acceptodds
Under review as a conference paper at ICLR 2027

ReqVS: Reliable Virtual Screening via Interaction Entropy

Abstract

Sequence-based virtual screening has emerged as a scalable approach for drug discovery by enabling compound prioritization directly from protein sequences without requiring 3D structures. However, its practical application remains limited by unreliable predictions on new targets and unexplored chemical space, largely due to the uncertainty introduced when inferring structure-dependent protein–ligand interactions from sequence information alone. Here, interaction uncertainty is quantified using an entropy-based estimator, while an ensemble mechanism integrates complementary information from heterogeneous sequence-based interaction predictors. Based on this, ReqVS is developed as a reliability-aware sequence-based virtual screening framework that prioritizes compounds maximizing predicted interaction scores while minimizing the uncertainty to produce robust compound rankings. Benchmark evaluations further show that ReqVS achieves performance comparable to structure-based methods while improving ranking robustness across targets. When applied to the challenging GPCR target HCAR1, ReqVS identified a high-affinity antagonist, representing the first experimentally validated HCAR1 antagonist discovered using sequence-based virtual screening. These results highlight reliability-aware sequence-based screening as a practical paradigm for structure-independent drug discovery.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Virtual Screen, AI for Science, Drug Discovery, Cheminformatics, Protein-Ligand interaction

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_mdhyOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes ReqVS, a reliability-aware, sequence-based virtual screening framework that integrates heterogeneous drug–target interaction (DTI) and drug–target affinity (DTA) predictors. It combines ensemble scores with a penalty based on averaged per-model binary predictive entropy. The authors report improved early enrichment on DUD-E and DEKOIS2.0 and present a prospective HCAR1 screening campaign involving approximately ten million compounds, followed by expert-guided filtering and cell-based evaluation of cpd60.

The problem is relevant, and the prospective component is potentially valuable. However, I have substantial concerns about the novelty and interpretation of the entropy estimator, the incremental benefit over ensemble scoring, and the pharmacological interpretation of the HCAR1 results.

Critical strengths
  1. A target-sequence-based approach can be useful when experimentally suitable receptor structures or functional states are unavailable.
  2. Attempting laboratory validation after a large-scale virtual screen is an important strength relative to benchmark-only studies.
Critical weaknesses
  1. Ensemble consensus, predictive entropy, and uncertainty-aware screening have substantial precedents, including SCORCH (McGibbon et al., 2023), SCORCH2 (Chen et al., 2025), and Moshawih et al. (2024). These methods differ from ReqVS, particularly in their structural inputs, but the authors should explain the specific algorithmic novelty beyond combining existing sequence-based predictors with an entropy penalty. Comparisons against straightforward consensus-ranking and uncertainty-penalized baselines are needed.

  2. Equations (7)–(8) average the individual models' predictive entropies, 1M∑j=1MH(pj)\frac{1}{M}\sum_{j=1}^{M}H(p_j), rather than measuring compound-specific disagreement between models. For example, predictions of 0.99 and 0.01 produce low average entropy despite strong disagreement. Consensus weighting considers model agreement at the weighting stage but does not resolve this distinction. Please compare the proposed estimator with ensemble predictive entropy, prediction variance, and mutual-information/Jensen–Shannon disagreement, and substantiate its proposed connection to uncertainty from unobserved binding structures.

  3. The incremental benefit of entropy regularization is modest. On DUD-E (Table 3), adding entropy changes mean EF0.1% from 41.64 to 42.65 and BEDROC from 0.454 to 0.459, while AUROC slightly decreases from 0.821 to 0.820. This suggests that ensembling contributes most of the gain. Please provide paired target-level significance tests or confidence intervals, sensitivity analysis for λ=0.2\lambda=0.2, and evidence that entropy meaningfully improves the top-ranked candidate list.

  4. Appendix B.2 uses library-dependent min–max normalization to convert DTA affinity scores into values treated as binding probabilities for Shannon entropy, although such scores are not necessarily calibrated probabilities. The consensus-derived weights may also favor models with shared errors. Please examine probability calibration, normalization and weighting alternatives, and the effect of synthetic-negative augmentation. Given known biases in DUD-E (Chen et al., 2019), stronger scaffold- and protein-homology-separated evaluations would support the generalization claims; comparisons with previously reported structure-based baselines should also be clearly distinguished from uniformly rerun experiments.

  5. The evidence for describing cpd60 as an HCAR1 antagonist is insufficiently clear. HCAR1 is mainly Gi/o-coupled, and receptor activation commonly reduces forskolin-stimulated cAMP. Appendix E does not clearly describe an agonist-challenge experiment using lactate or CHBA, so reduced cAMP alone does not establish antagonism. Please clarify the direction of the reported inhibition, include agonist-challenge and receptor-specificity controls, and report independent biological replicates beyond the three technical replicates. The GloSensor and HTRF IC50 values (14.7 ± 1.2 and 2.5 ± 1.2 μM) indicate assay-dependent functional potency, not directly measured high binding affinity; the phrase “high-affinity antagonist” should be supported or softened. The related HCAR1 screening study by Fan et al. (2026) should also be discussed without conflating its target-specific approach with ReqVS's sequence-only setting.

  6. The prospective HCAR1 results do not isolate the contribution of ReqVS. cpd60 was originally ranked 552nd and became first only after expert filtering for carboxyl and hydroxyl groups. Only one of the two filtered candidates was tested, alongside seven manually selected compounds, and the prospective screen used six rather than eight models. Consequently, the eight-compound rank–inhibition correlation cannot establish an unbiased top-K hit rate. Please report score-only and entropy-regularized rankings under the actual six-model configuration and distinguish the effects of computational ranking from expert selection.

