acceptodds
Under review as a conference paper at ICLR 2027

zPocket: Better Pocket Finding with Frozen Cofolding Features

Abstract

Identifying protein-ligand binding sites (pockets) is important for understanding protein function and developing drugs. However, no public method has leveraged cofolding models for novel pocket prediction; furthermore, this task lacks a suitable benchmark as existing benchmarks are within the training distribution of cofolding models. We therefore developed PocketWatch, a curated dataset of biologically relevant ligand pockets. The PocketWatch test set consists of unseen pockets stratified by sequence similarity, structural similarity, and crypticity to enable evaluation of generalization to novel and challenging pockets. Using this dataset, we develop zPocket, a method that trains pocket-prediction heads on top of frozen features from cofolding models. We show that zPocket heads outperform baseline methods across all stratifications. Our best zPocket model achieves overall per-target AUPRC of 0.757 compared to 0.626 for the strongest baseline, VN-EGNN, retrained on the PocketWatch split (0.544 when using the published weights); the widely used P2Rank reaches 0.503. We further show that zPocket can be used to outperform baseline methods on top- recall of predicted pocket centroids through clustering per-residue predictions or using zPocket to rescore centroids proposed by other models. Finally, we explore two applications of zPocket: crystallographic fragment screens and a proteome-scale scan for uncharacterized pockets on human proteins. Together these results suggest that zPocket heads built on cofolding trunk features can greatly advance ligand binding site prediction.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
binding site prediction, protein–ligand interactions, cofolding models, benchmark and temporal splits

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

This paper introduces zPocket, a method that predicts protein–ligand binding pockets using lightweight prediction heads trained on frozen representations from cofolding models. The authors also develop PocketWatch, a benchmark designed to evaluate generalization to novel binding pockets. On 1275 test proteins covering 1562 pockets, zPocket achieves a per-target AUPRC of 0.757, outperforming the several retrained baselines, such as 0.626 for VN-EGNN, and 0.503 for P2Rank. Combining zPocket with existing pocket detectors further improves pocket localization, reaching a top-(N+2) recall of 0.666. The authors also evaluate the method on 43 fragment-screening pockets and apply it to approximately 20000 human proteins, identifying 155 potentially novel pockets after filtering.

Critical strengths
  1. I think the methodology proposed by the authors is rational. Existing Cofolding models, trained via large-scale pretraining, may have already learned extensive knowledge of protein-small molecule interactions. Directly leveraging the internal representations of these models to predict binding pockets represents a reasonable and practically valuable research direction. Notably, the model can produce residue-level Pocket Probability without requiring explicit ligand input. Furthermore, the empirical results are good. After retraining VN-EGNN and UniSite on the same dataset, zPocket still exhibits clear advantages. Consistent observations across different frozen backbones also strengthen the credibility of the conclusions.

  2. I think the PocketWatch benchmark is potentially a more valuable contribution than the model design itself. Rather than simply improving prediction performance, the authors address an important question in the era of cofolding models: whether a model can truly identify novel binding pockets or is simply recognizing patterns already seen during pretraining. PocketWatch uses temporal splitting and stratifies the test set by sequence similarity, structural similarity, and crypticity, providing a clearer picture of model performance on pockets with different levels of difficulty. The authors also retrain several baselines and explore applications in fragment screening and proteome-scale pocket discovery. Overall, I appreciate the comprehensive experimental design and the effort to evaluate the model's generalization ability.

Critical weaknesses
  1. This is my main concern with the paper. Although the authors emphasize novel pocket prediction, 71.6% of the test pockets have high structural similarity to previously known binding sites. Performance drops substantially on pockets with little structural precedent, where the improvements over baselines become smaller and some confidence intervals overlap. Therefore, the overall AUPRC of 0.757 may not fully reflect the model's ability to discover genuinely novel pockets. The evaluation of cryptic pockets is also limited by the small sample size: only 40 of the 1,562 test pockets are labeled as cryptic, and just 20 of these lack holo-like conformations in the predicted structures. I believe more convincing evidence on genuinely novel and challenging cryptic pockets is needed to support the paper's main claims.

