acceptodds
Under review as a conference paper at ICLR 2027

MSF-HDiff: Molecular Semantic Field-Guided Multi-Agent Hierarchical Diffusion for Small-Molecule Antibiotic Discovery

Abstract

The continued rise in antimicrobial resistance creates an urgent need to accelerate antibiotic discovery. Artificial intelligence (AI) has shown great promise in accelerating small-molecule antibiotic discovery, specifically generative models that can design de novo small molecules. However, existing generative models often struggle to simultaneously satisfy molecular requirements while making targeted chemical modifications to optimize desired properties. To address this, we introduce MSF-HDiff, a closed-loop framework that combines multi-agent reasoning with hierarchically conditioned diffusion for de novo small-molecule antibiotic discovery. Our model features a novel framework that integrates hierarchical condition search, transferable molecular semantic fields for property-directed chemical guidance, and reference-based feasibility control during diffusion sampling. This enables iterative molecular design that balances global antibiotic objectives with local chemical optimization. MSF-HDiff outperforms other generative models in generating de novo small molecules with strong antibacterial activity, favorable molecular properties, and high structural reliability as shown in Figure 1.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Small-molecule antibiotic discovery, Hierarchical diffusion, Multi-agent systems, Molecular semantic field, Controllable molecular generation, Phenotype-driven design

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_M5gfOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes an approach for generating small molecule antibiotics. The paper highlights the limitations of antibacterial design objective, limited applicability of property targets for the new antibiotic design, issues with local property optimisation competing with global structural requirements of the molecule. To overcome these, three contributions are made using agents for integrating biological evidence, molecular semantic fields for property directed design and integration of property guidance with structural and compositional constraints.

Critical strengths
  1. The problem itself is very important.
  2. the paper highlights the issues with current approaches quite well.
Critical weaknesses
  1. the paper is hard to follow from exactly when the paper starts to introduce its own approach
  2. It is not clear how the paper is resolving the three key challenges it discusses in lines 068-079
Questions for authors
  1. Please explain how exactly your approach is overcoming the three challenges that you mentioned.

  2. How do you argue the necessity of each module? Can one do ablation studies?

  3. What exactly are molecular semantic fields? I understand a ten dimensional property vector is created but I guess that is the target for property prediction. what is the input ? Is that composed of the chemical words in context of the actual molecule? It is presently unclear from the description.

4 comments

Dear Reviewer M5gf,

Thank you for your thoughtful and constructive feedback. To address your concerns and questions, we have conducted additional experiments evaluating the contributions of individual mechanisms within the HDiff module through targeted ablation studies, including comparisons with a validity/C/F rejection-filter control. We have also investigated the effectiveness of semantic-window selection by comparing learned selection against random selection of executable semantic windows on held-out task. In addition to these experiments, we have revised the Introduction to more explicitly articulate how our approach addresses the three key challenges identified in the manuscript and clarified the proposed methodology and molecular semantic fields. We hope that these additional experiments and clarifications adequately address your concerns and improve the clarity of the manuscript.

W1 – Lack of clarity in the presentation of the proposed methodology

We have revised the Proposed Approach section to improve its clarity and organization, making it easier for readers to follow the transition from the background and motivation to the introduction of our proposed method. We have also clarified the descriptions of the key components and their respective roles to provide a more coherent explanation of our approach.

W2/Q1 – Unclear mapping between the three identified challenges and the proposed solutions

We have clarified how the proposed framework addresses the three challenges outlined in the Introduction below: The first main challenge is translating biological antibacterial objectives into actionable molecular design strategies, as biological objectives do not directly specify the molecular structures required to achieve them. To address this, we developed a closed-loop multi-agent framework that integrates biological evidence and computational feedback to guide molecular generation, coordinating global design objectives with the exploration and refinement of local structural features.

The second main challenge is understanding how combinations of local chemical features influence whole-molecule properties, as desired property targets do not specify which combinations of chemical features will produce the intended improvements. To address this, we introduced molecular semantic fields, learned from 228,539 molecular-pair records spanning ten physicochemical properties, to capture relationships between combinations of local structural features and changes in whole-molecule properties. These learned relationships enable the identification of structural modifications likely to achieve desired molecular property profiles.

The third main challenge is balancing local property optimization with broader structural and compositional requirements to ensure that generated molecules satisfy multiple design objectives simultaneously. To address this, we developed a hierarchical diffusion framework with feasibility control that integrates semantic-field-based property guidance with structural and compositional constraints during molecular generation, balancing desired property improvements with the generation of chemically feasible molecules. We have incorporated these clarifications into the revised Introduction to ensure it is clear how our novel model addresses these three challenges.

Q2 – Justification for individual modules and the need for component-wise ablation studies

To systematically evaluate the necessity and individual contributions of the proposed modules, we conducted comprehensive ablation studies at four levels: (i) Overall framework ablation, comparing MSF-HDiff against GPT-based and standard diffusion-based alternatives; (ii) Multi-agent planning ablation, independently removing the Librarian, Discovery, and Semantic Field agents; (iii) Hierarchical conditioning ablation, independently removing individual global and structural conditions, as well as their complete condition sets; and (iv) Hierarchical diffusion and feasibility-control ablation, comparing ON-only and OFF-only configurations, independently removing the Reference Value Network (RVN) and minimum-KL adjustment, and evaluating alternative constraint-enforcement strategies, including Hard Rejection Mode and post-generation filtering.

The ablation studies demonstrate that each component contributes complementary benefits to molecular generation. Removing planning agents or hierarchical conditions reduced predicted activity and molecular quality, with the Discovery Agent having the greatest impact on activity. Additionally, RVN and minimum-KL independently improved molecular quality, while MSF-HDiff outperformed Hard Rejection Mode and post-generation filtering in ranking score, independent composite score, and diversity. Overall, these results highlight the contributions of the proposed planning, conditioning, and reference-based feasibility mechanisms. We have detailed this all in Appendix K.

Q3 – Clarification of the definition, representation, and construction of molecular semantic fields

We thank the reviewer for identifying this lack of clarity. Molecular semantic fields (MSFs) are not simply ten-dimensional molecular property vectors; rather, they represent how local chemical substructures and their combinations contribute to a molecule's overall chemical meaning.

The concept draws on two analogies: in a physical field, local disturbances collectively produce an overall response, while in language, words acquire meaning through their neighbouring words and combinations within a sentence. Similarly, chemical words represent local structural units, the complete molecule forms a "chemical sentence," and semantic windows capture local combinations of chemical words within their molecular context. We define an MSF as a representation of how these co-occurring semantic-window combinations jointly shape the integrated chemical meaning of a molecule.

