acceptodds
Under review as a conference paper at ICLR 2027

CQFiner: Protein Context Query Refinement for Test-Time Scaling Resolution of Binding Site Detection

Abstract

Protein binding site prediction is one of the most fundamental problems in drug discovery and structural biology. A mainstream approach is to place anchors or query nodes at different locations, move and classify these queries, and finally use the coordinates of high-confidence queries as the predicted binding site locations. Resolution is a key determinant of predictive performance of such methods. However, in this work we demonstrate that the available resolution upper limit is constrained by GPU memory, over-smoothing, and suboptimal training dynamics. More importantly, these query-based methods fix the query number and distribution during training and inference, thus cannot benefit from higher spatial resolution at inference, or specific query sampling strategy according to user priors. Here we introduce Context Query Refiner (CQFiner). CQFiner models the protein as a fixed readonly context and formulates binding site prediction as an object detection task conditioned on this context, refining both query positions and confidence scores from coarse to fine. We demonstrate how object detection practices help the model benefit from higher resolution and how readonly protein context enables test time scaling and prior-based query sampling. Based on the architecture, CQFiner substantially outperforms state-of-the-art methods on both in-distribution and out-of-distribution benchmarks, such as COACH420, HOLO4K, etc.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
protein, binding site prediction, graph neural network

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

This paper presents CQFiner for protein binding-site detection. It combines local residue-to-query message passing, near-site supervision, a ViSNet coarse stage, and an EGNN refinement stage with fixed protein features. The read-only protein context allows the query count and sampling distribution to change at inference. Experiments show large gains on HOLO4K and COACH420, with further structural-robustness evaluation on AlphaFold-predicted structures, APOBind, and LIGYSIS.

Critical strengths
  1. Strong overall performance. CQFiner achieves substantial improvements on the main benchmarks, with gains in both binding-site localization and residue-mask prediction. These results provide strong empirical support for the proposed framework.

  2. Good generalization ability across different structure types. The model maintains competitive performance on AlphaFold-predicted and apo structures without dataset-specific tuning. This broadens its practical relevance beyond experimental holo structures.

Critical weaknesses
  1. The spatial-resolution argument is too broad. The paper does not sufficiently distinguish spatially sampled queries from learned detection queries, such as those used in UniSite. For spatial queries, adding more points can improve geometric coverage. For DETR-like queries, however, the query count controls the capacity to represent candidate pockets and does not directly correspond to spatial resolution. The motivation should therefore be scoped more carefully.

  2. Improved generalization is not clearly separated from stronger in-domain performance. CQFiner obtains higher absolute scores on most OOD benchmarks, but these results may largely reflect its stronger in-domain performance. UniSite also handles structural changes well and sometimes overtakes CQFiner: on HOLO4K, AP₀.₅ changes from 0.591 to 0.592 for UniSite when moving to AlphaFold structures, compared with a drop from 0.640 to 0.575 for CQFiner. The current evidence therefore supports strong performance across structure types, but does not clearly establish better generalization due to the proposed architecture. Since this work largely inherits UniSite’s data organization and task formulation, which differ substantially from those used in earlier approaches, it remains unclear how much of the observed generalization comes from the proposed architecture versus the inherited dataset design and training protocol.

Questions for authors
  1. Please clarify the scope of the spatial-resolution argument. In particular, how does it apply to learned DETR-like queries, such as those in UniSite, whose number does not directly correspond to spatial sampling density?

  2. What evidence shows that CQFiner generalizes better than UniSite, rather than mainly retaining its stronger in-domain performance? Table 2 suggests that UniSite shows smaller performance drops or even gains under some structural changes. Please discuss how these results affect the generalization claim.

  3. How much of the performance on OOD benchmarks can be attributed to the proposed architecture versus the data organization and task formulation inherited from UniSite?

3 comments
AuthorsOpenReview ↗

Reply of W1/Q1

The reviewer acknowledged CQFiner's performance advantage on experimental holo structures and its competitiveness on predicted and APO structures. However, the reviewer questioned the scope of the spatial-resolution argument and the sources of OOD performance (including UniSite's data organization and CQFiner's architecture). We first revised the main text to delimit the spatial-resolution argument for DETR-like query-based methods such as UniSite. Using UniSite, we then showed how CQFiner's components/designs can improve these models, with locality-preserving architecture and supervision tasks being key. Next, we analyzed CQFiner's OOD generalization across protein lengths and interaction patterns. Finally, we discussed UniSite's distinctive data organization, which appears to exhibit a clear distribution shift relative to earlier datasets.

W1 / Q1: How should the spatial-resolution argument apply to learned-query methods such as UniSite?

Q: The number of spatially sampled queries affects geometric coverage, whereas the number of learned queries in DETR-like methods such as UniSite mainly controls candidate-pocket capacity and does not directly correspond to spatial sampling density. Please clarify the scope of the spatial-resolution argument and how this idea can be applied to learned queries.

A: We thank the reviewer for the constructive comments. We will explicitly distinguish spatial queries with physical coordinates from purely learned queries. Spatial resolution directly corresponds to the former. For the latter, although CQFiner's approach cannot be applied directly, its design principles are easy to incorporate because each query still predicts a single site: each query primarily makes a local prediction, and more queries ensure greater spatial coverage and accuracy. To achieve this, we added spatial anchors to UniSite's DETR-style architecture, giving query initialization and mask prediction explicit spatial references. Experiments show that this version outperforms standard UniSite with the same number of queries, and increasing inference queries further improves pocket recall, demonstrating that this design idea can also be combined with DETR-like architectures.

Original UniSite learns a separate embedding for each query, so more queries mainly increase candidate capacity. We initialized queries with FPS-sampled residue features plus a shared learnable embedding, added CQFiner's distance gate to the final query–residue dot product, and supervised each query to predict its neighboring site's residue mask; these designs essentially fully reuse CQFiner's design (see Methods in the main text). Each query thus corresponds to a residue coordinate, primarily focuses on surrounding residues, and predicts the corresponding mask. Therefore, more queries essentially increase spatial coverage.

We compare this version with standard UniSite on the validation set below.