  2. While the overall research direction is promising, I feel that the proposed methodology offers limited technical novelty. Leveraging pair representations from pretrained structure prediction models and training lightweight prediction heads using existing architectures is a relatively straightforward approach. Similar ideas have also been explored in prior work, such as AF2Bind. Although zPocket achieves impressive results, most of the compared baselines follow different technical approaches, and comparisons with closely related methods that also leverage representations from pretrained structure prediction models are limited. The current ablations demonstrate the benefits of pair representations and different head architectures, but do not fully clarify where the improvements come from. I believe the authors should better highlight the unique technical contributions of zPocket and provide stronger comparisons to distinguish methodological improvements from the advantages of using powerful pretrained cofolding models.

Questions for authors
  1. Since leveraging pretrained pair representations for downstream prediction tasks is not entirely new, could the authors further clarify the main methodological innovations of zPocket? In particular, how much of the performance improvement comes from the pretrained cofolding representations versus the prediction head architecture? Additional ablations using alternative representations, such as predicted structures or distogram features, could help clarify this distinction.

  2. The authors compare against UniSite-1D but do not include UniSite-3D, which incorporates structural information and supports both residue-level and pocket-level predictions. In addition, AF2Bind, which also leverages pair representations from pretrained structure prediction models for binding-site prediction, is discussed in the paper but not included in the experiments. Could the authors clarify why these methods were excluded and, if feasible, provide additional comparisons? I believe these comparisons would help better establish the performance advantages and technical contributions of zPocket.

2 comments
AuthorsOpenReview ↗

Reply to Reviewer vwoU

Thank you for the detailed suggestions. The requested ablations and baselines are now in the paper. We also give the headline numbers here and cite the common response for the full tables.

W1: Novel and cryptic pockets. We agree our “novel” pocket set is dominated by structurally similar pockets, and our goal in creating the benchmark was to make performance as a function of structural similarity visible to the field. Excluding the high-coverage bin leaves 443 pockets, on which zPocket reaches 0.445 per-pocket AUPRC versus 0.337 for retrained UniSite-3D, the strongest baseline: a paired gain of +0.109 [95% CI: 0.091, 0.126]. In the hardest bin (0–20%; 133 pockets), zPocket leads the strongest baseline, DeepPocket, by +0.084 [0.036, 0.135], Holm-adjusted p = 1.5e-4. Common response §1 reports these paired tests alongside the marginal intervals. We also revised our introduction to surface this point, as obscuring it was never our intention.

We agree that the 40-pocket cryptic subset is too small for a strong generalization claim. Appendix B.9 further shows that much of the apparent crypticity effect is associated with predicted-fold accuracy. We therefore treat these results as preliminary and report the sample sizes explicitly. We also note that the existing works from the industrial labs that report performance on cryptic pockets do not report the size of this evaluation set, and we think one of our contributions is revealing the small number of pockets available for such evaluations, adding more context to the numbers reported from other labs.

W2 / Q1: Technical contribution and ablations. We know similar work using paired representations has been done, but we expand on it in several ways: we create a benchmark suitable for the new generation of models and show their gains in performance, and we show how these per-residue scores can be combined with discrete pocket-finding methods to better prioritize pockets, which is practically useful for the field beyond comparing per-residue metrics. We also added your requested analyses to better understand the contribution of the representations versus the prediction head. The requested distogram control reaches 0.701 per-target AUPRC versus 0.745 for pair features under the same pairformer2 architecture, a paired gain of +0.044 [95% CI: 0.037, 0.052]. A two-block sequence-transformer head reaches 0.633 on ESMC-6B embeddings despite roughly 2.5 times the parameters of the pair head. Even with mixing, sequence-based representations cannot match the structural pair representations, suggesting the pair features truly contain useful binding signal. Common response §3 gives the full ablations.