The ten semantic axes provide measurable coordinates for this whole-molecule meaning, covering physicochemical properties related to lipophilicity, polarity, hydrogen bonding, flexibility, and ionization. Importantly, this ten-dimensional vector is the target representing the global semantic state or change, not the sole model input.

The model input comprises the source SMILES, its ordered chemical-word sequence, its initial semantic state, and the set of paired semantic windows describing local structural changes. Each window retains its source–target chemical words and contextual positions. The complete window set is associated with a single ten-dimensional target representing the resulting global semantic change. This enables the model to learn how combinations of local structural changes, within the context of the complete molecule, influence overall molecular properties.

During generation, the Semantic Field Agent uses these learned relationships, alongside the objectives, reference molecules, and previous feedback, to select target-window combinations, which HDiff realizes through complete-molecule generation. We have clarified these definitions and the input–target distinction in Section 3.1.1 and Appendix D.

Dear Reviewer M5gf, Thank you for your thoughtful feedback. With the ICLR author–reviewer discussion closing today (November 18, AoE), we wanted to let you know that we have posted our response. We have addressed each of your comments and included the results of all the analyses you requested.

We appreciate your time, especially on short notice, and would be grateful if you could consider the new results as you finalize your assessment. If any questions remain, we would be happy to respond before the discussion closes.

Best, The Authors

Thank you very much for your response. I believe that with large amounts of changes and the poor readability of the paper in the original submission the paper would need a full review. therefore, I keep my score.

Reviewer_H69QOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes MSF-HDiff, a multi-agent hierarchical diffusion framework for de novo small-molecule antibiotic discovery. The framework combines three main components: (1) a multi-agent system that translates biological objectives and feedback into hierarchical molecular constraints, (2) molecular semantic fields learned from 228,539 molecular transformation pairs to guide local chemical properties, and (3) a conditional diffusion model with reference-based feasibility control to balance property optimization and structural validity.The authors evaluate the framework on E. coli and S. aureus, comparing it against five public baselines and several ablated variants. The reported results demonstrate high molecular validity, uniqueness, drug-likeness, and favorable predicted molecular properties. However, antibacterial activity is evaluated computationally rather than experimentally.

Critical strengths
  1. Integration of hierarchical molecular constraints with diffusion-based generation. The framework provides a structured mechanism for combining global compositional constraints, scaffold-level requirements, and local chemical-property guidance. The reference-based feasibility mechanism is particularly interesting, as it attempts to prevent semantic guidance from substantially degrading the probability of generating structurally qualified molecules. This provides a potentially useful approach to multi-objective molecular generation.

  2. Comprehensive computational evaluation of generated molecular properties. The authors evaluate both structural reliability and multiple molecular properties across two bacterial species. MSF-HDiff achieves high validity and uniqueness, with VUN rates of 100.00% and 99.13%, and PoseBusters validity of 98.20% and 98.30% for E. coli and S. aureus, respectively. The improvements in QED and synthetic accessibility relative to several baselines are also encouraging. These results suggest that the framework can generate chemically plausible molecules with favorable computationally predicted properties.

Critical weaknesses
  1. Limited evidence that the proposed framework improves actual antibacterial discovery. The primary concern is that the evaluation relies almost entirely on computational predictions rather than experimentally measured antibacterial activity.Although the authors introduce an independent pMIC predictor, this is another machine-learning model (a Morgan-fingerprint random forest), rather than an independent experimental assessment. It is unclear whether the predictor provides reliable estimates for newly generated molecules outside its training distribution.Furthermore, MSF-HDiff does not consistently achieve the highest predicted antibacterial activity. For example, Table 2 shows:
  • For E. coli, MSF-HDiff achieves an independent pMIC of 4.32, compared with 4.77 for MSF-Guided GPT.
  • For S. aureus, MSF-HDiff achieves 4.43, compared with 5.01 for MegaMolBART.
  • For E. coli, the independent composite scores of MSF-HDiff (62.02) and MSF-Guided GPT (62.01) are nearly identical. These results suggest that the improvements are driven primarily by better optimization of general molecular properties rather than consistent improvements in predicted antibacterial activity.Athough experimental validation may be beyond the scope of this work, stronger evidence of activity-predictor generalization and a more careful characterization of the activity–property trade-off are needed to substantiate the central claims.
  1. Insufficient experimental isolation of the proposed methodological contributions. The paper combines several existing ideas, including multi-agent planning, molecular transformation learning, conditional diffusion, and value-based sampling. While their integration is potentially useful, the specific contributions of molecular semantic fields and reference-based feasibility control are not sufficiently established. In particular, the following issues remain: First, the molecular semantic fields are constructed from local molecular transformations, but the diffusion model generates molecules from fully masked sequences. It is therefore unclear whether the selected semantic windows actually produce the intended local chemical changes or reliably control the target properties.Second, the proposed reference-based feasibility mechanism is a central technical contribution, but the experiments do not directly isolate its effectiveness against simpler alternatives, such as standard conditional diffusion, rejection sampling, or diffusion without the reference-value constraint. The reported structural metrics also do not directly establish satisfaction of the hierarchical structural conditions.Third, although generation-attempt budgets are matched, the comparison does not clearly account for the additional computational cost of multi-agent reasoning, repeated property evaluation, and reference-guided sampling. Overall, more targeted ablations and controlled comparisons are necessary to establish which components are responsible for the improvements and whether the added complexity is justified.
Questions for authors

Q1. How reliable are the antibacterial activity predictors on generated molecules? The main antibacterial evaluation relies on Chemprop and an independent random-forest predictor. Could the authors clarify how these predictors were trained and validated, including whether scaffold-disjoint or out-of-distribution testing was performed? Do the predictors share experimental training data? Given the potential for optimization against imperfect predictors, how do the authors establish that higher predicted pMIC corresponds to genuinely improved antibacterial activity?

Q2. What is the specific advantage of MSF-HDiff over MSF-Guided GPT? For E. coli, Table 2 shows nearly identical independent composite scores for MSF-HDiff (62.02) and MSF-Guided GPT (62.01), while MSF-Guided GPT achieves higher independently predicted pMIC (4.77 versus 4.32). Could the authors clarify which performance improvements justify the additional complexity of hierarchical diffusion? Are the differences statistically significant across independent runs?

Q3. Do molecular semantic fields actually provide transferable and controllable chemical guidance? The proposed semantic fields are learned from molecular transformation pairs, but the generator samples complete molecules rather than explicitly editing the original molecules. Could the authors provide quantitative evidence that selected semantic windows lead to the intended structural modifications and property changes? A comparison against simpler alternatives, such as random semantic-window selection or descriptor-based property guidance, would help establish the contribution of the learned semantic fields.

