acceptodds
Under review as a conference paper at ICLR 2027

ChemVerse: Benchmarking Active Learning for Virtual Screening of Million-Scale Libraries Across Docking Oracles and Biological Targets

Abstract

Virtual screening is widely used in early-stage drug discovery to prioritize promising molecules before costly experimental testing. Molecular docking enables this process at large scale, but exhaustively screening modern chemical libraries can become computationally prohibitive. Iterative, model-based selection offers an alternative by evaluating a limited subset with a docking oracle, training a surrogate model on the resulting scores, and using rapid inference to prioritize a much larger chemical space. However, it remains unclear which aspects of this setup most strongly determine the performance of adaptive virtual screening. We introduce ChemVerse, a benchmark comprising 389 million docking scores across seven protein targets and three chemical libraries ranging from molecules to molecules. Across four distinct docking oracles, seven biological targets, and three chemical libraries, complemented by publicly available DOCK3 screens, we systematically compare six molecular representations, three uncertainty-aware surrogate models, and four acquisition strategies. We find that the docking oracle strongly determines the learnability of its score landscape, producing substantially greater variation in optimization performance than molecular representation and uncertainty quantification. Three-dimensional representations perform consistently well across oracles, while uncertainty quantification has comparatively little effect. ChemVerse provides a systematic framework for understanding which components of ML-guided virtual screening generalize across oracles and which must be reconsidered as the screening environment changes.

Primary area:
datasets and benchmarks
Keywords:
Bayesian Optimization, Drug Discovery, Molecular Representation Learning

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_H2VwOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The authors introduce ChemVerse, a benchmark for docking-based Bayesian optimization in virtual screening, covering seven protein targets and three screening libraries, with 389 million docking scores. They systematically evaluate four axes: docking oracle, molecular representation used as input to the surrogate model, uncertainty quantification method, and acquisition function. They find that the docking oracle has the largest impact on optimization performance, followed by molecular representation and acquisition function, while the choice of uncertainty quantification method has relatively little effect.

Critical strengths

S1 Large-scale and comprehensive benchmark: The authors provide a valuable dataset of 389 million docking scores across multiple protein targets, chemical libraries, and docking oracles, which is a useful resource for the molecular ML community.

S2 Some interesting empirical findings: The benchmark provides useful insights into the relative importance of different components of BO-based virtual screening, particularly the strong influence of the docking oracle and the limited impact of the uncertainty estimation method.

Critical weaknesses

W1 Limited investigation of surrogate models and uncertainty quantification: Although the authors systematically compare molecular representations, uncertainty estimation methods, and acquisition functions, they use a single fixed surrogate model architecture. Given the importance of surrogate performance for BO, this limits the comprehensiveness of the benchmark and the generalizability of its conclusions, particularly regarding the limited impact of uncertainty quantification under the specific large-batch acquisition protocol.

W2 Limited ML contribution and relevance to ICLR: While the benchmark is sound and valuable for the molecular ML community, its methodological novelty is limited, and the main findings are largely unsurprising. Furthermore, as the largest effect on optimization performance comes from the choice of docking oracle rather than the ML methods, I find it difficult to view this primarily as a benchmark of AI/ML methods. In my opinion, the contribution is therefore not sufficiently significant for ICLR, and I lean toward rejection, primarily due to its limited ML contribution rather than concerns about the quality of the work.

