acceptodds
Under review as a conference paper at ICLR 2027

ForesightMol: Chemistry-Guided Retrieval-Augmented Generation for Structure-Based Drug Design

Abstract

Retrieval-augmented molecular generation uses existing molecules to guide ligand construction, but reference compatibility does not guarantee favorable downstream properties or chemically feasible growth. Ranking the resulting candidates solely by docking score may further limit the diversity of chemical alternatives retained under fixed experimental budgets. We propose ForesightMol, a chemistry-guided framework that integrates outcome-aware reference assessment, feasible molecular growth, and complementary candidate selection. A scientific agent leverages pretrained scientific knowledge to interpret retrieved evidence, prioritize references, and guide molecular growth, while paired short-horizon rollouts assess their downstream effects. Evidence from accepted references guides local growth decisions, while shared valence and connectivity constraints exclude infeasible extensions during both reference assessment and generation. A multi-objective set selection procedure then balances molecular quality, predicted binding compatibility, and structural diversity within a fixed delivery budget. Experiments on 100 CrossDocked2020 pockets show improvements of 29.4% in mean QED and 14.8% in scaffold diversity over the best baseline. Controlled analyses identify contributions from both generation and selection, supporting the framework's effectiveness in delivering diverse, high-quality molecular candidates.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Structure-based drug design, Molecular generation, Retrieval-augmented generation, LLM agents

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_ijjzOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes ForesightMol, a chemistry-guided retrieval-augmented framework for structure-based drug design that integrates reference intervention, chemically constrained molecular growth, and multi-objective candidate selection. Experiments on 100 CrossDocked2020 test pockets demonstrate improved molecular quality and scaffold diversity over Rag2Mol-G.

Critical strengths

1.Meaningful problem formulation.

The paper identifies an important limitation of retrieval-augmented molecular generation: reference compatibility does not necessarily imply favorable downstream growth. Evaluating references through their induced generation trajectories provides a more outcome-oriented alternative to conventional similarity-based retrieval.

2.Coordinated generation and selection.

ForesightMol integrates rollout-verified reference intervention, chemical feasibility constraints, and multi-objective candidate selection into a coherent framework. The ablation studies provide evidence that these components contribute differently to molecular quality and structural diversity.

3.Improved candidate quality and diversity.

Experiments on CrossDocked2020 demonstrate that ForesightMol improves molecular quality and scaffold diversity over Rag2Mol-G with docking-based selection. Crossed comparisons further support the contributions of both generation and selection.

Critical weaknesses
  1. Limited methodological novelty beyond the integration of existing techniques.

Although ForesightMol presents a coherent framework, its main components rely on established techniques, including LLM-based heuristic assessment, autoregressive molecular generation, short-horizon rollout evaluation, valence filtering, and multi-objective candidate selection. The theoretical analysis establishes the limitations of similarity-based utility prediction, but it does not demonstrate that the proposed LLM-guided intervention mechanism is uniquely necessary or superior to simpler alternatives. The authors should better isolate the methodological novelty of the framework and clarify whether the improvements arise from the proposed coordination strategy or from combining existing techniques.

  1. Lack of Clarity and Inconsistencies in the Methodology

Several methodological descriptions and notations are ambiguous or incomplete, hindering understanding and reproducibility. First, in Section 3.1 (page 4), the authors define that “P=(pi,hi)P = {(p_i , h_i)}, where pi∈Rp_i ∈ R and hih_i denote atom coordinates and features, including element and residue types”. It is unclear whether this is an atom-level or residue-level representation. The former implies atomic coordinates and features, whereas the latter implies residue features and representative coordinates, such as those of Cα atoms. The authors should clarify the representation. Second, in Equation (3), vv in HvtH^t_v and the function EE are undefined. The meaning of the backslash symbol “\” in Equation (5) is also unclear. Third, the definition of “rr” is inconsistent: page 4, line 193, introduces “Given a references rr, …”, whereas page 5, line 222, states “For each challenger rr, …”. The authors should clarify its meaning and use the notation consistently. Fourth, on page 5, line 229, the source, role, and significance of the “random input ξ” are not explained. Fifth, Equation (8) separates three inequality conditions with commas, leaving it unclear whether the feasible set is defined by their intersection or union. The authors should use ∩ or ∪, as appropriate, to clarify their logical relationship. Sixth, the first function PtrP_t^r in Equation (10) is undefined, and xtx_t in Equation (12) is insufficiently explained. Moreover, Equation (12) uses eje_j with j∈J(a)j \in J(a), whereas page 6, line 270, defines J(a)J(a) as the set of existing atoms connected to the new atom. If eje_j denotes atom jj itself, the notation should be clarified, as ej∈J(a)e_j \in J(a) would be more appropriate. Finally, page 6, line 285, mentions both “Vina Score” and “Vina Dock” evaluations. Appendix C (page 17, line 887) states that “Vina Score evaluates the generated pose, whereas Vina Dock evaluates the best energy…”. However, Vina Score is typically a numerical energy or affinity estimate in kcal/mol, rather than a pose itself. The authors should clarify the definitions, evaluation procedures, and differences between these metrics, particularly regarding their treatment of generated poses and binding energy.

