acceptodds
Under review as a conference paper at ICLR 2027

CoPocket: Pocket Identification Without Receptor Structures Using Co-Folding Representations

Abstract

Locating ligandable pockets without a known ligand is an early step in drug discovery. However, most pocket predictors need a receptor structure and are benchmarked on ligand-bound conformations that do not exist before a binder is found. Sequence-based methods avoid this requirement, although UniSite's sequence-only variant underperforms its structure-informed counterpart. Benchmarks also conflate new structures of known sites with newly discovered pockets, and compute residue-level and pocket-level metrics from different outputs. CoPocket predicts pockets from sequence and multiple sequence alignment, without receptor structures or a ligand. Its set-prediction decoder maps frozen co-folding residue and pair representations to scored masks, evaluated at both pocket and residue levels. We also build the ChronoPocket dataset, which consolidates over 230,000 PDB ligand contacts into 37,269 pockets and splits them by the date each pocket was first observed. Its ChronoPocket-Discovery subset holds out proteins with previously known sites and additional pockets first observed after the training cutoff. On the temporal test set, CoPocket reaches a pocket average precision of 0.430 at an intersection-over-union threshold of 0.5, compared with 0.230 for an ESM C-based sequence-only baseline. On UniSite test proteins absent from our training data, it outperforms the structure-based UniSite-3D (0.347 vs. 0.303). On ChronoPocket-Discovery, it recovers 20.5% of later-observed pockets in its top five predictions, a challenging reference for pocket discovery.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Protein pocket identification, Ligand-free prediction, Co-folding representations, Set prediction, Temporal benchmarking

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_5o3yOpenReview ↗
Rating
3
Confidence
2
Informativeness
3(+1)
Presentation
2
Soundness
3(+1)
Read review
Paper summary

This paper proposes CoPocket, a pocket identification framework using protein sequence and MSA with a co-folding pretrained representation, without the need for the structure. The authors also introduce ChronoPocket, a temporal benchmark that distinguishes newly observed pockets from repeated observations of known binding sites. Overall, the paper addresses a meaningful problem and demonstrates promising experimental results to support the main claim.

Critical strengths
  1. The problem is meaningful, predicting pockets without requiring receptor structures or known ligands is valuable for early-stage drug discovery.
  2. This work builds a large-scale dataset ChronoPocket for the pocket identification task with temporal evaluation and protein-centered labels.
  3. CoPocket outperforms the ESMC-based baseline on ChronoPocket and achieves better performance than UniSite on the evaluated external subset. The evaluation covers both pocket-level and residue-level metrics.
  4. The motivation in this paper is relatively clear, and the figure design is elaborate.
Critical weaknesses
  1. The description and the contribution of the co-folding pretrained model are still somewhat not clear. I understand that the co-folding model, such as IntelliFold V2-Flash, is pretrained on protein-ligand complexes and also the protein-only data, therefore, it may contain contact information for the protein input. The paper emphasizes the use of co-folding representations, but CoPocket only extracts protein-only representations from IntelliFold V2-Flash without ligand. And the reported improvements demonstrate the effectiveness of the overall CoPocket system but do not establish how much of the gain is attributable specifically to the pretrained co-folding representations. It remains unclear whether the effectiveness of these representations is specifically related to protein–ligand interaction knowledge acquired during co-folding pretraining, or primarily to general protein structural information. Although the paper compares the performance between IntelliFold V2-Flash and ESMC, a pure PLM, the contribution is not explicitly and directly evaluated.

  2. The proposed framework demonstrates promising results using pretrained representations from IntelliFold V2-Flash. However, it remains unclear whether the observed effectiveness generalizes to other co-folding backbones. Evaluating CoPocket with an additional pretrained co-folding model, such as Boltz-1 or Protenix, would further strengthen the core claim of this paper.

  3. The contributions of several proposed components, such as pairformer, are not independently evaluated. Moreover, it seems that removing the pair-derived attention bias does not consistently reduce performance. This diminishes the contribution of the method, considering that the significant information mainly comes from the use of a pretrained co-folding model.

  4. Although recovering later-observed pockets is an important contribution, the discovery evaluation does not include comparisons with existing methods. Without such baselines, it is difficult to assess the significance of CoPocket especially for discovering additional pockets.

Questions for authors
  1. The proposed framework is evaluated using only IntelliFold V2-Flash as its co-folding backbone. Testing additional co-folding models would help assess whether the approach generalizes across different co-folding pretrained representations.

  2. Some experimental results or analysis to clarify other weaknesses would be good, if possible.

7 comments

Dear Reviewer 5o3y, thank you very much for your time and insightful comments. In the revised manuscript, all updated and newly added content is highlighted in blue.

Below, we address each of your weaknesses and questions point by point.

W1. When extracting representations from ligand-free proteins alone, to what extent do the gains come specifically from co-folding pretraining rather than generic protein structural information or downstream modeling?