Q4. How much does reference-based feasibility control contribute independently? The Reference Value Network and minimum-KL sampling mechanism are presented as important methodological innovations. Could the authors isolate their contributions by comparing the full method against ON-only diffusion, OFF-only diffusion, and simpler validity-based filtering under comparable computational budgets? It would also be helpful to report calibration metrics for the Reference Value Network and actual satisfaction rates for the global and structural constraints, rather than relying primarily on general molecular validity metrics.

Q5. Are the baseline comparisons fair in terms of computational resources and optimization opportunities? The experiments use the same number of molecular generation attempts across models. However, MSF-HDiff additionally employs multi-agent planning, repeated property evaluation, and reference-guided sampling. Could the authors report the total computational costs, including GPU time, language-model calls, and property-predictor evaluations? Additionally, were the baseline models given comparable access to task-specific molecular evidence and opportunities for iterative optimization?

7 comments

Dear Reviewer H69Q,

Thank you for your thoughtful and constructive feedback. To address your concerns and questions, we have conducted additional experiments on independent MIC predictor validation, candidate ranking and activity–property trade-offs, ablation of the hierarchical diffusion mechanisms and feasibility controls, and end-to-end computational cost. Additionally, we have provided further clarification on the interpretation of our results and the contributions of the proposed framework. We hope that these new experiments and clarifications help address your concerns.

W1a – Experimental validation and activity-predictor generalization

We would like to clarify that our primary objective is the computational de novo generation of antibacterial candidates, rather than experimental validation. Our framework generates novel molecular structures without wet-lab testing during generation, while synthesis and antibacterial testing are subsequent steps. Importantly, our MIC predictors are trained on experimentally measured data, grounding their predictions in observed antibacterial activity.

To address concerns regarding predictor reliability and dependence on a single model, we conducted two additional analyses. First, we trained external-data pMIC predictors for E. coli and S. aureus using the same fingerprint-based random forest architecture and preprocessing pipeline as the original Independent pMIC Model. We benchmarked these predictors, alongside publicly available pMIC prediction models (SPARK-2503-ecol and SPARK-2503-saur), against the original Independent pMIC Model using experimental MIC measurements. The original model demonstrated the strongest agreement with experimental measurements across both species and all evaluation metrics (Table R1). Second, we reevaluated the generated molecules using the two alternative external predictors, recalculating predicted pMIC values and composite evaluation scores while keeping all other evaluation criteria and scoring weights unchanged. Despite variations in absolute predicted pMIC values and some rankings, MSF-HDiff consistently achieved the highest composite evaluation score across all three predictor configurations for both species (Table R2). These results indicate that the comparative advantage of MSF-HDiff is robust across different activity predictors and is not dependent on a single model or training dataset. Further details are provided in Appendix J.

Table R1.Performance comparison of the original independent pMIC model, external dataset trained pMIC model, and SPARK-2503 models for E. coli and S. aureus pMIC prediction.

MetricIndependent pMIC ModelExternal-Data pMIC ModelSPARK-2503 pMIC Model
E. coli
MAE0.60 ± 0.050.71 ± 0.050.83 ± 0.07
RMSE0.78 ± 0.050.92 ± 0.071.09 ± 0.08
R²0.43 ± 0.080.20 ± 0.11-0.11 ± 0.18
Pearson's r0.66 ± 0.060.57 ± 0.070.41 ± 0.09
Spearman's ρ0.55 ± 0.080.43 ± 0.080.28 ± 0.09
S. aureus
MAE0.54 ± 0.031.07 ± 0.081.74 ± 0.13
RMSE0.70 ± 0.041.33 ± 0.082.08 ± 0.13
R²0.57 ± 0.05-0.53 ± 0.23-2.77 ± 0.84
Pearson's r0.76 ± 0.030.23 ± 0.070.32 ± 0.06
Spearman's ρ0.75 ± 0.040.17 ± 0.110.33 ± 0.09

Table R2. Robustness of generated compound evaluation to independent MIC predictor selection.

Compound Generation ModelPredicted pMIC (↑)Composite Score (↑)
Independent pMIC ModelExternal-Data pMIC ModelSPARK-2503 pMIC ModelIndependent pMIC ModelExternal-Data pMIC ModelSPARK-2503 pMIC Model
E. coli
MSF-HDiff4.324.154.2662.0260.3861.62
LDMol4.153.993.6751.0649.4747.37
DrugDiff4.093.973.8548.8447.7247.21
DemoDiff4.214.073.7355.1753.8451.58
MegaMolBART4.634.274.1554.7851.5650.76
AutoMolDesigner4.374.103.9658.0455.5154.83
SyntheMol-RL4.524.113.7354.6450.9048.31
S. aureus
MSF-HDiff4.434.253.0263.5061.8854.47
LDMol4.294.062.0254.6452.6143.41
DrugDiff4.254.042.3548.9447.0538.97
DemoDiff4.284.062.6054.3752.4345.54
MegaMolBART5.014.253.6659.0452.0849.21
AutoMolDesigner4.404.182.7757.3255.3848.01
SyntheMol-RL4.614.152.7055.7251.4845.68

W1b – Activity–property trade-offs

We acknowledge that MSF-HDiff does not consistently achieve the highest predicted pMIC across all comparisons. As mentioned, MegaMolBART achieves higher independent pMIC than MSF-HDiff for S. aureus. However, this comes at the expense of other important properties, including substantially lower drug-likeness (QED: 0.37 vs. 0.71 for E. coli; 0.46 vs. 0.77 for S. aureus), poorer synthetic accessibility, and higher predicted toxicity. It also exhibits lower structural reliability and reduced novelty for S. aureus (87.40% vs. 100%). Consequently, MegaMolBART achieves lower independent composite scores for both E. coli (54.78 vs. 62.02) and S. aureus (59.04 vs. 63.50), despite its higher predicted pMIC. This showcases the multi-objective design of MSF-HDiff, which prioritizes balancing predicted antibacterial activity with broader molecular quality, drug-likeness, synthetic accessibility, toxicity, and structural reliability, rather than maximizing predicted activity in isolation.

Furthermore, MSF-Guided GPT is an ablation variant of MSF-HDiff rather than an independent baseline. Its comparable performance is expected, as it retains aspects of the underlying framework while omitting specific components. The nearly identical E. coli composite scores (62.02 vs. 62.01) indicate comparable overall candidate quality. We go ahead with MSF-HDiff because it still outperforms MSF-Guided GPT, explained further in Q2.