References

  1. McGibbon, M., et al. (2023). SCORCH: Improving structure-based virtual screening with machine learning classifiers, data augmentation, and uncertainty estimation. Journal of Advanced Research, 46, 135–147. https://doi.org/10.1016/j.jare.2022.07.001
  2. Chen, L., et al. (2025). SCORCH2: A Generalized Heterogeneous Consensus Model for High-Enrichment Interaction-Based Virtual Screening. Advanced Science, 12(42), e08318. https://doi.org/10.1002/advs.202508318
  3. Moshawih, S., et al. (2024). Consensus holistic virtual screening for drug discovery: a novel machine learning model approach. Journal of Cheminformatics, 16, 62. https://doi.org/10.1186/s13321-024-00855-8
  4. Chen, L., et al. (2019). Hidden bias in the DUD-E dataset leads to misleading performance of deep learning in structure-based virtual screening. PLOS ONE, 14(8), e0220113. https://doi.org/10.1371/journal.pone.0220113
  5. Lind, S., et al. (2026). Profiling of HCAR1 signalling reveals Gαi/o and Gαs activation without β-arrestin recruitment and the discovery of an allosteric agonist. British Journal of Pharmacology, 183(10), 2219–2235. https://doi.org/10.1111/bph.70312
  6. Fan, M., et al. (2026). HCAR1 antagonist screening based on boundary-selected negative sampling strategy and multi-level graph neural network. Computer Methods and Programs in Biomedicine, 278, 109262. https://doi.org/10.1016/j.cmpb.2026.109262
Questions for authors
  1. What specific new algorithmic principle does Eq. (8) introduce relative to standard ensemble predictive entropy, and what evidence connects it to uncertainty from latent structural configurations?
  2. Are the modest improvements over ReqVS-ES statistically robust? Would ensemble predictive entropy, disagreement-based uncertainty, or simple consensus rankings achieve similar results?
  3. How are the DTA normalization-derived values calibrated as probabilities? How sensitive are entropy scores and top-K rankings to library composition and model weighting?
  4. What agonist-challenge and receptor-specific experiments demonstrate that cpd60 antagonizes HCAR1 instead of activating Gi-associated signaling or interfering with the cAMP assay?
  5. How much of the prospective ranking gain remains without expert-guided filtering, and how does the claim of novelty relate to Fan et al. (2026)?
13 comments
AuthorsOpenReview ↗

Response 1

Thank you for your careful review and constructive feedback. We hope our responses have addressed your concerns and would be happy to clarify any remaining points. We will provide a revised manuscript shortly, incorporating the changes outlined in our responses.

Comment 1:

“W1: Ensemble consensus, predictive entropy, and uncertainty-aware screening have substantial precedents ... Comparisons against straightforward consensus-ranking and uncertainty-penalized baselines are needed.” & “Q1: What specific new algorithmic principle does Eq. (8) introduce relative to standard ensemble predictive entropy, and what evidence connects it to uncertainty from latent structural configurations?”

Re: We thank the reviewer for drawing our attention to these related studies. SCORCH combines predictors based on structural features and uses the standard deviation of model outputs to assess confidence. SCORCH2 further introduces weighted consensus scoring between two XGBoost models based on three-dimensional interaction features. Moshawih et al. integrate multiple screening scores and determine their weights based on training and validation performance. These methods require protein structures and pocket information, as well as molecular docking, which incurs additional computational time. These studies have advanced ensemble frameworks, and we will add a discussion of them to the Related Work section in the revised manuscript.

We would like to respectfully clarify that our methodological contributions include both the formulation of interaction entropy for sequence-based prediction and the unified integration of heterogeneous DTI and DTA models, as detailed below:

  • Interaction entropy. The interaction entropy is tailored to sequence-based models and motivated by an analysis of their underlying inference process. During inference, these models do not explicitly receive the three-dimensional complex conformation, binding pose, or intermolecular contact information for a given protein–ligand pair. These structural interactions are crucial to understanding molecular recognition and protein function [1], and predicting them is a central objective of approaches such as AlphaFold 3. We use mutual information to quantify the difference in predictive uncertainty with and without access to this interaction information, yielding a decomposition of predictive uncertainty into sequence-induced interaction uncertainty and intrinsic noise. As shown in Eq. (6), sequence-based models must implicitly account for possible structural interactions, giving rise to uncertainty. Based on this formulation, ReqVS uses Shannon entropy as a practical surrogate for this uncertainty (Eq. (8)).

  • Unified integration of DTI and DTA models. Our framework and consensus scoring scheme integrate heterogeneous DTI and DTA models independently of their underlying architectures. These model categories differ in both training objectives and output spaces: DTI models predict interaction propensity, whereas DTA models estimate continuous binding affinity. ReqVS brings their outputs onto a common numerical scale and combines information from both prediction tasks to improve candidate prioritization.

To clearly illustrate the robust gains in virtual screening performance provided by interaction entropy,we evaluated 53 combinations on DUD-E containing at least two DTI and two DTA models selected from the eight constituent models. ReqVS achieved higher EF0.1% than ReqVS-ES in 49 of these combinations. The balanced configurations—2 + 2 , 3 + 3 , and 4 + 4 —contain 36, 16, and 1 combinations, respectively.

DTI + DTACombinationsMedian EF0.1%Median EF0.5%Median EF1%Median BEDROCHigher EF0.1% combinations
2 + 23631.26 → 31.4225.21 → 25.8821.57 → 22.140.3568 → 0.363432/36
3 + 31636.74 → 37.9330.63 → 31.3125.66 → 26.210.4161 → 0.422016/16
4 + 4141.64 → 42.6533.51 → 34.2928.16 → 28.580.4540 → 0.45861/1

[1]Salentin S, Schreiber S, Haupt V J, et al. PLIP: fully automated protein–ligand interaction profiler[J]. Nucleic acids research, 2015, 43(W1): W443-W447.

