acceptodds
Under review as a conference paper at ICLR 2027

No Ligand Binds in Isolation: Competition-Aware Bioactivity Prediction

Abstract

Accurate prediction of drug-target bioactivity is central to early-stage drug discovery, yet existing deep learning models evaluate protein-ligand interactions in isolation, ignoring the competitive binding context that governs drug action in cellular environments. In radioligand displacement assays, observed bioactivity is modulated by the relative binding strength and spatial pose overlap between a test ligand and a competing radioligand within the same protein pocket—factors that conventional single-ligand models neglect. We present DrugComp, a competition-aware deep learning framework that explicitly encodes atom-level binding affinities and steric overlaps between competing ligands within shared protein binding pockets. DrugComp integrates seamlessly with diverse backbone architectures, including graph neural networks and transformers. To support competition-aware learning, we curate CompDB, a dataset of 10,675 radioligand displacement assays spanning 151 proteins, 181 radioligands, and 8,674 test ligands, enriched with docking-derived structural context. Evaluated in five-fold cross-validation, DrugComp consistently outperforms competition-blind baselines across all architectures, achieving up to 44.7% improvement in bioactivity prediction metrics. These results demonstrate that reframing bioactivity prediction as competitive-context modeling substantially improves predictive performance and better reflects the mechanistic basis of drug action.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Deep Learning, Drug Discovery, Protein-Ligand Interaction, Bioactivity Prediction, Ligand Competition

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_n5GnOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a novel competition-aware deep learning based framework for bioactivity prediction. Different from existing drug discover methods, which only consider the interactions between proteins and ligands, the proposed DrugComp also includes the effects caused by cellular environments and drug action. Additionally, a new dataset named CompDB is curated by the authors, which contains a large number of test ligands and radioligands for experimental validation.

Critical strengths
  1. The new dataset curated by the authors could be very beneficial to the machine learning community. Unlike conventional deep learning datasets that rely primarily on straightforward data annotation, this dataset incorporates substantial biological prior knowledge into its annotation, collection, and curation processes, enabling a more meaningful evaluation of whether the proposed model can effectively generalize to real-world biological downstream tasks.
Critical weaknesses
  1. The authors employed LLMs for certain information filtering and extraction tasks. My main concern is the potential risk of LLM hallucinations during this process. Given the sensitivity and complexity of biological data, even a small proportion of hallucinated or inaccurate information could significantly compromise the reliability and integrity of the entire dataset.
  2. The proposed method constructs a graph using radioligands and test ligands to identify potential test ligands that could competitively displace radioligands. While this approach is biologically reasonable, from a deep learning perspective, it may introduce a bias toward the radioligands already represented in the dataset, potentially limiting the model's generalization ability. In real-world drug discovery scenarios, however, such models are expected to generalize beyond known ligand interactions and identify previously unseen candidates. Therefore, I am concerned that the proposed graph construction strategy may lead the model to rely heavily on existing radioligand-associated patterns rather than learning more generalizable protein–ligand interaction mechanisms.
Questions for authors
  1. Regarding the first weakness, I would like to see the authors to provide more details regarding the LLM uses during dataset curation. Also the authors need to verify the accuracy of the information extracted by the LLM.
  2. Regarding generalization, I would like the authors to provide additional zero-shot evaluation cases to demonstrate that the proposed method has not overfitted to the protein–ligand pairs in the training set and can effectively generalize to previously unseen interactions.
4 comments

Response to Weakness 1 and Question 1: LLM extraction

We carefully considered the potential risk of LLM hallucinations during dataset curation and therefore designed a highly constrained prompt for radioligand extraction:

Given an assay description, identify and extract the radioligand (tracer) used in the experiment. Instructions:

  1. Extract only the radioligand explicitly mentioned in the assay description. The radioligand should be a small-molecule compound.
  2. If the radioligand is radio-labeled (e.g., [3H], [125I]), ignore the isotope label and output only the compound name.
  3. If no radioligand is explicitly mentioned, or its identity cannot be determined unambiguously, return an empty response.
  4. Output only the radioligand name without any explanations or additional text. Assay description: {assay_description}

In particular, we explicitly instruct the LLM to abstain from extraction when the assay description does not specify a radioligand. To further address the reviewer's concern, we conducted an additional manual verification, in which two annotators independently reviewed all 861 distinct descriptions from 10,675 assay records. After standardizing radioligand names, the annotators achieved 100% inter-annotator agreement (Cohen's κ=1.0\kappa = 1.0). Comparing the LLM outputs against these manually verified records, we found that 99.27% of the extractions were correct. The remaining 0.73% exclusively involved cases where no radioligand was explicitly specified in the assay description, but the LLM inferred a radioligand.