  1. Limited generalization evaluation.

The experiments are conducted on only 100 CrossDocked2020 test pockets, leaving it unclear whether the observed improvements generalize to substantially different protein targets and retrieval distributions. Evaluation on more diverse targets and under different retrieval settings would strengthen the evidence for the framework's generalizability.

  1. Inconsistent Tense Usage.

On page 2, line 64, the authors write “We observed that reference value…”, using the past tense while the surrounding text consistently uses the present tense. The authors should check and ensure consistent tense usage throughout this passage.

Questions for authors
  1. Methodological Novelty.

Can the authors clarify the core methodological novelty of ForesightMol and provide evidence that the LLM-guided intervention mechanism outperforms simpler heuristic or learned alternatives?

  1. Methodological Clarity.

Can the authors clarify the ambiguous notations and definitions in Equations (3), (5), (8), (10), and (12), the protein pocket representation, the role of random input xix_i, and the distinction between Vina Score and Vina Dock?

  1. Generalization and Practical Utility.

Can the authors evaluate ForesightMol on more diverse protein targets and retrieval settings, and provide further evidence of its practical utility, particularly given the deterioration in redocking scores and the lack of experimental validation?

7 comments
AuthorsOpenReview ↗

Response Part I

W1&Q1. Methodological novelty of ForesightMol and evidence of superiority of LLM-guided intervention against simpler heuristic or learned alternatives.

Thank you for the helpful suggestions. We clarify the core methodological novelty of ForesightMol below and present additional controlled experiments to evaluate the benefits of LLM-guided intervention over simpler or heuristic alternatives.

Existing LLM-assisted molecular design frameworks have successfully supported task planning and tool orchestration [1,2], yet how to translate pretrained chemical knowledge into actionable guidance during molecular generation remains underexplored [3]. ForesightMol addresses this gap by connecting LLM chemical knowledge with the evolving generation state through context engineering and verifying proposed interventions through short-horizon rollouts. This design enables chemical expertise to inform molecular growth, while evidence-grounded explanations support subsequent candidate prioritization. Specifically, our technical novelty comprises two aspects.

  • Thoroughly structured context engineering activates chemical knowledge of LLM. By investigating chemical knowledge, we structure the current generation state, reference descriptors, and growth diagnostics into chemical contexts, enabling the LLM to apply pretrained knowledge to reference prioritization and translate verified assessments into additional biases at the prediction heads for molecular growth.

  • Set-level portfolio optimization with evidence-grounded explanations. We formulate fixed-budget portfolio selection as a joint optimization of molecular quality, predicted binding compatibility, structural diversity, and pocket coverage. The LLM integrates input-design and generation decisions with selection evidence to provide a traceable rationale for each included molecule, helping researchers identify promising candidates for downstream evaluation.

Evidence supporting LLM-guided intervention. To isolate the contribution of LLM guidance and compare it with simpler alternatives, we evaluate six controlled configurations under the same experimental settings. The similarity-based selector ranks references by structural similarity to the current molecule, whereas the heuristic-based selector applies scoring rules that reward chemical compatibility and favorable pocket contacts while penalizing steric clashes and size mismatch. We evaluate the full generated candidate pools before portfolio selection to assess the effects of the generation-stage mechanisms.

MethodMean QED ↑Best QED ↑SA ↓Lipinski ↑Mol. div. ↑Scaf. div. ↑Vina Score ↓Vina Dock ↓
(A) Similarity-based reference selector0.5160.7903.6700.8230.8170.738−6.1873−8.3044
(B) Rollout-only0.5180.7633.6390.8260.8190.753−6.1998−8.3277
(C) Agent-only0.5430.7813.4180.8290.8060.750−6.2593−8.3801
(D) Heuristic-based reference selector0.5350.7583.4090.7980.7430.684−6.1593−8.2220
(E) No-guidance0.5630.7983.4650.8310.8240.770−6.2797−8.3276
(F) ForesightMol0.5860.8033.2440.8340.8250.775−6.3515−8.5306

As shown in the table above, the agent-only variant (C) improves mean QED and both Vina metrics over similarity-based (A) and heuristic-based (D) selection, even without rollout verification or prediction-head guidance. Rollout alone (B) yields limited gains, whereas combining it with LLM assessment (E) improves mean QED and diversity over (C). The full method (F) performs best across all reported metrics.

[1] Liddia: Language-based intelligent drug discovery agent. EMNLP. 2025.

[2] Rethinking chemical research in the age of large language models. Nat Comput Sci. 2025.

[3] From generalist to specialist: A survey of large language models for chemistry. COLING. 2025.

AuthorsOpenReview ↗

Response Part II

W2&Q2. Clarification of ambiguous notation and definitions in Section 3, including the protein pocket representation, reference notation, random input ξ\xi, acceptance conditions, and Vina Score and Vina Dock evaluation procedures.

Thank you for your careful reading and valuable comments. We address each point below.

  1. Unclear representation granularity in Section 3.1.

    P=(pi,hi)i=1NP\mathcal{P}=\\{(\mathbf{p}_i,\mathbf{h}_i)\\}_{i=1}^{N_P} represents NPN_P pocket heavy atoms. For each heavy atom ii, pi∈R3\mathbf{p}_i\in\mathbb{R}^3 gives its coordinates, and hi\mathbf{h}_i denotes its scalar feature vector, encoding its element type, the amino-acid type of its parent residue, and its backbone/side-chain status.

