acceptodds
Under review as a conference paper at ICLR 2027

CtrlMol: Controllable Molecular Generation for Drug Design via Flexible Condition-Adaptive Diffusion

Abstract

Conditional molecular generation is critical for drug design, where generated molecules need to satisfy varying design requirements. However, existing molecular generation methods lack flexibility in handling diverse reference conditions and controllability in adjusting generation toward desired molecular profiles, limiting their ability to meet some customized requirements. To address these limitations, we propose CtrlMol, a novel controllable molecular generation framework via flexible condition-adaptive diffusion model. CtrlMol achieves flexibility through Arbitrary-Count Condition Encoding that accommodates zero or more retrieved molecular and pocket conditions, followed by Self-Adaptive Modulation aggregating them into the denoising process in a unified manner. CtrlMol also achieves controllability through Independent Dual-Axis Classifier-Free Guidance that separately adjusts the guidance strengths of molecular and pocket conditions during inference, allowing controllable generation trade-offs between precision and diversity. Experiments on the TarPass benchmark demonstrate that CtrlMol consistently outperforms competing methods, achieving 28.8% improvement in target specificity and superior molecular property control. Case studies on a cardiovascular target ADRB2 further validate that the flexibility and controllability jointly enable effective adjustment of the generation distribution.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
drug design, molecular generation, controllable generation, flexible conditions, diffusion model

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_b7NuOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

CtrlMol addresses 3D molecular generation with variable numbers of reference ligands and target pockets. It uses a hybrid diffusion model, separate encoders for molecular and pocket conditions, pooled feature modulation, and two independently adjustable classifier-free guidance scales. The model is trained using ChEMBL and BindingDB, evaluated against 15 baselines on 18 TarPass targets, and examined through retrieval-count and guidance-scale sweeps on ADRB2. Its strongest reported result is improved agreement with the chemical distribution of known target actives, measured by FCD and molecular-property distances, while its docking scores remain behind the optimization-based baselines.

Critical strengths
  1. Variable, independently adjustable conditioning is a useful design objective. CtrlMol accommodates empty or multiple ligand and pocket reference sets and supplies separate guidance coefficients for the two sources. This directly addresses a practical mismatch between fixed-input conditional generators and heterogeneous information available in early drug discovery (Sections 3.1–3.4).

  2. The distributional results are competitive. In Table 1, CtrlMol obtains FCD Ref = 16.351 versus 22.982 for the strongest baseline on that metric, and Wass. Ref = 0.497 versus 0.629 for the best baseline. Tables 2–3 also show substantial changes in ADRB2 FCD as references and guidance scales are varied. These are substantive experimental observations, even though they need a more controlled interpretation.

Critical weaknesses
  1. The evaluation may reuse the chemical information it measures. Section 4.1 states that ChEMBL and BindingDB supply both training examples and retrieved conditions at inference, whereas TarPass evaluates proximity to approximately 1,000 known active compounds per target. The manuscript does not specify whether the evaluation actives, their close analogues, or their scaffolds are excluded from training and from the retrieved ligand sets. Nor does the common TarPass scoring protocol in Appendix B.3 establish that the baselines receive comparable target-active reference information. Consequently, the FCD improvement could partly reflect access to the reference chemical distribution rather than a generalizable gain in target-aware generation. This is an unresolved information-exposure and comparator question, not a demonstrated leakage finding.

  2. The new components have not been isolated from simpler alternatives. Tables 2 and 3 vary the number of reference conditions and the two guidance scales; neither removes or replaces ACCE, SACM, or IDA-CFG. At gM = gP = 1, adding one pocket to 100 ligand references changes ADRB2 FCD from 19.25 to 19.63 (Table 2), whereas at K = 300 the favorable gM = 2, gP = 1 setting improves upon gM = 2, gP = 0 (17.52 versus 19.19; Table 3). The pocket effect is therefore setting-dependent, and the sweeps alone do not establish the proposed mechanisms as the source of the headline gains. There is also a concrete formulation ambiguity: the two-branch average in Section 3.4.2 gives z + (γM ⊙ z + βM)/2 when the pocket branch is absent, rather than the full single-condition modulation claimed immediately below the equation. The implementation and the value added beyond simpler pooled conditioning and standard guidance need to be distinguished.

Questions for authors
  1. For each TarPass target, how many evaluation active molecules (or matching scaffolds) also occur in the training corpus and in the ligand sets retrieved at inference? Are molecules used to compute the target-reference FCD explicitly excluded from retrieval?