AuthorsOpenReview ↗

Response 2

Comment 2:

“W2: Equations (7)–(8) average the individual models' predictive entropies ... rather than measuring compound-specific disagreement ... Please compare the proposed estimator with ensemble predictive entropy, prediction variance, and mutual-information/Jensen–Shannon disagreement...” & “Q2: Are the modest improvements over ReqVS-ES statistically robust? Would ensemble predictive entropy, disagreement-based uncertainty, or simple consensus rankings achieve similar results?”

**Re:**To better address the reviewer's concerns, we evaluated alternative ranking and uncertainty-penalty strategies. Keeping the remaining screening settings fixed, we compared the proposed interaction-entropy penalty with no penalty, consensus ranking (rank averaging), and penalties based on Jensen–Shannon divergence (JSD), prediction variance, and absolute deviation. In these experiments, consensus ranking and disagreement-based penalties reduced screening performance, whereas ReqVS's interaction-entropy penalty consistently improved the reported screening metrics.

MethodDUD-E EF0.1%DUD-E EF0.5%DUD-E EF1%DUD-E BEDROCDEKOIS2.0 EF0.1%DEKOIS2.0 EF0.5%DEKOIS2.0 EF1%DEKOIS2.0 BEDROC
No penalty41.9233.7528.110.454317.2218.1818.120.5272
Consensus ranking27.4222.1618.070.307213.7813.4012.180.3626
JSD40.2832.9227.340.445317.2217.5417.610.5164
Variance41.5333.4727.930.451416.8418.1217.920.5234
Absolute deviation41.2933.1527.480.447517.2218.1217.670.5207
ReqVS42.6534.2928.580.458618.3718.6918.690.5331

We would like to clarify that interaction entropy is not intended to penalize disagreement between models. Instead, it penalizes the uncertainty of each individual model’s prediction for a given protein–ligand pair, as defined in Eqs. (7)–(8) (L201–215). Intuitively, for predictions with the same ensemble mean, interaction entropy penalizes ambiguous individual predictions near 0.5 while favoring more confident predictions of activity or inactivity near 1 or 0, respectively.

We will add results and discussion of interaction entropy to the appendix in the revised manuscript.

AuthorsOpenReview ↗

Response 3

Comment 3:

“W3: The incremental benefit of entropy regularization is modest ... while AUROC slightly decreases ... Please provide paired target-level significance tests or confidence intervals, sensitivity analysis ... and evidence that entropy meaningfully improves the top-ranked candidate list.”

Re: We thank the reviewer for raising this concern about detail. We want to respectfully clarify that the 0.001 decrease in AUROC more closely reflects global ranking performance, while virtual screening aims to filter a small number of molecules from a large chemical library. In this scenario, the EF and BEDROC (raise EF0.1% by 1.01 and BEDROC by 0.005 in DUD-E) can more directly point out the improvement in the early recognition of ReqVS [1].

To better illustrate the contribution of interaction entropy, we first performed paired target-level comparisons between ReqVS and ReqVS-ES and estimated 95% confidence intervals using 50,000 paired bootstrap resamples of targets. The mean EF0.1% gains were 1.007 [95% CI: 0.217, 1.868] on DUD-E and 1.531 [0.383, 3.062] on DEKOIS2, with maximum target-level gains of 18.76 and 31.00, respectively.

Benchmark (targets)ReqVS-ES EF0.1%ReqVS EF0.1%Mean gain [95% CI]Maximum target-level gain
DUD-E (102)41.6442.65+1.01 [0.22, 1.87]+18.76
DEKOIS2 (81)16.8418.37+1.53 [0.38, 3.06]+31.00

Second, we added a sensitivity analysis for λ\lambda and compared the evaluation metrics. For λ≤0.8\lambda \leq 0.8, interaction entropy can robustly improve EF0.1% and EF0.5%, indicating that the gains are not restricted to our chosen value.

λDUD-E EF0.1%DUD-E EF0.5%DUD-E EF1%DUD-E BEDROCDEKOIS2.0 EF0.1%DEKOIS2.0 EF0.5%DEKOIS2.0 EF1%DEKOIS2.0 BEDROC
041.9233.7528.100.454217.2218.1718.110.5272
0.0542.1233.8028.180.455717.9818.4318.240.5296
0.142.1234.0028.370.456918.3718.4318.330.5313
0.242.6534.2828.570.458618.3718.6918.680.5330
0.442.5134.5028.630.458217.6019.1418.620.5322
0.842.6434.3228.000.434718.3719.4518.590.5157
140.9531.3122.850.329417.2219.1917.470.4613

Finally, to quantify the benefit of interaction entropy for the top-ranked candidate lists, we held other settings fixed, comparing only the absence or presence of the entropy penalty (λ=0\lambda = 0 versus 0.20.2). For each target, we selected the same number of top-ranked candidates under both settings and counted the known actives retrieved. These improvements were obtained without increasing the number of selected candidates, showing better recovery of known actives within the same selection budget with the proposed interaction entropy.

BenchmarkTotal candidates / actives / selected candidatesPer-target budgetActives retrieved without entropyActives retrieved with entropy
DUD-E1,434,015 / 22,805 / 1,381Top-0.1%1,0331,055
DUD-E1,434,015 / 22,805 / 14,290Top-1%7,5647,723
DEKOIS2100,408 / 3,239 / 486Top-0.5%285293
DEKOIS2100,408 / 3,239 / 972Top-1%568586

[1]Truchon J F, Bayly C I. Evaluating virtual screening methods: good and bad metrics for the “early recognition” problem[J]. Journal of chemical information and modeling, 2007, 47(2): 488-508.