Q2: UniSite-3D and AF2Bind. Both are now included (common response §2), reaching 0.544 and 0.456 per-target AUPRC with published weights, versus 0.745 for zPocket. We also retrained UniSite-3D on PocketWatch, and it reaches 0.648 [0.631, 0.664], a gain of +0.103 [+0.094, +0.113] over its own published weights, which makes it the strongest baseline in Table 1. zPocket still leads it by 0.098 [0.089, 0.106]. Adding UniSite-3D was the right thing to do, and we are grateful for the suggestion. For AF2Bind, we also added a comparison on AF2Bind's own benchmark: 0.738 Recovery against 0.662 for AF2Bind (also described in common response §2).

The distogram control and the two missing baselines were the right things to ask for, and the paper is stronger for them. We hope these address your concerns and that you will consider raising your score. We would also be glad to discuss any other details further.

We’ve edited our reply above to lay out more clearly how this work extends prior work, to clarify our point about cryptic pockets, and to add how we revised the introduction to surface the point about structurally novel pockets sooner. With the discussion period closing tomorrow, we hope you will review our changes, and if our added experiments address your concerns, that you’ll consider raising your score.

Reviewer_tdN5OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes zPocket, a method that predicts ligand-binding pockets using internal representations from protein structure prediction models such as Protenix-v2 and ESMFold2. The pretrained models are kept frozen, and only a small prediction head is trained to predict pocket probabilities for each residue. The authors also introduce PocketWatch, a new benchmark with a time-based split to reduce overlap with the training data of existing models. The results show that zPocket achieves higher AUPRC than the strongest baseline, VN-EGNN. It also improves performance by reranking pocket predictions from other methods.

Critical strengths
  • The authors developed PocketWatch to distinguish previously known pockets from new pockets that may not have been seen during cofolding model training. This is an important contribution because it allows better evaluation of how well pocket prediction models generalize to new pockets.

  • zPocket achieves better performance than existing methods using pair representations from pretrained cofolding models. It also improves the ranking of pocket candidates predicted by other methods. This suggests that zPocket could be useful in real drug discovery pipelines.

Critical weaknesses
  • The paper does not compare zPocket with several important pocket prediction methods, including AF2Bind and DeepPocket. In particular, AF2Bind also uses internal representations from a pretrained protein structure prediction model. Without a direct comparison, it is difficult to judge the novelty and performance advantages of zPocket.

  • I think the generalization performance on new pockets is still limited. About 71.6% of the test pockets are structurally very similar to previously known pockets. However, the AUPRC drops to 0.347 for structurally novel pockets, and the improvement over baseline methods is not statistically clear. Therefore, I do not think the paper provides enough evidence that zPocket can reliably discover new pockets, which is one of its main goals.

Questions for authors
  • How does zPocket compare with recent methods such as AF2Bind on the same PocketWatch test set? In particular, does zPocket still outperform these methods on structurally novel pockets?
2 comments
AuthorsOpenReview ↗

Reply to Reviewer tdN5

Thank you — both requested baselines are now run, along with the paired statistical comparisons. Full protocols are in the common response; we summarize the results here.

AF2Bind and DeepPocket. We added all of the requested baseline methods: on all 1,275 PocketWatch targets, AF2Bind and DeepPocket reach 0.456 and 0.475 per-target AUPRC respectively, versus 0.745 for zPocket. On AF2Bind's own benchmark, zPocket reaches 0.738 Recovery versus 0.662 for AF2Bind; that benchmark overlaps the training distributions of ESMFold2, Protenix-v2, and AF2Bind, so we treat it as a measure of expressive capacity rather than generalization. Removing the 53 targets overlapping DeepPocket's training set slightly increases zPocket's paired gain in the lowest-coverage bin, from +0.084 to +0.088. Common response §2 details the protocols and overlap checks.

Novel-pocket performance. We addressed the issue of generalization to novel pockets by adding paired comparisons in the low-similarity buckets. In the lowest-coverage bin (0–20%; 133 pockets), zPocket reaches 0.327 per-pocket AUPRC versus 0.243 for DeepPocket, the strongest baseline: a paired gain of +0.084 [95% CI: 0.036, 0.135], Holm-adjusted p = 1.5e-4. Against AF2Bind in that same bin the gain is +0.113 [0.072, 0.156], adjusted p = 4e-5. Below 80% coverage (443 pockets), the gain over AF2Bind is +0.127 [0.104, 0.149]. Common response §1 reports the full analysis. These results establish relative improvements on the evaluated structurally novel pockets, while the low absolute scores remain a limitation. We also revised the introduction so that the small number of structurally novel pockets, and the across-the-board performance drop on them, is stated up front; we think building a benchmark with this stratification is a useful contribution in itself as it made this .

