acceptodds
Under review as a conference paper at ICLR 2027

Multi-Pose Structural Evidence for LLM Agent Molecular Optimization in Structure-Based Drug Design

Abstract

Structure based molecular optimization aims to refine candidate ligands with improved binding scores and molecular quality. Existing methods typically lack explicit modeling of interaction consistency across docking poses and verification of interaction changes after optimization. This paper proposes MEDD, an LLM agent framework for structural evidence guided molecular optimization. It aggregates protein ligand interactions across multiple docking poses into a unified structural evidence representation. This representation is then used to identify stable interactions and pose sensitive regions based on cross pose consistency. The LLM then formulates testable interaction hypotheses and translates them into constrained local molecular modifications. The resulting candidates are docked again to verify whether the hypothesized interaction changes are realized while key interactions are preserved. Experiments on CrossDocked2020 show that MEDD achieves improved docking scores while maintaining QED, SA, and MRR, and achieves a higher key interaction preservation rate and a higher proportion of structurally supported optimization hypotheses. These results support the use of explicit structural evidence for reliable and verifiable LLM guided molecular optimization.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Structure-Based Drug Design, Molecular Optimization, Large Language Model Agents, Molecular Docking, Structural Evidence

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

This paper proposes MEDD, an LLM-agent framework for structure-based molecular optimization using multi-pose docking evidence. The method aggregates protein–ligand interactions across multiple docking poses to distinguish stable contacts from variable optimization opportunities. An LLM agent then generates interaction hypotheses and constrained molecular modifications, followed by redocking and structural verification. Experiments on CrossDocked2020 demonstrate improvements over the initial MolCRAFT-generated molecules in docking scores, drug-likeness, synthetic accessibility, and molecular reasonability. The authors further provide component ablations and structural case studies to support the proposed framework.

Critical strengths

1. Well-motivated structural evidence integration.

The paper addresses a meaningful limitation of docking-guided molecular optimization: relying on a single docking pose may introduce unreliable structural guidance. Aggregating interactions across multiple poses is a reasonable strategy for reducing sensitivity to individual predicted configurations.

2. Interpretable hypothesis-driven optimization.

The explicit hypothesis–editing–verification workflow is conceptually appealing. Separating LLM reasoning from chemical transformations and structural evaluation improves the traceability of molecular modifications and provides a clearer interpretation of optimization decisions.

3. Comprehensive evaluation of molecular properties.

The experiments consider multiple complementary metrics, including Vina scores, QED, SA, MRR, and structural similarity. The ablation studies also suggest that multi-pose evidence, explicit hypotheses, and verification feedback contribute to optimization performance.

Critical weaknesses

W1. Limited methodological novelty beyond existing molecular optimization frameworks.

The main contribution primarily involves integrating multi-pose interaction fingerprints, LLM-based molecular editing, and docking-based verification into an iterative optimization workflow. These components are individually well established, while the proposed cross-pose consistency measure is essentially an interaction occurrence frequency. The paper does not sufficiently demonstrate a substantial methodological advance beyond their integration.

W2. Questionable assumptions underlying multi-pose evidence and potentially over-constrained molecular optimization.

The method assumes that interactions consistently observed across docking poses provide more reliable structural evidence. However, poses generated by the same docking algorithm may be highly correlated, and interaction frequency does not necessarily reflect physical interaction stability. More importantly, preserving stable contacts or their associated molecular motifs may unnecessarily restrict the chemical search space, since alternative functional groups or binding geometries can establish equally favorable or stronger interactions with the same protein residues. This potential exploration–exploitation trade-off is not adequately investigated and may partly explain MEDD's inferior performance relative to CIDD.

W3. Lack of sensitivity analysis for core design choices.

The number of docking poses, evidence thresholds (τs\tau_s and τv\tau_v), and refinement rounds are central hyperparameters that directly affect structural evidence construction and optimization behavior. The authors explicitly acknowledge that controlled sensitivity analyses were not performed. Without evaluating these choices, it is difficult to determine whether the reported improvements are robust or depend on specific experimental configurations.

W4. Insufficient quantitative validation of the structural verification mechanism.

A central claim is that MEDD verifies whether hypothesized interaction changes are realized while preserving important contacts. However, the experimental evaluation primarily emphasizes conventional molecular quality metrics and qualitative structural examples, without sufficiently detailed quantitative evidence on hypothesis verification success and interaction preservation. Furthermore, verification relies on the same docking-based framework used to generate the initial evidence, raising questions about whether it provides independent evidence of improved binding interactions.

W5. Missing computational efficiency analysis and budget-matched comparisons.

MEDD requires multiple docking poses, repeated structural analysis, and several rounds of LLM reasoning and molecular editing, potentially introducing substantial computational overhead. However, the paper does not report wall-clock runtime, LLM token consumption, the number of docking evaluations, or overall computational cost. Comparisons with CIDD and simpler optimization baselines under matched computational budgets would be important for assessing whether the additional complexity is justified, especially given that MEDD underperforms CIDD on the main endpoint metrics.

Questions for authors
  1. Structural evidence and search-space constraints: How do the authors account for correlations among docking poses when estimating interaction consistency? Furthermore, can the framework explore alternative functional groups or binding geometries that preserve or improve favorable interactions without retaining the original molecular motifs? Could the current preservation strategy limit optimization performance?