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

This paper presents ReqVS, an ensemble of protein sequence-based interaction classifiers and affinity regression models that adjusts compound rankings using a predictive entropy penalty. The method is evaluated on DUD-E and DEKOIS2.0 and applied to large-scale screening for HCAR1. Following expert filtering and candidate selection, the authors test compounds including cpd60 in cellular cAMP assays and use docking and molecular dynamics simulations to explore possible mechanisms of action.

Critical strengths

The paper addresses a practically relevant problem. Improving early enrichment in large-scale virtual screening without requiring a target’s 3D structure as model input is valuable for targets with limited structural information.

The ensemble achieves promising screening performance. ReqVS outperforms its individual component models on several early enrichment metrics across both benchmarks, supporting the practical value of combining their predictions.

Critical weaknesses

The entropy term provides only modest gains, with no evidence that they exceed statistical variation. On DUD-E, adding entropy to the same ensemble increases EF0.1% from 41.64 to 42.65, EF1% from 28.16 to 28.58, and BEDROC from 0.454 to 0.459, while AUROC decreases from 0.821 to 0.820. These small differences could reflect statistical variation. Most of the improvement over individual models is already achieved by the ensemble without entropy, so the overall gains cannot be attributed primarily to the proposed interaction entropy.

Key virtual screening baselines are missing. DrugCLIP and S²Drug directly address active compound retrieval and should be included, particularly S²Drug, which also uses protein sequence information. The current results do not establish an advantage over these methods. Comparisons under matched conditions are needed to assess both screening performance and computational efficiency.

The experiments do not establish target-specific antagonism, and manual selection complicates interpretation of screening success. cpd60 ranked 552nd before expert functional-group filtering, and the other seven tested compounds were manually selected. Their results therefore cannot be treated as the hit rate of the model’s ranking. The evidence mainly comes from cellular cAMP assays, without direct binding measurements, receptor-dependence controls, or a clear demonstration of blockade of a known agonist. For the Gi-coupled receptor HCAR1, the reported inhibition of forskolin-induced cAMP also needs to be reconciled with the antagonist claim. A cellular IC50 does not establish high binding affinity, and docking or molecular dynamics cannot substitute for experimental confirmation of antagonism.

Questions for authors

see the weaknesses

6 comments
AuthorsOpenReview ↗

Response 1

Thank you for your insightful comments and feedback, which have helped us improve our paper. Please find our responses below. Please let us know if any additional clarification is needed from us. We will provide a revised manuscript shortly, incorporating the corresponding changes described in our responses.

Comment 1:

“W1: The entropy term provides only modest gains, with no evidence that they exceed statistical variation ...On DUD-E, adding entropy to the same ensemble increases EF0.1%... while AUROC decreases...Most of the improvement ... achieved by the ensemble without entropy, so the overall gains cannot be attributed primarily to the proposed interaction entropy.”

Re: We sincerely appreciate the reviewer’s insightful comment. To assess the consistency of the entropy's benefits across model compositions, we evaluated 53 combinations on DUD-E containing at least two DTI and two DTA models selected from the eight constituent models. ReqVS (using entropy) achieved higher EF0.1% than ReqVS-ES in 49 of these combinations. The balanced configurations—2 + 2 , 3 + 3 , and 4 + 4 —contain 36, 16, and 1 combinations, respectively.

DTI + DTACombinationsMedian EF0.1%Median EF0.5%Median EF1%Median BEDROCHigher EF0.1% combinations
2 + 23631.26 → 31.4225.21 → 25.8821.57 → 22.140.3568 → 0.363432/36
3 + 31636.74 → 37.9330.63 → 31.3125.66 → 26.210.4161 → 0.422016/16
4 + 4141.64 → 42.6533.51 → 34.2928.16 → 28.580.4540 → 0.45861/1

Values are shown as ReqVS-ES → ReqVS.

We would like to respectfully clarify that virtual screening aims to select a small number of compounds from a large library for experimental validation. Therefore, modest improvements in early enrichment can be practically meaningful by lowering the expected cost per hit. In this scenario, AUROC measures discrimination across the entire library, whereas EF and BEDROC correspond more directly to the early-recognition objective [1].

At the same time, for the concern about the contribution of interaction entropy to the overall performance gains. We would like to clarify that our methodological contributions include both the formulation of interaction entropy for sequence-based prediction and the unified integration of heterogeneous DTI and DTA models, as detailed below:

  • Interaction entropy. Our formulation is tailored to sequence-based models and motivated by an analysis of their underlying inference process. In our latent structural formulation, sequence-based prediction must implicitly account for the possible structural configurations, giving rise to interaction uncertainty (Eqs. (3)–(5)). Interaction entropy provides a practical surrogate for incorporating this uncertainty into screening.

  • Unified integration of DTI and DTA models. Our framework and consensus scoring scheme integrate heterogeneous DTI and DTA models independently of their underlying architectures. These model categories differ in both training objectives and output spaces: DTI models predict interaction propensity, whereas DTA models estimate continuous binding affinity. ReqVS brings their outputs onto a common numerical scale and combines information from both prediction tasks to improve candidate prioritization.

We hope that the additional results and explanations help clarify how these two aspects work together to improve early enrichment.

[1]Truchon J F, Bayly C I. Evaluating virtual screening methods: good and bad metrics for the “early recognition” problem[J]. Journal of chemical information and modeling, 2007, 47(2): 488-508.

AuthorsOpenReview ↗

Response 2

Comment 2:

“W2: Key virtual screening baselines are missing... Comparisons under matched conditions are needed...”