We also updated our introduction to raise the point about the low number of structurally novel pockets, and how performance deteriorates on this set across all of the evaluated methods, while zPocket still outperforms existing methods on it. We think creating a benchmark with this stratification is a useful contribution in itself, since prior works report aggregate performance that obscures this limitation.

We hope this addresses the concern behind your score and that you will consider raising it; we are glad to run anything further during the discussion period.

We’ve edited our reply above for clarity and to add how we revised the introduction to surface the point about structurally novel pockets up front. With the discussion period closing tomorrow, please take a look, and if our new experiments address your concerns, we hope you’ll consider raising your score.

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

This paper presents zPocket, a protein-ligand binding site prediction framework that trains lightweight prediction heads on frozen representations from pretrained cofolding models, including Protenix-v2 and ESMFold2. The authors also introduce PocketWatch, a curated benchmark with temporal splits and stratifications based on sequence similarity, structural similarity, and pocket crypticity. Experimental results show that zPocket outperforms several existing pocket prediction methods, achieving a per-target AUPRC of 0.757 compared to 0.626 for the strongest retrained baseline.

Critical strengths

The paper addresses an important problem in structure-based drug discovery. The proposed approach is straightforward and effective, demonstrating the potential of pretrained cofolding representations for binding site prediction. PocketWatch is also a valuable contribution, particularly its consideration of structural similarity and potential data leakage. The experiments are generally comprehensive, including comparisons with retrained baselines, architectural ablations, and evaluations on challenging pockets.

Critical weaknesses
  1. The generalization ability to genuinely novel pockets remains unclear. As reported in Section 4.1, 71.6% of the test pockets have max_qcov ≥ 80%, indicating that most test pockets have close structural precedents. The per-pocket AUPRC drops substantially to 0.347 for max_qcov below 20%. Although the authors provide stratified evaluations, the overall performance is dominated by relatively familiar pockets. More detailed results and statistical comparisons on low-similarity pockets would help establish the generalization ability of zPocket.

  2. The methodological novelty is somewhat limited, as the proposed approach mainly combines frozen representations from existing cofolding models with lightweight prediction heads. In particular, AF2Bind has already explored using pretrained AlphaFold representations for binding site prediction. The baseline comparison could be strengthened by including AF2Bind.

  3. The current experiments do not fully separate the contributions of pretrained cofolding representations and prediction head architectures. Although different heads are compared, the baselines use different representations and architectures. A more controlled comparison using alternative protein representations with comparable prediction heads would help clarify the source of the performance improvements.

  4. The practical applications remain preliminary. The fragment screening evaluation includes only nine proteins and shows substantially lower performance than the main benchmark. The proteome-scale analysis identifies candidate novel pockets without experimental validation. Therefore, the practical effectiveness of zPocket for discovering genuinely new ligandable sites remains to be established.

Questions for authors
  1. Can the authors provide stronger evidence that zPocket significantly outperforms existing methods on genuinely novel pockets with low structural similarity?

  2. Can the authors provide more controlled experiments to distinguish the contributions of pretrained cofolding representations and prediction head architectures?

  3. What are the structural training data cutoffs for Protenix-v2 and ESMFold2? How do the authors ensure that the test pockets are unseen by the pretrained backbones?

1 comment
AuthorsOpenReview ↗

Reply to Reviewer VQgm

Thank you for the constructive questions, particularly the request to clarify training cutoffs. W1 through W3 overlap with the other reviews, so we summarize the key results here and cite the common response for the full analysis.