We tested both questions under an identical decoder architecture, 384-residue cropping schedule, MSA inputs, and training budget (5 epochs / 19,840 steps) via single-factor feature ablations and frozen-backbone replacements (Table R1 and Table R2).

Table R1. Single-factor representation and input ablations under matched downstream architecture and controlled training (ChronoPocket test n = 2,890, homology-filtered subset n = 1,546, and ChronoPocket-D n = 308). Nat. AUPRC denotes full-sequence per-residue macro AUPRC, and Disc.@5 denotes ChronoPocket-D Top-5 recall (%); bold marks the best value per column.

Model Variant (Matched Training)AP0.3\mathrm{AP}_{0.3}AP0.5\mathrm{AP}_{0.5}Filtered AP0.5\mathrm{AP}_{0.5} (n=1,546n = 1{,}546)DCAn\mathrm{DCA}_nNat. AUPRCDisc.@5 New / Nov.-low (%)
CoPocket (matched control, IntelliFold)0.5470.4350.3620.7270.70419.7 / 18.1
w/o pretrained pair features (Z(0)=0Z^{(0)} = 0)0.4540.3210.2120.6870.64410.6 / 8.9
Residue features only (S(0)S^{(0)}, w/ MSA, 2-layer Transformer adapter)0.4160.2570.2050.6870.6368.2 / 7.4
ESM C 600M residue features (seq. only, matched)0.4140.2810.1880.6840.6209.1 / 7.7
w/o MSA (query sequence only)0.1790.0810.0210.6350.3801.1 / 1.0
Residue features only, w/o MSA0.1080.0280.0080.5870.3500.8 / 0.8

First, as shown in Table R1 (Section 5.2, lines 356–370, Table 3; Appendix F.1, Table S8), keeping the MSA and all downstream modules fixed while zeroing out only the frozen co-folding pair representation Z(0)Z^{(0)} reduces AP0.5\mathrm{AP}_{0.5} from 0.435 to 0.321 (0.212 on the homology-filtered subset) and drops ChronoPocket-D recall to 10.6%. Removing the Pairformer layers and adapting residue states S(0)S^{(0)} with a depth-matched Transformer encoder further lowers AP0.5\mathrm{AP}_{0.5} to 0.257—comparable to the matched sequence-only language model ESM C 600M (0.281)—while withholding MSAs collapses performance to ≤0.081\le 0.081. This confirms that even without ligand inputs, MSA-conditioned pretrained pairwise features Z(0)Z^{(0)} supply the primary geometric prior (the adaptation of ligand-free co-folding states is detailed in Section 4.2, lines 202–208). Second, because pretraining data sources cannot be disentangled inside a single checkpoint, we replaced the frozen trunk—under identical (384,128)(384, 128) feature dimensions—with template-free AlphaFold2 monomer, which was pretrained exclusively on protein structures (Table R2, top two non-reference rows; Section 5.3, lines 410–455, Table 4; Appendix F.3, Table S11). AlphaFold2 monomer outlines broad surface cavities effectively (coarse overlap AP0.3=0.548\mathrm{AP}_{0.3} = 0.548 vs. 0.547, maintaining high Top-5 coverage in part by emitting more candidate masks), but falls short on strict pocket boundary precision (AP0.5\mathrm{AP}_{0.5}: 0.411 vs. 0.435), pocket center localization (DCCn\mathrm{DCC}_n: 0.471 vs. 0.489), full-sequence residue selectivity (Nat. AP 0.684 vs. 0.704; MCC 0.533 vs. 0.571), and ChronoPocket-D Top-1 prioritization (5.9%, matching sequence-only ESM C and trailing both co-folding trunks at 7.0%–8.0%). Thus, generic monomer folding suffices to locate broad surface clefts, whereas small-molecule co-folding pretraining further sharpens pocket contact boundaries and Top-1 site ranking.


W2 / Q1. Could other co-folding backbones such as Boltz-1 or Protenix be tested to assess cross-backbone generality?

Under the same (384,128)(384, 128) feature dimensions, 384-residue crop schedule, and 19,840-step training budget, we evaluated frozen Boltz-1 and Protenix v0.5.0 (Table R2; corresponding to Section 5.3, lines 410–455, Table 4 and Appendix F.3, Table S11).

Table R2. Cross-backbone replacements under an identical CoPocket decoder and training budget (ChronoPocket full test, homology-filtered subset n = 1,546, and ChronoPocket-D). Nat. AP denotes native per-residue macro AUPRC; bold marks the best value per column.