Re: We thank the reviewer for highlighting these relevant baselines. DrugCLIP and S²Drug formulate virtual screening as a retrieval task by aligning pocket and ligand representations in a shared embedding space. S²Drug further incorporates protein-sequence pretraining to improve screening performance, while its virtual screening model still requires three-dimensional pose and pocket information. We therefore reran DrugCLIP and report the additional comparisons below. The S²Drug does not provide pre-trained checkpoints, training data, or inference code, making reproduction infeasible at this time.

To provide a fair comparison under matched conditions, we reran DrugCLIP[1] on DUD-E under the following settings. We evaluated four settings:

  • Target HOLO structure available: experimental holo structures with known pockets (Holo),
  • Only sequence information available: AF2-predicted structures with reference-defined pockets (AF2+oracle), and AF2 structures with Fpocket candidates selected either by maximum overlap with the reference pocket (AF2+IoU) or by the highest Fpocket Pocket Score (AF2+pred).

To assess computational efficiency, we selected HDAC2, a DUD-E target close to the benchmark medians in both protein-sequence length and candidate-library size. All runtime measurements were performed on a system equipped with a single NVIDIA GeForce RTX 4090 GPU.

ScenarioMethod / settingEF0.1%EF0.5%EF1%BEDROCTime/s
Known experimental structureDrugCLIP Holo37.4932.0226.400.426040.5
No experimental structureDrugCLIP AF2+oracle31.1126.7122.300.3698251.7
No experimental structureDrugCLIP AF2+IoU19.2116.6813.710.2344254.1
No experimental structureDrugCLIP AF2+pred8.697.386.040.1046255.6
No experimental structureReqVS42.6534.2928.580.4586184.8

Notably, obtaining an experimental structure beforehand can require substantial time, which is not included in the reported inference time for the “Known experimental structure” scenario.

ReqVS achieved strong early enrichment without requiring target structures or pocket annotations. At the same time, DrugCLIP performance decreased when experimental structures were unavailable and pocket information remained unknown.

We will add these results and additional discussion on these baselines in the Appendix of the revised manuscript.

[1]Jia Y, Gao B, Tan J, et al. Deep contrastive learning enables genome-wide virtual screening. Science 391, eads9530[EB/OL].(2026)

AuthorsOpenReview ↗

Response 3

Comment 3:

“W3: The experiments do not establish target-specific antagonism, and manual selection complicates interpretation of screening success ... The evidence mainly comes from cellular cAMP assays, without direct binding measurements, receptor-dependence controls, or a clear demonstration of blockade of a known agonist ... A cellular IC50 does not establish high binding affinity, and docking or molecular dynamics cannot substitute for experimental confirmation of antagonism.”

Re: We thank the reviewer for this important comment. Both the GloSensor and HTRF dose-response assays were conducted in HCAR1-overexpressing HEK293 cells stimulated with 20 mM sodium lactate, a known HCAR1 agonist. Forskolin was used to elevate intracellular cAMP, whereas sodium-lactate activation of Gi-coupled HCAR1 suppressed this response. The concentration-dependent activity of cpd60 therefore reflects blockade of agonist-induced HCAR1 signaling rather than direct inhibition of forskolin-induced cAMP accumulation. The consistent responses across two mechanistically distinct cAMP platforms further reduce the likelihood of a platform-specific readout. We will supplement the Methods with the sodium lactate challenge and clarify the assay procedure in the revised manuscript.

Although the cellular IC50 is an operational measure of functional inhibitory potency rather than direct binding affinity [1], functional IC50 values have also been used to characterize HCAR1 inhibitors in previous studies. For example, He et al. identified reserpine as an HCAR1 inhibitor using a lactate-challenged GloSensor assay and reported a functional IC50 of 3.3 ± 1.4 μM [2]. Accordingly, the reproducibility of cpd60 across two assay platforms, together with the docking and molecular-dynamics results, supports its potential as an HCAR1 antagonist candidate without constituting definitive proof of direct binding.

For the question about expert filtering, we would like to respectfully clarify that expert-guided selection is commonly used in prospective virtual screening [3–5]. It incorporates additional information beyond interaction scores, including target-specific functional-group considerations, physicochemical properties, compound availability, and synthetic feasibility. These considerations complement computational rankings and help identify candidates that can be obtained and evaluated experimentally.

[1]Neubig, R. R., Spedding, M., Kenakin, T. & Christopoulos, A. International union of pharmacology committee on receptor nomenclature and drug classification. XXXVIII. Update on terms and symbols in quantitative pharmacology. Pharmacological Reviews 55, 597–606 (2003).

[2]He, J. et al. The lactate receptor HCAR1 drives the recruitment of immunosuppressive PMN-MDSCs in colorectal cancer. Nat Immunol 26, 391–403 (2025).

[3]Jia Y, Gao B, Tan J, et al. Deep contrastive learning enables genome-wide virtual screening. Science 391, eads9530[EB/OL].(2026)

[4]Lyu J, Kapolka N, Gumpper R, et al. AlphaFold2 structures guide prospective ligand discovery[J]. Science, 2024, 384(6702): eadn6354.

[5]Zhou G, Rusnac D V, Park H, et al. An artificial intelligence accelerated virtual screening platform for drug discovery[J]. Nature Communications, 2024, 15(1): 7761.