  2. Which baselines, if any, had access to the same retrieved active-ligand and pocket references as CtrlMol, and how were proposal counts and rejection/postprocessing budgets matched?

  3. Do the authors have comparisons that replace SACM with simple pooling/FiLM or remove either conditioning branch while keeping the same retrieved references? If not, can Tables 2–3 be described as parameter sweeps rather than component ablations?

  4. In Section 3.4.2, is the single-condition residual intentionally halved by the two-branch average, or is there an additional rescaling or masking step in the implementation? Please also clarify how zero-valued conditions map to zero modulation parameters.

2 comments
AuthorsOpenReview ↗

Response to Reviewer b7Nu (1/2)

Q1/W1: For each TarPass target, how many evaluation active molecules (or matching scaffolds) also occur in the training corpus and in the ligand sets retrieved at inference? Are molecules used to compute the target-reference FCD explicitly excluded from retrieval?

A1: We thank the reviewer for the question and weakness on data leakage. 90% of TarPass set occur in the training set, and 74% of TarPass set occur in the retrieval set. To investigate data leakage in CtrlMol, we kept the TarPass testing set unchanged to ensure fairness and used it as a blacklist, removing all blacklisted molecules from the training set to obtain CtrlMol-. Additionally, we removed all blacklisted molecules from both the training and retrieval sets to obtain CtrlMol--. We re-trained CtrlMol- and CtrlMol--, and re-evaluated the performance of CtrlMol- and CtrlMol-- addedin∗∗Table1∗∗added in **Table 1**. The target-aware advantage slightly decreased but remained consistently better than baselines. Compared with the second-best DeepBlock, CtrlMol- and CtrlMol-- were still improvements of 20% and 12% in FCD Ref, respectively.

We have appended the details of data exclusions at the end of Appendix B.1, and added the performance of data leakage evaluation in Table 1. Details in the Shared Concern 1 of General Response.

ModelValidityDiversityWass. Ref ↓↓Wass. Shift ↓↓Vina Score Med,˙↓Med\., ↓Ligand Efficiency ↑↑Sim. Ref ↑↑FCD Ref ↓↓FCD Shift ↓↓
DRAGONFLY110.6292.166-8.1920.2610.12727.1311.443
DeepBlock110.8842.685-7.7760.3390.12022.9821.313
SimpleSBDD0.9660.9870.8562.322-8.2300.3190.11239.0092.665
TamGen~10.9681.3464.243-6.8900.3570.10336.6482.078
DiffSBDD0.9680.9871.1703.957-8.0320.3260.09431.8242.093
DrugFlow0.8310.3990.6332.512-9.7760.3350.13032.1782.442
IPDiff~10.9401.9066.896-7.7180.4210.06950.2923.830
Lingo3DMol~10.9060.8223.408-9.0240.2850.12229.5412.131
MolCraft10.4980.7632.649-9.9870.3500.11238.3692.880
PocketFlow10.6921.4144.722-7.7950.4120.09736.9672.469
SurfGen10.9481.1224.126-9.5310.3770.11042.5623.245
TargetDiff~10.9681.2334.719-7.5830.3990.09644.5943.497
DrugFlow-PA0.6000.6370.9553.658-10.3740.4040.12842.8973.371
MolPilot110.7963.156-10.4830.3380.11941.4953.208
REINVENT4111.3925.592-10.0230.3670.10349.9213.890
CtrlMol110.4971.760-8.5100.3090.12816.3511.210
CtrlMol-110.6341.880-8.4750.3270.12618.3771.308
CtrlMol--110.6732.007-8.4700.3270.12520.2541.358
Random——0.695—-7.8350.2930.11626.984—
AuthorsOpenReview ↗

Response to Reviewer b7Nu (2/2)

Q2: Which baselines, if any, had access to the same retrieved active-ligand and pocket references as CtrlMol, and how were proposal counts and rejection/postprocessing budgets matched?

A2: We thank the reviewer for the question. None of the baselines had access to the same conditions as CtrlMol. The baselines support at most one pocket and at most one ligand as conditions, corresponding to K=0/1 and L=0/1 in Table 2. This setting cannot bring out the full advantages of CtrlMol.

Following the rule of TarPass Benchmark, we should generate 1,000 molecules for each target with each model. The count of "1,000 molecules" is directly subject to the output of each model, neither more or less. We did not apply any rejection/postprocessing.


Q3/W2: Do the authors have comparisons that replace SACM with simple pooling/FiLM or remove either conditioning branch while keeping the same retrieved references? If not, can Tables 2–3 be described as parameter sweeps rather than component ablations?