Frozen BackbonePretraining & CutoffAP0.3\mathrm{AP}_{0.3}AP0.5\mathrm{AP}_{0.5}DCCn\mathrm{DCC}_nNat. AP / MCCFiltered AP0.5\mathrm{AP}_{0.5} (n=1,546n = 1{,}546)Disc. Top-1 (New / Nov.)Disc. Top-5 New [95% CI]
ESM C 600M (ref.)Sequence MLM0.4140.2810.3910.620 / 0.5020.1885.9 / 5.69.1 [6.6, 12.0]
AlphaFold2 monomerProtein folding (Pre-2021-09-30)0.5480.4110.4710.684 / 0.5330.3675.9 / 5.122.6 [18.9, 26.8]
Boltz-1Co-folding (2021-09-30)0.5150.3970.4540.688 / 0.5620.3188.0 / 6.918.8 [15.4, 22.5]
IntelliFold V2-FlashCo-folding (cutoff undocumented in public docs)0.5470.4350.4890.704 / 0.5710.3627.0 / 6.419.7 [16.2, 23.6]

Replacing IntelliFold V2-Flash with frozen Boltz-1 (Section 5.3, lines 426–431, Table 4) substantially outperforms all 20 external baselines in Table 2 (strongest external baseline AP0.5=0.128\mathrm{AP}_{0.5} = 0.128) and the matched ESM C 600M control (0.281) across every metric, achieves the highest ChronoPocket-D Top-1 discovery recall (8.0% / 6.9%), and preserves 18.8% Top-5 recall (18.5% / 19.4% on the homology-filtered subset; Appendix F.3, Table S11 Panel B). This confirms that CoPocket transfers across co-folding trunks rather than depending on IntelliFold V2-Flash alone.

For Protenix v0.5.0, training under the identical unadapted protocol yielded validation AP0.3=2.6×10−5\mathrm{AP}_{0.3} = 2.6 \times 10^{-5}, so we excluded it from the unadapted comparison table (Section 5.3, lines 421–425). Specifically, Protenix's raw single-residue state S(0)S^{(0)} emitted outside its diffusion conditioner lacks trunk-specific normalization, lies at an extreme numerical scale, and exhibits near-collinear token directions, saturating downstream layers of a generic decoder; full output-scale diagnostics are documented in Appendix F.3 (lines 1467–1477).


W3 (+ Q2). Individual modules such as Pairformer lack standalone ablations, and removing pair-derived attention bias did not consistently degrade performance; does this suggest the downstream architecture contributes little?

Under matched initialization and training steps (19,840 steps), we completed full downstream ablations covering Pairformer refinement, the mask-quality head, decoder pair bias, and fixed-window inference controls (Table R3).

Table R3. Controlled downstream ablations and fixed-window inference comparisons at matched training steps (Table 3; extended metrics and threshold sweeps in Tables S8–S10). Bold marks the best value per column.

Downstream Variant / Inference ControlAP0.3\mathrm{AP}_{0.3}AP0.5\mathrm{AP}_{0.5}Filtered AP0.5\mathrm{AP}_{0.5} (n=1,546n = 1{,}546)DCAn\mathrm{DCA}_nNat. AUPRCDisc.@5 New / Nov.-low (%)
CoPocket (matched control)0.5470.4350.3620.7270.70419.7 / 18.1
w/o pair-derived attention bias (updated simplified)0.5530.4390.3700.7360.70521.1 / 19.9
w/o Pairformer refinement0.2770.1010.0910.6280.5544.7 / 4.1
w/o final quality head0.5180.3940.3330.7230.69919.9 / 18.4
Objectness-only ranking (fixed window)0.5260.403—0.7290.70419.5 / 18.1
Binary-union window fusion (fixed window)0.5470.391—0.7220.55817.1 / 15.3

Table R3 establishes the necessity of the downstream components and clarifies how pairwise information flows: (1) Removing only the two trainable Pairformer blocks while retaining (S(0),Z(0))(S^{(0)}, Z^{(0)}) and the decoder bias causes AP0.5\mathrm{AP}_{0.5} to drop from 0.435 to 0.101 (Nat. AUPRC to 0.554, discovery recall to 4.7%; Section 5.2, lines 371–378, Table 3, line 394; Appendix F.1, Table S8). Together with Table R1, both "Z(0)Z^{(0)} without Pairformer" (0.101) and "Pairformer without Z(0)Z^{(0)}" (0.321) trail the joint design (0.435), demonstrating that frozen co-folding pair features must be transformed by trainable Pairformer refinement. (2) Because those Pairformer blocks already fold pairwise couplings from ZZ into the residue state SS, the partner-averaged additive decoder bias adds no further gain; we report the bias-free variant achieving 0.439 AP0.5\mathrm{AP}_{0.5} (21.1% discovery) and formally adopt it as our simplified updated decoder (Section 5.2, lines 375–383, Table 3, line 396; Appendix F.1, line 1403). (3) The quality head calibrates cross-protein ranking (0.394 without it), while confidence-weighted probability fusion outperforms binary union on strict boundary precision (0.435 vs. 0.391) and unified residue AUPRC (0.704 vs. 0.558; Section 5.2, lines 405–409; Appendix F.2, Tables S9–S10).