W1c – Activity-predictor generalization and Characterization of the Activity–Property Trade-off

As described in our response to comment W1a, we conducted additional analyses to assess the generalization and robustness of our MIC predictors using independently trained predictors on external datasets and alternative pretrained SPARK-2503 models.

Regarding the activity–property trade-off, MSF-HDiff was designed for multi-objective molecular optimization rather than maximizing predicted antibacterial activity alone. It balances predicted activity, drug-likeness (QED), synthetic accessibility (SAS), toxicity, and structural reliability, achieving the strongest overall performance across these properties while maintaining strong predicted antibacterial activity. This demonstrates its ability to generate candidates with a balanced molecular profile relevant to drug development.

Furthermore, we conducted a reference-neighborhood analysis using experimentally measured E. coli activity data from SPARK to independently assess the target relevance of the structural regions explored by generated molecules. MSF-HDiff's reference neighborhood contains 57.35% active molecules, compared with 33.92% in the background, corresponding to approximately 1.69-fold activity enrichment. More details are found in Appendix N. Together, these findings connect multi-property evaluation with independent experimental reference evidence, supporting MSF-HDiff's ability to explore target-relevant structural regions while coordinating molecular properties.

W2a/Q3 – Transferability and controllability of molecular semantic fields

We added an additional ablation to test directly whether learned semantic-window selection and conditioning provide controllable guidance, and whether they do so better than simpler alternatives. We added a HDiff w/ Random W ablation, which retains the full pipeline (ON/OFF pathways, RVN, and minimum-KL adjustment) but replaces the agent-selected windows with randomly assigned ones. This allows us to isolate the contribution of the information contained in the selected windows, rather than simply the presence of conditioning. On the E. coli task, MSF-HDiff outperformed Random W on every metric except diversity (Table R8).

We also evaluated how frequently generated molecules satisfy each target condition. Under matched settings, MSF-HDiff exceeds Random W on every condition (Table R9). The largest gap occurs for the joint condition (66.7% vs. 43.3%), which provides a more stringent test of controllability. MSF-HDiff also exceeds ON-only Mode (39.0% joint), indicating that combining the ON and OFF pathways with importance-corrected mixing and feasibility control contributes to reliably satisfying multiple conditions simultaneously.

To further examine the directional effects of semantic conditioning, Figure A1(C) in the manuscript shows the shifts in cLogP and TPSA of generated molecules relative to the OFF-only baseline, which uses no semantic windows. With the learned windows, MSF-HDiff shifts cLogP downward and TPSA upward, consistent with the intended property changes. ON-only Mode shows the same trend. In contrast, Random W produces shifts centered near zero, while the OFF-only baseline is centered at zero by construction. These results suggest that learned semantic windows guide generation along specific, interpretable property directions that random windows do not reproduce. Further details are found in Appendix K.4.

Table R8. Comparison of MSF-HDiff and Random Semantic-Window Selection on Molecular Generation Performance.

MetricMSF-HDiffRandom W
Independent pMIC (↑)4.323.88
QED (↑)0.710.47
SAS (↓)2.873.70
Toxicity (↓)0.140.17
Ranking score (↑)63.5451.40
Independent composite score (↑)62.0251.32
Diversity (↑)0.860.91

Table R9. Target-Condition Satisfaction Rates of MSF-HDiff and Random Semantic-Window Selection on the E. coli Task.

ConditionMSF-HDiffRandom W
C96.1%74.3%
F73.0%49.9%
W77.4%71.4%
Joint {C, F, W}66.7%43.3%

W2b/Q4 – Ablation and validation of reference-based feasibility mechanisms

To address this concern, we expanded the E. coli ablation study to evaluate semantic conditioning, pathway contributions, and constraint-enforcement mechanisms. We compared OFF-only Mode (without semantic conditioning), ON-only Mode (ON pathway alone), and HDiff with Random (W) (random rather than agent-selected semantic windows). We also introduced w/o RVN and w/o Minimum-KL ablations to isolate the contributions of RVN-based feasibility enforcement and minimum-KL sampling, respectively. Finally, we evaluated Hard Rejection Mode, which retains only candidates meeting validity and reference-constraint criteria under the same computational time budget, as well as post-generation filtering as an alternative constraint-enforcement strategy. As shown in Table R5, MSF-HDiff benefits from the complementary contributions of semantic conditioning, dual-pathway sampling, RVN-based feasibility enforcement, and minimum-KL sampling. Additional results are provided in Appendix K.4.

Table R5. Ablation results of hierarchical diffusion module components and feasibility control in MSF-HDiff on the E. coli task.

ConfigurationDiversity (↑)Independent pMIC (↑)Oracle pMIC (↑)QED (↑)SAS (↓)Toxicity (↓)Ranking score (↑)Independent composite score (↑)
MSF-HDiff0.864.324.470.712.870.1463.5462.02
HDiff w/ Random W0.913.883.900.473.700.1751.4051.32
ON-only Mode0.874.304.440.573.280.1758.9157.66
OFF-only Mode0.894.144.090.343.040.1951.0751.51
w/o RVN0.874.264.390.633.170.1959.3058.09
w/o Minimum-KL0.884.324.430.583.300.1758.8157.86
Hard Rejection Mode0.794.074.190.453.050.1554.9253.94
Pose-generation filtering0.844.144.150.593.200.1756.9256.91
Reviewer_bjZ5OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

MSF-HDiff is a closed-loop system for de novo antibacterial small-molecule design that couples multi-agent planning with a conditioned absorbing-mask diffusion model over SELFIES. Three agents supply conditions at different timescales. A Librarian Agent assembles literature evidence, reference actives and user examples and fixes run-level global conditions, namely an allowed element set and a molecular-weight interval. A Discovery Agent sets a structural plan per stage, covering scaffold archetype, ring-system profile and motif topology across eleven categorical fields, together with property-level guidance. A Semantic Field Agent converts that guidance into one to four local target windows. The semantic-field idea is that local structural units influence whole-molecule properties in context: a database of 228,539 standardized molecular pairs links paired source and target windows around an unchanged core to a ten-axis property response, attributed to the complete paired-window group rather than additively to individual windows. The agent is a LoRA adapter trained in two phases, first on contextual property-response prediction and then on executable window selection with group-relative policy optimization against a batch reward. The diffusion module injects global and structural conditions into the backbone and semantic windows through gated residual adapters that can be switched off. A Reference Value Network estimates the probability that an intermediate state will decode into a valid molecule satisfying the global and structural conditions under window-free continuation, and reverse sampling solves a minimum-KL reweighting of the semantically conditioned candidate distribution subject to a feasibility bound derived from window-free candidates. Evaluation covers two bacterial targets, with structural reliability over three seeds under a fixed 1,000-attempt budget, molecular properties over the top 1,000 of 10,000 attempts, comparisons against five public generators and three study-specific ablations, a reciprocal-rank-sum aggregate, chemical-space analyses, and component ablations removing individual agents and condition blocks.