A3: We thank the reviewer for the question and weakness on ablation studies. Tables 2–3 appear to be parameter sweeps, but imply component ablations. As described in Section 3.4.2, the advantage of SACM over FiLM lies in adaptively supporting zero-condition input at any time, rather than higher performance. The ablation study on retrieval conditions (Table 2) demonstrates that SACM is necessary for near-optimal performance, which requires a zero-condition scenario K=0orL=0K=0 or L=0. In a similar way, the advantage of IDA-CFG over standard CFG or standard conditioning lies in supporting heterogeneous guidance scales. The ablation study on guidance scales (Table 3) demonstrates that IDA-CFG is necessary for optimal performance, which requires heterogeneous guidance scales.


Q4/W2: In Section 3.4.2, is the single-condition residual intentionally halved by the two-branch average, or is there an additional rescaling or masking step in the implementation? Please also clarify how zero-valued conditions map to zero modulation parameters.

A4:

We thank the reviewer for pointing out this ambiguity. The single-condition residual is not intentionally halved in our implementation. The 1/2 factor in Section 3.4.2 applies only when both condition branches are present.

In Section 3.4.2, it has been described the three scenarios. When both condition types exist, the modulation averages the two branches $z \leftarrow \frac{1}{2}\big[(1+\gamma_{\mathcal{M}}) \odot z + \beta_{\mathcal{M}} + (1+\gamma_{\mathcal{P}}) \odot z + \beta_{\mathcal{P}}\big]$.

When either condition type is absent, the corresponding parameters become zero, reducing to single-type modulation z←(1+γ)⊙z+βz \leftarrow (1+\gamma) \odot z + \beta.

When both are absent, it naturally degenerates to identity z←zz \leftarrow z.

Reviewer_oKHKOpenReview ↗
Rating
1
Confidence
4
Informativeness
3
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes CtrlMol, a molecular generation framework conditioned on variable numbers of reference ligands and pockets. It combines condition encoding with mean-pooled FiLM modulation and independently adjusts classifier-free guidance strengths for the two condition types. Experiments on TarPass and an ADRB2 case study evaluate reference-distribution fidelity and control over the generated distribution.

Critical strengths

Practical motivation. Supporting variable or missing ligand and pocket information within a single model addresses a relevant design requirement.

Promising distributional fidelity. Under the evaluated settings, generated molecules closely match the chemical and physicochemical distributions of reference ligands.

Critical weaknesses

The baselines do not adequately address the central claims. Despite emphasizing reference similarity and property-distribution matching, the paper lacks comparisons with pharmacophore-, scaffold-, and ligand-conditioned generators specifically designed for these objectives. Lower reference distances than pocket-only models do not establish an advance over relevant conditional generation methods.

The interpretation of the metrics is overstated. FCD and descriptor Wasserstein distance measure similarity to reference distributions; they do not directly establish biological target specificity or discovery of useful novel chemical space. In particular, describing an FCD reduction as a “28.8% improvement in target specificity” overstates its demonstrated meaning.

The practical benefit of additional conditioning information remains unclear. Despite access to both reference ligands and pockets, CtrlMol achieves weaker binding affinity than several pocket-based baselines. This is not inherently a failure, but the claimed exploration advantage requires evidence of a favorable novelty–coverage–affinity trade-off against relevant methods.

The contributions of the proposed components and the training setup are insufficiently established. Comparisons do not isolate the benefits of the proposed mean-pooling/FiLM and CFG extensions over simpler alternatives. The construction of reference sets for each training molecule, including exclusion of training targets and evaluation ligands from those sets, is also unclear.

Questions for authors
  1. Why were pharmacophore-, scaffold-, and ligand-conditioned generators not included as primary baselines? Are there results demonstrating advantages over relevant methods given the same reference information?

  2. What justifies interpreting FCD-based results as “target specificity”? Please distinguish reference-distribution fidelity from biological binding specificity.

  3. Given the weaker binding affinity, which results support a favorable trade-off among novelty, chemical-space coverage, and affinity?

  4. How are reference ligands and pockets selected for each training molecule? Please clarify how overlap and structural similarity with the molecule being reconstructed and with evaluation ligands are controlled.

  5. What evidence isolates the contributions of SACM and IDA-CFG relative to simple reference-embedding pooling and standard conditioning/CFG alternatives?

2 comments
AuthorsOpenReview ↗

Response to Reviewer oKHK (1/2)

We thank the reviewer for the constructive feedback. Below we address each point. We hope these responses address the concerns.

Q1/W1: Why were pharmacophore-, scaffold-, and ligand-conditioned generators not included as primary baselines? Are there results demonstrating advantages over relevant methods given the same reference information?