  2. Undefined notation in Equation (3).

    In Equation (3), EψE_\psi denotes the combined encoding and reference-interaction mapping, comprising the encoders and cross-context interaction module, with learnable parameters ψ\psi. It maps the current state sts_t and reference rr to the hidden representation Ht(r)H_t(r). The superscripts s\mathrm{s} and v\mathrm{v} indicate its scalar and vector feature components, respectively.

  3. Unclear backslash notation in Equation (5).

    In Equation (5), ∖\setminus denotes set difference. Thus, R(P)∖rt0\mathcal{R}(\mathcal{P})\setminus\\{r_t^0\\} is the retrieved reference set excluding the current default reference rt0r_t^0.

  4. Inconsistent use of reference notation rr.

    Throughout, rr denotes a retrieved reference molecule. Equation (3) considers any r∈R(P)r\in\mathcal{R}(\mathcal{P}), while “challenger” denotes its intervention role when selected by the agent from RtA⊆R(P)∖rt0\mathcal{R}_t^{\mathcal{A}}\subseteq\mathcal{R}(\mathcal{P})\setminus\\{r_t^0\\}.

  5. Unclear source and role of random input ξ\xi.

    Here, ξ\xi denotes the randomness in sampling atom and bond types from the generator’s predicted distributions. This randomness supports diverse molecular exploration but can affect consistency across rollouts. We therefore use the same random seed for paired default-reference and challenger rollouts to control sampling variability when evaluating the effect of reference choice.

  6. Unclear logical relationship among the conditions in Equation (8).

    The three conditions in Equation (8) must hold simultaneously: sufficient rollout utility gain, risk increase within tolerance, and agent confidence meeting the threshold. We have connected the corresponding sets with ∩\cap to make this conjunction explicit.

  7. Undefined and insufficiently explained notation in Equations (10) and (12).

    In Equation (10), pˉtr(a)\bar{p}_t^r(a) denotes the generator’s baseline action distribution over Bt(r)\mathcal{B}_t(r), conditioned on the current state sts_t and reference rr, before Agent-based reweighting.

    In Equation (12), xtx_t denotes the current partial ligand, and vj(xt)v_j(x_t) is the sum of the bond orders already incident to atom jj. The proposed bond adds bjb_j to this sum. We intend J(a)J(a) to contain the indices of existing atoms connected to the new atom, while eje_j denotes the element type of atom jj. Thus, j∈J(a)j\in J(a) is the appropriate indexing relation, and ν(ej)\nu(e_j) gives the corresponding valence upper bound. We have explicitly defined these quantities and replace “existing atoms” with “indices of existing atoms” in the definition of J(a)J(a).

  8. Unclear definitions and evaluation procedures for Vina Score and Vina Dock.

    Both metrics report a numerical binding-affinity estimate in kcal/mol using AutoDock Vina [1,2]. Vina Score evaluates the generated pose using --score_only, without changing its position or conformation. Vina Dock performs redocking, allowing changes in ligand position, orientation, and rotatable bonds, then reports the lowest score among poses returned under the specified search settings. Both use the same scoring function, and lower values indicate more favorable predicted binding.

We have incorporated these clarifications into the revised manuscript to improve the consistency of the notation and the clarity of the methodological definitions.

[1] AutoDock Vina: Improving the speed and accuracy of docking with a new scoring function, efficient optimization, and multithreading. J Comput Chem. 2010.

[2] AutoDock Vina 1.2.0: New docking methods, expanded force field, and Python bindings. J Chem Inf Model. 2021.

AuthorsOpenReview ↗

Response Part III

W3&Q3. Further evaluation of ForesightMol on protein targets that differ substantially from those in CrossDocked2020 is needed to establish how well the framework generalizes.

Thank you for the helpful suggestion. Although CrossDocked2020 is widely used to benchmark structure-based drug design methods, generalization to targets outside the training distribution is important for real-world drug discovery. We used MMseqs2 [1] to compare protein targets from the Binding MOAD dataset [2], following the test split provided by DiffSBDD [3], against those from CrossDocked2020 to identify targets with low sequence similarity. We excluded entries with overlapping receptor PDB IDs or sequence matches with at least 30% identity over at least 80% of the shorter chain. After deduplicating entries by the sequences of their primary ligand-contacting chains, we retained thirty targets. We then compared the methods in a zero-shot setting, using their CrossDocked2020-trained checkpoints on the selected Binding MOAD targets without fine-tuning.