  2. Hyperparameter sensitivity: How sensitive is MEDD to the number of docking poses, the stable/variable interaction thresholds, and the refinement budget? Can the authors provide sensitivity analyses for these key design choices?

  3. Structural verification: Can the authors provide quantitative results on hypothesis verification success rates and key-interaction preservation rates? Is there any independent structural assessment beyond the docking protocol used to construct the original evidence?

  4. Computational efficiency and comparison fairness: What are the average runtime, LLM token consumption, and number of docking evaluations per optimization task? How does MEDD compare with CIDD and simpler optimization baselines under matched computational budgets?

5 comments
AuthorsOpenReview ↗

Response to Q1: Structural Evidence and Search-Space Constraints

We thank the reviewer for these insightful questions. We clarify the motivation behind multi-pose evidence, the flexibility of our editing strategy, and its associated optimization trade-offs.

1. Multi-pose evidence and pose correlations.

The motivation for multi-pose analysis is to mitigate uncertainty in predicted binding orientations and potential pose variations, rather than relying on a single docking configuration. In our implementation, Vina generates up to 20 poses, and PLIP extracts their protein–ligand interactions. Contacts observed in every valid pose (support_count == valid_pose_count) are classified as stable, while partially supported contacts are treated as variable. These observations are aggregated into structural hypotheses and passed to DesignAgent as editing guidance.

We acknowledge that docking poses are correlated. Therefore, cross-pose consistency serves as heuristic structural evidence, not a calibrated stability probability or a collection of independent observations. Explicit pose clustering and correlation-aware weighting are not implemented in the current framework, and we do not claim otherwise.

2. Flexible exploration through hypothesis-guided editing.

MEDD does not assume that LLM agents possess stronger 3D structural perception than pocket-conditioned deep generative models. Instead, the initial generator provides a structure-informed molecule, while the agents incorporate chemical knowledge, interaction evidence, and iterative feedback to guide refinement.

Specifically, DesignAgent proposes chemical modifications based on stable/variable interactions and protected-region guidance; MolGenAgent generates candidate SMILES; and RDKit, Vina, and PLIP provide validity and structural evaluation. ReflectionAgent incorporates the resulting feedback, and ChooseAgent selects the final candidate.

Controlled editing is intended to make structural hypotheses traceable and testable, not to freeze the original molecular motifs. Preserving favorable interactions does not necessarily require retaining the same functional groups. The implementation allows scaffold and substituent modifications outside explicitly protected subgraphs, while contact loss generally informs subsequent decisions rather than serving as an unconditional rejection rule.

This flexibility is reflected in our supplementary results: 93.9% (921/981) of MEDD outputs underwent chemical structural changes, and only 13.46% retained identical Murcko scaffolds. Among 770 paired tasks where both methods performed chemical edits, MEDD achieved higher Contact Retention (+6.31 percentage points) and PLIF Retention (+6.42 percentage points). These results support structural flexibility alongside interaction preservation, although they do not directly establish the success of individual functional-group substitutions.

3. Preservation–optimization trade-off.

We agree that structural constraints can restrict exploration, particularly when substantially different binding geometries are preferable. Indeed, CIDD achieves greater unconstrained mean Vina improvement than MEDD (−0.8083 versus −0.6500 kcal/mol, 981 paired tasks).

However, MEDD targets evidence-guided refinement that balances molecular improvement with interaction preservation. Our supplementary evaluation shows:

MetricCIDDMEDD
Contact Retention ↑46.18%52.65%
PLIF Retention ↑42.36%48.98%
[email protected] ↑8.26%28.24%
[email protected] ↑3.67%13.97%
[email protected] ↑38.17%47.44%

Contact Retention, PLIF Retention, and SCIR are evaluated on 820 paired tasks with recoverable poses; SPIR and SPJIR use 981 paired tasks. These results demonstrate stronger structure-constrained improvement, rather than universal superiority in unconstrained optimization. We acknowledge that the specific effect of relaxing preservation constraints requires further controlled analysis.

Implementation and reproducibility.

To clarify these mechanisms, we provide the implementation details, relevant code locations, complete structured Agent prompt templates, and their dynamic input/output specifications in the supplementary materials. These cover multi-pose interaction aggregation, structural-hypothesis construction, protected-edit constraints, candidate generation, structural verification, and final selection.

Overall, MEDD uses multi-pose evidence to support chemically informed Agent decisions and hypothesis-guided refinement, while allowing molecular modifications and explicitly acknowledging the trade-off between preservation and unconstrained exploration.

AuthorsOpenReview ↗

Response to Q2: Hyperparameter Sensitivity

We thank the reviewer for this suggestion. Our configuration was chosen to maintain comparability with CIDD in LLM backbone and nominal candidate budget while balancing structural evidence, refinement quality, and cost. We further conducted 42 actual MEDD runs on six fixed tasks across seven configurations, varying one parameter at a time.

Hyperparameter definitions. MEDD contains three key hyperparameters:

1) Number of docking poses (K): The maximum number of Vina-generated binding poses used to construct multi-pose interaction evidence. A larger K provides more configurations for identifying recurring and pose-dependent interactions.

2) Refinement budget (M): The maximum number of candidate-generation attempts during hypothesis-guided refinement. It controls editing opportunities and associated LLM/docking costs.