A1: We thank the reviewer for the question and weakness on our central claim. CtrlMol explicitly claims the advantage of target awareness, and TarPass Benchmark also emphasizes the target awareness shortfall of current models (baselines). CtrlMol and baselines in TarPass Benchmark have the same central claim on target awareness.

As described in Section 2.2 about relevant methods, "pharmacophore-, scaffold-, and ligand-conditioned generators" may refer to current LBDD methods, which typically accept fixed types and fixed counts of conditions as input and lack target-aware functions. We agree that LBDD methods are relevant comparisons. However, they are not directly comparable to CtrlMol under TarPass because they generally require a specific condition type, whereas CtrlMol uses variable-size retrieved ligand sets and target information. Therefore, in this paper we followed the TarPass benchmark baselines to evaluate the target-awareness setting consistently. We have revised the manuscript to clarify that our empirical claim is relative to TarPass baselines, not to all possible ligand-conditioned LBDD methods. We also add this as a limitation and an important direction for future work.


Q2/W2: What justifies interpreting FCD-based results as “target specificity”? Please distinguish reference-distribution fidelity from biological binding specificity.

A2: We thank the reviewer for raising this important point. We agree that “target specificity” was misleading about binding affinity, and FCD only measures reference-distribution fidelity. In the TarPass Benchmark, FCD-based results are defined as "target awareness".

We have revised "target specificity" to "target awareness" throughout the manuscript. Details in the Shared Concern 2 of General Response.


Q3/W3: Given the weaker binding affinity, which results support a favorable trade-off among novelty, chemical-space coverage, and affinity?

A3: We thank the reviewer for the question and weakness on favorable trade-off. CtrlMol explicitly claims target awareness rather than binding affinity, so the relevant trade-off of interest is between target awareness and affinity ratherthannovelty–coverage–affinitytrade−offrather than novelty–coverage–affinity trade-off. CtrlMol is a de novo generation method oriented toward hit identification rather than lead optimization, where the emphasis is on target-aware exploration over binding affinity. Therefore, our favorable trade-off means prioritizing target awareness while keeping affinity adequate. Table 1 supports our favorable trade-off between target awareness and affinity.

We have revised the statement about the trade-off between target awareness and binding affinity in the abstract. Details in the Shared Concern 2 of General Response.*

AuthorsOpenReview ↗

Response to Reviewer oKHK (2/2)

Q4/W4: How are reference ligands and pockets selected for each training molecule? Please clarify how overlap and structural similarity with the molecule being reconstructed and with evaluation ligands are controlled.

A4: We thank the reviewer for the question and weakness on data exclusion and training details. For each training molecule, its molecular condition is only itself, its pocket condition is one of the paired pockets. In addition to the CFG mechanism, both molecular and pocket condition can be absent. Multi-to-one sampling only requires one-to-one training.

We clarify that 90% of TarPass set occur in the training set, and 74% of TarPass set occur in the retrieval set. To further investigate data leakage in CtrlMol, we kept the TarPass testing set unchanged to ensure fairness and used it as a blacklist, removing all blacklisted molecules from the training set to obtain CtrlMol-. Additionally, we removed all blacklisted molecules from both the training and retrieval sets to obtain CtrlMol--. We re-trained CtrlMol- and CtrlMol--, and re-evaluated the performance of CtrlMol- and CtrlMol-- addedin∗∗Table1∗∗added in **Table 1**. The target-aware advantage slightly decreased but remained consistently better than baselines.

We have appended the details of data exclusions at the end of Appendix B.1, and added the performance of data leakage evaluation in Table 1. Details in the Shared Concern 1 of General Response.

Table 1

ModelValidityDiversityWass. Ref ↓↓Wass. Shift ↓↓Vina Score Med,˙↓Med\., ↓Ligand Efficiency ↑↑Sim. Ref ↑↑FCD Ref ↓↓FCD Shift ↓↓
DRAGONFLY110.6292.166-8.1920.2610.12727.1311.443
DeepBlock110.8842.685-7.7760.3390.12022.9821.313
SimpleSBDD0.9660.9870.8562.322-8.2300.3190.11239.0092.665
TamGen~10.9681.3464.243-6.8900.3570.10336.6482.078
DiffSBDD0.9680.9871.1703.957-8.0320.3260.09431.8242.093
DrugFlow0.8310.3990.6332.512-9.7760.3350.13032.1782.442
IPDiff~10.9401.9066.896-7.7180.4210.06950.2923.830
Lingo3DMol~10.9060.8223.408-9.0240.2850.12229.5412.131
MolCraft10.4980.7632.649-9.9870.3500.11238.3692.880
PocketFlow10.6921.4144.722-7.7950.4120.09736.9672.469
SurfGen10.9481.1224.126-9.5310.3770.11042.5623.245
TargetDiff~10.9681.2334.719-7.5830.3990.09644.5943.497
DrugFlow-PA0.6000.6370.9553.658-10.3740.4040.12842.8973.371
MolPilot110.7963.156-10.4830.3380.11941.4953.208
REINVENT4111.3925.592-10.0230.3670.10349.9213.890
CtrlMol110.4971.760-8.5100.3090.12816.3511.210
CtrlMol-110.6341.880-8.4750.3270.12618.3771.308
CtrlMol--110.6732.007-8.4700.3270.12520.2541.358
Random——0.695—-7.8350.2930.11626.984—