Bold and italicized values indicate the best and second-best results within the same comparison group, respectively, with ties marked at the precision shown in the table.

Table R1. Prediction performance and inference-query scaling of standard UniSite and spatially anchored UniSite on the UniSite-DS validation set.

ModelInference query budgetAP₀.₃ (%)AP₀.₅ (%)Recall₀.₅ (%)Recall₀.₅@5 (%)
Standard UniSite3256.6637.3850.1549.14
Spatially anchored UniSite3259.8340.5955.1153.95
Spatially anchored UniSite6459.5640.6555.8254.15
Spatially anchored UniSite12858.8740.5457.2154.92
Spatially anchored UniSite20058.4340.3158.0854.92
Spatially anchored UniSite40057.4439.4758.5754.82

With the same 32 queries, the spatially anchored version improves AP₀.₃ by 3.17 percentage points, AP₀.₅ by 3.21 percentage points, and recall by 4.97 percentage points over standard UniSite. Increasing the number of queries to 200 with fixed weights raises recall from 55.11% to 58.08%, a further improvement of 2.97 percentage points and 7.94 percentage points above standard UniSite; when only the top five candidates are retained, recall is also 5.78 percentage points higher than standard UniSite. With 400 queries, recall reaches 58.57%. However, overall AP does not improve, possibly because of generalization problems introduced by multiple queries (the queries are not independent because of self-attention). These results show the following, although the DETR-like query design itself is unrelated to spatial resolution. Because each query still represents a site, it is easy to incorporate CQFiner's approach and improve performance by designing query features and supervision targets.

Accordingly, we have revised the paper to restrict the resolution argument to spatial resolution and the corresponding specific query types. These approaches combining learnable DETR-like queries with CQFiner's design have been added to the appendix.

AuthorsOpenReview ↗

Reply of W2/Q2

W2 / Q2: Is better OOD performance merely a continuation of higher ID accuracy?

Q: CQFiner's higher absolute OOD accuracy may mainly come from its higher ID accuracy. In Table 2, UniSite exhibits smaller performance drops under some structural changes, or even slight improvements. What evidence can distinguish higher baseline accuracy from stronger generalization ability?

A: We thank the reviewer for the constructive comments. First, we would like to clarify that CQFiner itself was not specifically designed for OOD settings, and our OOD evaluation is primarily intended to demonstrate CQFiner's accuracy. However, we did find that CQFiner has certain OOD advantages, mainly including generalization across interaction patterns and protein lengths.

Protein-length scaling: Unlike VNEGNN and UniSite, where each query interacts with the entire protein, queries in CQFiner interact only locally, which means that even when protein length increases, the messages received by each query in CQFiner remain essentially unchanged. In contrast, VNEGNN and UniSite experience distribution shifts and performance loss because the number of residues interacting with each query increases. This allows CQFiner to scale to greater protein lengths at test time while maintaining better accuracy. To illustrate this, we sort samples by the number of input residues and use the longest 10% as the OOD test set, splitting the remaining samples into training and ID validation sets at a ratio of 8∶1, for an overall ratio of 8∶1∶1. We first train UniSite and then select CQFiner weights based only on ID validation performance, such that its in-distribution validation AP₀.₅ is lower than UniSite's. After fixing the weights, we evaluate on the test set consisting of the longest 10% of proteins. The results are quite striking. Even when CQFiner and UniSite have comparable [email protected] on the ID dataset (proteins of lengths comparable to those in the training set), CQFiner has an advantage of approximately 10 percentage points on the OOD test set, demonstrating CQFiner's length scalability.

Table R2. Comparison of protein-length OOD generalization with approximately matched ID validation accuracy.

ModelID validation AP₀.₅ (%)Long-protein OOD AP₀.₅ (%)
UniSite44.0311.90
CQFiner42.5022.94

Interaction-pattern generalization: The full analysis shows that CQFiner's advantage over UniSite is more pronounced for pockets with lower similarity to the training set. We use PLIP (PLIP 2021: expanding the scope of the protein–ligand interaction profiler to DNA and RNA) to extract protein–ligand interactions, take the 12,653 sites from all 8,348 training proteins as the reference, compute each evaluation site's maximum similarity to training pockets, and then compare the two models' prediction quality on that site. All three evaluation sets show a consistent trend. For example, in the group with similarities of 0.25–0.50, CQFiner's IoU@5 advantages are 8.23, 10.31, and 8.75 percentage points, respectively; when similarity reaches above 0.75, the two models perform similarly. This shows that CQFiner's prediction advantage also extends to pockets whose interaction patterns are less similar to those in the training set. We believe this may be because CQFiner's GNN-based backbone models interactions more effectively than DETR's Transformer.

Table R3. Pocket-interaction similarity to the complete training set and CQFiner's relative prediction advantage.

Evaluation setSimilarity to the training setNumber of sitesCQFiner IoU@5 (%)UniSite IoU@5 (%)Difference (percentage points)
UniSite validation set[0.00, 0.25)6042.3134.73+7.58
UniSite validation set[0.25, 0.50)33055.9147.68+8.23
UniSite validation set[0.50, 0.75)98968.2262.69+5.53
UniSite validation set[0.75, 1.00]52676.2974.99+1.30
COACH420[0.00, 0.25)1055.4538.98+16.48
COACH420[0.25, 0.50)7161.8451.53+10.31
COACH420[0.50, 0.75)10470.0762.55+7.52
COACH420[0.75, 1.00]2873.9175.79−1.87
HOLO4K-sc[0.00, 0.25)3551.1745.58+5.59
HOLO4K-sc[0.25, 0.50)28959.2650.51+8.75
HOLO4K-sc[0.50, 0.75)94168.2166.74+1.47
HOLO4K-sc[0.75, 1.00]35275.9777.11−1.14
AuthorsOpenReview ↗

Reply of Q3

Q3: How much of the OOD performance comes from CQFiner's architecture, and how much from UniSite's data organization?

Q: CQFiner inherits UniSite's data organization and task formulation, which differ substantially from those of earlier methods. How can the contributions of model architecture and the training-data scheme to OOD performance be distinguished?