3) Stable-contact threshold (τ): The minimum fraction of valid docking poses supporting an interaction for it to be classified as stable. The formal setting τ=1.0 requires an interaction to appear in every valid pose, while lower thresholds classify more contacts as stable, potentially strengthening preservation guidance.

The formal configuration is K=20, M=5, and τ=1.0. We evaluated K∈{5,10,20}, M∈{3,5,7}, and τ∈{0.5,0.75,1.0}, using the same initial molecules, protein pockets, and LLM model identifier across configurations.

Sensitivity analysis results

ConfigurationMean ΔVina ↓Success ↑Contact Retention ↑LLM CallsDocking CallsTime (s)
K=5−0.75450.0%0.80237.06.0380.6
K=10−1.61033.3%0.83536.46.0404.9
K=20 (default)−1.77050.0%0.89037.45.8360.1
M=3−1.26133.3%0.83523.04.0262.8
M=7−1.35916.7%0.82052.27.4437.1
τ=0.5−1.49966.7%0.75138.25.6351.7
τ=0.75−1.27350.0%0.88437.45.8358.8

Mean ΔVina, Contact Retention, and resource consumption are calculated over completed tasks, while Success is evaluated over all six predefined tasks per configuration.

1. Docking pose count (K). Increasing K from 5 to 20 improves mean ΔVina from −0.754 to −1.770 kcal/mol and Contact Retention from 0.802 to 0.890, while maintaining comparable LLM usage. This suggests that additional pose evidence may support refinement. Although Success is not monotonic, K=20 gives the strongest combined affinity improvement and retention on this small task subset, without more candidate-generation calls.

2. Refinement budget (M). Reducing M from 5 to 3 decreases average LLM calls from 37.4 to 23.0 and runtime from 360.1 to 262.8 seconds, but also reduces Success from 50.0% to 33.3% and increases the risk of failing the minimum valid-candidate requirement. Conversely, increasing M to 7 substantially increases LLM calls (52.2) and runtime (437.1 seconds), without improving conventional Success (16.7%). The additional cost of M=7 does not yield consistent gains across metrics. Thus, M=5 offers a practical balance while matching CIDD's nominal attempt budget.

3. Stable-contact threshold (τ). The default τ=1.0 emphasizes interactions consistently observed across all valid poses. Relaxing τ to 0.5 increases Success from 50.0% to 66.7%, but reduces Contact Retention from 0.890 to 0.751 and yields a less favorable mean ΔVina (−1.499 versus −1.770 kcal/mol). At τ=0.75, Contact Retention remains relatively high (0.884), without improving other primary outcomes. These results indicate a trade-off in selecting stable-contact evidence: lowering τ labels more contacts as stable, but does not necessarily improve their final retention. Given the six-task sample, the trend should be interpreted cautiously.

Overall, K=20, M=5, and τ=1.0 reflect a resource-aware, comparability-oriented design rather than post-hoc selection for one metric. These small-scale sensitivity results are consistent with the intended balance among evidence quality, refinement performance, and cost.

Detailed results and configurations are provided in the supplementary materials.

AuthorsOpenReview ↗

Response to Q3: Structural Verification

We thank the reviewer for this question. We conducted a retrospective analysis of formal MEDD records to quantify hypothesis realization, key-interaction retention, and post-hoc structural consistency, without modifying the original optimization procedure.

1. Quantitative hypothesis verification.

We introduce the Hypothesis Realization Rate (HRR) to quantify the proportion of proposed structural hypotheses supported by the final selected molecule's interaction evidence. Unlike endpoint Success, HRR assesses whether proposed interaction changes appear in the final predicted structures.

Specifically, MEDD considers two types of hypotheses:

1) Stabilize variable interactions: A hypothesis is verified when the targeted interaction exhibits higher occupancy across valid final docking poses than in the initial multi-pose evidence.

2) Preserve stable interactions: A hypothesis is verified when the targeted interaction remains present across all valid final poses.

Interactions are matched by interaction type and protein chain/residue identity. Importantly, improved Vina scores alone do not constitute successful hypothesis realization.

HRR is calculated as:

HRR=NVerifiedNVerified+NNotVerified\mathrm{HRR} = \frac{N_{\mathrm{Verified}}}{N_{\mathrm{Verified}} + N_{\mathrm{NotVerified}}}

From 19,948 hypotheses across 836 formal tasks, we obtained:

OutcomeCount
Verified6,695
Not Verified13,224
Not Assessable29
HRR33.61%
95% CI[32.43%, 34.87%]

The 95% CI uses 10,000 pocket-clustered bootstrap resamples. The 29 non-assessable hypotheses were excluded because final structural evidence was missing or invalid.

Additionally, 788 of 834 assessable tasks (94.48%) contained at least one verified hypothesis. This indicates that some proposed interaction objectives are realized in final predicted structures. HRR is not experimental confirmation or a per-iteration verification success rate.

2. Key-interaction preservation.

We further quantified Stable Interaction Retention (SIR), defined as the proportion of initially identified Stable/Protected interactions retained across all valid final poses.

The formal records contained 141 stable interactions across 18 tasks, of which 92 were preserved, yielding a strict SIR of 65.25% (92/141), with a pocket-clustered 95% CI of [42.64%, 100.00%]. Under a less restrictive criterion requiring preservation in at least one final pose, retention increased to 87.94% (124/141).