Upon further inspection, we found that these inferred compounds are commonly used radioligands in relevant assays and are therefore plausible candidates. Although they cannot be regarded as experimentally verified radioligands for the corresponding assays, we believe such candidates can still be useful in practical applications of our framework. Specifically, large-scale bioactivity prediction often involves protein–ligand pairs without prior wet-lab measurements or assay descriptions. In such scenarios, plausible radioligands can be introduced to construct the competitive binding context required by DrugComp, enabling its application to previously untested protein–ligand pairs.

Response to Weakness 2 and Question 2: Generalisation

We provide additional zero-shot evaluations to assess generalization to unseen targets and ligands.

Practical assay setting. In competitive binding assays, radioligands are typically established tracers reused across experiments, whereas test ligands are the new candidates being screened. Introducing a new radioligand requires additional experimental characterization and validation [1]. This is reflected in CompDB-GPCR, where each target is associated with an average of 71 distinct test ligands but only 2 radioligands. Thus, generalization to new candidates using established radioligands represents a primary practical application of DrugComp.

Zero-shot generalization. To further assess generalization, we evaluated DrugComp under a target-disjoint split, restricting test samples to those whose test ligands were also absent from training. Thus, both targets and test ligands are unseen during training.

BackboneBaselineDrugCompRMSE gain
DTIGN1.4221.380+0.042
GEMS1.2571.238+0.019
HypSeek1.3201.256+0.064

DrugComp consistently improves RMSE across all three backbones, suggesting that its benefits extend to previously unseen target–ligand combinations rather than relying solely on memorized interactions.

[1] Hulme, Edward C., and Mike A. Trevethick. "Ligand binding assays at equilibrium: validation and interpretation." British journal of pharmacology 161.6 (2010): 1219-1237.

We sincerely thank the reviewer for the careful review and constructive feedback. We have provided point-by-point responses, including additional analyses and clarifications, and would greatly appreciate your feedback on whether these adequately address your concerns. If so, we hope you might consider revisiting your overall assessment.

Thanks for the authors' response. I think the rebuttal has fully addressed my questions. Therefore I'm willing to raise my score.

Thank you for your positive feedback and for raising your rating. We sincerely appreciate your time in reviewing our paper and responses.

Reviewer_Nh8KOpenReview ↗
Rating
4
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary
  • The authors of this work focus on the competitive binding context when looking at protein ligand interaction networks - something that is usually overlooked in modern ML based classifiers.
  • The proposed solution explicitly encodes steric overlaps and atom level binding affinities between competing ligands within shared protein binding pockets. The authors also provide CompDB with ~10k radioligand displacement assays.
  • The main question is "not only whether a compound binds, but whether it binds strongly enough, and in a compatible pose, to win the competition."
  • Instead of treating each protein-ligand pair as an isolated observation, the authors learn directly from radioligand displacement assays, where measured bioactivity arises from competition between a test ligand and a radioligand within the same binding pocket.
Critical strengths
  • Rich set of node features + inter-ligand edges to capture what is going on in the protein so downstream ML models have good context.
  • One of few papers that still show results averaged over 5 fold cross validation.
  • Interesting ablation experiments covering various modeling choices as well as competitor importance
Critical weaknesses

Not critical weaknesses, just average weaknesses:

  1. Nitpicky - Figure 2 placement can be improved - it goes 1, 2, 3, 5, 4 currently in a clockwise direction.
  2. Not a major issue, but it would be nice to see what radioligand identity extraction would look like with a different LLM since only 100 randomly sampled assays were manually reviewed.
Questions for authors
3 comments

We sincerely thank the reviewer for the positive and encouraging assessment of our work and provide detailed responses to both points below.

Response to Weakness 1

We have swapped the labels of panels (iv) and (v) to ensure a consistent clockwise order and updated the corresponding text and caption accordingly.