Q5/W4: What evidence isolates the contributions of SACM and IDA-CFG relative to simple reference-embedding pooling and standard conditioning/CFG alternatives?

A5: We thank the reviewer for the question and weakness on contributions of the proposed sub-modules. The contributions of SACM and IDA-CFG lie in flexible controllability. Unlike FiLM, SACM adaptively supports zero-condition input at any time. IDA-CFG supports heterogeneous guidance scales, whereas standard CFG supports only homogeneous guidance scales and standard conditioning uses just a single fixed guidance scale. The ablation study on retrieval conditions (Table 2) demonstrates that SACM is necessary for near-optimal performance, which requires a zero-condition scenario K=0orL=0K=0 or L=0. The ablation study on guidance scales (Table 3) demonstrates that IDA-CFG is necessary for optimal performance, which requires heterogeneous guidance scales.

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

The paper addresses conditional molecular generation when the available conditions are heterogeneous and variable in number: in practice a designer may have zero, one, or many retrieved active ligands and/or one or more binding pockets, possibly unpaired. Existing methods assume fixed condition types and counts at training time and cannot adjust condition strength at inference. CtrlMol contributes three components: (i) Arbitrary-Count Condition Encoding, which accepts zero or more molecular and pocket conditions; (ii) Self-Adaptive Condition Modulation, which aggregates heterogeneous conditions of different types and counts into a unified modulation of the denoising network; and (iii) Independent Dual-Axis Classifier-Free Guidance (IDA-CFG), which exposes separate guidance scales for the molecular axis (g_M) and the pocket axis (g_P) so that the fidelity–diversity trade-off can be tuned at inference. Experiments are on the TarPass benchmark over 18 targets against 15 baselines, reporting validity, Uniq@1k diversity, descriptor distances to reference ligands (Wasserstein over 15 descriptors), Vina score, ligand efficiency, and target-specificity metrics (Tanimoto similarity, FCD to reference and to a random background). Ablations vary the retrieved condition counts (K, L) and the two guidance scales, with a case study on ADRB2.

Critical strengths
  1. The problem framing is well grounded in actual drug-discovery workflow. Most conditional-generation papers assume a fixed conditioning configuration, which is not how a medicinal chemist works: you retrieve whatever ligands and structures exist and want the generator to degrade gracefully as the evidence thins out. Decomposing the requirement into "flexible in what you can condition on" (ACCE + SACM) and "controllable in how strongly you follow it" (IDA-CFG) is a sensible factorization, and the two-axis guidance is a natural interface for a practitioner who wants to trade novelty against resemblance separately for the ligand axis and the pocket axis.

  2. The ablations are informative about the mechanism, not just the endpoint. The K/L sweep shows a clear saturation behaviour (FCD decreasing with K and plateauing near ~19 once K ≈ 100), and when both conditions are absent (K = 0, L = 0) the model falls back to a drug-like distribution (FCD 9.96 to the random background) rather than failing. The guidance-scale grid (Table 3) shows an interior optimum at g_M = 2, g_P = 1 with degradation at higher scales, which is the behaviour one would hope for from a CFG-style knob and supports the claim that the two axes act semi-independently.

Critical weaknesses
  1. The headline claim is measured on a metric that is not the design objective, and the objective-aligned column contradicts the abstract. The reported "28.8% improvement in target specificity" is an improvement in FCD to the target's reference ligands (16.351 vs. 22.982 for DeepBlock). But a distributional distance to the reference ligands rewards producing molecules that resemble what was already retrieved; with K up to 300 retrieved molecules injected as conditions, low FCD is partly a statement about how strongly the model copies its own conditioning set, not about target-specific binding. Meanwhile the paper's own Table 1 shows CtrlMol's Vina score trailing all three optimization-based baselines, and the text concedes this ("lags behind all three optimization-based methods"). The abstract should lead with that trade-off rather than with a 28.8% figure on a proxy whose relationship to affinity is unestablished. Without an experiment decoupling "resembles the retrieved set" from "binds the target" — e.g. FCD computed against held-out ligands never used as conditions — the main claim is not supported.