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

ReqVS is a reliability-aware framework for sequence-based virtual screening. Its stated premise is that a sequence-only predictor must implicitly marginalize over unobserved protein-ligand structural configurations, so the residual uncertainty can be written as a conditional mutual information between the interaction outcome and the latent structure, and approximated by the Shannon entropy of the predictive distribution. The screening objective subtracts a weighted entropy term from the predicted interaction score, with a trade-off coefficient set to 0.2. The implementation ensembles eight heterogeneous sequence models, four drug-target interaction classifiers and four drug-target affinity regressors, all retrained from scratch on a curated ChEMBL v36 set with benchmark ligands and proteins removed and with a cross-cluster negative-augmentation scheme. Drug-target interaction outputs are used directly as probabilities and affinity outputs are min-max normalized across the screening library. Per-model binary entropies are computed from these probabilities, model weights are estimated without labels by comparing each model against the ensemble mean using pseudo precision, recall and F1, and scores and entropies are then aggregated with category-level weights before ranking. Benchmarks cover all 102 DUD-E targets and all 81 DEKOIS2.0 targets, with enrichment factors, AUROC, BEDROC and average rank of top actives, compared against eight constituent sequence models, framework ablations, and a set of structure-based baselines. A prospective campaign screens a ten-million-compound commercial library against HCAR1, applies an expert carboxyl-plus-hydroxyl filter to the top 1,000, and tests one resulting candidate plus seven manually chosen compounds in cAMP assays; the top candidate shows micromolar activity, followed by docking, molecular dynamics and MM/PBSA analysis.

Critical strengths
  1. Benchmark coverage is complete rather than selective: all 102 DUD-E and all 81 DEKOIS2.0 targets, with five metrics and per-target heatmaps (L256-269, L1080-1187).
  2. All eight constituent sequence models are retrained from scratch on one curated corpus, with benchmark ligands and proteins removed before training, which removes 22.48 percent of classification and 19.63 percent of regression records (L702-712).
  3. The prospective campaign is real wet-lab validation on a target with no antagonist structures, confirmed in two orthogonal cAMP assays (L446-452), which is rare in this literature.
  4. The authors disclose that the deployed screen used a six-model instantiation while the benchmarks use eight, and report the retrospective audit rather than quietly presenting the eight-model system (L1214-1216).
Critical weaknesses

1.The entropy term does not measure uncertainty. Since scores and entropies are aggregated with the same category-level weights (L918-940), the final objective equals a weighted average of g(p) = p - 0.2 H(p) over models, and g is strictly increasing in p above about 0.007, so the method is an ensemble of fixed monotone recalibrations of each model's probability, not a penalty on disagreement. 2.Below that crossover the objective is decreasing in p, so among very low-scoring compounds a confident non-binder outranks a borderline one, which is the opposite of the intended behaviour. 3.The ablations are consistent with this: entropy adds about one point over the score-only ensemble (DUD-E EF 0.1 percent 42.65 against 41.64, DEKOIS EF 0.5 percent 18.69 against 18.24) and lowers AUROC on both benchmarks, 0.820 against 0.821 and 0.825 against 0.828 (L1008-1044). 4.The error bound is vacuous: the surrogate error is shown to equal the conditional entropy of the outcome given structure and is then bounded by one bit (L672-694), but the surrogate itself is a binary entropy bounded by one bit, so the bound says nothing; the equations also mix natural-log entropy with a base-two bound. 5.Affinity outputs become probabilities by min-max normalization over the screening library (L222-226, L852-863), so a compound's reported uncertainty depends on which other compounds are screened alongside it, and the quantity has no probabilistic meaning. 6.The consensus weighting is circular and then discarded: pseudo precision, recall and F1 are computed against the ensemble mean (L884-900), rewarding conformity rather than complementarity, and the weight table then collapses the per-model scores into two category-level averages so no per-model reliability estimate survives. 7.Structure-based baselines are taken from a different paper (L254-262) while the paper's own metrics are computed without considering multiple poses (L300-303), so the headline comparison with Glide, EquiScore and the rest is across protocols rather than matched. 8.The prospective evidence is thinner than presented: the only ReqVS-selected tested compound ranked 552nd raw and reached 53.29 percent inhibition, but manually chosen compounds at raw ranks 1.88 million and 4.58 million reached 35.71 and 35.26 percent (L1228-1241), and the reported correlation rests on eight points with two leverage outliers.

Questions for authors
  1. Given that scores and entropies share category-level weights (L918-940), do you agree that the objective reduces to a weighted ensemble of a fixed monotone transform of each model's probability? If so, how should interaction entropy be interpreted, and what would a genuine disagreement-based measure, such as the entropy of the ensemble mean minus the mean entropy, give instead?
  2. The ablation tables show entropy adding about one point of enrichment while lowering AUROC on both benchmarks (L1008-1044). What is the sensitivity of every metric to the trade-off coefficient across a range, given that 0.2 was chosen on empirical knowledge (L912-917)?
  3. DUD-E is known for decoy bias that a ligand-only model can exploit. What does a protein-blind control reach, that is, each constituent model and ReqVS with the target sequence removed or randomized? Without this, an EF 0.1 percent of 42.65 is hard to attribute to interaction modelling.
  4. Were the structure-based numbers recomputed under your single-pose, no-docking protocol, or taken as reported (L254-262, L300-303)? A matched-protocol rerun on even a subset of targets would make the competitive claim interpretable.
  5. Since affinity outputs are min-max normalized over the library (L852-863), how do the reported entropies and rankings change when the same target is screened against libraries of different size and composition?
  6. A model that simply reproduced the ensemble mean would score perfectly under the pseudo precision, recall and F1 of the weighting step (L884-900). How do the resulting weights differ from uniform, and what happens to the benchmark results under uniform weighting?
  7. Of the eight compounds tested against HCAR1, only one came from the ReqVS-plus-filter pipeline; how were the seven manually selected compounds chosen, and what would the inhibition rates look like for, say, twenty compounds drawn at matched intervals down the ranked list?
  8. Under the eight-model configuration used for all benchmarks, cpd60 falls to the top 34 by filtered rank (L1214-1216). Would the benchmarked system have selected it, and what does the eight-model top-two look like?
9 comments
AuthorsOpenReview ↗

Response 1

Comment 1:

“W1: The entropy term does not measure uncertainty ... the method is an ensemble of fixed monotone recalibrations of each model's probability, not a penalty on disagreement.” & “Q1: ... how should interaction entropy be interpreted, and what would a genuine disagreement-based measure, such as the entropy of the ensemble mean minus the mean entropy, give instead?”