Minor weaknesses and suggestions

  • W3 Overview figure An overview figure summarizing the benchmark and its experimental dimensions would help readers understand the overall setup.
  • W4 Active learning vs. Bayesian optimization The title refers to active learning (AL), but the paper actually focuses on Bayesian optimization (BO). While the two are closely related, their objectives differ: AL typically aims to improve predictive accuracy, whereas BO aims to identify high-performing candidates. The authors themselves acknowledge this distinction in Section 2.1 and explicitly state that BO better describes their setting. I therefore find the use of AL in the title, and occasionally elsewhere in the paper (e.g., the Discussion), somewhat misleading and would suggest using BO consistently.
  • W5 Missing details Important methodological details, such as the surrogate architecture and the representations used in different experiments, should be more clearly stated in the main paper.
  • W6 Aggregate metrics The definitions of some reported metrics, particularly the "median best-to-worst AUC spread" and "seed-subtracted recovery AUC," should be clarified. Not sure I understand either of these.
  • W7 Presentation of results Understandably, many results are presented in the appendix, including comparisons across several experimental axes, which sometimes makes the results challenging to follow. At the same time, the three recovery metrics in Figures 1–3 seem to consistently tell the same story. I would therefore suggest showing only one of them in the main paper and using the freed-up space to present more of the important comparisons currently relegated to the appendix.
  • W8 Random acquisition baseline A random baseline could help contextualize recovery performance, especially for difficult docking landscapes.
Questions for authors
  • Q1 AI-based docking Why are modern AI-based docking methods like DiffDock excluded? Their inclusion would also strengthen the in-scope nature of this paper for an AI conference.
  • Q2 Cross-oracle comparability This is mentioned in the appendix, but should probably be mentioned in the main paper too: the molecule pools between oracles vary somewhat due to different failing complexes. Docking failures are excluded independently for each oracle, resulting in potentially different effective molecular pools. Would the conclusions hold on a shared pool?
  • Q3 Replicates How are the five independent replicates aggregated in the main figures, and how much variability is there between runs?
6 comments

We thank the reviewer for recognizing the scale and value of ChemVerse, its empirical insights, and the soundness of the benchmark. We address the concerns below.

W1. Surrogate architecture and generalizability of the uncertainty findings

We thank the reviewer for raising this point. Holding the neural-network architecture constant was a deliberate experimental design choice, allowing us to systematically isolate the contributions of molecular representations, uncertainty estimation methods, and acquisition strategies without confounding effects from architectural variations. This controlled comparison is central to the scientific rigor of ChemVerse. Our findings demonstrate that MC Dropout, SWAG, and last-layer Laplace yield remarkably similar recovery performance across the evaluated conditions. We do not suggest that these conclusions extend universally to all surrogate architectures or acquisition protocols. Importantly, we have also expanded our experiments to include a Gaussian-process surrogate on HTS-2M, providing an additional comparison beyond neural networks (see response to Reviewer aQvD above). We will clarify the scope of our conclusions while emphasizing the value of these controlled comparisons.

W2. ML contribution and relevance to ICLR

We respectfully but strongly disagree with the assessment that ChemVerse has limited relevance to ICLR. The ICLR 2027 Call for Papers explicitly identifies "Datasets and Benchmarks," "Optimization," "Probabilistic Methods," "Representation Learning," and "Applications to Physical Sciences" as research areas within its scope. ChemVerse lies precisely at the intersection of these areas. The contribution of ChemVerse is not another neural-network architecture, but a systematic, large-scale investigation of fundamental machine-learning questions: Which molecular representations are most effective? How much does uncertainty estimation improve optimization? Which acquisition strategies are most sample-efficient? And how do these conclusions depend on the objective landscape? Addressing these questions across multiple docking oracles, protein targets, and chemical spaces provides insights that cannot be established through isolated optimization experiments. Furthermore, the observation that oracle choice can influence optimization performance more strongly than individual ML components is itself an important finding, rather than a weakness. It highlights the critical dependence of machine-learning performance on the underlying objective landscape and the necessity of benchmarking across diverse conditions. There is substantial precedent for such contributions at leading ML conferences, including Gao et al. (NeurIPS 2022), Kristiadi et al. (ICML 2024), and Xie et al. (ICLR 2026). We therefore respectfully maintain that ChemVerse is firmly within ICLR's stated scope and that its systematic evaluation of representation learning, probabilistic modeling, and sample-efficient optimization constitutes a substantive ML contribution.

Minor weaknesses and suggestions (W3-8)

We thank the reviewer for these suggestions that improve clarity of our paper.