MethodMean QED ↑Best QED ↑SA ↓Lipinski ↑Mol. div. ↑Scaf. div. ↑Vina Score ↓Vina Dock ↓
Pocket2Mol0.6500.8803.0820.9890.8270.521−3.1379−6.7003
MolCRAFT0.6470.8762.5060.9220.8790.664−3.9248−6.3149
TargetDiff0.5950.8703.4110.9230.9040.744−2.2568−6.1686
Rag2Mol-G0.6630.8593.2030.9460.7970.666−4.1774−7.3941
ForesightMol0.6900.8912.3070.9470.8050.709−4.2553−7.6031

As shown in the table above, ForesightMol achieves the best results on five of eight metrics among the evaluated methods, including mean and best QED, SA, Vina Score, and Vina Dock. Relative to Rag2Mol-G, ForesightMol improves molecular quality and scaffold diversity while also achieving a more favorable redocking score. Together, these results provide empirical evidence that the benefits of ForesightMol extend to targets with low sequence similarity to those in CrossDocked2020 under zero-shot evaluation.

[1] MMseqs2 enables sensitive protein sequence searching for the analysis of massive data sets. Nat Biotechnol. 2017.

[2] Binding MOAD (mother of all databases). Proteins. 2005.

[3] Structure-based drug design with equivariant diffusion models. Nat Comput Sci. 2024.

[4] Pocket2Mol: Efficient molecular sampling based on 3D protein pockets. ICML. 2022.

[5] MolCRAFT: Structure-based drug design in continuous parameter space. ICML. 2024.

[6] 3D equivariant diffusion for target-aware molecule generation and affinity prediction. ICLR. 2023.

[7] Rag2Mol: Structure-based drug design based on retrieval augmented generation. Brief Bioinform. 2025.

W4. Inconsistent tense usage on page 2.

Thank you for pointing this out. We have changed “We observed” to “We observe” to match the surrounding text. We have also proofread the entire manuscript to improve clarity and consistency.


We sincerely thank you for your thoughtful comments and careful review, which have helped strengthen our manuscript. If our responses and additional evidence have satisfactorily addressed your concerns, we respectfully ask that you consider increasing your score. We would be pleased to provide any further clarification and greatly appreciate the time and effort you have devoted to evaluating our work.

Reviewer_3vDcOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes ForesightMol, a chemistry-guided retrieval-augmented framework for structure-based molecular generation. The method combines three components: (1) an LLM-based scientific agent that evaluates and prioritizes retrieved reference molecules, (2) short-horizon matched rollouts to verify whether an alternative reference improves downstream molecular growth, and (3) valence/connectivity-aware generation followed by multi-objective candidate-set selection. The reported results on 100 CrossDocked2020 pockets show substantial improvements in QED, synthetic accessibility, and scaffold diversity over Rag2Mol-G with conventional docking-based selection.

Critical strengths
  1. The central idea is highly interesting and potentially impactful. Using pretrained scientific knowledge to guide molecular growth through reference assessment is an original and promising direction.

  2. The separation between LLM proposal, rollout-based verification, and programmatic deployment is well motivated. The LLM is not given direct authority to deploy an intervention, which is a sensible design choice.

  3. The ablation studies and structural analyses provide evidence that both reference intervention and candidate selection contribute to the reported gains.

Critical weaknesses

The main concern is the fairness and interpretability of the experimental comparison.

The combined Agent–rollout module gives ForesightMol two additional advantages that are not available to the end-to-end baselines.

First, the Agent is explicitly asked to assess drug-likeness, synthetic tractability, trajectory risk, and related chemical properties, creating a potential information-leakage or objective-alignment issue because the final evaluation also reports QED, SA, and Lipinski compliance. Second, rollout introduces an additional inference-time search and evaluation loop: alternative references are used to generate short-horizon continuations, these continuations are compared, and the result is used to accept, reject, or reweight subsequent generation. The competing methods are evaluated as ordinary generators without an equivalent counterfactual generation-and-selection procedure. Therefore, the comparison is asymmetric in both information access and inference-time computation. The reported gains should not be attributed solely to a better molecular generator unless the authors demonstrate that they persist under matched search and evaluation budgets.

Moreover, the improvements are not uniformly dominant across metrics. Although ForesightMol improves QED and scaffold diversity under the selected protocol, it yields less favorable mean Vina Docking scores than the Rag2Mol-G plus Docking baseline, and the paper does not report confidence intervals or statistical tests establishing that the observed differences are robust.

Questions for authors
  1. How is the additional rollout computation accounted for in the comparison? Are the baselines given the same number of candidate continuations, branch evaluations, selection operations, docking calls, and total inference-time compute? What are the results when the same rollout-based search and selection procedure is applied to all baseline generators?

  2. Can the contributions of Agent and rollout be separated? The current “reference intervention” appears to combine Agent proposals, rollout verification, reference switching, and post-acceptance guidance. Did the authors evaluate Agent-only, rollout-only, no-guidance, and non-LLM/rule-based evaluator variants? Without these controls, how can the individual importance of Agent reasoning and rollout-based search be assessed?

  3. How robust are the reported improvements under a fair evaluation protocol? Do the gains persist when all methods use the same candidate-set selector and when evaluation is performed on metrics not exposed to the Agent or selector? Please also report confidence intervals or statistical tests, particularly given that the improvement in QED and scaffold diversity is accompanied by less favorable Vina Docking scores.

4 comments
AuthorsOpenReview ↗

Response Part I

Q1. How is the additional rollout computation accounted for in the comparison? Are the baselines given the same number of candidate continuations, branch evaluations, selection operations, docking calls, and total inference-time compute? What are the results when the same rollout-based search and selection procedure is applied to all baseline generators?