We report the strict criterion as primary to match MEDD's definition. Since qualifying contacts occur in only 18 tasks, SIR has limited coverage and a wide CI. CIDD SIR cannot be recovered comparably from its historical records.

To provide a broader comparison, we also performed a unified PLIP 3.0.1 evaluation on 820 paired MEDD–CIDD tasks:

MetricCIDDMEDDPaired difference (95% CI)
Contact Retention ↑46.18%52.65%+6.47 pp [4.30, 8.64]
PLIF Retention ↑42.36%48.98%+6.61 pp [4.60, 8.64]

These results favor MEDD in predicted interaction retention under a common protocol. Unlike SIR, these measures cover broader contact patterns in both methods.

3. Independent structural assessment.

We separately re-evaluated persisted MEDD and CIDD structures using a unified PLIP 3.0.1 protocol, without relying on Agent decisions or Vina scores to judge interaction preservation.

However, these structures retain coordinates from the original docking workflow. This is a separate post-hoc interaction assessment, not validation by an independent docking engine, molecular dynamics, or experimental structures.

Overall, the retrospective evidence gives HRR = 33.61% and strict SIR = 65.25%; the 820-task PLIP comparison also favors MEDD in Contact and PLIF Retention. These results support hypothesis-guided refinement at the level of predicted structures, without establishing physical binding mechanisms. Full definitions and statistics are included in the supplementary materials.

Reviewer_TjaWOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

MEDD is an LLM agent framework for structure-based molecular optimization. It aggregates protein–ligand interactions across multiple docking poses, marks stable contacts and variable regions, then lets the LLM propose interaction hypotheses and local edits. On CrossDocked2020, it improves docking scores while keeping QED, SA and MRR, with higher key interaction preservation than baselines.

Critical strengths

Multi-pose interaction evidence is a clear idea: it uses consistency across poses instead of trusting a single docking pose.

Critical weaknesses
  1. MEDD doesn't outperform CIDD on any metric in Table 1.
  2. The core parts of MEDD (multi-pose evidence and the explicit interaction hypothesis) barely move the Vina score in the ablation (–7.93 vs –7.81 and –7.93 vs –7.91).
  3. Important implementation details are missing: the paper never says which LLM backend was used.
  4. In Figure 4, the 2D structures of the refined molecule have obvious chemical errors: a triple bond between C and O, and a nitrogen with five bonds.
Questions for authors
  1. Many numbers in Table 1 are identical to those in the CIDD paper, but some key metrics for MolCRAFT and CIDD (Vina, MRR, etc.) differ quite a bit from the original papers. Can the authors explain the differences?
  2. Could the authors report the Vina score, QED, and related properties for the two molecules shown in Figure 4?
4 comments
AuthorsOpenReview ↗

Response to Q1: Baseline Discrepancies and Conventional Performance

We thank the reviewer for raising these concerns. We clarify the evaluation protocol and provide additional evidence for MEDD's contribution.

1. Evaluation protocol.

To minimize confounding effects from initial molecules, protein pockets, and LLM backends, our direct MEDD–CIDD comparison uses identical MolCRAFT-generated starting molecules, matched pockets, the same deepseek-chat backend, and unified evaluation criteria. This differs from the original CIDD study, which used GPT-4o and a different initial-molecule selection. Consequently, the locally evaluated MolCRAFT/CIDD results are not expected to exactly reproduce the original published aggregates. We acknowledge that this distinction was insufficiently explained.

2. Conventional performance.

We acknowledge that CIDD achieves stronger unconstrained molecular-property improvements. On 981 strictly matched tasks, CIDD achieves a larger mean Vina improvement (−0.8083 versus −0.6500 kcal/mol) and higher joint improvement rate (44.34% versus 40.06%).

However, MEDD pursues a complementary objective: improving molecular properties while retaining important structural features. As reported in Table 1, the MolCRAFT reference and MEDD results are −7.162 versus −7.896 for Vina, 0.508 versus 0.551 for QED, 0.684 versus 0.736 for normalized SA, 65.65% versus 86.95% for MRR, and 11.82% versus 29.97% for Success. These descriptive reference values are distinct from strictly paired improvements. Within MEDD's actual trajectories, median Vina changes from −7.335 to −8.097, QED from 0.470 to 0.525, and normalized SA from 0.703 to 0.766.

3. Additional structure-preservation evidence.

We conducted supplementary analyses on 981 matched tasks across 99 pockets, including 820 pairs with valid interaction evaluations.

MetricCIDDMEDDDifference
ECFP4 Similarity ↑0.2840.460+0.176
Contact Retention ↑46.18%52.65%+6.47 pp
PLIF Retention ↑42.36%48.98%+6.61 pp
[email protected] ↑8.26%28.24%+19.98 pp
[email protected] ↑3.67%13.97%+10.30 pp
[email protected] ↑38.17%47.44%+9.27 pp

ECFP4 Similarity measures similarity to the starting molecule; Contact/PLIF Retention measure preservation of predicted protein–ligand interactions. [email protected] requires improved Vina (ΔVina < 0) and ECFP4 similarity ≥ 0.5; SPJIR additionally requires ΔQED ≥ 0 and ΔSA ≥ 0. [email protected] requires improved Vina and Contact Retention ≥ 0.5.

