solely to additional sampling or proposal distillation.

2. Isolating the benefit of the search strategy.

We further compare independent sampling, unguided sequential editing, greedy feedback, and MH under the same generation budget. We vary the number of candidates (KK) used for final selection to control for selection-pool size. We evaluate 250 GeLLMO-Mistral lead–task instances.

KIndependentSequentialGreedyMH (ours)
10.05090.06980.06870.0815
40.14470.19330.18850.2478
80.20710.27480.26870.3214
120.24430.32740.31230.3924
160.27430.36360.34930.4254
200.29230.39290.37900.4631

MH achieves the highest SOR across all candidate budgets. At K=20K=20, it reaches 0.4631, compared with 0.3929 for unguided sequential editing, 0.3790 for greedy feedback, and 0.2923 for independent sampling.

Overall, sequential editing improves over independent sampling, while MH provides additional gains over unguided and greedy refinement. These improvements persist under matched generation budgets, candidate-retention rules, and final selection sizes, demonstrating gains beyond additional sampling.

AuthorsOpenReview ↗

Rebuttal (Part 2)

W3. Implementation of the Hastings Correction

We establish that validity-conditioned sampling satisfies the exact MH equilibrium condition, while molecular preprocessing introduces a tractable approximation whose benefits are demonstrated by ablations.

1. Effective proposal distribution and state-dependent normalization.

Following Section 4, let qθ(⋅∣x,P)q_\theta(\cdot\mid x,P) denote the raw LM proposal conditioned on molecule xx and focused prompt P=Pϕ(t)P=P_\phi(t). Let V\mathcal V denote the set of valid, oracle-scoreable outputs.

Our implementation retries invalid generations without advancing the chain. Under validity-only conditioning, the effective proposal becomes

qθvalid(x′∣x,P)=qθ(x′∣x,P)Zθ(x,P),q_\theta^{\mathrm{valid}}(x'\mid x,P)=\frac{q_\theta(x'\mid x,P)}{Z_\theta(x,P)},

where

Zθ(x,P)=Pr⁡x′∼qθ(⋅∣x,P)(x′∈V).Z_\theta(x,P)=\Pr_{x'\sim q_\theta(\cdot\mid x,P)}(x'\in\mathcal V).

Zθ(x,P)Z_\theta(x,P) is the state-dependent probability of generating a valid molecular edit. Thus, validity retries condition the LM proposal on chemical admissibility, concentrating the search budget on valid candidates.

2. Proposition: Interpretation of the implemented Hastings correction.

Eq. (6) expresses the Hastings correction using qθvalidq_\theta^{\mathrm{valid}}, whereas our implementation evaluates the acceptance ratio 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\\},

where UnU_n is the step-specific utility.

Proposition. For fixed PP and UnU_n, under validity-only conditioning, the implemented acceptance rule satisfies the exact MH equilibrium condition for

π~n,P(x)∝eUn(x)Zθ(x,P).\boxed{\widetilde{\pi}_{n,P}(x)\propto e^{U_n(x)}Z_\theta(x,P).}

Proof. Although validity retries introduce the state-dependent normalization Zθ(x,P)Z_\theta(x,P), this factor cancels when the probability flow is evaluated under π~n,P\widetilde{\pi}_{n,P}:

\begin{aligned}&\widetilde{\pi}_{n,P}(x)q_\theta^{\mathrm{valid}}(x'\mid x,P)\alpha_n(x,x')\\\\&\propto\min\left\\{e^{U_n(x)}q_\theta(x'\mid x,P),e^{U_n(x')}q_\theta(x\mid x',P)\right\\}\\\\&=\widetilde{\pi}_{n,P}(x')q_\theta^{\mathrm{valid}}(x\mid x',P)\alpha_n(x',x).\end{aligned}

The middle expression is symmetric in xx and x′x', establishing the equilibrium condition. Thus, validity retries and raw-probability acceptance form an exact MH procedure under validity-only conditioning, with a feasibility-aware target proportional to eUn(x)Zθ(x,P)e^{U_n(x)}Z_\theta(x,P).

3. Mapping generated strings to molecular states.

Our implementation applies a deterministic mapping h(z)h(z) that extracts, canonicalizes, and removes salts from generated strings zz. Since multiple strings can map to the same molecule x′x', the effective molecular proposal becomes

qH(x′∣x,P)=∑z∈V\h(z)=x′qθ(z∣x,P)Zθ(x,P).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)}.

The implementation evaluates raw LM probabilities for canonical representations rather than summing over equivalent strings. This provides a tractable approximation to the molecular proposal distribution, while the validity-only formulation retains the exact equilibrium result above. The equilibrium condition applies to fixed prompts and utilities, rather than the entire adaptive refinement procedure.

4. Empirical evaluation of molecular preprocessing.

We ablate both preprocessing components.

Validity retries. We compare the shipped protocol (On), which redraws invalid generations without advancing the chain, against consuming the step on an invalid generation (Off). All other settings remain identical, using the frozen q0q_0 proposal.

BackboneSOR (On/Off)SSOR (On/Off)Sim (On/Off)
GeLLMO-Mistral0.448/0.4120.172/0.1520.710/0.755
GeLLMO-Llama0.500/0.3800.172/0.1320.670/0.775
mCLM-4B0.295/0.2950.071/0.0810.746/0.752

Validity retries improve SOR for both GeLLMO backbones, particularly GeLLMO-Llama (0.380 to 0.500), demonstrating the benefit of allocating refinement steps to valid molecular proposals.

Salt stripping. We evaluate its contribution across representations.