W1 / Q1: Generalization. The submission reported marginal bootstrap confidence intervals; the revision adds paired tests. After excluding the dominant high-coverage bin, zPocket reaches 0.445 per-pocket AUPRC on 443 pockets, versus 0.337 for the strongest baseline, retrained UniSite-3D: a paired gain of +0.109 [95% CI: 0.091, 0.126]. In the hardest bin (0–20% coverage; 133 pockets), it scores 0.327 versus 0.243 for DeepPocket, the strongest baseline there: +0.084 [0.036, 0.135], Holm-adjusted p = 1.5e-4. Across the full comparison family, 124 of 126 tests reach significance after correction (common response §1).

W2: AF2Bind and novelty. On PocketWatch, zPocket reaches 0.745 per-target AUPRC versus 0.456 for AF2Bind. On AF2Bind's own benchmark, a single frozen cofolding representation achieves Recovery of 0.738, versus 0.662 for AF2Bind reading the AF2 pair representation. We note the scope of that second result: AF2Bind's benchmark overlaps the training distribution of ESMFold2, Protenix-v2, and AF2Bind, and its splits control only for redundancy within the benchmark itself, so it measures expressive capacity rather than generalization. That is the gap PocketWatch is built to measure, and we read the two results together as the paper's thesis rather than as independent wins.

W3 / Q2: Representation versus head. Holding the readout fixed, Protenix-v2's single representation scores 0.474 per-target AUPRC against 0.669 for its pooled pair representation; for ESMFold2 the corresponding comparison is 0.611 against 0.639. The size of that gap is trunk-dependent, and the stronger evidence is architectural: under the same pairformer2 head, ESMFold2 pair features reach 0.745 versus 0.701 on distogram logits, and a two-block sequence transformer on ESMC-6B embeddings reaches 0.633 despite roughly 2.5 times the parameters of the pair head. These controls support a substantial contribution from the pair representation without claiming to eliminate every architectural difference; common response §3 gives the details and the seven-trunk comparison (0.733–0.757).

Q3: Training cutoffs and unseen pockets. The checkpoints used are Protenix-v2 with a structural training cutoff of 2021-09-30, and ESMFold2 with a reported cutoff of September 2021. We did not use the later Protenix-v2 checkpoint trained through 2025-06-30. These cutoffs are now explicit in the methods; the PocketWatch test period begins on 2024-01-01.

A test chain must contain a pocket site first observed on or after 2024-01-01, and sites observed before that date are masked from scoring. Structural novelty is measured against a Foldseek database restricted to the 182,401 RCSB entries released on or before 2021-09-30. These checks address experimental structural overlap; they do not establish the provenance of undated distillation data, and RF3, one of the seven trunks in Table A10, publishes its cutoff without an exact day. The five additional trunks, with documented structural cutoffs no later than September 2021, score 0.733–0.746 under the same head recipe (Appendix A.5; Table A10).

W4: Applications. We agree that these remain exploratory. The proteome scan generates hypotheses, not experimentally validated hits. The revised fragment-screen evaluation includes 12 proteins and 35 scored sites, adding PDB group depositions and requiring label-transfer quality to pass on both trunks' predicted folds. We retain the lower performance on this task as a limitation and release the pocket set and prediction tool for further evaluation.

The cutoffs you asked about are now explicit in the methods, and the generalization analysis is considerably stronger for your having pressed on it. We would welcome your view on whether anything further would help, and if these revisions address your concerns we hope you will consider raising your score.

Comments on the paper

3 comments
AuthorsOpenReview ↗

Common response (1/2)

We thank all three reviewers. The revised paper adds paired statistical comparisons, the requested baselines, representation controls, and explicit trunk training cutoffs. Below we address the shared concerns, followed by replies to individual questions. We have also uploaded an updated draft of the manuscript; the results are final, though we are still tightening some of the surrounding prose.

Change of main trunk. After our initial submission, Protenix-v2's weight license became more restrictive, so we revised the paper to focus on ESMFold2, which remains open-weight. As a result, most numbers reported in our paper have shifted slightly: the headline per-target AUPRC is now 0.745 rather than 0.757, and the lowest-coverage per-pocket AUPRC is 0.327 rather than 0.347. Furthermore, baselines requiring input structures now use ESMFold2 folds. We made this adjustment to facilitate reproduction and maximize access, but all Protenix-v2 results are still available in the Appendix and remain as additional evidence of the zPocket approach. Unless specified otherwise, results below use ESMFold2.