A: We thank the reviewer for the constructive comments. We have added a control in which CQFiner is trained on the scPDB data used by VN-EGNN (scPDB contains 4,499 training complexes, while UniSite-DS contains 8,348 training proteins). Overall, we find that the training-data scheme affects different test sets differently. CQFiner trained on scPDB achieves AP₀.₅ and localization performance on COACH420 comparable to those obtained with UniSite-DS training, but performance declines on HOLO4K and the UniSite validation set. For the other OOD datasets, the UniSite-DS-trained model performs better on AF-HOLO4K-sc and CryptoBench, while the scPDB-trained model performs slightly better on APObind. Overall, the model trained on scPDB performed comparably to the model trained on UniSite-DS on the out-of-distribution (OOD) datasets, but substantially worse on the UniSite-DS validation set. These results suggest that UniSite-DS differs markedly from existing public benchmarks in how its data are organized. One possible explanation is that UniSite maps multiple binding sites onto a single structure, which may introduce structural discrepancies and misalignment of ligand coordinates. Owing to these differences in data organization, evaluation on these OOD datasets may disadvantage models trained on UniSite-DS.

Table R4. Internal validation and external test results for CQFiner trained on UniSite-DS or per-complex scPDB. Each cell lists the results of UniSite/scPDB training. The internal validation results of the UniSite-DS model retain the corresponding validation records from the main table in the paper.

Evaluation datasetProteinsAP₀.₃ (%)AP₀.₅ (%)DCC (%)DCA (%)
UniSite validation set2,05366.42 / 50.6851.67 / 35.5161.45 / 48.6374.59 / 62.92
COACH42028777.75 / 72.3956.65 / 56.2264.29 / 63.9982.74 / 82.74
HOLO4K-sc1,63375.84 / 69.5459.36 / 52.8672.66 / 65.1785.35 / 77.57
AF-COACH42025471.94 / 69.5347.02 / 48.2963.12 / 60.8082.39 / 75.75
AF-HOLO4K-sc1,33774.31 / 68.3552.52 / 46.8069.31 / 64.5883.51 / 77.87
APObind holo10,65053.31 / 56.8922.00 / 23.1354.30 / 57.6771.76 / 72.57
APObind apo10,65049.47 / 51.1514.80 / 15.4247.66 / 49.4965.69 / 65.83
LIGYSIS2,77526.27 / 19.2414.51 / 10.3621.13 / 17.3432.78 / 28.13
CryptoBench21435.28 / 30.5416.22 / 13.3037.85 / 33.6452.34 / 49.07
Reviewer_CdBSOpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

CQFiner is a query-based method for protein-ligand binding-site prediction. It samples queries near the protein surface, uses local residue-to-query connections, and supervises coordinate offsets only for queries near annotated sites, following an object detection paradigm. A ViSNet backbone predicts coarse displacements, followed by EGNN refinement against fixed protein representations. Removing query-to-residue and query-to-query communication makes individual query predictions independent of the other queries, improving memory usage. A distance-gated head additionally predicts site-specific residue masks. The method reports substantial improvements in localization and mask AP on COACH420 and HOLO4K, with additional evaluation on AlphaFold structures, APOBind, and LIGYSIS. Separate experiments investigate inference-time query scaling and oracle-guided query sampling.

Critical strengths

• The paper addresses an important problem and proposes a sensible combination of surface sampling, local geometric refinement, and reusable protein representations.

• The main reported improvements are substantial. DCC@4 increases from 0.586 to 0.767 on HOLO4K and from 0.545 to 0.735 on COACH420 relative to VN-EGNN, alongside improvements in mask AP.

• Independent queries provide a clear architectural rationale for changing query counts and sampling distributions at inference.

• Local connectivity substantially reduces training memory in the reported measurements, from 15.69 to 1.85 GiB at 200 queries.

Critical weaknesses

Insufficient documentation of data separation and baseline comparability limits verification of the substantial gains. The manuscript states that it follows the default UniSite splits and selects checkpoints using UniSite validation AP0.5. However, it does not provide a sufficiently explicit account of the exact training, validation, and test cohorts, all training data used, or their relationships to the external evaluation datasets. The baseline descriptions leave related uncertainties: whether published weights were used or models were retrained, which training datasets and splits each baseline used, which results were recomputed or taken from previous publications, and how configurations and post-processing were selected. Query budgets also differ, with CQFiner using 200 spatial queries and the original UniSite configuration using 32 learned site queries as default for example. Reported improvements are important, and their interpretation depends on establishing that they are not explained by train-test contamination, differences in training-data exposure, or evaluation configurations. I have not identified evidence that leakage occurred; the concern is that the information provided is insufficient to independently verify data separation and attribute the gains confidently. Clarifying this issue would have the greatest effect on my recommendation.

Questions for authors
  1. Exact data and splits: Please identify the complete training, validation, and test datasets used for CQFiner (and the other baseline models). Was UniSite-DS the sole supervised training dataset? Were any additional training stages, datasets, or pretrained weights used beyond the frozen ESM embeddings? Please also identify the exact ESM checkpoint.
  2. Train–test separation: Please clarify which train–test exclusion and similarity controls were inherited from the UniSite protocol and how they were applied in these experiments, including to APOBind and LIGYSIS. Please confirm the exact evaluation subsets, including whether HOLO4K denotes HOLO4K-sc, and provide the split and exclusion files. Were any additional overlap checks performed? If pocket-level sequence or structural similarity was not assessed, please clarify the resulting scope of the generalization claims.
  3. Baseline provenance and settings: For each baseline, dataset, and metric, were the reported numbers obtained by rerunning the model or taken directly from the original papers or another benchmark publication? For reused numbers, please identify the source and confirm that the evaluation subsets, annotations, and metric implementations are comparable. For rerun models, please specify whether published weights/default settings were used or the model was retrained, including training data, splits, pretraining, inputs/embeddings, query counts, post-processing, and parameter-selection procedures. Please provide the configurations and evaluation scripts needed to reproduce these comparisons.
  4. Comparable outputs and evaluation: Some baselines do not directly predict site-specific residue masks. Please explain how each method’s native outputs were converted into the evaluated centers and masks, and whether these conversions were selected using validation or test data. How much of the AP-IoU advantage reflects learned mask prediction rather than improved localization?