Thank you for the valuable suggestions. Rollout introduces a modest additional computational cost, which we examine from two perspectives.

  • Computational complexity. For QQ reference-assessment queries, LL challengers plus the default reference, rollout horizon HH, and beam width BrB_r, rollout requires at most Q(L+1)[1+(H−1)Br]Q(L+1)[1+(H-1)B_r] additional growth-step evaluations, scaling as O(Q(L+1)HBr)O(Q(L+1)HB_r). With Q≤3Q\leq3, L=2L=2, H=2H=2, and Br=8B_r=8, this permits up to 81 additional evaluations per pocket, approximately 10% of the observed average generation workload.

  • Empirical overhead. We use Rag2Mol-G as the matched baseline because it shares the same generator backbone. Our rollout mechanism verifies reference choices during retrieval-augmented autoregressive growth, so applying it to diffusion-based generators or optimization methods would not provide a directly comparable test of its contribution. Under the shared experimental protocol described in Section 4.1, rollout added approximately 6% more growth-step evaluations. The baseline and full method required 576.6 and 615.9 seconds per 100 completed molecules, respectively, corresponding to a runtime increase of 6.8%. To assess whether the gains arise from additional search alone, we also evaluated a rollout-only variant. The results are reported in the table below.

MethodMean QED ↑Best QED ↑SA ↓Lipinski ↑Mol. div. ↑Scaf. div. ↑Vina Score ↓Vina Dock ↓
(A) Baseline0.5160.7903.6700.8230.8170.738−6.1873−8.3044
(B) Rollout-only0.5180.7633.6390.8260.8190.753−6.1998−8.3277
(C) ForesightMol0.5860.8033.2440.8340.8250.775−6.3515−8.5306

The results indicate that adding rollout alone to the baseline yields limited improvements (A vs. B). This aligns with its intended role as a second-stage verification mechanism that evaluates LLM proposals before deployment, helping limit the effects of unreliable suggestions and potential hallucinations on molecular growth.

Q2. Can the contributions of Agent and rollout be separated?

Thank you for your insightful suggestion. We have conducted more fine-grained ablations, including Agent-only, rollout-only, no-guidance, and heuristic-based reference selection, to assess the contribution of each mechanism. The heuristic-based selector applies scoring rules that reward chemical compatibility and favorable pocket contacts while penalizing steric clashes and size mismatch. And agent-only disables rollout verification and growth guidance, while no-guidance retains Agent proposals and rollout verification but disables prediction-head biases.

MethodMean QED ↑Best QED ↑SA ↓Lipinski ↑Mol. div. ↑Scaf. div. ↑Vina Score ↓Vina Dock ↓
(A) Baseline0.5160.7903.6700.8230.8170.738−6.1873−8.3044
(B) Rollout-only0.5180.7633.6390.8260.8190.753−6.1998−8.3277
(C) Agent-only0.5430.7813.4180.8290.8060.750−6.2593−8.3801
(D) Heuristic-based reference selector0.5350.7583.4090.7980.7430.684−6.1593−8.2220
(E) No-guidance0.5630.7983.4650.8310.8240.770−6.2797−8.3276
(F) ForesightMol0.5860.8033.2440.8340.8250.775−6.3515−8.5306

As shown in the table above, Agent-only improves mean QED, SA, and both Vina metrics over the baseline (A vs. C). Combining Agent proposals with rollout verification further improves mean and best QED and both diversity measures over Agent-only (C vs. E). This combination also outperforms heuristic-based selection in QED, diversity, and both Vina metrics (D vs. E). Finally, adding post-acceptance growth guidance improves every reported metric (E vs. F), with the full framework achieving the best results across all metrics. These comparisons support the distinct and complementary contributions of Agent proposals, rollout verification, and growth guidance.

AuthorsOpenReview ↗

Response Part II

W & Q3. How robust are the reported improvements under a fair evaluation protocol? Do the gains persist when all methods use the same candidate-set selector and when evaluation is performed on metrics not exposed to the Agent or selector? Please also report confidence intervals or statistical tests.

Thank you for your careful reading and valuable feedback. We address your concerns regarding (1) potential information leakage, (2) evaluation fairness, (3) molecular quality and scaffold diversity at matched docking performance, and (4) statistical uncertainty through confidence intervals.

First, we reduce potential information leakage by providing anonymized molecular descriptors and keeping the LLM’s delivery-stage explanations separate from candidate selection. As described in Appendix D.5, during generation the LLM assesses reference candidates without being prompted to optimize the reported evaluation metrics. At the delivery stage, it provides explanations only after the candidate set has been finalized.

Second, we assess evaluation fairness by comparing both full candidate pools and delivery sets obtained under shared selection procedures. Table 1 in the manuscript reports generation-stage performance before final selection. We further added MolCRAFT [1] and DecompOpt [2], evaluating their full pools and applying both Dock ranking and S–D selection uniformly across generators.