1. Generalization to structurally novel pockets

We agree that the aggregate score obscures the difficulty of structurally novel pockets. We believe the relevant question is whether zPocket improves on baselines within these less familiar subsets, which we discuss in detail below. Our goal in stratifying by structural similarity was to make the deteriorated performance on structurally novel pockets visible. Previous benchmarks and reported results focused on aggregate performance, and we believe this clearer picture of where the task stands is an important contribution in its own right; we updated our introduction to surface this point so that it does not appear buried.

Figure 1C originally showed marginal confidence intervals for each method. Their overlap does not test the difference between methods evaluated on the same pockets. We now report bootstrap intervals for paired mean differences and two-sided Wilcoxon signed-rank tests across the full 126-comparison family (14 strata/subsets × nine baselines; Appendix B.2).

Removing the dominant, highest-coverage bin leaves 443 pockets with max_qcov < 80%. zPocket achieves 0.445 per-pocket AUPRC on this subset:

BaselineBaseline AUPRCPaired gain over baseline [95% CI]Holm-adjusted p
UniSite-3D, retrained0.337+0.109 [+0.091, +0.126]6e-32
P2Rank0.326+0.120 [+0.097, +0.143]1e-19
AF2Bind0.319+0.127 [+0.104, +0.149]1e-23
DeepPocket0.309+0.136 [+0.109, +0.165]7e-22
VN-EGNN, retrained0.282+0.164 [+0.142, +0.186]8e-46
UniSite-1D, retrained0.265+0.180 [+0.160, +0.200]3e-47

Of those 443 pockets, the 133 least similar to anything pre-cutoff (0–20% coverage). zPocket achieves 0.327 per-pocket AUPRC and leads every baseline:

BaselineBaseline AUPRCPaired gain over baseline [95% CI]Holm-adjusted p
DeepPocket0.243+0.084 [+0.036, +0.135]1.5e-4
P2Rank0.232+0.095 [+0.053, +0.141]5e-4
UniSite-3D, retrained0.230+0.097 [+0.066, +0.131]4e-8
AF2Bind0.215+0.113 [+0.072, +0.156]4e-5
UniSite-1D, retrained0.172+0.155 [+0.118, +0.195]6e-12
VN-EGNN, retrained0.170+0.158 [+0.116, +0.201]5e-11

For the submitted Protenix-v2 configuration, the corresponding P2Rank comparison is 0.347 versus 0.254, paired gain +0.094 [+0.052, +0.137], Holm-adjusted p = 5e-5 within that trunk's own comparison family.

Overall, 124 of 126 comparisons reach significance after correction, and zPocket has a higher mean in every cell. Both nonsignificant comparisons occur in the 20–40% bin (70 pockets): P2Rank, +0.022 [−0.022, +0.072]; and AF2Bind, +0.044 [+0.008, +0.085], both adjusted p ≥ 0.17. The same analysis for the Protenix-v2 configuration leaves a single nonsignificant cell, the same P2Rank comparison in the 20–40% bin. These results support improvement on low-similarity pockets while leaving substantial room to improve absolute performance. The full comparisons appear in Tables A6–A8.

AuthorsOpenReview ↗

Common response (2/2)

2. Requested baseline comparisons

We have added AF2Bind, DeepPocket, and UniSite-3D. We also retrained UniSite-3D on PocketWatch under the same harness we used for UniSite-1D, and the retrained UniSite-3D is now our strongest baseline for the per-residue pocket-prediction task. All methods use the same PocketWatch test targets, masks, and scoring protocol; methods requiring input structures receive the same ESMFold2 folds.