  • W3: We will add an overview figure summarizing the benchmark dimensions and campaign workflow.,
  • W4: We use “Bayesian optimization” consistently in the title and wherever referring to our optimization objective.
  • W5: The main text explicitly refers readers to the appendix for the surrogate architecture and detailed experimental configurations, following standard practice for reporting implementation details. We will make these references more prominent to improve accessibility..
  • W6: The metrics are standard within the field (e.g., Gao et al. (NeurIPS 2022)), but we will provide an explicit definition of the median best-to-worst AUC spread, including the factors held fixed and the comparisons entering the median, and provide the seed-subtracted recovery-AUC formula with its normalization and integration interval in the appendix.
  • W7: Figures 1–3 address distinct methodological questions: molecular representation, uncertainty approximation, and acquisition strategy, respectively. Within each figure, the three recovery thresholds reflect different downstream screening priorities.
  • W8: We have added several additional baseline experiments. For details, please refer to our response to Reviewer aQvD above.

Q1. Inclusion of AI-based docking methods

We appreciate the suggestion, but an important technical distinction is necessary. DiffDock primarily predicts protein-ligand binding poses, and its confidence score reflects confidence in the predicted pose rather than binding affinity. Consequently, DiffDock alone does not provide the scoring oracle required for our Bayesian optimization benchmark. To directly address the reviewer's suggestion, we performed additional experiments using DiffDock-generated poses on LSD-10K, followed by rescoring with Vina, Vinardo, and Glide. Importantly, the Bayesian optimization results with Vina rescoring closely resemble those obtained using conventional Vina docking, preserving the principal recovery trends and molecular representation rankings (see Figure). These observations are consistent with the independent findings of Vu et al. (J. Chem. Inf. Model., 2025), who systematically evaluated DiffDock-L and Vina pose generation with different scoring functions across 43 protein targets. Their study demonstrated that the choice of scoring function can have a greater influence on virtual-screening performance than the pose-generation method. We will include our additional results in the revised manuscript.

In addition to these results, the submitted manuscript already includes an AI-based affinity oracle: Boltz-2 co-folding and affinity prediction on LSD-10K, described in Sec 3.3 and Appendix A.1.6. Exhaustive co-folding at the million-molecule scale is computationally prohibitive, motivating this smaller-scale evaluation. We will discuss these AI-based docking oracles more explicitly in the main text.

Q2. Comparability on a shared molecular pool

We quantified the valid fraction of each library, i.e. the molecules that yield valid scores from all docking oracles for each receptor. These “failure filters” are applied in all benchmark experiments.

ReceptorHTS-2MLSD-5M with DOCK3LSD-5M without DOCK3
5HT2A92.93%32.92%93.22%
AmpC91.78%95.27%95.27%
D490.01%65.10%91.41%
EP4R76.50%56.94%83.05%
SERT91.39%68.57%94.19%
Sigma275.33%66.73%84.40%
Mpro85.29%59.03%91.94%

The four common oracles share substantial portions of both libraries: 75.33–92.93% for HTS-2M and 83.05–95.27% for LSD-5M. Our comparisons of featurizers, BNN approximations, and acquisition functions are conducted on identical eligible pools for each oracle. Thus, differences in pool membership between oracles do not confound these within-oracle comparisons. In the main text, we acknowledge this limitation when interpreting differences across docking oracles.

Q3. Aggregation and variability across replicates.

For each experimental condition, we compute seed-subtracted recovery AUC separately for five independent runs and average these within each receptor. Each bar then averages the seven receptor means with equal weight. Error bars indicate ±1 standard error across receptor means.

To quantify run-to-run variability separately, we calculate the sample standard deviation of the five AUCs within each condition. For the featurizer comparisons using Laplace neural networks:

LibraryMedian run-to-run SD (AUC percentage points)Interquartile range
HTS-2M1.040.59–2.07
LSD-5M0.640.37–1.12
Enamine-10M0.710.39–1.38
LSD-10K3.302.03–5.36

These summaries span receptors, docking programs, featurizers, and the displayed top-k thresholds. Run-to-run variability is typically around one AUC percentage point or less on the large libraries and higher on LSD-10K.

We appreciate the reviewer's recognition of ChemVerse as a sound and valuable benchmark. Given that the principal reservations concern its ML significance and relevance to ICLR, we hope our clarifications demonstrate that ChemVerse directly addresses fundamental questions in probabilistic modeling, representation learning, and sample-efficient optimization, all explicitly within ICLR's scope. Together with the additional experimental evidence, we respectfully ask the reviewer to reconsider the overall assessment. If concerns remain, we would welcome further clarification on what aspects of the ML contribution are considered insufficient to warrant a more favorable recommendation.