Molecular metrics use 981 pairs, whereas interaction-based metrics use 820 structurally evaluable pairs. Importantly, MEDD's Contact Retention advantage persists among 770 pairs where both methods chemically modify the starting molecule (+6.31 pp).

4. Why interaction preservation matters.

Improvements in computational docking scores do not necessarily establish stronger experimental binding. Prior research has highlighted limitations in docking-pose reliability and the importance of physical plausibility checks (Buttenschoen et al., Chemical Science, 2024).

Although structure-conditioned generators incorporate three-dimensional information, subsequent LLM-based editing does not automatically preserve the structural features of the initial molecule. Relying solely on a single docking pose or scalar score may overlook important interaction changes.

MEDD therefore aggregates evidence across multiple poses, distinguishes stable contacts from editable regions, and incorporates hypothesis-guided editing and post-editing verification. Our supplementary results support improved structural preservation, rather than experimentally verified binding superiority or a quantified realization rate for individual hypotheses.

In summary, MEDD does not aim to outperform CIDD on every conventional metric. Its contribution lies in structure-informed molecular refinement, with measurable advantages in interaction retention and structure-constrained optimization.

AuthorsOpenReview ↗

Response to Q2: Molecular Properties in Figure 4

We thank the reviewer for this suggestion. The molecular properties of the initial and optimized molecules in Figure 4 are summarized below.

MetricInitial MoleculeOptimized Molecule
Vina (kcal/mol) ↓−4.852−8.445
QED ↑0.5270.830
Normalized SA ↑0.4400.790
MRR ↑01
Success01

The optimized molecule achieves a more favorable Vina score, higher drug-likeness (QED), improved synthetic accessibility (normalized SA), and better molecular reasonability (MRR). It also satisfies the predefined joint success criteria, whereas the initial molecule does not.

This example illustrates MEDD's ability to improve multiple molecular properties through structure-guided refinement. Nevertheless, it represents an individual case rather than evidence of general superiority, and the Vina scores are computational predictions rather than experimentally measured binding affinities.

AuthorsOpenReview ↗

Response to the Three Weaknesses

We thank the reviewer for the constructive feedback. We address the ablation results, missing LLM configuration, and chemical structure error below.

1. Limited Vina improvements in the ablation study.

We acknowledge that the original ablation figure did not clearly illustrate the contributions of individual components, particularly given the modest differences in mean Vina. However, Section 4.4 and Appendix Table 4 additionally report MRR and Success Rate, which provide a more comprehensive evaluation.

Results on the aligned 200-task subset are summarized below:

ConfigurationVina ↓MRR ↑Success ↑
Full MEDD−7.930986.08%28.0%
w/o Multi-pose−7.812282.14%14.5%
w/o Hypothesis−7.913082.73%14.5%
w/o Edit Specification−7.902884.26%18.0%
w/o Verification−7.918082.73%14.0%

All configurations contain 200 valid outputs. Vina and MRR are averaged over valid outputs; Success is calculated over all 200 tasks using the joint criteria: Vina < −8.18 kcal/mol, QED > 0.25, normalized SA > 0.59, and MRR = 1.

Although mean Vina changes are modest, removing any component consistently reduces MRR and Success. Without multi-pose evidence, Success drops from 28.0% to 14.5%; without hypothesis generation or verification, it drops to 14.5% and 14.0%, respectively. Removing edit specifications reduces Success to 18.0%.

These results suggest that the components contribute more to satisfying multiple optimization criteria jointly than to maximizing a single property. Small changes in mean Vina can coexist with larger changes in threshold-based Success. We agree that the original figure did not communicate this distinction clearly; the Appendix provides the more informative comparison.

2. Missing LLM backend and configuration.

We acknowledge that omitting the LLM configuration was an oversight. Our intention was to emphasize MEDD's architectural contribution rather than advantages attributable to a particular LLM, but this does not justify omitting essential reproducibility details.

In the main experiments, both CIDD and MEDD used DeepSeek deepseek-chat through https://api.deepseek.com.

MEDD's six roles (InteractionAgent, DesignAgent, ExperienceAgent, MolGenAgent, ReflectionAgent, and ChooseAgent) shared one LLM client, differentiated by role-specific prompts. Temperature was 0.7 for all roles except MolGenAgent (1.0). The MEDD configuration was verified against 27,113 historical LLM calls. The exact server-side model checkpoint is not recoverable from the API records.

The anonymized implementation, complete prompt templates, and experimental configurations are included in the supplementary materials.

3. Chemically incorrect bonds in Figure 4.

We thank the reviewer for identifying the erroneous C–O triple bond and overvalent nitrogen, and apologize for this visualization error.

The problem arose during bond-order reconstruction in the docking-pose conversion pipeline: Vina PDBQT was converted to SDF using Open Babel 3.2.1 and subsequently loaded using RDKit.

Specifically, the affected region is the hydroxyl-substituted alkynyl fragment attached to the nitrogen-containing heterocycle. The original SMILES fragment N3C#CO specifies a C≡C triple bond, with single bonds connecting the fragment to the ring nitrogen and terminal hydroxyl oxygen.

Because PDBQT does not fully preserve bond-order information, Open Babel incorrectly inferred the molecular graph during conversion, producing an apparent C≡O bond and an overvalent nitrogen in the adjacent heterocyclic region.