  2. The validity/diversity/specificity numbers jointly look too good, and the protocol behind them is not verifiable from the main text. CtrlMol reports validity = 1.0, Uniq@1k = 1.0, and the lowest FCD to reference ligands, i.e. perfect sample-level validity and diversity alongside the strongest distributional match to a specific reference set. Each of these is individually achievable, but their co-occurrence at the extreme is unusual and needs a protocol-level explanation: over how many molecules is validity computed, what counts as a valid molecule, and is Uniq@1k computed within a target or pooled across all 18 targets? In addition, the 18 targets are averaged into a single number in Table 1 with no per-target dispersion, so it is impossible to tell whether the FCD advantage is uniform or driven by a few easy targets — which matters because the ADRB2 case study, where the guidance sweep is performed, is only one of the 18. Non-3D baselines being included in the same table as 3D pocket-conditioned methods (with the affinity comparison explicitly acknowledged as apples-to-oranges) further muddies the headline ranking.

Questions for authors
  1. Can you compute the target-specificity metrics (FCD Ref, Sim. Ref) against held-out reference ligands that were never available to the model as conditions? This would separate "resembles the retrieved set" from "captures the target's chemical space". If the advantage disappears under that protocol, how should the 28.8% claim be restated?

  2. Please give the per-target breakdown (mean ± std over the 18 targets) for Validity, Uniq@1k, FCD Ref, and Vina Score, and state the exact definition and sample size used for validity and for Uniq@1k (within-target or pooled).

  3. How do you reconcile validity = 1.0 and Uniq@1k = 1.0 with the lowest FCD to a specific reference set? A short protocol paragraph plus the per-target numbers would settle this.

  4. Given that binding affinity is the quantity a medicinal chemist ultimately cares about, what is the recommended operating point (g_M, g_P, K, L) when affinity is prioritized over resemblance, and what affinity is achieved there? The guidance sweep in Table 3 is reported only in FCD.

  5. For the K/L ablation, conditions are retrieved for the ADRB2 case. Are the retrieved ligands excluded from the evaluation set, and does the saturation at K ≈ 100 depend on retrieval similarity (e.g. does it hold when retrieved ligands are weaker analogues)?

3 comments
AuthorsOpenReview ↗

Response to Reviewer rGCU (1/3)

We thank the reviewer for the constructive feedback. Below we address each point. We hope these responses address the concerns.

Q1/W1: Can you compute the target-specificity metrics FCD Ref, Sim\. Ref against held-out reference ligands that were never available to the model as conditions? This would separate "resembles the retrieved set" from "captures the target's chemical space". If the advantage disappears under that protocol, how should the 28.8% claim be restated?

A1: We thank the reviewer for the question. To further investigate data leakage in CtrlMol, we kept the TarPass testing set unchanged to ensure fairness and used it as a blacklist, removing all blacklisted molecules from the training set to obtain CtrlMol-. Additionally, we removed all blacklisted molecules from both the training and retrieval sets to obtain CtrlMol--. We re-trained CtrlMol- and CtrlMol--, and re-evaluated the performance of CtrlMol- and CtrlMol-- addedin∗∗Table1∗∗added in **Table 1**, which contains computation of the target-aware metrics FCD Ref, Sim\. Ref against held-out reference ligands. The target-aware advantage slightly decreased but remained consistently better than baselines. Compared with the second-best DeepBlock, CtrlMol- and CtrlMol-- were still improvements of 20% and 12% in FCD Ref, respectively.

We have appended the details of data exclusions at the end of Appendix B.1, and added the performance of data leakage evaluation in Table 1. Details in the Shared Concern 1 of General Response.

We have changed all the "28.8%" to the advantage of target awareness throughout the manuscript. We have revised the statement about the trade-off between target awareness and binding affinity in the abstract. Details in the Shared Concern 2 of General Response.