Reviewer_Wq85OpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces CoPocket, a protein pocket prediction method that uses frozen cofolding representations and a set prediction decoder to identify multiple binding pockets from protein sequences and MSAs, without requiring receptor structures at inference. The authors also construct ChronoPocket, a temporal benchmark containing 37,269 pockets, designed to distinguish newly discovered binding sites from repeated observations of known sites. On the ChronoPocket test set, CoPocket achieves a pocket AP of 0.430 at IoU 0.5, outperforming the ESM C baseline (0.230) and P2Rank (0.074). It also outperforms UniSite-3D on an external test subset (0.347 vs. 0.303). On the ChronoPocket-Discovery subset, CoPocket achieves a top-five recall of 20.5% for newly observed pockets.

Critical strengths
  1. The research addresses a problem of practical significance, and the experimental results are strong. The ability to predict multiple independent binding pockets without requiring receptor coordinates is a direction with clear practical value. On the ChronoPocket test set, CoPocket achieves a Pocket AP of 0.430, significantly outperforming the 0.230 score of the ESM-C variant. It also surpasses UniSite-3D on the UniSite external test set.

  2. ChronoPocket-Discovery offers a meaningful perspective for evaluation. Rather than simply treating newly published protein-ligand structures as "new" pockets, the authors explicitly distinguish between sites that existed previously and additional sites observed for the first time at a later date. This design allows for a more targeted assessment of a model’s capacity to recover newly identified binding sites, providing genuine value as a benchmark.

Critical weaknesses
  1. CoPocket primarily combines existing techniques, including pretrained cofolding representations, Pairformer blocks, and set prediction. While the overall approach is practical and achieves promising results, I feel that the methodological novelty is relatively limited. The paper does not clearly demonstrate what new technical contributions it offers beyond integrating these existing components.

  2. The current ablation studies do not adequately establish the contributions of the key components. For example, removing the Pair-derived Attention Bias does not reduce Pocket AP (0.430 vs. 0.434), while the contributions of Pair Representations and trainable Pairformer refinement are not directly evaluated. Moreover, the ESM C baseline differs in both backbone and training settings. Without more controlled ablations, it is difficult to determine which design choices drive the performance improvements and whether the proposed components provide meaningful benefits.

Questions for authors
  1. The proposed approach mainly combines existing techniques, including cofolding representations, Pairformer blocks, and set prediction. Could the authors further clarify the key methodological innovations of CoPocket beyond integrating these existing components?

  2. The current ablation experiments do not clearly establish the contributions of the key components. In particular, removing the pair-derived attention bias does not reduce Pocket AP (0.430 vs. 0.434). Could the authors provide more controlled ablations to isolate the contributions of pair representations and trainable Pairformer refinement, ideally using the same backbone and training settings?

  3. ChronoPocket-Discovery is an interesting contribution, but the reported top-five recall of 20.5% lacks comparisons with existing methods. Could the authors provide additional baseline results on this subset? Also, given that the cofolding backbone's pretraining cutoff is not verified, how can the authors assess whether these later-observed pockets are genuinely novel to the model?

4 comments

Dear Reviewer Wq85, thank you very much for your time and thoughtful feedback. In the revised manuscript, all updated and newly added content is highlighted in blue.

Below, we address each of your weaknesses and questions point by point.

W1 / Q1. CoPocket primarily integrates existing techniques such as co-folding representations, Pairformer blocks, and set prediction; could the authors further clarify its key methodological innovations beyond component integration?

We clarify this explicitly in the revised Section 1 (line 095) and the rewritten Section 4 (lines 194–280). Although CoPocket retains the general "one-to-one set prediction + frozen trunk representations" paradigm, it is not a plug-and-play assembly of off-the-shelf components: feeding frozen co-folding representations directly into a standard Set Decoder yields near-failure performance (AP0.5=0.101\mathrm{AP}_{0.5} = 0.101, substantially below a matched sequence-only ESM C 600M baseline at 0.281). To enable multi-pocket detection without receptor structures or candidate ligands, CoPocket introduces four targeted designs, each independently validated in controlled ablations (Section 5.2, Table 3; summarized in Table R1):