Minor comments

  • In 2.1 authors claim that having more queries and higher spatial resolution should lead to better performance. They mention coverage and I agree recall metrics might benefit but this might come at a precision and false positive cost. The motivating analysis in figure 1 does not show precision metrics.
  • for figure 1b query redundancy, cosine similarity and effective rank are quite difficult to interpret in absolute terms. I have no idea what 0.71 similarity means. It would be more interpretable if compared between queries that are close to a site vs queries that are far from a site. Also, I assume the number of sites is very small compared to the number of queries, so most queries are not close to any site and might therefore be very similar to each other and queries that do get close to a site might be close to the same site, making them similar to each other as well. Comparisons among background, same-site, and different-site queries, together with changes across layers, would make Figures 1b and 4a more interpretable. The loss allocation to me is very hard to interpret and "model wastes capacity on hard queries" is very speculative. this is not necessarily pathological distribution of losses.
  • Figure 1a query-count sweep. The offset baseline is evaluated at only three query counts, spanning 8–32 queries. This provides limited evidence for a general conclusion about its inability to benefit from denser queries. A broader sweep, where computationally feasible, would strengthen the motivating analysis.
  • The 32-query model’s improvement with additional inference queries is useful but modest and saturates around 200 queries. Appendix E states that these experiments use random surface sampling, whereas default benchmark inference uses farthest-point sampling. Does the scaling benefit persist under the default sampler and for the default 200-query model? Accuracy–latency comparisons, including training directly at the larger budget, would better establish practical benefits.
  • Could you add random surface-query baselines for DCC/DCA and appropriate chance or random-ranking baselines for AP, using comparable prediction budgets and post-processing?
  • Table 2 includes only UniSite and VN-EGNN, although P2Rank and GrASP provide stronger baseline DCA results on some main benchmarks. Including additional representative methods would strengthen the robustness assessment.
  • in section 3.2, I think it is very sensible to have local queries and to not have outgoing edges from queries back to residues and it logically helps with test time scaling and prior injection. I am curious if this comes at a cost for cryptic pockets or allosteric pockets. Are there other datasets the authors could evaluate on for these particular cases?
3 comments
AuthorsOpenReview ↗

Reply of Q1-Q3

We thank the reviewer for the detailed comments. We have added traceable descriptions of the data splits and model sources, and supplemented the corresponding experiments as requested by the reviewer. In addition, we found that the original dataset split did indeed contain data leakage, which we have corrected. Specifically: the default split provided by UniSite-DS did not exclude proteins from the test sets (COACH420 and HOLO4K) as claimed in the UniSite paper, which would overestimate CQFiner's performance. Therefore, in the revised paper, we excluded protein structures contained in the test sets as required by UniSite, and fully retrained all baseline models that showed an advantage to ensure a fair comparison. The paper has therefore been comprehensively updated. The authors apologize for the incorrect dataset-processing procedure and the misleading experimental results in the previous version of the paper.

Q1/W0: Complete training data, training stages, and pretrained weights

UniSite-DS is the only supervised training dataset for the main models, and the task networks are trained from scratch using frozen ESM features. The exact model used is facebook/esm2_t33_650M_UR50D, with 1,280-dimensional residue features; no additional training stages, datasets, or pretrained weights are used.

Q2: Splits, exclusion rules, test coverage, and overlap checks

UniSite officially provides two splits: the 0p9 split defines the training and validation sets of UniSite-DS, while the benchmark split records the samples in the test sets. Therefore, excluding the test samples from the 0p9 split gives the final training and validation sets without data leakage. The specific data-filtering steps are shown below, ultimately retaining 12,653/3,099 training/validation sites. We have added the detailed dataset-processing procedure to the Appendix of the paper.

Table R1. Data-filtering procedure for the training and validation sets.

Processing stepTraining proteinsValidation proteins
Official UniSite 0.9 split9,2172,293
Intersect each list with the official train_benchmark list to exclude proteins also present in the test sets8,3802,061
Remove sites with annotated pocket residues lacking coordinates, then remove proteins with no usable sites8,3482,053
Remove sites with empty original ligands8,3482,053

HOLO4K in the paper refers to HOLO4K-sc. APObind and LIGYSIS are not included in the COACH/HOLO exclusion rules above. We emphasize that, in the revised paper, all baselines are retrained, validated, and tested on unified dataset splits (including OOD evaluation), so the comparison is fair.

Q3: Baseline sources, training configurations, and comparison fairness

The baseline values in the original Table 1 were cited from the UniSite paper; in the revised paper, the models are retrained on the same training/validation protein lists, with unified external-test lists, labels, and scoring. In addition, none of the rerun models uses released weights (all are trained from scratch). All baselines use their official default training configurations, except VNEGNN. Because VNEGNN is prone to training instability at the official learning rate of 1e-3, we reduced it to 1e-4 and found that this led to substantially better performance than that reported in the UniSite paper.

Bold and italicized values indicate the best and second-best results within the same comparison group, respectively, with ties marked at the precision shown in the table.

Table R2. Common-retraining results and results reported in the UniSite paper. Each cell lists AP₀.₃ / AP₀.₅ / DCC / DCA, in %. Only the common-retraining rows are included in the rankings.

MethodSourceCOACH420HOLO4K-sc
P2RankUniSite paper61.88 / 26.18 / 46.43 / 74.1160.11 / 26.25 / 53.00 / 81.88
P2RankCommon retraining62.46 / 23.80 / 50.00 / 77.9860.22 / 24.15 / 53.84 / 81.60
GrASPUniSite paper71.50 / 49.14 / 48.51 / 76.2066.68 / 41.26 / 51.31 / 74.16
GrASPCommon retraining71.85 / 47.16 / 50.30 / 77.3870.91 / 43.48 / 58.10 / 83.43
VN-EGNNUniSite paper26.37 / 11.38 / 54.46 / 75.3026.06 / 13.46 / 58.61 / 69.99
VN-EGNNCommon retraining49.64 / 22.31 / 58.63 / 82.1447.85 / 18.67 / 61.33 / 77.53
UniSite-3DUniSite paper71.96 / 39.77 / 47.02 / 73.8170.91 / 54.46 / 57.16 / 78.79
UniSite-3DCommon retraining66.45 / 32.54 / 37.80 / 69.3567.22 / 47.18 / 53.37 / 76.54
CQFinerCommon retraining77.75 / 56.65 / 64.29 / 82.7475.84 / 59.36 / 72.66 / 85.35

CQFiner retains its advantage with the same training membership and test protocol. Descriptions of all baselines have also been provided in the Appendix of the revised paper.