Thank you for the detailed response and additional experiments.

I acknowledge your point regarding ICLR's scope and agree that systematic empirical benchmarks can constitute valuable ML contributions without introducing new algorithms. My original characterization of the work as outside ICLR's scope was therefore too strong. My remaining concern is primarily about the significance and generalizability of the ML insights.

Regarding the surrogate architecture, I agree that keeping it fixed enables controlled comparisons between representations. However, different representations may benefit from different architectures or hyperparameter configurations. Consequently, the results establish their relative performance under the chosen surrogate configuration rather than their general suitability for BO. The additional Gaussian-process experiments are a welcome step toward addressing this limitation.

I also appreciate the clarifications regarding DiffDock, cross-oracle comparability, and replicate variability.

Finally, while the proposed revisions appear to address several of my minor comments, the manuscript has not yet been updated, so I cannot assess their implementation. Could you please upload the revised manuscript, ideally accompanied by a version with the changes highlighted as a supplementary file?

Reviewer_aQvDOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper creates an enormous dataset of docking scores and fits surrogate models to them to study black box optimization algorithms, finding that the biggest axis of variability is the docking oracle itself (i.e. the objective function).

Critical strengths

The main strength of the paper is the dataset: as far as I am aware it is the largest docking dataset in its class.

Critical weaknesses

Data availability: the dataset does not seem to be available, and the paper does not specify whether it will be made available. Without the dataset I think this paper loses its biggest contribution (see other weaknesses below).

Data quality: to reach a large data size the authors seemed to have compromised docking quality. For example, Vina's exhaustiveness is set to 1 (down from a default of 8), only one stereoisomer was tried for Glide, etc. All this is in the context of docking data already being a pretty poor approximation of binding (i.e. very "noisy"). The paper doesn't explicitly make any claims about the quality of the docking data compared to ground truth binding affinity, but I think this needs to be a factor in evaluating the value of the dataset. After all, the whole reason to be interested in docking (as opposed to any other complex program that outputs a number for a molecule) is that it mirrors something people care about in domains like drug discovery. So my worry is that this dataset might be too noisy to be really useful.

Limited baselines: the authors mainly test one type of model, with different input representations and uncertainty quantification techniques. I think the paper would highly benefit from the following baselines:

  • Cheap model-free methods like random selection and diversity only selection (like rdkit minmax picker). These methods essentially set the performance floor.
  • GP-based method (for example Tanimoto kernel GP). Arguably all of the BNN approximations tested are cheap ones that tend to underestimate uncertainty. A GP makes stronger modelling assumptions but will likely not underestimate uncertainty so much.
  • Thompson sampling: a BO method with a natural "batch" mode

Scientific gap: model calibration. Unless I missed something, it looks like model performance was only judged via metrics that use the mean prediction, not the uncertainty. It is possible that poor uncertainty calibration (e.g. overconfidence or underconfidence) is driving the results to be, e.g., too greedy-like or too random-like uniformly. It would be good to show some metrics (like Gaussian NLL) that give a sense of how good the calibration is. Poor calibration is not grounds for rejection in my opinion (it might mean that the oracle function is just hard to learn), but it is important to help understand the results.

Questions for authors

Most important question is the dataset release plan.

7 comments

We thank the reviewer for the detailed and thoughtful feedback.

Data availability

The dataset, along with training and evaluation pipelines, will be released as part of a publicly available Python module upon the paper’s acceptance.

Data quality

We thank the reviewer for raising this important point. Vina's computational cost scales approximately linearly with its exhaustiveness parameter. At the scale of millions to billions of molecules, increasing exhaustiveness from 1 to 8 would increase computational requirements approximately eightfold, making comprehensive benchmarking prohibitively expensive. Importantly, we have now performed additional experiments to directly evaluate whether this computational trade-off affects our conclusions. We re-docked the entire 2.1-million-molecule HTS library against AmpC at exhaustiveness 8. Despite the eightfold increase in computational effort, docking scores remained highly correlated (Spearman R = 0.94, Pearson R = 0.95), with 98.6% of the original top-1,000 candidates retained within the top-10,000 at exhaustiveness 8. Crucially, repeating the Bayesian optimization campaigns with both settings changed the mean top-k scores by at most 0.05 kcal/mol. These results demonstrate that our choice of exhaustiveness substantially reduces computational cost without materially affecting the benchmark conclusions. We will include these validation experiments in the revised manuscript