Response to Weakness 2

Following the reviewer's suggestion, we re-extracted radioligand identities for all assay descriptions in CompDB using Qwen2.5-7B-Instruct, achieving 97.6% agreement with the original Llama-3 annotations after normalizing isotope notation (e.g., 125I-labeled endothelin 1 vs. [125I]endothelin 1). We further validated the extraction accuracy through independent manual review by two annotators of all distinct assay descriptions, achieving 100% inter-annotator agreement (Cohen's κ=1.0\kappa = 1.0) and confirming that 99.27% of the LLM extractions were correct.

Looks good to me, thank you for the changes!

Thank you for your positive feedback. We appreciate your time in reviewing our paper and responses.

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

The authors introduce DrugComp, a deep learning framework for ligands binding in the presence of competing binders. This is a hugely important problem which, as the authors describe, is largely ignored in the vast literature of binding affinity prediction. The authors also contribute a dataset of competition-aware binding called CompDB, of great importance to future work in this line of work.

Critical strengths

The problem

The authors aim to address a very important and largely ignored problem in the field. As they say in the title, "no ligand binds in isolation". Data scarcity is probably the main reason people have not tried to address this issue with deep learning earlier; therefore, the contribution of CompDB is very useful and significant. Generally, bringing attention to this problem is beneficial to the field as a whole.

The model

The authors aim to solve this with a model that takes as input a triple of pocket atoms, ligand atoms, and radioligand atoms. Whilst this might be in some ways problematic (see weaknesses), it is an original approach.

Critical weaknesses

Main Weakness 1: Inconsistency about the poses

The way the docking is done for the dataset creation is described inconsistently throughout the paper. On the other hand, the authors say that “both test ligand and radioligand are docked into the same protein structure”, but on the other hand, they say that docking is done with Boltz-2. Boltz-2 is a cofolding model, meaning it takes as input a protein sequence (not a structure), and produces a cofolded 3D structure. Therefore, Boltz-2 will produce a different protein conformation for the test ligand and the radio ligand. This inconsistency has downstream consequences:

  • The inter-ligand edges are distances between two poses generated in two different receptors
  • The atom-level Vina features are computed against different protein coordinates.
  • The perhaps more complex dilemma. Because we are calculating the displacement, one of two things can happen: 1 if Boltz-2 cofolding adapts the pocket to each ligand, the ligand's coordinates are only relative to the receptor it was cofolded with, so moving it into a different receptor's frame is meaningless. 2 if, on the other hand, Boltz-2 returns essentially the same pocket regardless of ligand (which is the most likely, as MSA and template play the biggest role in predicting the protein structure, and those don't change). In that case, the situation is better (the inconsistency above disappears), but now Boltz is also likely to place all ligands in the same binding pocket, especially for GPCRs, meaning the difference in placement is effectively always zero; thus this feature carries no information.

Main Weakness 2: Training

There is basically no information throughout the paper about how this model is trained: What is the loss function, training curves... The most concerning aspect about this is the training split. As I understand from Appendix F.3 (and this really should be in the main text), the split is random. This means that there will be extreme data leakage, which explains the high model performance. The only leakage prevention is to ensure the same triplet (R, L, P) is not in training and testing, but that is not nearly enough. In fact, we could have a pair ligand-protein appear in training and test, with only the radioligand being different, which is very problematic. There is extensive evidence in the field that a much stricter split by pocket-ligand similarity is required. Table 4 hints at this: DrugComp's advantage becomes much less significant (in fact, it is not significant at one sigma) with group-disjoint splits, and I suspect a split by pocket-ligand similarity would reduce this advantage further.

Main Weakness 3: Memorisation

  • I worry that the model might just be memorising the protein corresponding to the radioligand. Usually, for a given protein, the same radioligand is used, so the model could simply be learning that the combination of a given protein and radioligand is "correct", and be ignoring the 3D structures. In fact, the data only contains 181 tracers across 151 proteins. Also, the fact that the models that improve the most in the results are the worst ones at handling proteins makes me think this could be the case, as well as the fact that testing on proteins not seen in training makes the improvement go down significantly. One way to check this would be to remove the 3D structure from the model, replace it with just a label, and see how much of the improvement we still see.