MethodPer-target AUPRC on 1,275 targets
zPocket (ESMFold2, pairformer2)0.745 [0.729, 0.761]
UniSite-3D, retrained on PocketWatch0.648 [0.631, 0.664]
VN-EGNN, retrained on PocketWatch0.598
UniSite-1D, retrained on PocketWatch0.595
VN-EGNN, published0.559
UniSite-3D, published0.544 [0.528, 0.561]
UniSite-1D, published0.531
P2Rank0.477
DeepPocket, published0.475 [0.457, 0.493]
AF2Bind, published0.456 [0.441, 0.471]

AF2Bind. We validated our implementation against its released predictions (median per-protein Spearman correlation 0.991 across 12 proteins). On PocketWatch, zPocket's paired per-target gain is +0.289 [+0.274, +0.304]. AF2Bind is evaluated with published weights, without PocketWatch retraining.

We also evaluated our heads on AF2Bind's own ten-fold benchmark, using its published splits, Recovery metric, and inverse-redundancy weighting (Table A9). Both methods read a single frozen pair representation, so this is a like-for-like comparison.

MethodRecovery on AF2Bind's ten folds
zPocket (Protenix-v2, pairformer2)0.738 ± 0.010
zPocket (ESMFold2, pairformer2)0.694 ± 0.012
AF2Bind (AF2-pair)0.662

Both of our heads also exceed the strongest variant AF2Bind reports (0.690), which adds ESM2 and ESM1-IF embeddings to the pair features. We report against AF2Bind (AF2-pair) as an apples-to-apples comparison of predictors based on co-folding features to the one based on AF2 features. Ten-fold means ± population standard deviation; Table A9 gives the full ladder.

DeepPocket and UniSite-3D. Using published weights, UniSite-3D's structure encoder improves on published UniSite-1D by only +0.013. However, when retrained on PocketWatch, UniSite-3D reaches 0.648 [0.631, 0.664], a paired gain of +0.103 [+0.094, +0.113] over its own published weights. This makes UniSite-3D (PW) the strongest baseline we are able to construct, but zPocket still leads it by 0.098 [0.089, 0.106], paired over the same 1,275 targets, with the lead holding on every metric in Table 1 (AUROC 0.951 vs 0.907, Rp 0.682 vs 0.608).

DeepPocket performs similarly to P2Rank overall and is the strongest baseline in the lowest-coverage bin (§1). We also checked DeepPocket's scPDB training overlap: 53 test targets (4.2%) share a UniProt accession with its training set. Removing them preserves the advantage over DeepPocket and slightly increases the paired gain: +0.084 to +0.088 in the lowest-coverage bin and +0.136 to +0.139 below 80% coverage (Appendix B.5).

3. Representation versus prediction head

Comparisons with existing methods alone cannot distinguish how much of the improvement comes from the cofolding representation versus the prediction head.

Holding the readout architecture fixed. With the same MLP architecture, Protenix-v2's single representation scores 0.474 per-target AUPRC, versus 0.669 for its pooled pair representation, which uses no pair mixing. For ESMFold2, the corresponding comparison is 0.611 for ESMC-6B embeddings versus 0.639 for pooled pair features (Tables 1 and A5).

Alternative representations and a larger head. The new controls in Table 2 and Appendix B.4 give:

RepresentationHeadPer-target AUPRC
ESMFold2 pair tensorpairformer20.745
ESMFold2 distogram logitspairformer20.701
ESMC-6B per-residue embeddingsTwo-block sequence transformer0.633
ESMC-6B per-residue embeddingsMLP0.611

The distogram result shows that predicted geometry carries substantial signal, while the pair tensor improves AUPRC by 0.044 [0.037, 0.052]. The sequence transformer has approximately 2.5 times as many parameters as the pair head; it improves over the MLP by 0.022 [0.018, 0.026] but remains 0.112 behind the pair head. Thus, adding cross-residue mixing and more parameters to this per-residue representation does not close the gap.

Finally, the same pairformer2 recipe scores 0.733–0.757 across seven cofolding trunks (Table A10), above the strongest retrained baseline at 0.648. Head architecture also matters, but gains diminish: adding the second pairformer block improves ESMFold2 AUPRC by 0.008. Together, these controls support a substantial contribution from the pair representation without claiming to eliminate every architectural difference.

Dear Reviewers,

This is a gentle reminder to review the authors’ response at your earliest convenience.

Thank you for your time and consideration.

AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.