On the topic of stereoisomers, we would like to clarify that Schrödinger’s LigPrep, when used with the -s 1 setting, preserves explicitly specified stereochemistry and limits enumeration only at unspecified stereocenters. We verified this directly by comparing -s 1 and -s 8, obtaining identical docking scores for 497/497 LSD molecules and 487/488 HTS molecules with fully specified stereochemistry. Importantly, 97.3% of LSD molecules and 61.0% of HTS molecules have fully specified stereochemistry. Our analysis indicates that expanded stereoisomer enumeration could affect the docking scores of approximately 1.1% and 20.8% of the respective libraries. Furthermore, this approach is consistent with established ML docking benchmarks: DOCKSTRING assigns unspecified stereocenters randomly but consistently, while MolPAL generates a single 3D structure per SMILES without exhaustive stereoisomer, tautomer, or protonation enumeration. We have included these additional validation results in the appendix and clarified the molecular preparation protocol in the Methods section.

These additional validation results have been included in the appendix, and the molecular preparation protocol has been clarified in the Methods section.

Limited baselines

We agree that additional baselines help contextualize the performance of the evaluated neural surrogates. As suggested by the reviewer, we have extended the HTS-2M experiments to include Gaussian-process surrogates, uniform random acquisition, and diversity-based selection using a MaxMin picker. The accompanying Figure extends Figure 2 with these baselines, enabling direct comparison with MC Dropout, SWAG, and Laplace.

Thompson sampling: Exact Thompson sampling from a GP-type posterior requires sampling from a joint posterior distribution over all candidate molecules. For libraries containing millions of molecules, this is computationally infeasible because memory requirements scale $ \mathcal{O}(n^2) $ and computational costs scale O(n3)\mathcal{O}(n^3) with library size. While approximate sampling methods can reduce these costs, they depend strongly on the surrogate model and the uncertainty quantification technique. For example, approaches developed for GPs cannot be directly applied to Bayesian neural networks, which require different approximations. These differences would introduce additional sources of variation, making it difficult to compare Thompson sampling fairly across the different surrogate models in our benchmark.

Model Calibration

We thank the reviewer for this important suggestion. We have measured calibration across the matrix. The following table reports uncertainty calibration metrics (median over 28 cells, i.e. 4 docking engines x 7 receptors) on the HTS library (UniMol-v2 embeddings; LaplaceNN with Bayesian ridge on identical splits).

n_trainRMSER2mean sigmaNLL95% cov.ridge NLLridge cov.
1,0000.7040.480.1976.90.461.060.95
2,0000.6960.470.13017.30.301.000.95
5,0000.7050.510.06886.90.160.970.96
10,0000.7010.510.045201.70.110.950.96

While predictive accuracy (R2 and RMSE) are largely unaffected by the number of training examples, uncertainty calibration increases strongly with the number of training examples. This suggests that, especially at the later stages of an optimization campaign, acquisition functions that use the predictive sigma converge towards the behavior of fully greedy optimization. This is consistent with our observations, and we will discuss this explicitly in the paper.

We have modified the paper to include the Gaussian NLL, interval coverage and miscalibration area as complementary quantifications of model performance, and have substantially revised the corresponding discussion. We believe that this analysis strengthens the conclusions that can be drawn from our benchmarks.

In light of our commitment to publicly release the ChemVerse dataset, together with the additional docking validation experiments, expanded baselines, and uncertainty calibration analyses, we hope that we have addressed the reviewer's principal concerns and respectfully ask for reconsideration of the rating.

Reviewer_aQvDOpenReview ↗

Follow up questions

Thanks for the detailed rebuttal message and the extra experiments. These results are in the right direction, but I have some follow up questions before potentially revisiting my score.