  1. Co-folding pairwise priors and trainable Pairformer geometric adaptation (Section 4.2, lines 202–208): Frozen co-folding outputs (S(0),Z(0))(S^{(0)}, Z^{(0)}) are optimized for 3D coordinate generation rather than residue-level pocket segmentation. We introduce two trainable Pairformer blocks to jointly refine (S(0),Z(0))(S^{(0)}, Z^{(0)}) for pocket boundary prediction, and formally remove the redundant decoder pair-derived attention bias once controlled ablations confirmed it was unnecessary (Section 5.2, Table 3).
  2. Decoupled soft-IoU quality calibration for ranking (Section 4.3–4.4, lines 224–245): Unlike conventional set decoders that rank solely by presence probability objectness (oko_k), CoPocket independently predicts each mask's crop-local soft IoU quality uku_k (detached from the mask gradient path) and ranks by ck=okukc_k = o_k u_k, suppressing high-objectness fragmented false positives in cross-protein ranking (Table 3; Table S10).
  3. Confidence-weighted sliding-window probability fusion and unified residue projection (Section 4.5, lines 270–280): To handle quadratic pairwise memory on long sequences, inference processes 384-residue overlapping windows. Across overlaps, we replace hard binary union with confidence-weighted probability fusion, and derive continuous per-residue scores ri=max⁡g(c~gp~gi)r_i = \max_g(\tilde c_g \tilde p_{gi}) directly from the same pocket set so instance segmentation and residue ranking remain mutually consistent (Appendix F.2, Tables S9–S10).
  4. ChronoPocket / ChronoPocket-D: a first-observation-dated, de-duplicated discovery benchmark (Section 3, lines 135–191; Appendix B, Table S1): We cluster repeated complex observations per protein by earliest PDB release year and exclude all 681 cross-period proteins from decoder training, cleanly separating genuine novel-site discovery from repeated re-observation of known pockets.

Table R1. Targeted failure modes addressed by CoPocket's four core innovations and their individual quantitative contributions under matched-step (19,840-step) controlled ablations (Section 5.2, Table 3 and Table 2; Nat. AUPRC denotes full-sequence per-residue macro AUPRC, and Disc.@5 denotes ChronoPocket-D Top-5 recall).

Core InnovationProblem AddressedControlled Ablation (Full →\rightarrow Removed or Naive Alternative)ΔAP0.5\Delta\mathrm{AP}_{0.5}Δ\Delta Nat. AUPRCΔ\Delta Disc.@5 (New)
1a. Trainable Pairformer adaptationFrozen co-folding states cannot be decoded directly by Set Decoder0.435 →\rightarrow 0.101 (w/o Pairformer refinement)-0.334-0.150-15.0 pp
1b. Co-folding pairwise prior Z(0)Z^{(0)}Single-residue features alone degrade to sequence-LM performance0.435 →\rightarrow 0.321 (Z(0)=0Z^{(0)} = 0) / 0.257 (S(0)S^{(0)} only)-0.114 / -0.178-0.060 / -0.068-9.1 / -11.5 pp
2. Decoupled soft-IoU calibration (ck=okukc_k = o_k u_k)objectness-only ranking misranks fragmented masks across proteins0.435 →\rightarrow 0.394 (retrained w/o quality head) / 0.403 (oko_k only)-0.041 / -0.032-0.005 / 0.000+0.2 / -0.2 pp
3. Probability fusion & unified projectionBinary union degrades mask boundaries and unified residue ranking0.435 →\rightarrow 0.391 (binary union fusion, fixed-window control)-0.044-0.146-2.6 pp
4. ChronoPocket-D (low-overlap novel sites)Existing benchmarks conflate new sites with re-observed known sitesNovel-low@5 (Table 2 released): 19.9% vs. best baseline 10.7%——+9.2 pp

Table R1 demonstrates that each component resolves a concrete failure mode and yields a distinct, measurable improvement beyond simple module integration.


W2 / Q2. The current ablation studies do not adequately isolate key components; in particular, removing pair-derived attention bias did not lower Pocket AP (0.430 vs. 0.434), and the ESM C baseline differed in backbone and training settings. Could the authors provide more controlled ablations under the same backbone and training configuration?

Under an identical backbone interface, 384-residue cropping schedule, MSA inputs, and training budget (seed 42, 5 epochs / 19,840 steps), we independently retrained all single-factor ablations alongside an architecturally matched ESM C 600M baseline (Table R2; corresponding to the revised Section 5.2, lines 356–408, Table 3 and Appendix F.1, Table S8). Regarding the decoder pair-derived attention bias, our controlled retraining confirms the same finding: removing the cross-attention pair bias maintains or slightly improves every evaluation metric (Table 3, line 396; Table S8, line 1403). Accordingly, we formally adopt the bias-free decoder as our simplified updated design.

Table R2. Controlled ablations and fixed-window inference controls trained for an identical 19,840 steps (evaluated on ChronoPocket test n = 2,890, homology-filtered subset n = 1,546, and ChronoPocket-D n = 308). Bold marks the best value per column.