Critical strengths

The system is documented to a depth that would allow reimplementation: notation and condition definitions (L756-917), six algorithms covering the overall loop, semantic-field preparation, agent training and inference, and diffusion training and sampling (L918-1241), and explicit reward construction (L1566-1619). Attributing the ten-axis property response to the complete paired-window group rather than additively to individual windows is a defensible modelling choice for context-dependent edits (L219-237). The generation budget counts invalid and duplicate attempts in the denominator and batches are only initiated when the remaining budget allows a full batch (L216-218, L1404-1412), which is stricter accounting than is common. Structural reliability is reported as mean and standard deviation over three seeds (L391-412), and the component ablation covers each agent and each condition block separately rather than only the full-versus-none contrast (L2052-2105). The ablation table discloses trade-offs rather than hiding them, including configurations where removing a component raises diversity while degrading other properties (L2052-2105).

Critical weaknesses

1.The "independent" evaluation is not independent: the oracle is a message-passing network used in the loop and the independent score swaps in a random forest on fingerprints trained on the same actives (L1976-1990), with all other terms and weights unchanged (L2004-2012), so the two composite scores differ only in predictor family and share the same training data and applicability domain. 2.Molecular results are the top 1,000 of 10,000 attempts (L379-383), selected by the composite score that is then reported, which favours the one method that optimizes that score during generation; the paper does not state whether baselines were ranked by the same function or their own. 3.Several headline metrics are saturated or definitionally weak: novelty is absence of the canonical SMILES from ChEMBL and is 100 percent for every method (L1890-1943), so the VUN gap to the diffusion ablation, 100.00 against 43.30 on one target, is a uniqueness collapse in the ablation rather than a gain from the method (L391-412). 4.The abstract's aggregate figure uses reciprocal-rank summation across metrics (L029-047, L430-433) that are not independent, since VUN is a product of three of the columns and end-to-end validity is a product of two more, so one underlying success is counted several times. 5.Table 2 and the ablation table carry no error bars at all, yet differences of a few points are read as effects; on the independent composite for one target the proposed model scores 62.02 against 62.01 for the variant that replaces the diffusion module with a language model, which the text itself describes as comparable (L385-389, L414-429). 6.The gains appear to come from the planning layer rather than the diffusion module: adding planning and guidance to the language-model generator raises its independent composite from 53.02 to 62.01 and its validity from 87.07 to 100.00 (L369-389), and the two agent-only variants diverge inconsistently between the two targets. 7.No ablation isolates the Reference Value Network or the minimum-KL sampling, since the diffusion ablation removes both at once and collapses (L362-368); the exponential-tilting procedure is never compared against simple rejection of infeasible candidates. 8.The recursive feasibility relationship is stated under exact transition probabilities, exact value estimates, consistent support and exact optimization (L1880-1889), so it is a property of an idealization and gives no bound for the deployed sampler. 9.The method is positioned for antibiotic discovery but omits the synthesizability-constrained antibiotic generators it cites (L108-120); the public baselines are general-purpose generators plus one antibacterial design tool.

Questions for authors

Were the baselines' 10,000 candidates ranked by your composite score or by their own criteria before taking the top 1,000 (L379-383)? Please also report the full 10,000-candidate distributions, since top-decile selection under the reported objective is the single largest source of advantage. The oracle and the independent predictor are trained on the same actives (L356-361, L1976-1990). Can you provide a genuinely held-out check, for example a predictor trained on a time-split or scaffold-split holdout, or external activity data not used anywhere in the pipeline? Were the species-specific predictors trained only on molecules active below the stated threshold (L356-361)? If no inactives were included, how is the regression expected to rank de novo molecules far outside that set? Can you report Table 2 and the ablation table with mean and standard deviation over the same three seeds used for structural reliability? Several conclusions rest on differences of under one point, including a reported tie on the independent composite (L385-389). What does the full model add over the language-model variant with planning and guidance, given 62.01 against 62.02 on one target and a reversal on the other (L369-389)? An ablation that keeps the diffusion backbone while removing only the Reference Value Network, and separately only the minimum-KL step, would answer this. How does the minimum-KL reweighting compare against simply rejecting candidates whose estimated feasibility falls below the window-free baseline, at matched compute (L1836-1889)? Diversity is lowest for the proposed model and rises when components are removed (L414-429, L2052-2105). What evidence, beyond the in-house scorers, indicates that the narrowed region is the productive one rather than the scorer's preferred one? Would you add a synthesizability-constrained antibiotic generator as a baseline (L108-120), since the paper's framing is antibiotic discovery and synthetic accessibility is one of the four ranking terms?

6 comments

Dear Reviewer bjZ5,

Thank you for your thoughtful and constructive feedback. To address your concerns and questions, we have conducted additional experiments on the independent MIC predictor, compound candidate ranking and activity–property trade-offs, uncertainty estimates and ablation of HDiff mechanisms, and comparison with the SyntheMol-RL baseline. In addition, we have clarified the abstract's aggregate performance metric, the complementary contributions of the planning and diffusion modules, and the theoretical assumptions underlying the recursive feasibility relationship. The additional results provide a more comprehensive evaluation of the model, offering further insights into its predictive reliability, ranking robustness, the contributions of individual model components, and performance relative to existing methods. We hope these new experiments and clarifications adequately address your concerns.

W1/Q2 - The independent evaluation of the oracle is not truly independent

To address this concern, we performed additional benchmarking using externally sourced experimental MIC data and publicly available pretrained pMIC prediction models. First, we trained an External-Data pMIC Model with the same fingerprint-based random forest architecture on external data that doesn’t overlap with our independent pMIC model. Additionally, we also evaluated publicly available SPARK-2503 models for both bacterial species. The original Independent pMIC Model achieved the strongest agreement with experimental MIC measurements across all metrics for both E. coli and S. aureus, supporting its selection as the activity predictor within our framework (Table R1).

We also investigated whether the choice of activity predictor influenced the comparative evaluation of generated compounds. We reevaluated the top 1,000 generated compounds from MSF-HDiff and each baseline model using both alternative external predictors, retaining the original generated compound sets and recalculating their predicted pMIC values and composite evaluation scores. All other evaluation criteria and scoring weights were kept unchanged. MSF-HDiff consistently achieved the highest composite evaluation score across all three predictor configurations for both E. coli and S. aureus. Although the absolute predicted pMIC values and composite scores varied across predictors, the relative advantage of MSF-HDiff over all baseline methods was maintained (Table R2).