Full candidate pools

GeneratorMean QED ↑Best QED ↑SA ↓Lipinski ↑Mol. div. ↑Scaf. div. ↑Vina Score ↓Vina Dock ↓
MolCRAFT0.5150.8363.8130.7410.8280.777−5.7860−7.5859
DecompOpt0.4800.6573.9230.7260.6240.764−5.8389−8.1485
Rag2Mol-G0.5180.7903.5700.8230.8170.738−6.1873−8.3044
ForesightMol0.5770.7953.2460.8300.8280.776−6.3427−8.5342

Delivery sets

GeneratorSelectorMean QED ↑Best QED ↑SA ↓Lipinski ↑Mol. div. ↑Scaf. div. ↑Vina Score ↓Vina Dock ↓
MolCRAFTDock0.5280.7404.2820.6570.8610.981−7.1527−9.5364
MolCRAFTS–D0.6070.7673.9210.9110.8650.993−7.1169−8.8786
DecompOptDock0.4020.5544.0110.7230.7610.972−6.6112−8.9823
DecompOptS–D0.4440.5703.9290.9020.7540.985−6.5938−7.9257
Rag2Mol-GDock0.4940.6413.7550.7490.7230.860−6.8353−9.8336
Rag2Mol-GS–D0.5330.6673.6620.9370.7580.922−6.6249−8.8802
ForesightMolDock0.5130.6873.6330.7500.7630.874−6.9381−9.9103
ForesightMolS–D0.6390.7823.5600.9260.8120.987−6.8634−9.0062

As shown in the tables above, ForesightMol retains gains with these additional baselines. Before selection, it achieves the highest mean QED and Lipinski pass rate and the lowest SA score. Under the shared S–D selector, it also achieves the highest mean and best QED and the lowest SA score.

Third, we assess whether molecular quality and diversity gains persist at matched or better redocking performance. We constrained the S–D selector to match or improve upon the mean Vina Dock score of Rag2Mol-G with Dock ranking on each evaluated pocket.

GeneratorSelectorMean QED ↑Best QED ↑SA ↓Lipinski ↑Mol. div. ↑Scaf. div. ↑Vina Score ↓Vina Dock ↓
Rag2Mol-GDock0.4940.6413.7550.7490.7230.860−6.8353−9.8336
ForesightMolDock0.5130.6873.6330.7500.7630.874−6.9381−9.9103
ForesightMolS–D0.6390.7823.5600.9260.8120.987−6.8634−9.0062
ForesightMolDock-constrained S–D0.5690.7373.7390.8470.7650.971−7.3965−9.8768

In this paired comparison, ForesightMol still achieves higher mean and best QED, lower SA, and greater molecular and scaffold diversity, while maintaining or improving both Vina metrics on average.

Finally, we assess statistical uncertainty using 95% percentile bootstrap confidence intervals for paired differences across pockets in Table 2. With both generators using the same S–D selector, the intervals for mean QED (+0.106 [0.073, 0.147]), SA (−0.102 [−0.191, −0.028]), scaffold diversity (+0.065 [0.036, 0.103]), and Vina Score (−0.239 [−0.438, −0.067] kcal/mol) all exclude zero. The mean Vina Dock difference is −0.126 kcal/mol, with a 95% CI of [−0.348, +0.062] kcal/mol.

[1] MolCRAFT: Structure-based drug design in continuous parameter space. ICML. 2024.

[2] DecompOpt: Controllable and decomposed diffusion models for structure-based molecular optimization. ICLR. 2024.

AuthorsOpenReview ↗

Response Part III

We sincerely thank you for the careful and constructive evaluation of our work. We have provided detailed, point-by-point responses to all the weaknesses and questions raised. We hope that our clarifications adequately address your concerns, and we would be very happy to engage in further discussion should any additional questions arise.

Dear Reviewer 3vDc,

Thank you for your time and helpful comments on our paper. We have uploaded the revised manuscript with all changes marked. Our response provides further analysis of rollout overhead, additional ablations, and evaluation fairness.

With the author–reviewer discussion period ending in about two days, we would appreciate your feedback on whether our rebuttal addresses your concerns. We welcome further discussion and would be happy to clarify any remaining questions.

Thank you again for your time and help in improving our paper.

Best regards,

The authors

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

This paper presents ForesightMol for protein pocket-conditioned 3D molecular generation and candidate selection. The method uses an LLM to propose reference molecules and growth guidance, validates reference interventions through paired short-horizon rollouts, and enforces valency and connectivity constraints during generation. It then selects a fixed-size candidate set based on molecular properties, predicted binding performance, and structural diversity. The paper evaluates both generated molecules and final selected sets on 100 test pockets from CrossDocked2020.

Critical strengths

The paper addresses a practically relevant problem. It considers both generation quality and candidate selection under a limited experimental budget, with a clear focus on delivering a set of molecules with favorable properties and complementary structures.

Reference selection includes an explicit validation step. Alternative references proposed by the LLM are accepted only after paired short-horizon rollouts from the same intermediate state. This reduces reliance on the LLM’s judgment alone and ties reference selection to observed generation outcomes.

Several computational metrics show improvements. Beyond QED, SA, and diversity in the final selected sets, the method also improves properties of the unfiltered candidate pools and some measures of geometric validity. These results support the framework’s effectiveness under the tested conditions.