  1. On the data quality, you reported that "98.6% of the original top-1,000 candidates retained within the top-10,000 at exhaustiveness 8". That's a slightly apples-to-oranges comparison- it says only ~1% of the top 1k got bumped down to rank 10k+. Can you report:

    • Spearman in top 1000 only
    • Overlap between old and new top 1000 (e.g. Jaccard similarity between these sets)
  2. Thompson sampling / GP baselines: I don't accept the argument that this is not comparable or infeasible. The random feature / Matheron method from this paper is basically standard now and allows for sampling in O(n)O(n) time. I recognize there is probably not enough time to run this baseline though.

    Can you also provide more detail about the GP baselines? What kernel, etc?

  3. Calibration: this is an incredibly useful experiment, and to me it confirms that your implementations are poorly calibrated. Unfortunately, most approximate inference methods tend to underestimate uncertainty, and the LaplaceNN seems to be doing this. If the same is true on the other methods, I would interpret this as "approximate inference methods cause the models to be overconfident and not sufficiently explore". This is one of the reasons why a GP baseline would be very helpful: with exact GP inference the posterior uncertainty should not collapse.

    Could you potentially correct this calibration by, for example, adding a "nugget" term (of epistemic uncertainty)? You could fit this with MAP with a strong prior away from 0. I'm sure Claude/other AI can help fill in the details.

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

ChemVerse introduces a valuable benchmark to evaluate machine-leanring-guided VS which contributes to the design and usage of the docking-based VS. It evaluates 6 molecular representations, 3 approaches, and 4 acquisition strategies with 3 large docking score datasets. The paper provided comprehensive results across representation, uncertainty, surrogate models, etc.

Critical strengths
  • the main idea of the work is very practical. Learning which acquisition methods is more effective can contribute to the efficiency of the whole decision process.
  • organization of this paper is clean and clear. Authors provided all necessary information to understand the field and the goal of this research.
Critical weaknesses
  • there are statements without actual result support (i.e. in section 5.3, MC Dropout shows numerical advantage, but difference is hard to assess from Figure 2).
  • figures can have better organization. For example, in figure 4, it's better to have one plot with 6 representations with different labels. The combined plot does not help me understand the result and disucssion.
  • the results are comparing the computational score, which does not completely equal to biological result. Since the main goal is to improve efficiency of drug discovery, it's better to have biological validation in the future.
Questions for authors
  • in 5.4, we can see that top 0.1% AUC across all methods are lower than top 1000. How do recovery performance and method rankings depend on the ratio between the query budget and the size of the target set?
  • more on the size of tested dataset, what is batch size's impact on the evaluation result?
5 comments

We appreciate the reviewer’s constructive comments and hope that the clarifications below address these concerns and justify a higher assessment of our work.

Numerical advantage of MC Dropout (Sec 5.3)

We thank the reviewer for this observation. The central finding of Section 5.3 is not the superiority of MC Dropout, but rather the remarkable similarity in performance across the three uncertainty estimation methods: MC Dropout, SWAG, and last-layer Laplace. After averaging across molecular representations, the largest difference in recovery AUC between these methods is less than one percentage point across the evaluated settings. The reference to MC Dropout's "small numerical advantage" reflects a modest difference in the observed means, rather than a claim of statistically significant superiority. We recognize that these small differences are difficult to discern in Figure 2. To make this comparison more transparent, we provide the numerical recovery AUC values for both molecular libraries in the tables below and will incorporate them into the revised manuscript's appendix. We will also clarify the wording in Section 5.3 to emphasize the consistency of performance across uncertainty estimation methods.

Table 1. HTS-2M: Mean normalized seed-subtracted recovery AUC (%) across six molecular representations and seven receptors. Higher is better; bold indicates the highest observed mean for each docking score and top-k setting.

Docking ScoreTop-kMC DropoutSWAGLaplace
Uni-Dock Vina10063.13963.36963.647
100046.80446.79646.949
0.1%41.69141.64041.778
AutoDock Vina10053.15353.18353.159
100040.19040.23840.139
0.1%35.33435.35135.244
Vinardo10049.85449.51249.770
100036.50936.27236.466
0.1%31.97931.78531.917
Glide10010.68211.55811.437
10007.5797.8767.808
0.1%6.8207.0376.976

Table 2. LSD-5M: Mean normalized seed-subtracted recovery AUC (%) across six molecular representations and seven receptors. Higher is better; bold indicates the highest observed mean for each docking score and top-k setting.