Importantly, the original SMILES passed RDKit parsing and sanitization. The error originated from docking-pose conversion rather than molecular generation.

We corrected the procedure by preserving the valid SMILES-derived molecular graph and mapping only the docking coordinates onto it, avoiding incorrectly inferred SDF bond orders. The corrected Figure 4 is provided in the supplementary fig directory.

Overall, the complete ablation results clarify MEDD's contribution to joint molecular optimization, while the verified LLM settings and corrected molecular representation address the concerns regarding reproducibility and chemical validity.

Thank you for the response. I still have concerns about the novelty, the reported performance, and the contribution of individual ablation components. I’ll keep my original score.

Additionally, the revised Figure 4 still has several issues. In Interactions 1 and 2, the geometry around the alkyne group in the refined molecule is not linear. In Interaction 3, a carbon atom and the oxygen atom in the terminal tert-butanol group overlap. These issues should be fixed in a main-text figure.

More critically, the authors confirm that the refined molecule contains the C#CO fragment (a ynol, with a hydroxyl group directly attached to an alkyne carbon). Ynols are highly unstable species that can tautomerize to ketenes, which can subsequently react with water to form carboxylic acids. So the docking in Figure 4 isn’t meaningful since the refined molecule wouldn’t bind in that form. I would suggest using a different refined example.

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

The paper studies molecular refinement given a protein pocket and an initial ligand. Its framework, MEDD, aggregates interactions across docking poses into stable contacts and regions available for modification. An LLM proposes local interaction hypotheses and constrained edit specifications, chemistry tools realize candidates, and the same docking pipeline checks the targeted changes and retained contacts before feeding the results into subsequent rounds. Experiments refine MolCRAFT-generated molecules on CrossDocked, compare with CIDD and other methods, and examine component ablations and editing magnitude. MEDD improves docking scores and molecular quality over the initial molecules and generally makes more localized changes than CIDD, while CIDD remains stronger on the main endpoint metrics. The authors argue that multi-pose evidence and post-editing verification support more controlled and inspectable molecular refinement.

Critical strengths
  1. The method addresses a practical issue: an improved docking score does not establish that an intended interaction has appeared. Connecting cross-pose evidence, explicit hypotheses, and post-editing checks makes it possible to relate design decisions to subsequent structural outcomes. Separating tool-derived measurements from LLM interpretation and planning is also well motivated.

  2. The experimental setup and Appendix A specify fixed initial candidates, shared pockets, and a common docking protocol, including docking and conformer-generation settings. Selecting initial candidates before refinement helps reduce outcome-dependent selection of starting molecules and provides a clearer basis for before-and-after comparisons.

  3. The component ablations cover the main stages, and Appendix D explains task alignment and the success-rate denominator while acknowledging trade-offs across property metrics. Editing-magnitude analyses and structural examples in both the main text and appendix help explain system behavior. The authors also openly acknowledge CIDD's stronger performance on the main endpoint metrics.

Critical weaknesses
  1. The experimental setup and Appendix F do not identify the LLM or its version. Appendix C provides input/output fields, design priors, and feedback schemata, but not the complete prompts actually used or sufficient material to reconstruct role-specific messages and tool calls. These schemata therefore do not reproduce the agent's behavior. The candidate-realization module, stable-contact thresholds, and verification acceptance rules are also unspecified. These details directly determine candidate generation and selection, so the actual configuration, prompts, and execution rules should be supplied.

  2. The abstract claims higher key-interaction preservation and a higher proportion of structurally supported hypotheses, but neither the main text nor the appendix reports aggregate results, comparison targets, or precise calculation rules for these quantities. The structural analyses mainly show before-and-after pocket examples, while the ablation tables report endpoint properties and editing magnitude. This does not yet substantiate the central empirical claim. The authors should use existing trajectory records to report the corresponding aggregate results and clarify how structural support is determined from tool outputs and which steps depend on LLM interpretation.

  3. The comparison with CIDD does not sufficiently establish the trade-off between benefits and resource use. MEDD makes more localized edits, whereas CIDD achieves stronger main endpoint properties, and their backbones and local-edit restrictions are not explicitly aligned. The framework includes multiple reasoning stages and repeated docking. Although the appendix specifies refinement-round limits and docking settings, it does not report actual LLM calls, input/output tokens, candidate evaluations or docking calls, or task runtime. The resource cost of improved locality therefore remains unclear, as does the contribution of method design versus search budget. Per-task resource statistics supported by existing records and explicit comparison conditions would help resolve this concern.

Questions for authors
  1. Please identify the LLM and version actually used, state whether the roles share a model, and provide the complete prompts or the relevant code locations. What candidate-realization module, stable-contact thresholds, and verification acceptance rules were used?

2.Are MEDD and CIDD comparable in backbone, local-edit restrictions, and search budget? Can existing logs provide per-task LLM calls and token usage, candidate evaluations or docking calls, and runtime, including how failures and retries are accounted for? If some statistics were not recorded, please specify the budget settings that remain verifiable.

  1. Please explain the exclusion rules between the initial tasks and the shared valid-task set, and provide the terminal-outcome counts from existing records. The initial-molecule Vina mean also differs between the main table and the distribution analysis; does this arise from sample coverage, reevaluation, or a labeling error?
4 comments
AuthorsOpenReview ↗

Response to Q1: LLM Configuration, Agent Prompts, and Verification Rules

We thank the reviewer for this suggestion. We have prepared the anonymized MEDD source code, complete agent prompts, and experimental configurations as supplementary materials, covering the following implementation details.