Critical weaknesses

The LLM’s independent contribution remains unclear. The current comparisons bundle LLM proposals, rollout validation, and subsequent growth guidance, making it difficult to separate chemical reasoning from additional search and programmed constraints. Useful controls would include non-LLM reference proposals with the same rollout validation, and separate ablations of reference proposals and growth guidance.

Generation decisions lack traceable examples. The LLM examples mainly explain candidate value using final QED, diversity, and geometry metrics. They do not show how chemical judgments about a partially generated molecule change reference selection, growth actions, and final outcomes, leaving limited process-level evidence for the claimed use of pretrained chemical knowledge.

Baseline coverage is limited. The experiments omit relevant generators such as MolCRAFT and property optimization methods such as DecompOpt. Since candidate selection accounts for part of the gains, different generators should be compared using the same selection strategy, evaluation protocol, and comparable budgets, with results reported for both full candidate pools and final selected sets.

The final selected set has worse redocking scores. Compared with Rag2Mol-G plus Dock ranking, the full method worsens mean Vina Dock from −9.833 to −9.006 kcal/mol, a difference of 0.827 kcal/mol, while only marginally improving Vina Score. The gains in QED and diversity therefore come at the expense of redocking performance. As QED and SA are explicit selection objectives, their improvement alone does not establish the LLM’s chemical design ability. Further comparisons should assess properties and diversity at matched docking performance or under an explicit binding-score constraint.

Questions for authors

see the weaknesses

4 comments
AuthorsOpenReview ↗

Response Part I

W1. The LLM’s independent contribution remains unclear. Useful controls would include non-LLM reference proposals with the same rollout validation, and separate ablations of reference proposals and growth guidance.

Thank you for this thoughtful suggestion. We conducted controlled ablations using the same generator to distinguish the contributions of LLM reference proposals, rollout validation, and post-acceptance growth guidance. All metrics below are computed on the full candidate pools before portfolio selection.

MethodMean QED ↑Best QED ↑SA ↓Lipinski ↑Mol. div. ↑Scaf. div. ↑Vina Score ↓Vina Dock ↓
(A) Similarity-based reference selector0.5160.7893.6700.8230.8170.738−6.1873−8.3044
(B) Rollout-only0.5180.7633.6390.8260.8190.753−6.1998−8.3277
(C) Agent-only0.5430.7813.4180.8290.8060.750−6.2593−8.3801
(D) No-guidance0.5630.7983.4650.8310.8240.770−6.2797−8.3276
(E) ForesightMol0.5860.8033.2440.8340.8250.775−6.3515−8.5306
  • LLM reference proposals. Structured chemical contexts describe the current intermediate, candidate references, and growth diagnostics, enabling the LLM to apply pretrained chemical knowledge to reference prioritization. Agent-only (C) selects references directly without rollout validation and improves mean QED, SA, and both Vina metrics over the backbone’s similarity-based selector (A). Under the same rollout validation procedure, LLM proposals (D) also outperform similarity-based proposals (B) in mean QED, SA, and both diversity metrics, supporting a contribution beyond additional search.

  • Rollout validation. Short-horizon rollouts verify whether a proposed reference improves prospective growth under shared chemical constraints. Rollout-only (B) compares the default reference with other high-similarity references and yields limited gains over (A). Combining this verification with LLM proposals (D) improves mean QED and both diversity metrics over Agent-only (C), showing how prospective validation complements LLM reference decisions.

  • Growth guidance. Verified LLM assessments are translated into bounded biases at the generator’s prediction heads. No-guidance (D) retains LLM proposals and rollout validation while disabling these biases. Enabling them in full ForesightMol (E) further improves all eight reported metrics, demonstrating the additional contribution of guidance beyond reference selection.

W2. Generation decisions lack traceable examples.

Thank you for this insightful comment. ForesightMol automatically retains the structured molecular evidence provided to the LLM and its assessments at each intervention, ensuring traceability of reference decisions. These records also inform the final delivery rationale. The “Reference Decision” paragraph in Section 4.2 reports how often interventions occur, and Figure 5 illustrates differences in the generated molecules.

Furthermore, we replayed both branches from the same partially generated molecule to observe how LLM-guided intervention affects subsequent molecular growth. In the 5NGZ chain A pocket, the LLM assessed a partial ligand containing 10 heavy atoms and proposed switching from the default reference to a challenger after weighing substructure overlap, pocket contacts, hydrogen-bonding features, and a moderate size penalty. The LLM assigned the proposal a confidence of 0.85. In matched rollouts, the challenger achieved a higher estimated utility than the default reference (0.8936 vs. 0.8753), with no recorded increase in risk. The system therefore accepted the switch and applied the four prediction-head adjustments. After these adjustments, the recorded trajectory of one ForesightMol product showed the addition of two carbon atoms to close a ring, followed by two oxygen atoms to form two carbonyl groups. This structure was absent from the Rag2Mol-G products generated from the same intermediate. Across successfully scored products from this pocket, ForesightMol improved mean Vina Score from −4.815 to −5.030 kcal/mol and mean Vina Dock from −6.230 to −6.429 kcal/mol relative to Rag2Mol-G.

We will include this case study in the appendix. The trace connects the LLM’s assessment of structured chemical evidence with rollout-based acceptance and subsequent changes in molecular growth.