Docking ScoreTop-kMC DropoutSWAGLaplace
Uni-Dock Vina10073.07572.69972.859
100061.63861.30061.430
0.1%49.40449.06849.173
AutoDock Vina10069.64969.07769.157
100058.68058.19158.112
0.1%46.72346.35446.278
Vinardo10062.65562.62462.512
100050.38350.18450.213
0.1%39.55839.38339.398
Glide10018.27018.34118.378
100014.89214.94714.918
0.1%12.08212.09312.055
DOCK310021.96921.56721.708
100020.65020.47820.441
0.1%19.75519.62219.615

Organization of Figure 4

Thank you for this suggestion. The separate panels were intended to show that the relationship between predictive accuracy and recovery holds within individual representations, rather than arising only from differences between representations. However, we agree that the current layout might make the overall comparison difficult to follow. Thus, we have revised the figure so that it distinguishes the six representations.

Docking scores versus biological outcomes

We agree that experimental biological validation is ultimately essential. Importantly, the value of ultra-large virtual screening has already been established through multiple studies demonstrating experimentally validated, potent inhibitors. Shoichet and colleagues showed that expanding a library from 99 million to 1.7 billion molecules doubled hit rates and improved potency (Liu et al., Nat. Chem. Biol., 2025). Other landmark studies, including Lyu et al. (Nature, 2019), VirtualFlow (Gorgulla et al., Nature, 2020), and AdaptiveFlow (Cecchini et al., Nat. Biotechnol., 2026), further demonstrate the experimental success of ultra-large virtual screening. The next challenge is computational scale. With chemical spaces now extending, exhaustive docking becomes prohibitively expensive. ChemVerse addresses the fundamental ML question of which molecular representations, uncertainty estimation, and acquisition strategies most efficiently identify top-ranked candidates across different docking oracles, protein targets, and chemical spaces. Thus, this work focuses on accelerating the computational prioritization step within an experimentally validated drug discovery paradigm. Further biological validation is important, but falls outside the scope of this methodological benchmark.

Dependence on the ratio between query budget and target-set size

We thank the reviewer for this important question. Recovery performance depends on both the query budget and the number of top-ranked molecules targeted for recovery. At a fixed query budget, increasing the target-set size requires identifying more molecules to achieve the same fractional recovery. Consequently, the absolute recovery AUC can differ between top-1000 and top-0.1%, particularly as the latter scales with library size. Importantly, our results demonstrate that the relative performance of acquisition strategies remains consistent across the three evaluated thresholds (top-100, top-1000, and top-0.1%). Greedy and UCB exhibit comparable performance and consistently outperform LogEI and LogPI in mean recovery AUC.These thresholds also reflect different practical screening objectives, from identifying a small set of candidates for experimental testing to generating larger shortlists for subsequent evaluation. We will clarify this distinction and the influence of the query-budget-to-target-set ratio in the revised manuscript.

Impact of acquisition batch size

We thank the reviewer for raising this important point. Batch size introduces a fundamental trade-off between sample efficiency and computational cost. At a fixed query budget, smaller batches enable more frequent surrogate updates, allowing newly acquired information to guide subsequent selections. Larger batches reduce retraining overhead but may introduce redundancy among selected candidates.This behavior has been investigated by Graff et al. (Chem. Sci., 2021), who demonstrated the importance of acquisition batch size and sequential model updates in molecular screening. In ChemVerse, we deliberately maintain a consistent batch-size protocol to isolate the effects of molecular representations, uncertainty estimation methods, and acquisition strategies. Thus, our comparisons remain controlled, although the optimal batch size may depend on library size and computational resources. We will clarify this distinction in the revised manuscript.

We hope these clarifications and additional results address your concerns, and we respectfully ask you to consider increasing your rating in light of the strengthened evidence and presentation.

Dear Reviewer cxfE,

Thank you again for your constructive feedback. Following your suggestions, we have clarified the numerical differences between uncertainty approximations, improved the figures, and expanded the discussion of biological relevance, recovery thresholds, and batch size in the revised manuscript.

As the Author–Reviewer Discussion period draws to a close, we would appreciate your feedback on whether our response and revisions address your concerns. We would be happy to clarify any remaining points.

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.