Table 1

ModelValidityDiversityWass. Ref ↓↓Wass. Shift ↓↓Vina Score Med,˙↓Med\., ↓Ligand Efficiency ↑↑Sim. Ref ↑↑FCD Ref ↓↓FCD Shift ↓↓
DRAGONFLY110.6292.166-8.1920.2610.12727.1311.443
DeepBlock110.8842.685-7.7760.3390.12022.9821.313
SimpleSBDD0.9660.9870.8562.322-8.2300.3190.11239.0092.665
TamGen~10.9681.3464.243-6.8900.3570.10336.6482.078
DiffSBDD0.9680.9871.1703.957-8.0320.3260.09431.8242.093
DrugFlow0.8310.3990.6332.512-9.7760.3350.13032.1782.442
IPDiff~10.9401.9066.896-7.7180.4210.06950.2923.830
Lingo3DMol~10.9060.8223.408-9.0240.2850.12229.5412.131
MolCraft10.4980.7632.649-9.9870.3500.11238.3692.880
PocketFlow10.6921.4144.722-7.7950.4120.09736.9672.469
SurfGen10.9481.1224.126-9.5310.3770.11042.5623.245
TargetDiff~10.9681.2334.719-7.5830.3990.09644.5943.497
DrugFlow-PA0.6000.6370.9553.658-10.3740.4040.12842.8973.371
MolPilot110.7963.156-10.4830.3380.11941.4953.208
REINVENT4111.3925.592-10.0230.3670.10349.9213.890
CtrlMol110.4971.760-8.5100.3090.12816.3511.210
CtrlMol-110.6341.880-8.4750.3270.12618.3771.308
CtrlMol--110.6732.007-8.4700.3270.12520.2541.358
Random——0.695—-7.8350.2930.11626.984—
AuthorsOpenReview ↗

Response to Reviewer rGCU (2/3)

Q2/Q3/W2: Please give the per-target breakdown (mean ± std over the 18 targets) for Validity, Uniq@1k, FCD Ref, and Vina Score, and state the exact definition and sample size used for validity and for Uniq@1k within−targetorpooledwithin-target or pooled. How do you reconcile validity = 1.0 and Uniq@1k = 1.0 with the lowest FCD to a specific reference set? A short protocol paragraph plus the per-target numbers would settle this.

A2/A3: We thank the reviewer for the question about per-target performance. We have added per-target performance in Appendix B.9. We have stated the exact definitions in Appendix B.2.

The protocol is as follows. CtrlMol checks chemical validity and rationality itself, and will not output invalid molecules. Following the rule of TarPass Benchmark, we generated 1,000 molecules for each target. These molecules must be valid from CtrlMol ∗∗Validity=10˙∗∗**Validity = 1\.0** and indeed non-duplicated ∗∗Uniq@1k=10˙∗∗**Uniq@1k = 1\.0** without post-processing. Then, we evaluated the performance within each target, and finally averaged across the 18 targets. Validity and Uniq@1k are basic performance of generative models, and indeed independent of target awareness (the lowest FCD to a specific reference set).

TargetValidityDiversityWass. Ref ↓↓Wass. Shift ↓↓Vina Score Med,˙↓Med\., ↓Ligand Efficiency ↑↑Sim. Ref ↑↑FCD Ref ↓↓FCD Shift ↓↓
5-HT2A110.2402.031-8.7980.3360.1259.2301.480
BCL2110.5671.277-7.3510.2530.11624.7501.250
BRAF110.4661.249-10.6300.3180.15020.7001.000
BRD4110.4921.431-7.1780.2560.11314.4901.060
BTK110.6051.204-8.84570.3010.11512.8701.280
Beta2AR110.8743.570-8.6670.3670.13017.5201.080
DRD2110.3422.178-9.3160.3410.1309.5601.540
HDAC6110.2982.118-6.7840.2550.11122.9101.980
HIV-RT110.3481.789-8.9690.4070.11321.5901.160
JAK2110.8422.081-8.2730.3330.11615.9801.400
MEK1110.6371.765-8.6480.3250.14818.5201.040
NAMPT110.2441.502-10.0030.3330.15111.0801.100
PI3Kα110.5001.912-7.8560.2950.12311.4301.240
PPARα110.6751.222-8.9150.2790.14223.3601.000
PRMT5110.3581.595-8.7010.2890.14116.3301.000
ROCK1110.2711.845-7.9830.3040.13010.2801.170
RARα110.4511.567-7.7730.2940.12319.4701.000
TYK2110.7421.346-8.4970.2800.13414.2501.000
Mean110.4971.760-8.5100.3090.12816.3511.210
STD000.1940.5420.9260.0390.0134.8910.250
AuthorsOpenReview ↗

Response to Reviewer rGCU (3/3)

Q4/W1: Given that binding affinity is the quantity a medicinal chemist ultimately cares about, what is the recommended operating point gM,gP,K,Lg_M, g_P, K, L when affinity is prioritized over resemblance, and what affinity is achieved there? The guidance sweep in Table 3 is reported only in FCD.