Re: We would like to respectfully clarify that interaction entropy is not intended to penalize disagreement between models. Different from disagreement-based measures such as JSD, it penalizes the uncertainty of each model’s prediction for a given protein–ligand pair, as defined in Eqs. (7)–(8) (L201–215). Intuitively, for predictions with the same ensemble mean, interaction entropy penalizes ambiguous individual predictions near 0.5 while favoring more confident predictions of activity or inactivity near 1 or 0, respectively. By ensembling the uncertainty (interaction entropy) of heterogeneous models, we can imitate the uncertainty of missing structure from different perspectives. We agree with your observation regarding the mathematical form of the objective. At the same time, we would like to emphasize that our formulation is tailored to sequence-based models and motivated by an analysis of their underlying inference process. During inference, these models do not explicitly receive the 3D complex conformation or intermolecular contact information for a given protein–ligand pair—information that approaches such as AlphaFold 3 aim to predict. These structural interactions are crucial to understanding molecular recognition and protein function [1]. Sequence-based models must therefore infer the relevant interaction information implicitly, giving rise to uncertainty (Eq. [3] and Eq.[5]).

To address the reviewer’s question, we added controlled comparisons of alternative penalty terms, together with a consensus-ranking baseline. These include a no-penalty control and four disagreement-based alternatives: Consensus ranking (uniform averaging of per-model percentile ranks), Jensen–Shannon divergence (JSD), defined as the entropy of the ensemble mean minus the mean entropy; prediction variance; and mean absolute deviation. In these experiments, the disagreement-based penalties generally reduced screening performance relative to the no-penalty control, whereas the proposed interaction entropy consistently improved performance across all reported metrics.

MethodDUD-E EF0.1%DUD-E EF0.5%DUD-E EF1%DUD-E BEDROCDEKOIS2.0 EF0.1%DEKOIS2.0 EF0.5%DEKOIS2.0 EF1%DEKOIS2.0 BEDROC
No penalty41.9233.7528.110.454317.2218.1818.120.5272
Consensus ranking27.4222.1618.070.307213.7813.4012.180.3626
JSD40.2832.9227.340.445317.2217.5417.610.5164
Variance41.5333.4727.930.451416.8418.1217.920.5234
Absolute deviation41.2933.1527.480.447517.2218.1217.670.5207
ReqVS42.6534.2928.580.458618.3718.6918.690.5331

We will added these comparisons and further clarification of the interpretation of interaction entropy in Appendix.

[1]Salentin S, Schreiber S, Haupt V J, et al. PLIP: fully automated protein–ligand interaction profiler[J]. Nucleic acids research, 2015, 43(W1): W443-W447.

Comment 2:

“W2.Below that crossover the objective is decreasing in p, so among very low-scoring compounds a confident non-binder outranks a borderline one, which is the opposite of the intended behaviour. "

Re: We would like to respectfully clarify that ReqVS is designed to prioritize a small number of compounds for experimental validation from a large screening library. The key practical question is therefore whether the decreasing region affects top-K selection under a limited experimental budget.

To examine this, we kept the screening pipeline fixed and flattened the decreasing low-probability branch by replacing each model's contribution g(p)g(p) with g(max⁡{p,p∗})g(\max\{p,p_\ast\}). This makes the contribution function non-decreasing while leaving the branch above the threshold unchanged. The top-0.1% candidate lists remained unchanged for 101 of the 102 DUD-E targets, with only one selected compound replaced for the remaining target. The top-0.1% lists remained unchanged for all 81 DEKOIS2.0 targets. These results indicate that this boundary behavior has little impact on early screening results under a limited experimental budget.

AuthorsOpenReview ↗

Response 2

Comment 3:

“W3: ... entropy adds about one point over the score-only ensemble ... and lowers AUROC on both benchmarks ...” & “Q2: ... What is the sensitivity of every metric to the trade-off coefficient across a range, given that 0.2 was chosen on empirical knowledge ...?”

Re: We have added a sensitivity analysis for λ\lambda and compared the evaluation metrics. For λ≤0.8\lambda \leq 0.8, interaction entropy can robustly improve EF0.1% and EF0.5%, indicating that the gains are not restricted to our chosen value.

λDUD-E EF0.1%DUD-E EF0.5%DUD-E EF1%DUD-E BEDROCDEKOIS2.0 EF0.1%DEKOIS2.0 EF0.5%DEKOIS2.0 EF1%DEKOIS2.0 BEDROC
041.9233.7528.100.454217.2218.1718.110.5272
0.0542.1233.8028.180.455717.9818.4318.240.5296
0.142.1234.0028.370.456918.3718.4318.330.5313
0.242.6534.2828.570.458618.3718.6918.680.5330
0.442.5134.5028.630.458217.6019.1418.620.5322
0.842.6434.3228.000.434718.3719.4518.590.5157
140.9531.3122.850.329417.2219.1917.470.4613

Compared with ReqVS-ES in Tables 3–4, ReqVS's AUROC decreases by 0.001 on DUD-E and 0.003 on DEKOIS2.0, while EF0.1% increases by 1.01 on DUD-E and EF0.5% increases by 0.45 on DEKOIS2.0.