AuthorsOpenReview ↗

Reply of Q4-Q7

Q4: How are different outputs converted? Do the AP gains come from masks or localization?

The AP gains under strict IoU thresholds largely come from learned pocket shapes, while localization ability is measured separately by DCC/DCA. To illustrate this, we fix CQFiner's mask-candidate centers and scores and replace only the learned masks with 9 Å geometric masks around the same centers.

Table R3. AP comparison between learned masks and 9 Å geometric masks with the same centers (in %).

DatasetLearned-mask AP₀.₃ / AP₀.₅9 Å geometric-mask AP₀.₃ / AP₀.₅
COACH42077.75 / 56.6577.28 / 30.60
HOLO4K-sc75.84 / 59.3675.52 / 29.22

AP₀.₃ remains similar while AP₀.₅ drops substantially, indicating that learned masks improve fine-grained pocket boundaries. The output conversions for each method are as follows. These implementations all follow the designs in the original UniSite paper.

Table R4. Localization centers and residue-mask conversion for each method.

MethodLocalization centersResidue masks
CQFinerQuery centers after scoring and spatial NMSNative aggregation of learned masks
UniSiteOfficial mask-atom/convex-hull centersNative site masks
VN-EGNNNative predicted centers and clusteringResidues with any heavy atom within ≤9 Å of a center
P2RankNative pocket centersNative residue_ids
GrASPNative atom clustering and convex-hull centersClustered atoms mapped to residues

Q5: Do more queries increase false positives?

When the number of returned candidates is fixed, more queries improve both recall and precision; when all candidates are returned, there is a precision cost. On the same 200 proteins from each of the two benchmarks, with the default model and FPS sampling fixed, increasing inference queries from 32 to 200 gives the following results.

Table R5. Recall and precision when inference queries increase from 32 to 200 (in %).

DatasetRecall@5Precision@5Recall@AllPrecision@All
COACH42068.51 → 74.0416.10 → 17.4068.94 → 77.455.45 → 2.56
HOLO4K-sc69.63 → 75.5618.80 → 20.4070.74 → 78.526.44 → 2.87

The metrics use one-to-one site matching at IoU≥0.5. Therefore, the practical value of scaling queries can be seen in finding more true pockets while still returning only 5 candidates.

Q6: Distinguishing feature similarity for background, same-site, and different-site queries

After grouping, high similarity remains between different sites and is not caused solely by the large number of background queries or duplicate predictions for the same site. For the original offset model in Figure 1b, the overall cosine similarity of 0.711 breaks down into 0.964 for the same site, 0.883 for different sites, and 0.708 for background. With local supervision and then the addition of a local graph, different-site similarity decreases to 0.620 and 0.549, respectively.

For the supplementary grouped analysis of Figure 5a, we fix 200 internal validation proteins, assign queries whose initial positions are within 10 Å of ligand heavy atoms to the nearest site, and treat the remaining queries as background. The final-layer results are shown below. These results maintain the original conclusion that full connectivity and All Query supervision lead to stronger over-smoothing, whether considering same-site, different-site, or background queries. In particular, in the FullConn-All Query variant, the similarity between sites and background is substantially stronger, reaching 0.7, indicating that severe over-smoothing has made even the query features corresponding to background and sites similar.

Table R6. Final-layer feature cosine similarity for different query groups.

ConfigurationSame siteDifferent sitesBackgroundSite and background
Local + NearQuery0.7880.5910.5200.269
FullConn + NearQuery0.7920.6050.5220.310
FullConn + AllQuery0.8570.7440.7980.728

Q7: Is loss allocation sufficient to demonstrate “wasted capacity”?

We agree that the proportion of loss alone is insufficient to determine whether capacity is wasted. We have changed this wording to directly observable facts: queries initialized far away account for more supervision terms and contribute a larger share of the displacement loss. The performance effect of local supervision is demonstrated by the Full + Near versus Full + All comparison.

AuthorsOpenReview ↗

Reply of Q8-Q12

Q8: The offset baseline needs a broader query range

We have added independent training with 64 queries, and observed no further improvement when increasing from 32 to 64 queries. The figure has been updated accordingly, while retaining the original COACH420 (287 cases) and HOLO4K-sc (1,629 cases) evaluation cohorts.

Table R7. Localization performance of independently trained offset baselines with different numbers of queries (in %).

Number of training queriesCOACH420 DCC / DCAHOLO4K-sc DCC / DCA
865.18 / 83.3368.70 / 80.57
1663.99 / 80.0668.70 / 82.17
3264.88 / 85.7172.03 / 85.83
6460.42 / 84.2367.01 / 83.62

Q9: Does the default 200-query model still benefit with FPS? What about accuracy and latency?

The default model trained with 200 queries still benefits under FPS, and the gains carry over to a limited candidate list. With fixed weights, we obtain the following results on the same 200 proteins from each benchmark. Overall, thanks to CQFiner's design of independent queries, both FPS and random sampling can be used without significantly affecting model performance.

Table R8. Accuracy and runtime of the default model with different numbers of inference queries. AP₀.₅ and Recall@All are in %, and lower runtime is better.

DatasetNumber of inference queriesAP₀.₅Recall@AllMean inference time
COACH4203255.6768.9446.49 ms
COACH42020057.7777.4586.19 ms
HOLO4K-sc3255.3770.7442.99 ms
HOLO4K-sc20058.5578.5289.07 ms

Q10: Random surface-query and random-ranking baselines

We have added random localization under the same prediction budget and postprocessing, as well as random ranking with fixed masks. On 200 proteins from each of the two benchmarks, using 200 unlearned random surface queries, random scores, and the same center NMS and top-n protocol, the results over 10 repetitions are as follows.