Other weaknesses

  • In Table 2, the standard deviations are very large, so pretty much none of the seen improvements is statistically significant. Also, the text says that "the full model achieves the best performance across all evaluated architectures", which, looking at the table, is simply not true
  • It seems strange to leave aspects like related work or training to the appendix, but to spend more than half a page on the equations from the vina features, which are not original contributions
Questions for authors
  • How often do we have the same protein, same ligand, and two different radioligands in the data? When that does happen, how big is the difference?
  • If the radioligand’s 3D subgraph is replaced by a learned embedding of its identity alone, how much of the improvement survives?
  • What is the loss function?
  • Have you tried ablations not including GCR, e.g. Baseline + LBS?
10 comments
AuthorsOpenReview ↗

Response to Main Weakness 1: Inconsistency about the poses

We thank the reviewer for the careful review and constructive feedback. We provide point-by-point responses below. Changes in the revised manuscript are highlighted in blue for ease of review.

We acknowledge that the current description may be misleading and have clarified the methodology in Section 3 and Section 4.1 of the revised manuscript. Specifically, Boltz-2 independently co-folds the test ligand and radioligand with the same protein sequence, yielding two ligand-bound receptor conformations. We then compute the per-atom receptor displacements between the two predicted structures, obtaining median values of only 0.31 Å for AR, 0.17 Å for SLC, and 0.50 Å for GPCR, indicating limited structural variation between the predicted receptors. This observation is broadly consistent with previous findings of relatively small backbone variations between ligand-bound structures (median RMSD of 0.43 Å) [1].

Importantly, we align the predicted receptor structures and adopt the radioligand-bound receptor conformation as the common reference for computing inter-ligand distances and atom-level Vina features. The test ligand poses are transformed into this shared coordinate frame while preserving their predicted geometries. This ensures that all measurements within the same (target, radioligand) pair share a fixed receptor conformation and radioligand pose, with only the test ligand varying.

We further clarify that placing ligands within the same binding pocket does not imply identical binding poses or negligible spatial differences. To verify this, we analyze all 13,963 competitive graphs in CompDB-GPCR and find that the median centroid distance between test ligands and competing radioligands is 3.21 Å. These results demonstrate substantial variation in the relative spatial arrangements of competing ligands, even within the same binding pocket, indicating that inter-ligand geometry provides nonconstant and meaningful spatial information for modeling ligand competition.

[1] Clark, Jordan J., et al. "Inherent versus induced protein flexibility: Comparisons within and between apo and holo structures." PLoS computational biology 15.1 (2019): e1006705.

AuthorsOpenReview ↗

Response to Main Weakness 2: Training

We provide additional training details and analyses to assess potential data leakage and generalization.

Training details. We employ MSE loss on standardized pIC50pIC_{50} values. For HypSeek, we retain the additional hyperbolic contrastive loss as specified in the original paper for both the competition-blind baseline and DrugComp. We have also added training curves (Figure 4) showing training loss, validation RMSE, and test RMSE across five folds for all three backbones. We can see that DrugComp achieves lower training loss, while its validation and test RMSE remain below those of the baseline throughout training, further demonstrating its effectiveness. Detailed optimizer settings, learning rate schedules, and other training configurations are provided in Appendix F.4, with architecture hyperparameters for all seven backbones reported in Table 8. The loss function is now explicitly described in Section 5.1 of the revised manuscript.

Data leakage. We would like to clarify that DrugComp and all competition-blind baselines use exactly the same dataset and identical train/test splits. Therefore, any potential data overlap affects both equally and does not introduce an unfair advantage for DrugComp. As already reported in Table 4 of the original submission, we evaluated all methods under three more stringent splitting strategies: ligand-disjoint, scaffold-disjoint, and target-disjoint splits. To further address the reviewer's concern about potential data leakage in random splitting, we analyze the overlap between the training and test sets and find that only 0.8% of test samples share the same (target, ligand) pair with the training set. We further excluded these overlapping samples and re-evaluated both DrugComp and the competition-blind baselines. The results show that DrugComp consistently maintains its improvements over the baselines, further supporting the robustness of our conclusions. The splitting protocol is now explicitly described in Section 3 of the revised manuscript, with the group-disjoint splits detailed in Appendix F.3.