We would like to respectfully clarify that these metrics capture different aspects of screening performance. AUROC measures discrimination across the entire library, whereas practical virtual screening typically selects only a small number of top-ranked compounds for experimental validation. EF measures the concentration of active compounds within a specified top-ranked fraction relative to random selection; a higher EF therefore means that more actives are retrieved within the same selection budget. BEDROC uses exponential weighting to assign greater importance to active compounds appearing near the top of the ranked list. EF and BEDROC thus correspond more directly to the practical goal of concentrating active compounds among the limited number of candidates that can be experimentally tested [1]. The observed improvements indicate better early recognition despite a small reduction in overall discrimination.

We will add these experiments and the corresponding discussion to the Appendix in the revised manuscript.

[1]Truchon J F, Bayly C I. Evaluating virtual screening methods: good and bad metrics for the “early recognition” problem[J]. Journal of chemical information and modeling, 2007, 47(2): 488-508.

Comment 4:

“Q3: DUD-E is known for decoy bias that a ligand-only model can exploit. What does a protein-blind control reach, that is, each constituent model and ReqVS with the target sequence removed or randomized?...”

Re: We thank the reviewer for raising this important concern. To address it, we conducted experiments in which we randomly replaced 20%, 40%, or 100% of the residues in the target sequences during model inference, then reranked the compounds and recalculated EF.

Sequence conditionEF0.1%EF0.5%EF1%BEDROC
Original sequence42.6534.2928.580.4590
20% replacement26.5220.5016.740.2810
40% replacement9.867.085.760.1070
100% replacement0.830.971.000.0210

Screening performance progressively declined as the proportion of randomized residues increased. When 100% of the residues were replaced, performance approached that of random selection, with EF values close to 1, where EF = 1 represents the expected enrichment under random selection.

AuthorsOpenReview ↗

Response 3

Comment 5:

“W4: The error bound is vacuous...but the surrogate itself is a binary entropy bounded by one bit, so the bound says nothing; the equations also mix natural-log entropy with a base-two bound. ”

Re: We thank the reviewer for this insightful observation. We acknowledge that the original one-bit bound is relatively loose, as it follows directly from the maximum entropy of a binary outcome and does not provide a tight characterization of the residual uncertainty.

To further characterize this residual uncertainty, we introduce a conditional bound based on Fano's inequality [1]. Specifically, let g∗(S)g^{\ast}(S) denote the optimal structure-conditioned binary classifier, with Bayes error Pe∗=P(g∗(S)≠Y)P_e^{\ast}=P(g^{\ast}(S)\neq Y). Then, the residual conditional entropy satisfies H(Y∣S)≤he(Pe∗)H(Y\mid S)\leq h_e(P_e^{\ast}), where heh_e is the binary entropy function with natural logarithms. This bound provides a tighter characterization when the structure-conditioned classification error is sufficiently small. For illustration, the corresponding upper bounds under different error levels are shown below:

Pe∗P_e^*Upper bound
0.010.0560
0.020.0980
0.050.1985
0.100.3251
0.200.5004
0.500.6931

We further provide empirical motivation for the bounded residual uncertainty based on the reproducibility of protein–ligand affinity measurements. For example, [2] summarized reported repeatability errors of approximately limited log units for pKi measurements. These findings suggest that protein–ligand affinity measurements exhibit a certain degree of experimental reproducibility, providing empirical motivation for considering a relatively limited residual uncertainty under consistent measurement conditions.

We will revise Appendix A to incorporate this conditional entropy bound, clarify the assumptions underlying its interpretation, and distinguish the mathematical bound from the empirical motivation. We will also unify the logarithm convention throughout the theoretical analysis and implementation. Together, these revisions clarify that the latent-structure formulation motivates sequence-conditioned predictive entropy as a practical uncertainty surrogate, while the tightness of its approximation depends on the residual uncertainty and the underlying modeling assumptions.

[1] Fano R M. Fano inequality[J]. Scholarpedia, 2008, 3(10): 6648.

[2] Hahn D F, Bayly C I, Boby M L. Best practices for constructing, preparing, and evaluating protein-ligand binding affinity[J]. Living Journal of Computational Molecular Science, 4 (1), 2022.

Comment 6:

“W5: ... a compound's reported uncertainty depends on which other compounds are screened alongside it, and the quantity has no probabilistic meaning.” & “Q5: ... how do the reported entropies and rankings change when the same target is screened against libraries of different size and composition?”

Re: To better assess sensitivity to library size and composition, we conducted additional experiments on DUD-E. We randomly sampled 10%, 50%, and 80% of the active compounds and decoys, with 20 repetitions at each sampling fraction. We compared screening performance using the original min–max parameters (original) with that obtained by recalculating the normalization parameters for each sampled library (recalculate). We do not report EF0.1% or the DEKOIS2.0 results because, at 10% sampling, the top-0.1% selection on DUD-E contains only 1.38 compounds per target on average, while EF0.1%, EF0.5%, and EF1% on DEKOIS2.0 each select only one compound per target. These small selection sizes make the corresponding estimates statistically unreliable.

Retained fractionEF0.5%EF1%EF5%
10%-recalculate34.87 ± 1.3228.99 ± 1.0310.38 ± 0.23
10%-original34.96 ± 1.4028.93 ± 1.0210.37 ± 0.22
50%-recalculate34.38 ± 0.4328.55 ± 0.3010.38 ± 0.05
50%-original34.39 ± 0.4228.55 ± 0.3010.38 ± 0.05
80%-recalculate34.37 ± 0.1628.60 ± 0.1210.39 ± 0.03
80%-original34.38 ± 0.1628.60 ± 0.1210.39 ± 0.03
100%34.2928.5810.40

Our method maintained stable screening performance across the tested library sizes and compositions.

This reply continues in the next comment due to the character limit.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.