Table R9. Localization performance of random surface queries and CQFiner (in %).

DatasetRandom-query DCC / DCACQFiner DCC / DCA
COACH4200.94±0.48 / 3.32±1.3465.96 / 82.13
HOLO4K-sc0.89±0.58 / 3.30±0.8173.70 / 87.41

When the learned masks are fixed and candidate scores are randomly shuffled within each protein, the 200-query model achieves AP₀.₅ of 3.33±0.53% and 3.60±0.43% on COACH420 and HOLO4K-sc, respectively (20 repetitions, mean±standard deviation).

Q11: Adding P2Rank and GrASP to the structural-robustness comparison

We have added P2Rank and GrASP trained on the common training set, and the results show that different methods have advantages under different structural conditions. The following uses 200 common proteins selected in advance for each condition, except LIGYSIS, which uses all 2,775 proteins and 6,882 sites, with APObind holo/apo kept paired; all three methods have completed evaluation. Each cell lists AP₀.₃ / DCC / DCA.

Table R10. Performance of the three methods under different structural conditions (in %).

Structural conditionCQFinerP2RankGrASP
AF-COACH42074.42 / 65.68 / 85.1756.32 / 42.37 / 73.3166.56 / 46.19 / 79.24
AF-HOLO4K-sc77.53 / 76.61 / 86.2956.48 / 51.61 / 79.0370.50 / 61.29 / 82.66
APObind holo61.69 / 52.00 / 72.0054.81 / 48.50 / 67.5066.19 / 50.50 / 78.00
APObind apo53.92 / 47.50 / 64.5039.05 / 35.00 / 59.0049.10 / 39.00 / 63.50
LIGYSIS26.27 / 21.13 / 32.7816.83 / 15.05 / 28.9719.51 / 14.97 / 25.59

Q12: Do local queries affect the identification of cryptic or allosteric pockets?

We have added CryptoBench results, comparing against the released UniSite-3D weights on 196 common proteins, showing that CQFiner can also identify some cryptic pockets. Despite the loss in accuracy, CQFiner still outperforms UniSite-3D. This evaluation concerns cryptic pockets; allosteric pockets have not been evaluated separately.

Table R11. Prediction performance on the common CryptoBench evaluation set.

ModelAP₀.₅ (%)DCC (%)DCA (%)
CQFiner14.5237.2451.53
UniSite-3D13.5526.0242.35
Reviewer_SVZuOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces CQFiner, a query-based framework for protein–ligand binding-site prediction that treats the protein as a read-only context. It combines local residue-to-query message passing and localized displacement supervision with a coarse-to-fine architecture: ViSNet extracts geometric features and predicts coarse query displacements, followed by EGNN-based refinement using fixed protein features. The model outputs pocket centers, confidence scores, and site-specific residue masks. Query independence enables inference with different query counts and sampling distributions. Evaluation covers COACH420, HOLO4K, APOBind, AlphaFold-predicted structures, and LIGYSIS, with substantial improvements reported on the main benchmarks. Ablations and controlled experiments examine query scaling, oracle-prior sampling, feature diversity, loss allocation, and memory consumption.

Critical strengths
  1. A coherent architecture that enables flexible query-based inference. Local connectivity and localized displacement supervision address the computational and optimization costs of refining distant queries, while the read-only protein context prevents queries from modifying protein or other query representations. This gives a clear architectural basis for changing query counts and sampling distributions at inference. The controlled experiments support this property, making the contribution consequential beyond improvements under a single fixed query configuration.

  2. Substantial empirical improvements on the main benchmarks. CQFiner improves DCC@4 over the strongest included baseline by 18.1 percentage points on HOLO4K and 19.0 points on COACH420, alongside improvements in residue-mask AP. The gains across both localization and mask-quality metrics provide meaningful evidence of improved pocket prediction. Evaluation on apo and AlphaFold-predicted structures, together with component ablations, further strengthens the empirical case for the overall design.

Critical weaknesses
  1. The mechanisms underlying the main gains are not sufficiently isolated. The “FullConn + All Queries” ablation simultaneously changes graph connectivity and displacement supervision. Although it supports the combined design, it does not establish how much each change contributes to accuracy, feature diversity, or optimization behavior. Separating these factors is important because the paper’s contribution includes a mechanistic explanation of why existing query-based approaches fail to benefit from higher resolution.

  2. The practical significance of test-time scaling remains incompletely established. Increasing inference queries from 32 to 800 improves AP₀.₅ by 2.85 percentage points and largely saturates around 200 queries, using a separate training configuration and evaluation subset. The paper does not establish the corresponding accuracy–latency trade-off, while prior-guided sampling is evaluated using ground-truth site locations. These experiments support robustness to changes in query sampling, but provide more limited evidence for the practical benefits of the headline scaling capability.

Questions for authors

Would a fair comparison with AF2BIND be feasible on the same benchmark?

3 comments
AuthorsOpenReview ↗

Reply of W1

The reviewer acknowledged CQFiner's flexibility in changing query number and distribution at inference time and its substantial benchmark improvements. However, the reviewer raised concerns about component ablations of full connectivity and local supervision, the practical significance of test-time scaling, and comparison with AF2Bind. We first conducted more complete ablations to explain why we treat full connectivity and all-query supervision as a whole and how supervision determines problems such as over-smoothing; we then explained test-time scaling's practical value through protein-length scaling and recall; finally, we compared against AF2Bind, which still supports CQFiner's performance advantage.

W1: FullConn + All Query changes both connectivity and supervision scope, making their respective contributions difficult to distinguish

Q: The “FullConn + All Query” ablation changes both graph connectivity and displacement supervision. Although it supports the combined design, it does not determine how much each change contributes to accuracy, feature diversity, or optimization behavior. Distinguishing these factors is important because the paper's contributions include a mechanistic explanation of why existing query-based methods fail to benefit from higher resolution.