Test set (rows/fold)ModelDTIGN RMSEGEMS RMSEHypSeek RMSE
All (1,990–1,992)Baseline0.8950.8571.039
DrugComp0.8610.7950.945
No overlapping target–ligand pair (1,975–1,977)Baseline0.8920.8531.036
DrugComp0.8590.7940.941

Chemical similarity. To further evaluate DrugComp's generalization, we stratified test samples by their nearest training ligand using ECFP4 Tanimoto similarity. The mean test RMSE across three random seeds is reported below (lower is better):

TanimotoRows/foldModelDTIGNGEMSHypSeekMean
0.7–1.01,714Baseline0.8510.8111.0070.890
DrugComp0.8400.7660.9260.844
0.5–0.7229Baseline1.0161.0431.1151.058
DrugComp0.9920.9931.0651.017
0.3–0.537Baseline1.3021.4281.4191.383
DrugComp1.2761.3241.2951.299
0.0–0.38Baseline1.4721.5841.7121.589
DrugComp1.4421.2831.6311.452

DrugComp achieves lower RMSE across all similarity ranges for all three backbones, with larger mean RMSE reductions in the two lower-similarity bands (0.084 and 0.137) than in the two higher-similarity bands (0.046 and 0.041). This suggests that its benefits are not limited to test ligands highly similar to those in the training set.

AuthorsOpenReview ↗

Response to Main Weakness 3: Memorisation

We agree that a given target is often assayed using a fixed radioligand. This reflects standard practice in competitive displacement assays rather than an artefact of our data curation. Using a consistent tracer helps ensure comparability across compounds tested against the same target. Importantly, the test ligand varies substantially within each (target, radioligand) pair, and this variation cannot be explained by simply memorising the pair identity.

The pair does not determine the ground truth bioactivity value. Within a given (target, radioligand) pair, different test ligands produce a wide range of bioactivity values. The eight largest pairs in CompDB-GPCR each span 3.84–6.05 log units, while their mean pIC50 values differ by only approximately 1.6 log units:

TargetRadioligandMeasurementspIC50 rangeSpanMean
GPR44 (CRTH2), human (Q9Y5Y4)[3H]PGD26994.21–10.005.797.26
CCR2, human (P41597)[125I]MCP16374.66–9.524.867.19
NK1 receptor, human (P25103)[125I]substance P4295.13–10.195.068.08
Cannabinoid CB1, rat (P20272)[3H]CP-559404055.40–9.243.847.66
Cannabinoid CB1, human (P21554)[3H]CP-559403525.01–9.704.697.51
Prostanoid EP1, human (P34995)[3H]PGE22975.09–9.894.807.50
Prostanoid EP3, human (P43115)[3H]PGE22934.46–9.364.907.66
Cannabinoid CB2, human (P34972)[3H]CP-559402694.59–10.646.056.51

All eight distributions share the pIC50 interval 5.40–9.24, which contains 78–100% of the measurements within each pair. Thus, pair identity provides limited information about an individual measurement compared with the substantial variation across test ligands. Even pairs sharing a radioligand exhibit overlapping distributions: [3H]CP-55940 is used for rat CB1, human CB1, and human CB2, with means of 7.66, 7.51, and 6.51, respectively, while [3H]PGE2 is used for EP1 and EP3, with means of 7.50 and 7.66.

A direct test of pair memorisation. We evaluated a pair-mean predictor that assigns each test measurement the mean pIC50 of its (target, radioligand) pair in the training fold. Under the random split, 99.1% of test measurements have their pair represented in training, allowing the predictor to use the corresponding pair mean. For the remaining 0.9% with unseen pairs, we use the overall training mean. The pair mean minimises training squared error among constant predictions for each pair, making this a direct benchmark for the proposed memorisation mechanism.

BackboneMemorisation oracleCompetition-blind baselineDrugComp
CheapNet1.0400.8910.869
DTIGN1.0400.8950.861
EGNN1.0400.9740.901
GEMS1.0400.8570.795
GIGN1.0400.9040.869
HypSeek1.0401.0390.945
T-ALPHA1.0400.9810.902

Because the pair-mean predictor depends only on the data and fold assignments, its RMSE is identical across backbones. It performs worse than DrugComp for all seven backbones, with RMSE gaps of approximately 0.095–0.245, and is no better than the competition-blind baselines, which do not use radioligand information at all. These results demonstrate that memorising the (target, radioligand) pair cannot account for DrugComp's predictive performance, independently of backbone capacity.

Target–radioligand relationships. The mapping is not one-to-one. Across 151 targets and 181 radioligands, the full dataset contains 334 distinct (target, radioligand) pairs. Among these, 73 radioligands (40%) are used across multiple targets, while 90 targets (60%) are assayed with multiple radioligands, collectively covering 84% of the measurements. Thus, neither target identity nor radioligand identity uniquely determines the other.