Model Variant (Matched Control Protocol)AP0.3\mathrm{AP}_{0.3}AP0.5\mathrm{AP}_{0.5}Filtered AP0.5\mathrm{AP}_{0.5} (n=1,546n = 1{,}546)DCAn\mathrm{DCA}_nNat. AUPRCDisc.@5 New / Novel-low (%)
CoPocket (matched control)0.5470.4350.3620.7270.70419.7 / 18.1
w/o pair-derived attention bias (updated simplified)0.5530.4390.3700.7360.70521.1 / 19.9
w/o pretrained pair features (Z(0)=0Z^{(0)} = 0)0.4540.3210.2120.6870.64410.6 / 8.9
Residue features only (S(0)S^{(0)}, w/ MSA, 2-layer Transformer adapter)0.4160.2570.2050.6870.6368.2 / 7.4
ESM C 600M residue features (seq. only, matched)0.4140.2810.1880.6840.6209.1 / 7.7
w/o MSA (query sequence only)0.1790.0810.0210.6350.3801.1 / 1.0
Residue features only, w/o MSA0.1080.0280.0080.5870.3500.8 / 0.8
w/o Pairformer refinement0.2770.1010.0910.6280.5544.7 / 4.1
w/o final quality head0.5180.3940.3330.7230.69919.9 / 18.4
Objectness-only ranking (fixed window)0.5260.403—0.7290.70419.5 / 18.1
Binary-union window fusion (fixed window)0.5470.391—0.7220.55817.1 / 15.3

From Table R2, these controlled experiments establish three key conclusions:

  • Co-folding pair priors Z(0)Z^{(0)} and MSAs are the primary drivers of performance (Section 5.2, lines 356–370): With the downstream decoder held fixed, zeroing Z(0)Z^{(0)} causes substantial drops in both detection and discovery; adapting single-residue features S(0)S^{(0)} with a Transformer encoder falls back to the level of the matched sequence-only ESM C 600M baseline, and removing MSAs causes severe degradation across both trunks, confirming that MSA-conditioned pairwise features carry essential binding geometry.
  • Trainable Pairformer adaptation is indispensable, whereas the decoder pair bias is redundant (Section 5.2, lines 371–383): Removing the two trainable Pairformer blocks while keeping frozen (S(0),Z(0))(S^{(0)}, Z^{(0)}) causes the steepest collapse, showing that co-folding states must undergo Pairformer adaptation. Because those Pairformer blocks already fold pairwise couplings into the residue representation SS, adding an extra partner-averaged pair bias in the decoder is redundant and can be removed for a cleaner, higher-performing model.
  • The quality head and probability fusion improve cross-protein ranking and boundary precision, respectively (Section 5.2, lines 405–409; Appendix F.2, Tables S9–S10): Soft-IoU quality calibration mainly suppresses fragmented false positives in cross-protein ranking, while confidence-weighted probability fusion improves both strict pocket boundaries and unified per-residue ranking.

Q3a. ChronoPocket-D's 20.5% Top-5 recall lacks comparisons with existing methods; could the authors provide baseline results on this subset?

We evaluated 20 published pocket-prediction baselines on the full ChronoPocket-D test set (308 proteins; 527 period-new sites and 392 strictly residue-novel sites) under a unified IoU≥0.5\mathrm{IoU} \ge 0.5 hit criterion, reporting recall across ranking budgets together with 10,000-resample paired protein-bootstrap 95% confidence intervals (Table R3; corresponding to Section 5.1, Table 2, p. 7, with all 21 methods and full ranking-depth sweeps in Appendix E.1, Table S4, p. 22 and Appendix E.3, Table S6, p. 24).

Table R3. Recall (%) across ranking budgets on ChronoPocket-D with 10,000-resample protein-bootstrap 95% CIs (CoPocket denotes the released checkpoint). Low-ov.@5 requires Top-5 predicted masks to have IoU <0.3< 0.3 with all previously known pockets on the protein; †\dagger marks DeepPocket's clean recalculated summary recall (per-protein bootstrap records unavailable); bold indicates the best value per column.

MethodPeriod-New Top-1Period-New Top-5 [95% CI]Top-10AllLow-ov.@5Strictly Novel Top-5 [95% CI]Novel Low-ov.@5
fpocket1.96.1 [4.1, 8.3]10.430.64.66.1 [3.8, 8.8]5.6
pyKVFinder1.35.5 [3.6, 7.6]10.615.94.25.9 [3.6, 8.4]5.4
P2Rank3.410.2 [7.6, 13.1]13.316.18.311.0 [7.9, 14.5]10.7
DeepPocket5.111.2$^\dagger$13.716.98.312.2$^\dagger$10.7
PUResNetV2.04.96.6 [4.5, 9.0]6.66.64.26.1 [3.7, 8.9]5.6
AF2BIND0.40.6 [0.0, 1.3]0.60.60.40.3 [0.0, 0.8]0.3
GPSite1.11.7 [0.7, 2.9]1.71.71.11.5 [0.5, 2.9]1.5
Seq2Pocket1.13.4 [1.9, 5.2]3.63.62.83.8 [2.0, 5.9]3.8
UniSite-1D2.84.7 [2.9, 6.9]5.35.52.53.1 [1.5, 4.9]2.6
UniSite-3D4.06.3 [4.2, 8.6]6.66.84.04.3 [2.4, 6.5]4.3
CoPocket8.720.7 [17.1, 24.6]25.030.715.720.7 [16.4, 25.4]19.9

Two main conclusions follow from Table R3: CoPocket nearly doubles the Top-5 discovery recall of the strongest structure-based baselines (DeepPocket and P2Rank, achieving a +10.4 pp paired gain over P2Rank) and substantially outperforms sequence and set-based baselines (UniSite-1D/3D; full comparison in Section 5.1, lines 336–354). Moreover, this discovery advantage holds after filtering predictions overlapping with known historical sites (Low-ov.@5 / Novel Low-ov.@5), as well as on the <40% homology-filtered subset and bilateral homology-filtered head-to-head comparisons across all 20 baselines (Table 2, p. 7; Appendix E.2, Table S5, p. 23).