BackboneSOR (On/Off)
GeLLMO-Mistral0.448/0.448
GeLLMO-Llama0.500/0.500
mCLM-4B0.292/0.000

The benefit of salt stripping is model-dependent. The GeLLMO-Llama and GeLLMO-Mistral backbones predominantly produce usable SMILES representations, so preprocessing does not change their success rates. In contrast, mCLM-4B produces molecule-block outputs that can contain salts, making normalization essential: without salt stripping, SOR collapses to 0.000, whereas enabling it restores SOR to 0.292. Overall, validity retries admit an exact MH equilibrium interpretation, while the ablations demonstrate the importance of molecular normalization, particularly for mCLM-4B, without affecting the GeLLMO backbones.

AuthorsOpenReview ↗

Rebuttal (Part 3)

W4. Overall Resource Efficiency

We provide a detailed resource comparison with C-MORAL, accounting for the cost of generating search trajectories, oracle evaluations, and the resulting training data.

StageReMedIC-MORAL
Training-time generations~105K actual (~160K maximum)~847K
Training-time oracle calls~91K actual~847K reward evaluations
Training data55,599 trajectory-derived examples100,000 training examples
Training optimizationOne shared adapter trained through supervised distillation301 iterations × 2 PPO epochs × 8 minibatches = 4,816 optimizer steps per task
Adapters trained1, shared across 10 tasks10, trained separately per task

The comparison shows that ReMedI requires approximately 9.3× fewer training-time oracle evaluations (91K versus 847K) and 8.1× fewer generations (105K versus 847K), while producing a single transferable adapter instead of ten task-specific policies. ReMedI uses supervised distillation rather than C-MORAL's per-task GRPO optimization.

To clarify the dataset sizes, 500 refers to the lead–task instances used for trajectory generation (100 per task across five training tasks), whereas 3,883 refers to the unique training leads pooled across C-MORAL's 100,000 training examples.

Overall, these results demonstrate that ReMedI's data efficiency extends to lower trajectory-generation and oracle-evaluation costs, while producing a single adapter that can be reused across multiple optimization tasks.

AuthorsOpenReview ↗

Rebuttal (Part 4)

W5 Independent validation of molecular improvements

We added two alternative property evaluators [3,4] and Murcko scaffold [5] analysis to examine whether the improvements transfer beyond the optimization oracle and to distinguish property success from structural preservation.

Property-success gains persist under alternative evaluators. We rescored frozen outputs using ADMETlab [4] and the mCLM’s PropertyEvaluator [3], neither of which participated in ReMedI’s optimization or distillation. No candidates were regenerated or reselected. We evaluate matched source–task instances, covering both IND and OOD tasks.

Across both alternative evaluators, ReMedI q1-MH outperforms C-MORAL on SOR and SSOR for both GeLLMO backbones. Under ADMETlab, SOR increases from 19.05% to 31.83% for Llama and 27.57% to 35.09% for Mistral; SSOR increases from 2.51% to 8.02% and 4.01% to 6.27%, respectively. Under mCLM’s PropertyEvaluator, SOR increases from 24.00% to 38.00% for Llama and 30.00% to 42.00% for Mistral; SSOR increases from 6.50% to 20.00% and 7.00% to 19.00%, respectively.

Since ADMETlab and mCLM’s PropertyEvaluator support only a subset of the properties available in ADMET-AI, we restrict evaluation to tasks for which all required properties can be evaluated: BPQ, ELQ, BCMQ, and HLMPQ for ADMETlab, and BPQ and HLMPQ for mCLM’s PropertyEvaluator.

ADMETlab — BPQ, ELQ, BCMQ, HLMPQ.

BackboneMetric ↑C-MORALq0-baseq0-MHq1-baseq1-MH
GeLLMO-LlamaSOR (%)19.0522.8131.8326.3231.83
SSOR (%)2.513.264.514.768.02
GeLLMO-MistralSOR (%)27.5719.8027.8226.8235.09
SSOR (%)4.013.015.015.016.27
mCLM-4BSOR (%)9.178.5018.5014.5023.00
SSOR (%)0.830.251.001.004.00

mCLM’s PropertyEvaluator — BPQ, HLMPQ.

BackboneMetric ↑C-MORALq0-baseq0-MHq1-baseq1-MH
GeLLMO-LlamaSOR (%)24.0025.5037.0031.5038.00
SSOR (%)6.506.0014.0011.5020.00
GeLLMO-MistralSOR (%)30.0024.0031.0029.0042.00
SSOR (%)7.004.009.008.5019.00
mCLM-4BSOR (%)16.829.5021.5018.5036.00
SSOR (%)0.001.003.004.0013.00

We agree that fingerprint similarity alone does not establish scaffold preservation or biological potency. We therefore added RDKit Bemis–Murcko scaffold analysis [3], assessing both scaffold similarity and exact retention across all ten tasks. For q1-MH, mean scaffold similarities are 0.469/0.508 for Llama, 0.499/0.534 for Mistral, and 0.472/0.481 for mCLM-4B, for IND/OOD respectively. This deviates very little from similarities reported in the paper.

Together with the SOR and SSOR improvements under ADMETlab and mCLM’s PropertyEvaluator, we show that ReMedI’s performance gains are not confined to a single evaluator and are reproduced when the same frozen outputs are assessed using different property oracles.

[1] Xie et. al., MARS: Markov Molecular Sampling for Multi-objective Drug Discovery

[2] Ahn et. al., Guiding Deep Molecular Optimization with Genetic Exploration

[3] Edwards et. al., mclm: A modular chemical language model that generates functional and makeable molecules

[4] Fu et. al., ADMETlab 3.0: an updated comprehensive online ADMET prediction platform enhanced with broader coverage, improved performance, API functionality and decision support.

[5] Greg Landrum. RDKit: Open-source cheminformatics. https://www.rdkit.org.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.