Together, these experiments demonstrate that MSF-HDiff's comparative advantage is robust across activity predictors and not dependent on the original predictor or training data. This is detailed further in Appendix J.

Table R1.Performance comparison of the original independent pMIC model, external dataset trained pMIC model, and SPARK-2503 models for E. coli and S. aureus pMIC prediction.

MetricIndependent pMIC ModelExternal-Data pMIC ModelSPARK-2503 pMIC Model
E. coli
MAE0.60 ± 0.050.71 ± 0.050.83 ± 0.07
RMSE0.78 ± 0.050.92 ± 0.071.09 ± 0.08
R²0.43 ± 0.080.20 ± 0.11-0.11 ± 0.18
Pearson's r0.66 ± 0.060.57 ± 0.070.41 ± 0.09
Spearman's ρ0.55 ± 0.080.43 ± 0.080.28 ± 0.09
S. aureus
MAE0.54 ± 0.031.07 ± 0.081.74 ± 0.13
RMSE0.70 ± 0.041.33 ± 0.082.08 ± 0.13
R²0.57 ± 0.05-0.53 ± 0.23-2.77 ± 0.84
Pearson's r0.76 ± 0.030.23 ± 0.070.32 ± 0.06
Spearman's ρ0.75 ± 0.040.17 ± 0.110.33 ± 0.09

Table R2. Robustness of generated compound evaluation to independent MIC predictor selection.

Compound Generation ModelPredicted pMIC (↑)Composite Score (↑)
Independent pMIC ModelExternal-Data pMIC ModelSPARK-2503 pMIC ModelIndependent pMIC ModelExternal-Data pMIC ModelSPARK-2503 pMIC Model
E. coli
MSF-HDiff4.324.154.2662.0260.3861.62
LDMol4.153.993.6751.0649.4747.37
DrugDiff4.093.973.8548.8447.7247.21
DemoDiff4.214.073.7355.1753.8451.58
MegaMolBART4.634.274.1554.7851.5650.76
AutoMolDesigner4.374.103.9658.0455.5154.83
SyntheMol-RL4.524.113.7354.6450.9048.31
S. aureus
MSF-HDiff4.434.253.0263.5061.8854.47
LDMol4.294.062.0254.6452.6143.41
DrugDiff4.254.042.3548.9447.0538.97
DemoDiff4.284.062.6054.3752.4345.54
MegaMolBART5.014.253.6659.0452.0849.21
AutoMolDesigner4.404.182.7757.3255.3848.01
SyntheMol-RL4.614.152.7055.7251.4845.68

W2/Q1 - Top-1,000 selection introduces potential bias in favour of the proposed method

The original Table 2 evaluates the quality of the 1,000 highest-ranked candidates each method produces within a budget of 10,000 attempts (each model was ranked by their own scoring function). In MSF-HDiff, this budget also supports a search over generation conditions. Earlier batches test structural and semantic conditions, and their results, including lower-scoring candidates, inform subsequent batches. This search performs part of the task-specific optimization at inference time, with model weights held fixed. For baselines that sample directly after training, the generation budget serves a different role. Within the rebuttal period, we therefore focused our additional experiment on whether conditions identified by the search yield reproducible performance on fresh samples. For the E. coli task, we fixed the five condition sets that contributed the most candidates to the original top 1,000. In each of three independently seeded runs, HDiff made 2,000 generation attempts per condition set. The new table reports property metrics without score-based selection, summarized as mean ± SD across runs, alongside the corresponding unranked baseline results. This SD measures sampling variability under the five fixed conditions. We omit diversity from this table because fixing five condition sets restricts the structural contexts HDiff explores, which would confound a diversity comparison with baselines sampled under their usual protocols. As seen in Table R3, MSF-HDiff achieved an independent composite score of 61.16 ± 1.14, outperforming all evaluated baseline methods in terms of mean composite score. These results have been added to Appendix I.

Table R3. Molecular property evaluation without score-based selection. Mean ± standard deviation across three independently seeded generation runs.

ModelIndependent pMIC (↑)Oracle pMIC (↑)QED (↑)SAS (↓)Toxicity (↓)Ranking score (↑)Independent composite score (↑)
MSF-HDiff4.25 ± 0.064.37 ± 0.080.72 ± 0.022.92 ± 0.100.16 ± 0.0162.37 ± 1.3961.16 ± 1.14
LDMol4.10 ± 0.043.96 ± 0.080.35 ± 0.013.83 ± 0.150.19 ± 0.0249.45 ± 1.4150.12 ± 1.02
DrugDiff4.03 ± 0.073.78 ± 0.120.37 ± 0.034.86 ± 0.180.16 ± 0.0146.43 ± 0.7147.98 ± 0.85
DemoDiff4.16 ± 0.023.91 ± 0.070.47 ± 0.013.82 ± 0.080.14 ± 0.0152.02 ± 0.9053.87 ± 0.44
MegaMolBART4.66 ± 0.104.70 ± 0.120.38 ± 0.054.08 ± 0.220.17 ± 0.0455.62 ± 2.8955.14 ± 2.63
AutoMolDesigner4.34 ± 0.064.04 ± 0.080.51 ± 0.033.23 ± 0.200.16 ± 0.0254.35 ± 0.1756.95 ± 0.29
SyntheMol-RL4.55 ± 0.084.23 ± 0.110.61 ± 0.043.79 ± 0.240.37 ± 0.1251.89 ± 2.0055.40 ± 2.20

W3 - Several headline metrics are saturated or weakly defined

We agree that novelty and VUN capture different aspects of molecular generation: novelty measures exact molecular-identity overlap with ChEMBL, whereas VUN measures the yield of valid, distinct, database-absent molecules per generation attempt.

The generator-replacement ablation (MSF-Guided Diff), which retains multi-agent planning and hierarchical guidance but replaces HDiff with standard diffusion without a Reference Value Network or reference-guided sampling, directly evaluates this capability. The results demonstrate a substantial loss of uniqueness following this replacement. For E. coli, under the same budget of 1,000 generation attempts, MSF-HDiff achieves 100.00% VUN compared with 43.30% for MSF-Guided Diff, corresponding to more than twice the yield of distinct, valid, database-absent candidates. This is detailed in Appendix K.1.