A4: We thank the reviewer for the question. We acknowledge that the guidance sweep in Table 3 reports only FCD. Figure 5 in Appendix B.8 showed the top-10 generated molecules with high binding affinity correspond to the recommended operating point ∗∗gM=2,gP=1,K=300,L=1∗∗**g_M=2, g_P=1, K=300, L=1** in Table 3.

CtrlMol mainly claims target awareness rather than binding affinity. We agree that CtrlMol is not skilled in affinity (Table 1), so the case study is not designed and recommanded around affinity-prioritized generation. Following the suggestion in Weakness 1, we have revised the statement about the trade-off between target awareness and binding affinity in the abstract.


Q5: For the K/L ablation, conditions are retrieved for the ADRB2 case. Are the retrieved ligands excluded from the evaluation set, and does the saturation at K ≈ 100 depend on retrieval similarity e\.g\. does it hold when retrieved ligands are weaker analogues?

A5: We thank the reviewer for the question about condition saturation. As in Answer 1, CtrlMol-- retrieved ligands excluded from the evaluation set, and CtrlMol/Table 3 did not. But CtrlMol and CtrlMol-- showed similar performance. The saturation issue has not been discussed in this work. Beyond retrieval similarity, retrieval priority may also affect saturation, since retrieving weakly related ligands can push the generated molecules away from the main distribution or even out of distribution.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Genernal Response

Dear Area Chair and Reviewers,

We thank all reviewers for their constructive feedback. In response to their comments, we have revised the manuscript and uploaded the revised version.

We have read the reviews closely and identified two shared concerns that require clarification and correction, primarily:

ii data leakage among training, testing and retrieval sets;

iiii CtrlMol claims 28.8% improvement in target specificity using FCD.

In this general response, we summarize our responses to these shared concerns and the main changes we are making to the manuscript.

Shared Concern 1: data leakage among training, testing and retrieval sets.

We first clarify relationships among training set, testing set and retrieval set in CtrlMol. The training set consists of ChEMBL and BindingDB, both of which are fully public and widely used databases. The testing set is the TarPass Benchmark, a published dataset. The retrieval set can be retrieved from anywhere by users, and is retrieved from ChEMBL in this work. The training, testing and retrieval sets are operationally independent of one another. Because the training set is derived from such widely used and fully public databases, it is less practical to discuss data leakage with respect to the baseline models and reference sets in the testing benchmark. Training, sampling, and evaluating CtrlMol in this manner is reasonably fair and sound.

To further investigate data leakage in CtrlMol, we kept the TarPass testing set unchanged to ensure fairness and used it as a blacklist, removing all blacklisted molecules from the training set to obtain CtrlMol-. Additionally, we removed all blacklisted molecules from both the training and retrieval sets to obtain CtrlMol--. We evaluated the performance of CtrlMol, CtrlMol- and CtrlMol-- (added in Table 1). The target-aware advantage slightly decreased but remained consistently better than baselines. Compared with the second-best DeepBlock, CtrlMol- and CtrlMol-- were still improvements of 20% and 12%, respectively.

We have appended the above discussion at the end of Appendix B.1, and added the performance of data leakage evaluation in Table 1.

Shared Concern 2: CtrlMol claims 28.8% improvement in target specificity using FCD.

We have realized that "target specificity" may cause confusion with binding affinity. Since our main results follow the TarPass Benchmark, we adopt "target awareness" instead of "target specificity," consistent with the terminology used in the benchmark. The benchmark measures target awareness using FCD reference−distributionfidelityreference-distribution fidelity. *We have ****revised "target specificity" to "target awareness" ***throughout the manuscript.

Following the investigation of data leakage, CtrlMol indeed demonstrated the advantage of target awareness, but the claim of "28.8% improvement" seems inappropriate and overstated. We have changed all the "28.8%" to the advantage of target awareness throughout the manuscript.

CtrlMol is a de novo generation method oriented toward hit identification rather than lead optimization, where the emphasis is on target-aware exploration over binding affinity. Therefore, our favorable trade-off is prioritizing target awareness while keeping affinity adequate. We have changed our claim in the abstract, "CtrlMol consistently outperforms competing methods, achieving superiority in target awareness and molecular property control with adequate binding affinity."

We also have added a detailed CtrlMol performance for each target in Appendix B.9 to support our claim.

We appreciate the reviewers’ comments, which helped us clarify the evaluation protocol, correct overstated claims, and better position the contribution of CtrlMol.

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.