In light of the full benchmark results (Table 2 and Appendix E.1, Table S4), we would also like to note an important contextual observation: because existing baselines were trained on different datasets and pocket-definition conventions while ChronoPocket evaluates all methods under a unified ligand-free protocol and consolidated pocket annotations, several baselines score poorly on pocket-level detection metrics (even under the looser AP0.3\mathrm{AP}_{0.3} threshold) yet achieve acceptable performance on per-residue classification metrics (such as Nat. AUPRC / AUROC). We find that this arises because methods trained with per-residue classification objectives typically rank binding residues near the top, but lack an explicit mechanism to organize predicted binding residues into coherent, distinct pocket instances (Section 5.1, lines 336–354). This observation motivated our adoption of UniSite's set-prediction training formulation; however, our experiments also indicate that current strategies for converting point/residue predictions into set predictions remain suboptimal, and we hope to explore more effective set-prediction formulations in future work.


Q3b. Given that the co-folding backbone's pretraining cutoff is not verified, how can the authors assess whether these later-observed pockets are genuinely novel to the model?

We fully agree on the importance of distinguishing genuinely unseen sites from potential pretraining exposure. In public documentation, the earlier IntelliFold-1 technical report (arXiv:2507.02025) specifies a PDB training cutoff of 2021-09-30, whereas the subsequent release note for IntelliFold V2-Flash states that PDB data were re-processed with an updated pipeline without re-specifying an explicit cutoff date in the public repository or model card. Therefore, we conservatively retain the "undocumented / unverified cutoff" disclosure in Section 6 (lines 512–520) and Appendix B.3 (p. 16).

Without relying on assumptions about the backbone cutoff date, we assess novelty through leak-free decoder supervision with low-overlap filtering and a quantitative MMseqs2 sequence-homology audit between training and evaluation splits:

  1. Zero-leakage decoder isolation and low-overlap constraints: Decoder training excludes all 681 cross-period proteins (Section 3, line 166), ensuring the prediction head never observes any ChronoPocket-D protein or its historical binding sites; meanwhile, Low-ov.@5 and Novel Low-ov.@5 (Table R3) filter out predictions overlapping known historical pockets (IoU≥0.3\mathrm{IoU} \ge 0.3), confirming that recall is not driven by re-identifying previously known sites.
  2. MMseqs2 sequence-homology analysis between training and validation/test sets: We aligned all validation and test sequences against the training set with MMseqs2, quantifying both full-sequence and local binding-site homology distributions (search parameters and per-bin local site-conservation statistics are reported in Appendix B.3, p. 16 and Appendix E.4, Figure S1 & Table S7, pp. 24–25). The audit shows that ChronoPocket is not dominated by close homologs: over one-fifth of both validation and test proteins have no significant training hit (No hit), whereas high-identity proteins (>90%) account for only ~10%. Binned test performance of CoPocket is summarized in Table R4.

Table R4. Sample counts and CoPocket performance on the ChronoPocket test set (2,890 proteins / 7,084 pockets) stratified by maximum training sequence identity (Appendix E.4, Table S7, which also reports per-bin binding-site identity and training-label overlap). Metrics: Proj. AP denotes pocket-projected per-residue macro AUPRC, and Nat. AP / ROC denote native full-sequence per-residue macro AUPRC and AUROC.