Furthermore, the additional external baseline, SyntheMol-RL, achieves 100% validity and novelty on both targets but exhibits VUN values of 54.00% and 63.67%, respectively. This demonstrates that VUN remains informative when validity and novelty reach their respective ceilings, and that the distinct-candidate yield advantage of MSF-HDiff extends beyond comparisons against its own ablations.

W4 - The abstract's aggregate score double-counts correlated metrics

Figure 1 provides a multi-criterion performance summary, with evaluation criteria rather than independent statistical events as the units of aggregation. Reciprocal-rank summation places metrics with different units and scales on a common relative-performance scale, making statistical independence unnecessary for this descriptive aggregation. Component and end-to-end metrics serve complementary purposes: validity, uniqueness, and novelty assess distinct aspects of generation, while VUN captures their combined yield per generation attempt. Similarly, individual structural checks assess specific aspects of molecular quality, whereas end-to-end PoseBusters validity evaluates successful structure construction against all required checks. Reporting both distinguishes individual performance differences from overall success in satisfying combined requirements. Thus, including both component-level and composite metrics intentionally reflects an evaluation scheme that considers both individual criteria and overall pipeline yield. We therefore interpret the reciprocal-rank summary as a compact descriptive comparison, rather than an independent statistical measure of overall performance. The underlying metric values, reported individually in Tables 1–2, remain the primary basis for assessing performance differences.

W5/Q4 - Key performance comparisons lack uncertainty estimates

Reporting mean ± SD for Table 2 requires capturing variability across the full generation and selection pipeline, not generation alone. As discussed in W2/Q1, MSF-HDiff adaptively searches generation conditions across its 10,000 attempts, with earlier batches informing subsequent ones. Thus, each independent run would need to repeat the condition-search process and top-1,000 selection. The same applies to the ablation results, requiring the corresponding evaluation procedure to be repeated for each model variant.

To further investigate the reproducibility of the identified generation conditions, we conducted an additional experiment using fixed conditions, as discussed in W1/Q2. For E. coli, we selected the five condition sets contributing the most candidates to the original top 1,000. Property metrics were evaluated without score-based selection across three independent seeds and compared with unranked baseline results. MSF-HDiff outperformed all evaluated baselines. Further details in Appendix I.

W6 - The observed gains appear to originate primarily from the planning layer

We agree that planning contributes substantially to generation performance. However, planning and HDiff serve complementary roles: planning determines which structural and semantic conditions to explore, whereas HDiff determines how these conditions are jointly realized within a generated molecule. HDiff achieves this through hierarchical conditioning, semantic adapters, and reference coordination during denoising.

We performed additional ablation experiments comparing the full MSF-HDiff framework against MSF-Guided Diff and MSF-Guided GPT, which replace the HDiff generation module with alternative diffusion- and GPT-based generators while retaining the planning and guidance framework. We further performed component-level ablations by removing the Reference Value Network (RVN), which helps preserve structural requirements under semantic guidance, and Minimum-KL regularization, which coordinates molecular property objectives during generation. The results (Table R4) show that the full model achieves the highest joint condition adherence among the compared generator configurations. Together, these findings demonstrate that HDiff contributes beyond planning by improving the coordinated realization of structural and semantic conditions and balancing molecular properties during generation.

Table R4. Ablation analysis of the planning and hierarchical diffusion components in MSF-HDiff on the E. coli task. Condition adherence rates (%) are reported for the individual conditions (C, F, and W) and their joint satisfaction.

ConfigurationC (%)F (%)W (%)Joint adherence (%)Independent pMICQEDSASToxicity
MSF-HDiff96.173.077.466.74.320.712.870.14
MSF-Guided Diff84.957.349.937.04.280.533.050.26
MSF-Guided GPT98.465.776.662.74.770.663.000.28
HDiff w/ Random W74.349.971.443.33.880.473.700.17
ON-only Mode86.659.070.139.04.300.573.280.17
OFF-only Mode91.368.3N/AN/A4.140.343.040.19
w/o RVN86.658.777.653.84.260.633.170.19
w/o Minimum-KL93.166.653.047.34.320.583.300.17
Hard Rejection Mode86.165.644.839.94.070.453.050.15

Comments on the paper

3 comments

Dear Area Chair,

We sincerely appreciate your time, careful evaluation, and thoughtful consideration. To facilitate your decision and provide a clear overview, we summarize the following:

  1. Core motivation of the paper
  2. The reviewers’ assessments
  3. Additional analyses completed during rebuttal to address the reviewers’ concerns

(1) Core Motivation and Novelties of Our Paper

Our work introduces MSF-HDiff, a novel framework for closed-loop, multi-agent small-molecule antibiotic discovery that integrates molecular semantic fields with hierarchically conditioned diffusion. Existing AI-driven molecular generation approaches often struggle to translate desired molecular properties into targeted chemical modifications while maintaining structural feasibility and satisfying broader molecular requirements. MSF-HDiff addresses these challenges by integrating hierarchical condition search, transferable molecular semantic fields, multi-agent reasoning, and reference-based feasibility control into a unified generative framework. These innovations enable property-directed molecular optimization, coordinated multi-objective generation, and the iterative design of structurally reliable antibiotic candidates.

The key contributions of MSF-HDiff include:

  • Closed-loop, multi-agent antibiotic discovery: Integrates biological evidence, specialized agent reasoning, and computational feedback to iteratively guide molecular generation.
  • Molecular semantic fields for property-directed design: Learns relationships between local chemical features and whole-molecule properties using a database of 228,539 molecular-pair records.
  • Hierarchical diffusion with feasibility control: Coordinates global molecular requirements with local property guidance while maintaining structural and compositional feasibility.
  • Transferable chemical guidance: Enables learned local chemical modifications to guide de novo molecular generation across different molecular contexts.

(2) Reviewer Ratings and Assessments

ReviewerOriginal RatingComments
bjZ52• Recognized the methodological depth, reproducibility, and comprehensive component ablations of MSF-HDiff.
• Highlighted the integration of molecular semantic fields and hierarchical feasibility control as interesting contributions.
• Raised concerns regarding evaluation independence, top-1,000 selection bias, metric redundancy, lack of statistical uncertainty in molecular property results, and insufficient isolation of the Reference Value Network and minimum-KL sampling.
• Requested stronger evidence of semantic-window effectiveness, predictor generalization, diversity–activity trade-offs, and comparisons against synthesizability-constrained antibiotic generators.
H69Q2• Appreciated the integration of hierarchical molecular constraints, semantic-field guidance, and reference-based feasibility control.
• Recognized the strong computational results in molecular validity, uniqueness, and drug-likeness.
• Raised concerns about the lack of experimental antibacterial validation, generalization of activity predictors, and limited evidence that semantic windows produce intended chemical modifications.
• Requested more targeted ablations, comparisons with simpler feasibility-control methods, structural-condition satisfaction rates, and computational cost analyses.
M5gf1• Recognized the importance of the antibiotic discovery problem and the relevance of the challenges identified in the manuscript.
• Requested clearer explanations of the individual modules and their necessity through ablation studies.
• Requested clarification of the definition, inputs, and underlying representation of molecular semantic fields.