A: We thank the reviewer for the constructive comments. We combine the two in ablations because they substantially affect training only when present together. Experiments show that using FullConn without All Query (long-range displacement supervision) makes the model spontaneously lower distant-edge attention scores, yielding results nearly identical to a cutoff graph. Full connectivity determines how much information each query receives, but supervision primarily determines whether it is used and causes over-smoothing or training-dynamics problems. We therefore treat the two as a whole, but have also added experiments separating them, as follows.

“FullConn+All Query” means using a fully connected graph between residues and queries and applying displacement supervision to all queries. In contrast, CQFiner's default configuration uses a cutoff-based local graph, in which each query receives messages only from neighboring residues; at the same time, only queries already near a site are supervised. The two are not entirely independent: without a fully connected graph, displacement supervision cannot be applied to all queries. This is because each query receives messages only from nearby residues and therefore cannot predict a displacement toward a distant site, making training meaningless in this setting. In fact, this is one of the most important reasons why VNEGNN uses a fully connected graph, because VNEGNN must predict long-range displacements; on the other hand, consider the case where a fully connected graph is used but displacement supervision is applied only to local queries. We conducted additional ablation experiments, as shown in Table R1. The results show that CQFiner's performance does not change significantly. The over-smoothing metrics in this setting are also similar to those of the local graph. This is because, when only queries already near a site are supervised, the model spontaneously learns to make greater use of edges between queries and neighboring residues while reducing the weights of edges to more distant residues. This result can also be demonstrated by ViSNet's inherent attention scores: the attention scores of local connections clearly decay faster with edge length. In summary, FullConn and All-Query are a coordinated design following VNEGNN's practice, and the two are not entirely independent. Although full connectivity leads to substantially higher GPU memory consumption, it is not a sufficient cause of performance degradation; problems such as over-smoothing and training loss imbalance occur and significantly reduce performance only when the model uses a fully connected graph and the training task also forces each node to use information from distant atoms.

Bold and italicized values indicate the best and second-best results within the same comparison group, respectively, with ties marked at the precision shown in the table.

Table R1. Ablation experiments on full connectivity and the scope of the supervision task

ConfigurationValidation AP₀.₅ (%) ↑Mean cosine similarity of query features ↓Peak training GPU memory (GiB) ↓
LocalConn + Near Query51.670.4691.85
FullConn + Near Query50.730.48315.30
FullConn + All Query41.820.78215.29

Table R2. Attention scores for edges of different lengths (absolute SiLU modulation before the final cosine envelope, normalized by the scores for 0–10 Å)

Edge lengthFullConn + Near QueryFullConn + All Query
<10 Å1.0001.000
10–20 Å0.6490.715
20–40 Å0.3920.489
40–60 Å0.3150.481
AuthorsOpenReview ↗

Reply of W2

W2: The practical significance of test-time scaling remains insufficiently established

Q: Increasing the number of inference queries from 32 to 800 improves AP₀.₅ by 2.85 percentage points, with performance largely saturating around 200 queries, using a separate training configuration and evaluation subset. The paper does not establish the corresponding accuracy–latency trade-off, while prior-guided sampling is evaluated using ground-truth locations. These experiments support robustness to changes in query sampling, but provide limited evidence for the practical benefits of the scaling capability highlighted in the title.

A: We thank the reviewer for the constructive comments. We apologize for not explaining the practical value of test-time scaling in sufficient detail, and describe several practical applications of this capability below. We added pocket-recall and protein-length OOD experiments to demonstrate the value of both scaling the number of queries and scaling the size of protein inputs. Increasing inference queries improves pocket recall by approximately 8 percentage points, at an additional cost of about 40–46 ms per protein; moreover, CQFiner has stronger protein-length generalization than UniSite. With comparable in-distribution validation accuracy, CQFiner's AP₀.₅ on longer OOD proteins remains 11.04 percentage points higher than UniSite's. None of these experiments relies on ground-truth pocket locations to guide inference.

(1) Protein-length scaling. CQFiner inherently generalizes across protein lengths. This is because, regardless of protein length, each query receives messages only from neighboring residues rather than all residues. In contrast, in many current state-of-the-art models, such as UniSite and VNEGNN, each query interacts with the entire protein sequence, and therefore suffers from domain-generalization problems and loses accuracy when protein length changes, similar to the context-extension problem encountered by LLMs (considering UniSite's Transformer architecture). To illustrate this, we selected the longest 10% of samples by input residue count as the test set, and split the remaining shorter proteins into training and validation sets at an 8:1 ratio. Next, we deliberately controlled validation accuracy to ensure comparable accuracy for UniSite and CQFiner, and added VNEGNN results on the same split. Finally, we evaluated all three models on the OOD test set to accurately assess their generalization ability. The results show that CQFiner still has the strongest generalization ability. These results demonstrate CQFiner's ability to generalize to proteins longer than those in the training set.

Table R3. Protein-length OOD generalization.

ModelID validation AP₀.₅ (%)Long-protein OOD AP₀.₅ (%)
VNEGNN16.176.56
UniSite44.0311.90
CQFiner42.5022.94

VNEGNN uses 10 neighbors, 8 queries, and the final weights from epoch 110.

(2) Improved recall. Increasing the number of queries allows more true pockets to be found. AP is affected by candidate coverage, false positives, and ranking, so we have added a recall metric that directly measures pocket coverage. With the default model weights and FPS sampling fixed, the results on the same 200 proteins from each of the two test sets are as follows.

Table R4. Effect of the number of inference queries on pocket recall.

DatasetNumber of inference queriesAP₀.₅ (%)Pocket recall, IoU≥0.5 (%)
COACH4203255.6768.94
COACH42020057.7777.45
HOLO4K-sc3255.3770.74
HOLO4K-sc20058.5578.52

Increasing queries from 32 to 200 improves recall by 8.51 and 7.78 percentage points, respectively. Even when only the top five pockets are retained for each protein, recall still improves by 5.53 and 5.93 percentage points, respectively. Therefore, the gains from additional queries carry over to a limited candidate list, providing direct value for pocket screening where missed pockets need to be reduced. In contrast, when the trained weights of VN-EGNN with 10 neighbors are fixed and the number of queries increases from 8 to 64, pocket recall on the two sets decreases by 3.40 and 4.44 percentage points, respectively, reflecting the different responses of the two architectures to changes in the number of inference queries. In practice, increasing CQFiner from 32 to 200 queries increases mean inference time from 46.49 ms to 86.19 ms and from 42.99 ms to 89.07 ms, respectively. Considering the recall gains and their subsequent saturation, 200 queries is a practical accuracy–runtime operating point in the current experiments.