Radioligand substitution. Table 3 provides further evidence that the model uses radioligand-specific information. Across 1,266 matched assays per fold, replacing the original radioligand with an alternative for the same target increases RMSE by 3.5–6.3%, with statistically significant differences. Because the target and test ligand are held fixed, this degradation demonstrates that predictions are sensitive to radioligand identity rather than being determined solely by the target.

Unseen targets. The smaller improvement under target-disjoint splits is consistent with the greater difficulty of out-of-distribution generalisation. Importantly, the competition-blind baselines are evaluated on identical splits and face the same distribution shift. Therefore, the reduced gain on unseen targets does not establish that the improvements observed under random splits arise from pair memorisation.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Discussion

We sincerely thank all three reviewers and the Area Chair for their time and constructive feedback. We summarise below how the concerns have been addressed.

Reviewer Feedback

Reviewer Nh8K (Rating 4, Soundness 3) questioned Figure 2's panel ordering and the reliability of LLM-extracted radioligand identities. We corrected the figure and independently verified the extraction using a second LLM. The reviewer confirmed:

"Looks good to me, thank you for the changes!"

Reviewer n5Gn (Rating 3, increased from 2; Soundness 2) raised concerns about LLM hallucinations and potential memorisation of training radioligands. We provided the extraction prompt, manual audit, independent LLM verification (99.27% agreement), and additional zero-shot evaluations. The reviewer confirmed:

"the rebuttal has fully addressed my questions"

Reviewer brtG (Rating 1, Soundness 1) raised concerns about pose consistency, missing methodological details, and radioligand memorisation. After our first response, the reviewer acknowledged:

"The pose description is clear, and the fact that the receptor barely moves between the two cofolded structures resolves the inconsistency I raised. The training details and loss are now in the paper, and the performance claim has been corrected. The paper is in better shape than it was."

We resolved the pose-consistency concern, added the training details and group-disjoint evaluations, and corrected the performance claim. We also conducted the exact radioligand identity ablation proposed by the reviewer.

Additional Evidence Addressing Remaining Concerns

1. Tracer swaps and competitive context

DrugComp models the competitive binding context of each assay rather than directly predicting affinity shifts caused by tracer swaps. Each test ligand forms a distinct competitive graph, even with the same target and tracer.

To assess whether tracer information matters, we performed counterfactual substitutions across 1,266 matched assays per fold. Replacing the assayed tracer with another real tracer increased RMSE by 4.7%, 6.3%, and 3.5% across three backbones, demonstrating that the models use tracer-specific information.

2. Radioligand identity ablation

We extended the reviewer's proposed experiment to all seven backbones. Every backbone showed the same ordering:

Competition-blind baseline → Radioligand identity → Radioligand 3D graph

RMSE decreased at each step. Importantly, identity embeddings already encode competition-related information, supporting our central claim that competitive context improves prediction. Explicit 3D geometry provides additional, backbone-dependent benefits. We have clarified this distinction in the manuscript.

3. Statistical significance

We accepted the reviewer's criticism of our original pooled tests and replaced them with more appropriate analyses:

  • Paired t-tests across five folds: 6/7 backbones achieve p < 0.05.
  • Multi-seed evaluation: DrugComp improves upon the competition-blind baseline in 73/75 paired fold–seed comparisons across five backbones and three seeds.
  • Target-level bootstrap across 136 GPCR targets: 95% confidence intervals exclude zero for all seven backbones.

These results provide stronger evidence of consistent improvements across backbones, seeds, and targets.

Kindly Requesting

We have now addressed the substantive concerns raised by all three reviewers through corrections, clarifications, and additional experiments. Two reviewers have explicitly acknowledged that their questions were addressed, with one increasing their rating.

We respectfully ask Reviewer brtG and the Area Chair to evaluate the paper based on the revised manuscript and complete rebuttal evidence, particularly the expanded seven-backbone identity ablation, counterfactual tracer evaluation, and corrected statistical analyses.

We believe these revisions substantially strengthen the scientific validity, clarity, and empirical support of our contribution.

Acknowledgements

We are grateful to all three reviewers, particularly Reviewer brtG for the exceptionally careful critique. The proposed identity ablation helped us distinguish the contributions of competition awareness and explicit 3D geometry, while the statistical criticism led to more rigorous evaluations.

These suggestions have materially improved our manuscript, and we sincerely appreciate the reviewer's efforts.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.