AuthorsOpenReview ↗

Response Part II

W3. Baseline coverage is limited. The experiments omit relevant generators such as MolCRAFT and property optimization methods such as DecompOpt. Different generators should be compared using the same selection strategy, evaluation protocol, and comparable budgets, with results reported for both full candidate pools and final selected sets.

Thank you for this valuable suggestion. We added MolCRAFT [1] and DecompOpt [2] to broaden the comparison. We evaluated both baselines on pockets from the same CrossDocked2020 test split using a shared evaluation protocol. We report results for both the full candidate pools, as in Table 1 of the manuscript, and delivery sets of up to 10 molecules per pocket. For the delivery sets, we apply both Vina Dock ranking and our Score–Dock-aware portfolio selector to each generator’s pool, as in Table 2.

Full candidate pools

GeneratorMean QED ↑Best QED ↑SA ↓Lipinski ↑Mol. div. ↑Scaf. div. ↑Vina Score ↓Vina Dock ↓
MolCRAFT0.5150.8363.8130.7410.8280.777−5.7860−7.5859
DecompOpt0.4800.6573.9230.7260.6240.764−5.8389−8.1485
Rag2Mol-G0.5180.7903.5700.8230.8170.738−6.1873−8.3044
ForesightMol0.5770.7953.2460.8300.8280.776−6.3427−8.5342

Delivery sets

GeneratorSelectorMean QED ↑Best QED ↑SA ↓Lipinski ↑Mol. div. ↑Scaf. div. ↑Vina Score ↓Vina Dock ↓
MolCRAFTDock0.5280.7404.2820.6570.8610.981−7.1527−9.5364
MolCRAFTS–D0.6070.7673.9210.9110.8650.993−7.1169−8.8786
DecompOptDock0.4020.5544.0110.7230.7610.972−6.6112−8.9823
DecompOptS–D0.4440.5703.9290.9020.7540.985−6.5938−7.9257
Rag2Mol-GDock0.4940.6413.7550.7490.7230.860−6.8353−9.8336
Rag2Mol-GS–D0.5330.6673.6620.9370.7580.922−6.6249−8.8802
ForesightMolDock0.5130.6873.6330.7500.7630.874−6.9381−9.9103
ForesightMolS–D0.6390.7823.5600.9260.8120.987−6.8634−9.0062

As shown in the tables above, among the compared methods, ForesightMol achieves the highest mean QED and Lipinski pass rate and the lowest SA before selection. With the shared S–D selector, it achieves the highest mean and best QED and the lowest SA. Its Vina Dock score is more favorable than Rag2Mol-G’s both before selection and under either shared selector. Across all four generators, S–D improves QED, SA, and Lipinski compliance at the cost of less favorable redocking scores relative to Dock ranking. We have included these results in the revised manuscript.

[1] MolCRAFT: Structure-based drug design in continuous parameter space. ICML. 2024.

[2] DecompOpt: Controllable and decomposed diffusion models for structure-based molecular optimization. ICLR. 2024.

W4. Further comparisons should assess properties and diversity at matched docking performance or under an explicit binding-score constraint.

Thank you for this insightful comment. We assess the LLM’s contribution to generation through controlled ablations and comparisons on the full candidate pools, before final selection. The delivery-set comparisons address a separate question about what each generator can deliver under a shared selection strategy and budget. To test whether the property and diversity gains persist without sacrificing redocking performance, we applied an explicit constraint to the S–D selector. For each pocket, the selected set of 10 molecules must have a mean Vina Dock score no higher than that of the Rag2Mol-G Dock Top-10 on the same pocket.

GeneratorSelectorMean QED ↑Best QED ↑SA ↓Lipinski ↑Mol. div. ↑Scaf. div. ↑Vina Score ↓Vina Dock ↓
Rag2Mol-GDock0.4940.6413.7550.7490.7230.860−6.8353−9.8336
ForesightMolDock0.5130.6873.6330.7500.7630.874−6.9381−9.9103
ForesightMolS–D0.6390.7823.5600.9260.8120.987−6.8634−9.0062
ForesightMolDock-constrained S–D0.5690.7373.7390.8470.7650.971−7.3965−9.8768

As shown in the table above, under the Vina Dock constraint, ForesightMol improves mean and best QED, synthetic accessibility, Lipinski compliance, and both molecular and scaffold diversity over Rag2Mol-G with Dock ranking, while achieving more favorable Vina Score and Vina Dock. These results show that its gains in molecular quality and diversity persist without sacrificing redocking performance.

AuthorsOpenReview ↗

Response Part III

We greatly appreciate your insightful and helpful comments, as they will undoubtedly help us improve the quality of our article. If our response has successfully addressed your concerns and clarified any ambiguities, we respectfully hope that you consider raising the score. Should you have any further questions or require additional clarification, we would be delighted to engage in further discussion.

Dear Reviewer XtvK,

Thank you for your time and helpful comments on our paper. We have uploaded the revised manuscript with all changes marked.

With the author–reviewer discussion period ending in about two days, we would appreciate your feedback on whether our rebuttal addresses your concerns. We welcome further discussion and would be happy to clarify any remaining questions.

Thank you again for your time and help in improving our paper.

Best regards,

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.