In summary, these experiments highlight the benefits of ensuring independence among queries and requiring each query to focus only on a local region while increasing spatial resolution, which not only improves recall but also allows protein size to scale.

AuthorsOpenReview ↗

Reply of W3

W3: Is a fair comparison with AF2BIND possible?

Q: Is a fair comparison with AF2BIND feasible under the same benchmark?

A: We thank the reviewer for the constructive comments, and have added comparisons between CQFiner and AF2Bind on COACH420 and HOLO4K-sc. It should be emphasized that AF2Bind is essentially similar to linear probing, prioritizing strong interpretability and high training efficiency, so its accuracy may be lower than that of other approaches that primarily prioritize accuracy. Using AF2BIND's official released weights and pocket-clustering procedure, CQFiner achieves higher AP₀.₃ and AP₀.₅ on both test subsets.

Table R5. Comparison between AF2BIND and CQFiner on the same valid subsets.

DatasetAF2BIND AP₀.₃ / AP₀.₅ (%)CQFiner AP₀.₃ / AP₀.₅ (%)
COACH42051.32 / 6.7178.05 / 56.86
HOLO4K-sc47.67 / 3.8276.39 / 60.36

Note that the two models use the same valid subsets here: 284/287 proteins and 333 sites for COACH420, and 1,575/1,633 proteins and 2,067 sites for HOLO4K-sc. The remaining 61 cases in which AF2BIND inference failed were not included in this subset comparison.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Update of the paper

We first sincerely apologize to the reviewers, area chairs, and readers for the insufficient checks of data splits, data processing, and experimental comparisons in the original manuscript, and for the resulting inaccurate results and additional reviewing burden. We should have thoroughly checked these issues before submission, and we take responsibility for them.

Following Reviewer CdBS's comments, we identified the following issues when re-examining the data-splitting procedure. First, we previously used the default training/validation split of UniSite-DS, but this default split retained test proteins that should have been excluded, which would overestimate model performance. Second, some baseline values in the original manuscript were directly cited from existing papers, and their training conditions were not sufficiently aligned with those of our model. In addition, our label conversion and metric definitions for some OOD data were inconsistent with UniSite's, leading to slight differences in accuracy. These issues have been corrected, and the corresponding implementation details have also been added to the Appendix of the paper. To address these issues, we corrected the data-filtering procedure according to the official benchmark lists, ultimately used a common split containing 8,348 training proteins and 2,053 validation proteins, retrained CQFiner, UniSite-3D, VN-EGNN, P2Rank, and GrASP, and unified the test lists, evaluation labels, and scoring protocols used in the comparisons. The sources of retained literature results have been separately marked, clearly distinguishing them from the results obtained through common retraining. We have accordingly updated the results of the main benchmark, OOD, and ablation experiments, along with the corresponding performance comparisons and analyses in the main text. The relevant values and margins of improvement in the original manuscript should be superseded by those in the revised manuscript.

Alongside these corrections, we have also added the following content in response to the reviewers' comments:

  • Extended the motivation experiments to an independently trained 64-query setting and added Local+Near, Full+Near, and Full+All comparisons to further analyze the joint effects of connectivity and supervision scope.
  • Added pocket recall and inference time for test-time scaling, as well as protein-length OOD experiments and a PLIP-based interaction-similarity analysis.
  • Clarified the scope of the spatial-resolution argument and added experiments on spatially anchored UniSite to the appendix.
  • Provided detailed descriptions in the appendix of data filtering, label and center definitions, baseline sources and training configurations, prediction-output conversion, and the scope of each evaluation set.

We thank the reviewers for their careful review and specific suggestions, which helped us identify and correct these issues. By clearly describing the errors, the correction process, and the scope of the updates, we hope to make the revised experimental results and the basis for comparison clearer and easier to verify.

AuthorsOpenReview ↗

Summary of the discussion

We sincerely thank the Area Chair and reviewers for their careful evaluation and constructive feedback. Below, we summarize the key discussions and revisions during the rebuttal period.

Reviewer SVZu raised concerns about the insufficient isolation of individual module contributions in our ablations, the practical significance of test-time scaling, and the comparison with AF2Bind. We clarified why only combining the two modules (fully connected graphs with supervision over all queries) leads to issues such as over-smoothing, and added ablations isolating each component. Additional experiments on protein-length OOD generalization and recall demonstrated CQFiner's greater robustness to longer proteins compared with SOTA methods such as UniSite, as well as improved recall with increasing query counts. New comparisons with AF2Bind further confirmed CQFiner's performance advantages.

Reviewer CdBS focused on dataset splitting, baseline configurations, and experimental details. Upon re-examination, we confirmed a data leakage issue and documented our data processing pipeline in detail in the rebuttal, public comment, and revised appendix. To ensure fair comparisons, we independently reproduced all baselines, achieving performance comparable to or better than the originally reported results (e.g., VNEGNN). With these revisions, CQFiner achieves improvements of 4.93/12.18/11.33/1.92 and 5.90/9.49/5.65/0.60 percentage points on HOLO4K-sc and COACH420, respectively ([email protected]/[email protected]/DCC/DCA).

Reviewer rTni questioned the applicability of our spatial resolution argument and the sources of OOD generalization. To address the lack of explicit spatial coordinates in DETR-style queries, we extended CQFiner's design to a DETR-style framework using residue-based query initialization, achieving improved performance and demonstrating its broader applicability. Further experiments showed that even when controlling for in-distribution accuracy, CQFiner maintains superior OOD generalization across both protein length and pocket interaction patterns. We also analyzed differences in data organization between UniSite and earlier datasets such as scPDB, identifying substantial distribution shifts that help explain the generalization difficulties on these earlier benchmarks.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.