1. LLM and Agent Configuration. The formal experiments used DeepSeek deepseek-chat (https://api.deepseek.com), confirmed by 27,113 recorded LLM calls. All roles shared the same LLM backend, using role-specific prompts rather than separately trained models. The recorded API model identifier was deepseek-chat, without a separately pinned model revision. The temperature was 0.7 for all roles except MolGenAgent (1.0).

MEDD includes six functional roles: InteractionAgent analyzes protein–ligand interactions; DesignAgent proposes molecular modifications; ExperienceAgent summarizes previous attempts; MolGenAgent generates molecular structures; ReflectionAgent compares modifications and structural feedback; and ChooseAgent selects the final candidate. Complete structured prompts, dynamic input/output specifications, and corresponding implementations are included in the supplementary materials.

2. Candidate Realization. MEDD follows multi-pose evidence construction, structural hypothesis generation, modification planning, molecular generation, structural verification, and selection.

The evidence and hypothesis modules (src/evidence/, src/hypothesis/) organize docking interactions into structural guidance. DesignAgent produces modification instructions, which MolGenAgent directly converts into SMILES using the shared LLM (src/generation/molgen_adapter.py), without an additional molecular generator.

Generated SMILES are validated using RDKit's Chem.MolFromSmiles(). Invalid outputs trigger bounded regeneration, while unsuccessful proposals are discarded. The formal protocol allows up to five candidate proposals and requires at least three valid, structurally evaluated candidates before final selection (src/generation/candidate_generation.py).

3. Stable-Contact Criteria. MEDD requests up to 20 Vina docking poses and extracts protein–ligand interactions using PLIP. A contact is classified as stable only when

supportcount(c)=validposecount\mathrm{support_count}(c)=\mathrm{valid_pose_count}

corresponding to 100% occupancy across valid poses.Contacts with partial support are classified as variable (src/evidence/interaction_representation.py).

Stable contacts are incorporated into protected-edit constraints, including mapped atoms, bonds, and charges, while variable contacts identify potential editing regions (src/editing/). These constraints guide molecular modification rather than requiring the final molecule to preserve every original interaction.

4. Verification and Candidate Acceptance. Each candidate undergoes the following checks:

1) Chemical and structural validity: Invalid SMILES, failed docking, missing valid Vina/interaction results, explicitly incompatible protected graphs, or infeasible geometry lead to candidate rejection.

2) Interaction verification: Updated docking/PLIP evidence is compared with the proposed structural hypothesis. Stable-contact preservation is categorized as PASS (preserved or improved occupancy), SOFT_DEGRADATION (reduced occupancy), or SEVERE_LOSS (complete loss). These outcomes inform subsequent reasoning and selection rather than serving as an unconditional hard acceptance threshold.

3) Final selection: After obtaining at least three valid evaluated candidates, ChooseAgent selects the final molecule using candidate SMILES, docking scores, interaction evidence, and structural feedback. Its output must match an existing candidate. Invalid selections trigger up to six attempts, followed by a random candidate fallback if no valid selection is returned.

Importantly, the formal MEDD workflow uses evidence-guided candidate selection, without imposing additional hard Vina, QED, or SA acceptance thresholds. The thresholds used to calculate the paper's final Success Rate are applied during evaluation, not candidate generation.

The supplementary package MEDD_Anonymous_Code.zip contains the relevant source modules, complete prompts, formal configurations, and verification logic, providing the implementation details needed to reproduce these decisions.

AuthorsOpenReview ↗

Response to Q2: Comparability, Search Budgets, and Computational Costs

We thank the reviewer for emphasizing experimental comparability and computational transparency. We audited the original implementation and logs and conducted an additional 20-pocket runtime benchmark.

1. Backbone and Editing Constraints. Both MEDD and CIDD refine MolCRAFT-generated molecules. Across 981 strictly paired tasks, we verified identical initial SMILES, protein pockets, and task IDs, ensuring matched starting conditions.

However, their editing strategies differ. MEDD uses multi-pose structural evidence to guide local modifications, whereas CIDD follows its original refinement strategy. MEDD produces smaller mean absolute heavy-atom changes (2.103 vs. 4.019) and higher initial-to-final ECFP4 similarity (0.4603 vs. 0.2842). Thus, our comparison uses matched inputs and unified evaluation, rather than identical editing restrictions.

2. Verifiable Search Budgets. The formal MEDD configuration specifies up to 20 docking poses, five candidate-generation attempts, and at least three valid evaluated candidates for selection. Vina uses exhaustiveness = 8, one CPU thread, and seed = 20250916. These parameters are documented in the implementation and supplementary configurations.

Identical starting conditions do not imply identical search expenditures. We therefore do not claim strictly equal LLM, candidate-evaluation, or docking budgets.

3. Measured Computational Costs. We benchmarked both methods on 20 matched pocket–molecule tasks using deepseek-chat, retaining all 40 method–task records. For the 18 tasks with recorded wall-clock measurements for both methods, the per-task averages are:

MetricCIDDMEDD
Wall-clock time (s)432.36283.09
LLM calls35.6135.56
Input tokens116,15692,665
Output tokens42,56114,994
Docking calls5.675.44

The paired wall-clock difference (MEDD − CIDD) is −149.28 s, with a 95% CI of [−197.24, −95.93] s.