Overall, the reviewers recognized the potential of MSF-HDiff's integrated multi-agent and hierarchical diffusion framework. They highlighted the methodological depth, modular design, and promising computational results. Their feedback primarily focused on requests for clearer methodological explanations and additional analyses to strengthen the validation of individual components, predictor reliability, and evaluation robustness. All points raised have been addressed through detailed clarifications and substantial new experiments, and these updates have been incorporated into the revised manuscript.

(3) Additional Analyses Completed

Based on the review discussion, here is a table summarizing the new experiments added and concerns addressed:

New Experiment / AnalysisWhat Was DoneMajor ResultReviewer Comments Addressed
Independent MIC Predictor ValidationBenchmarked the original independent pMIC predictor against an externally trained RF model and pretrained SPARK-2503 models using experimental MIC measurements.The original independent predictor showed the strongest agreement with experimental MIC measurements across all metrics for both species.bjZ5 W1/Q2; H69Q W1a, W1c, Q1
Alternative Predictor RobustnessRe-evaluated the top 1,000 generated compounds from MSF-HDiff and baselines using two alternative external activity predictors, keeping other scoring criteria unchanged.MSF-HDiff achieved the highest composite score across all three predictor configurations for both species, demonstrating robustness to predictor selection.bjZ5 W1/Q2; H69Q W1a, W1c, Q1
Unranked Generation EvaluationGenerated compounds under five fixed E. coli condition sets across three independent seeds and evaluated all generated compounds without score-based selection.MSF-HDiff achieved an independent composite score of 61.16 ± 1.14, outperforming all evaluated baselines in mean composite score.bjZ5 W2/Q1, W5
Activity–Property Trade-off AnalysisCompared predicted pMIC, QED, SAS, toxicity, structural reliability, novelty, and composite scores across generated compounds.MSF-HDiff achieved stronger overall composite performance, demonstrating a more balanced molecular quality profile rather than maximizing predicted antibacterial activity alone.H69Q W1b, W1c, Q1
HDiff Component AblationsCompared ON-only, OFF-only, Random W, w/o RVN, and w/o Minimum-KL configurations, alongside alternative generator configurations.Results demonstrated complementary contributions from semantic conditioning, dual-pathway sampling, RVN-based feasibility enforcement, and minimum-KL sampling.bjZ5 W6/W7; H69Q W2b/Q4; M5gf Q2
Feasibility-Control ComparisonsCompared MSF-HDiff against Hard Rejection Mode and post-generation filtering as alternative constraint-enforcement strategies.MSF-HDiff outperformed both alternatives in ranking score, independent composite score, and molecular diversity.bjZ5 W7; H69Q W2b/Q4; M5gf Q2
Semantic-Window SelectionCompared learned semantic-window selection against randomly assigned executable semantic windows using the same generation framework.MSF-HDiff outperformed Random W on every evaluated molecular property metric except diversity.H69Q W2a/Q3; M5gf Q2
Target-Condition Satisfaction and Property ControlEvaluated individual and joint condition satisfaction rates and changes in cLogP and TPSA relative to the OFF-only baseline.MSF-HDiff achieved higher satisfaction across every target condition than Random W, including joint satisfaction of 66.7% vs. 43.3%. Learned windows also produced the intended directional property shifts.H69Q W2a/Q3; M5gf Q2
SyntheMol-RL BaselineCompared MSF-HDiff against SyntheMol-RL using structural reliability and molecular property metrics.MSF-HDiff demonstrated more balanced performance across molecular diversity, toxicity, and overall molecular properties.bjZ5 W9/Q8
End-to-End Computational CostCompared resource consumption for generating 10,000 molecules using MSF-HDiff and MSF-Guided GPT, including GPU usage, API calls, and token consumption.MSF-HDiff reduced external API calls by 92.5% and output token consumption by 97.2%, at the expense of increased local computation and elapsed time.H69Q W2c/Q5, Q2
Module-Level Computational CostQuantified resource consumption across individual MSF-HDiff modules.Provided a breakdown of computational requirements and the resource contributions of individual modules.H69Q W2c/Q5
Experimental Reference-Neighborhood AnalysisCompared the top 1,000 generated molecules from each model against experimentally characterized E. coli reference compounds and assessed active molecule and Bemis–Murcko scaffold enrichment.MSF-HDiff achieved 1.69-fold molecule-level enrichment and 1.89-fold scaffold-level enrichment in experimentally active reference compounds relative to the SPARK background.bjZ5 Q7; H69Q W1a/Q1

We have also clarified the following points:

Clarification / RevisionWhat Was AddressedReviewer Comments
MIC Predictor Training DataClarified the species-specific training dataset composition, including active and inactive compound counts and activity thresholds.bjZ5 Q3
Abstract Aggregate Performance MetricClarified the rationale for reciprocal-rank aggregation and explained the distinction between individual evaluation criteria and composite metrics.bjZ5 W4
Theoretical Feasibility AssumptionsDistinguished the theoretical feasibility relationship under idealized assumptions from the practical RVN-estimated constraint enforced during deployment.bjZ5 W8
Planning Versus Diffusion ContributionsClarified the complementary roles of multi-agent planning and HDiff in selecting design conditions and realizing them during molecular generation.bjZ5 W6
Baseline FairnessExplained that publicly available baselines retained their original optimization mechanisms and were evaluated using the same generation budget, while acknowledging differences in computational requirements.H69Q W2c/Q5
Three Challenges and Proposed SolutionsExplicitly mapped the three challenges in the Introduction to the corresponding multi-agent planning, molecular semantic field, and hierarchical diffusion components.M5gf W2/Q1
Methodology PresentationReorganized the Proposed Approach section to improve the explanation, flow, and roles of individual components.M5gf W1
Molecular Semantic Field DefinitionClarified the definition, representation, construction, input–target distinction, and role of semantic windows in molecular generation.M5gf Q3

We thank the reviewers for their thoughtful feedback. The ratings in Section 2 remain as originally submitted, and we have yet to receive follow-up comments on the additional analyses presented in our rebuttal. We would be grateful if you could weigh this new evidence, summarized in Section 3, alongside the initial reviews when forming your recommendation. Thank you for your time and consideration.

Sincerely,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.