These results indicate lower observed MEDD latency in this benchmark. However, token statistics were recovered from different logging sources, and docking-call counts do not reflect individual docking durations. We therefore do not claim universally superior computational efficiency.

4. Failures, Retries, and Missing Records. Across all 20 tasks, CIDD produced 17 valid outputs (two timeouts, one scientific failure), while MEDD produced 18 (one timeout, one scientific failure). All failures remain in the original denominators, and available resource records are retained.

The formal MEDD protocol disabled API-level retries; bounded regeneration of invalid SMILES is handled separately within MolGenAgent. The supplementary benchmark records task-level LLM calls, token usage, docking attempts, runtime, and execution status, including failed attempts where available.

One CIDD timeout lacks an exact runtime; its lower bound is retained without imputation. Complete docking durations, fully aligned candidate-level evaluation histories, and reliable historical API costs are unavailable. Consequently, strict total-budget equivalence cannot be established.

5. Interpretation. CIDD achieves stronger unconstrained endpoint improvements, whereas MEDD provides greater structural preservation and higher structure-constrained improvement rates. The additional benchmark suggests that these benefits do not necessarily require greater wall-clock time, although its limited size and incomplete cost records restrict generalization.

We appreciate the reviewer's suggestion, which has helped us clarify the distinction between matched experimental inputs, method-specific editing constraints, and computational budget fairness.

AuthorsOpenReview ↗

Response to Q3: Task Exclusion Criteria, Terminal Outcomes, and Vina Score Discrepancy

We thank the reviewer for pointing out these ambiguities. We clarify the task-selection procedure, report the terminal outcomes from existing records, and distinguish the Vina statistics used in the main experiments and subsequent analyses.

1. Task Selection and Exclusion Rules.

We constructed a fixed evaluation set of 1,000 MolCRAFT-generated initial molecules across 100 protein pockets, with 10 molecules per pocket. MEDD and CIDD were evaluated using the same initial molecules and corresponding protein pockets.

A task was included in the strictly paired comparison only when both methods produced a valid final molecule with the required evaluation results. Tasks were not excluded based on whether their Vina, QED, or SA scores improved. Failed tasks remain in the original denominator when reporting completion rates.

The existing records provide the following terminal outcomes:

OutcomeCIDDMEDD
Initial tasks1,0001,000
Valid completed tasks992 (99.2%)981 (98.1%)
Tasks without valid final results819
Shared valid tasks981981

All 19 unsuccessful MEDD tasks were recorded as SCIENTIFIC_FAILURE. The eight CIDD tasks did not yield valid final results; their specific terminal categories cannot be further established from the available records.

The resulting 981 strictly matched tasks span 99 protein pockets. Their initial SMILES, pocket identities, and task IDs were verified to match exactly. Of the 19 tasks excluded from the paired analysis, 11 have valid CIDD results but no valid MEDD result, while eight lack valid results from both methods.

Thus, the paired analysis uses the intersection of valid completed tasks, rather than independently selecting favorable outputs for either method.

2. Clarification of the Vina Statistics.

We acknowledge that the manuscript does not sufficiently distinguish the initial-molecule statistics reported in different analyses.

Table 1 reports a MolCRAFT reference mean Vina score of −7.162 kcal/mol and a final MEDD mean of −7.8956 kcal/mol.

The distribution analysis reports initial and final Vina medians of −7.335 and −8.097 kcal/mol, respectively, rather than means.

The strictly paired analysis (981 tasks) gives a mean Vina change of −0.6500 kcal/mol, a median paired change of −0.441 kcal/mol, and 757/981 (77.2%) molecules with improved Vina scores.

The ablation analysis separately reports −7.931 kcal/mol for full MEDD and −7.812 kcal/mol without multi-pose evidence. These are ablation-specific endpoint means, not initial-molecule statistics.

These quantities serve different purposes. In particular, the difference between the Table 1 MolCRAFT mean and the distribution-analysis median should not be interpreted as a discrepancy between two estimates of the same statistic. Similarly, the mean paired Vina improvement is calculated from within-task initial-to-final differences and should not be confused with a final-score mean.

However, the initial Vina mean for the 981-task matched set has not yet been independently recomputed from the per-task records. Therefore, the available aggregate statistics do not establish how much of any remaining difference between initial-score means is attributable to sample coverage versus docking reevaluation, or whether a baseline label requires correction. We do not attribute the discrepancy exclusively to any of these causes. The manuscript requires clearer labeling of the reference baseline, paired-task statistics, and ablation results.

3. Reporting Clarification.

We will explicitly distinguish the MolCRAFT reference population, the 981-task matched evaluation set, and the ablation experiments, and identify whether each reported Vina statistic is a mean, median, or paired change. We will also clarify the respective sample denominators and evaluation sources.

These clarifications identify the remaining reporting ambiguity while preserving the recorded terminal outcomes and the independently verified paired comparison.

Thank you for your clarifications. However, I remain somewhat conservative regarding the paper's novelty and overall contribution to the field. I will make my final recommendation after taking the other reviewers' comments and discussions into consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.