Sequence Identity BinProteins / PocketsAP0.3\mathrm{AP}_{0.3} / AP0.5\mathrm{AP}_{0.5}DCAn\mathrm{DCA}_nProj. APNat. AP / ROC
No hit681 / 1,4110.397 / 0.2490.6270.5080.581 / 0.890
(0, 20]%1 / 11.000 / 1.0001.0000.8000.985 / 0.999
(20, 30]%273 / 4980.563 / 0.4210.7430.6040.690 / 0.927
(30, 40]%416 / 9200.529 / 0.3910.7250.6220.701 / 0.932
(40, 50]%396 / 9470.639 / 0.5050.7800.6830.759 / 0.934
(50, 60]%277 / 7270.598 / 0.4950.7680.7000.774 / 0.939
(60, 70]%213 / 4870.644 / 0.5490.8090.7180.785 / 0.942
(70, 80]%161 / 4260.539 / 0.4380.7320.6760.749 / 0.928
(80, 90]%183 / 7010.620 / 0.5320.7860.7120.779 / 0.927
(90, 100]%289 / 9660.541 / 0.4580.7390.6930.771 / 0.935
  • Stability across similarity bins and zero-homology generalization (Table R4 and Figure S1, pp. 24–25): Across the sequence-identity bins from (20, 30]% to (90, 100]%, CoPocket's performance remains stable rather than relying on high sequence similarity; on the zero-homology No hit stratum, CoPocket still substantially outperforms representative structure and sequence baselines (Figure S1), indicating cross-homology generalization rather than memorization of training sites.
  • Strict homology filtering and bilateral de-overlap verification (Table 2, p. 7 and Appendix E.2, Table S5 Panel A, p. 23): After removing test proteins sharing ≥40%\ge 40\% sequence identity over ≥90%\ge 90\% coverage with CoPocket's training set (excl. overlap, n=1,546n = 1{,}546 in Table 2), CoPocket maintains solid detection accuracy and novel-site discovery, and preserves consistent advantages over all 20 baselines when test homologs of each baseline's own training set are jointly excluded.

Finally, we transparently document the remaining boundary conditions in Section 6 (lines 512–520) and Appendix B.3 (p. 16) (unrestricted MSA database dates and retrospective union labels across temporal periods).


Reviewer_nZ8oOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The authors present a method for prediction of ligand binding sites from purely a sequence of a protein.

The input to their model is an amino acid sequence and the output of their model is a mask indicating which residues (positions) correspond to a binding site.

There can be many binding sites in a protein. They predict top 32 sites.

The presentation is not clear with respect to symbols, tensor shapes and a not so informative figure. Though there are visuals in the figure, the paper is very hard to read even for those well experienced in this field. I am listing down my notes as I tried to understand.

I have provided reviews in three parts A - for formulation, B - for data set and C for benchmark studies.

Part A The process involves three or four steps. Let us say a query protein sequence is (L,) string of amino acids (call it xq).

Step 1) perform multiple sequence alignment (MSA) of the xq against a database of protein sequences to result in (M,L,V) (for a reviewer's conceptual clarity) where V is the vocabulary of amino acids and gaps. xq_msa = msa(xq,database)

Step 2) This xq_msa object is then fed into a sliding-window segmentor to result in (M,N,V) objects xq_slide_i; where N is the slide length and stride is 192.

It is not clear if sliding window is applied before MSA to the actual sequence or after MSA to the aligned object.

Step 3) This xq_slide_i object is passed to IntelliFold which results in intermediate (N,384) and (N,N,128) tensors.

The description is scattered across the manuscript and in appendix, there is no clear flow.

Step 4) These intermediate items (N,384) and (N,N,128) are passed on through other modules to predict binding site masks for each residue (N,N) and (N,2) objectness and confidence scores (intermediate tensors)

Step 5) Among them top 32 were selected as learnable queries (32, 384) (xq_slide_i_query)

It is not clear this is for a slide or across several slides

Step 6) xq_slide_i_query is sent to other modules to predict (32,N) sites and (32,) confidence scores.

Part B

I understand that they are tracking chronological development of the same structure over years and do a careful inclusion of its binding sites in training, validation and test sets. However there can be several parallel discoveries of the same structure in multiple organisms, in which case sequence similarity thresholds come into picture espectially around binding site regions. The data set curation based on similarity between train and test sets will be useful.

Part C

Regarding benchmark studies, there are other methods like Unisite 1D (instead of 3D), Seq2Pocket, AlphaFold + DeepPocket or FPocket, AF2Bind which should have been compared against.

Critical strengths
  1. Ligand binding site prediction from protein sequence alone
  2. An interesting architecture that considers residue features, residue pair features, objectness scores and several binding site proposals as masks
Critical weaknesses
  1. Not clear flow of information - symbols, tensor shapes, module level inputs and outputs are not clear; details are scattered across the main as well as supplementary sections; while figures attempt to provide a big picture, it introduces several black box steps
  2. The benchmark studies could have been rigorous to include methods mentioned in paper summary
  3. A dive deep into working and non working cases is required with a deeper reasoning
Questions for authors
  1. What could be the performance against Alpha Fold + DeepPocket and FPocket; Unisite 1D; Seq2Pocket
  2. A dive deep reasoning on their working and non working cases?
  3. Quantitive understanding of train, validation, test set sequences in terms of their sequence similarities
4 comments

Dear Reviewer nZ8o, thank you very much for your time and for tracing our pipeline line by line to offer such detailed feedback. In the revised manuscript, all updated and newly added content is highlighted in blue.

Below, we address each of your comments and questions in order.

Part A / W1. Notation, tensor shapes, and module-level inputs/outputs are unclear in the manuscript: is the sliding window applied before or after MSA? How do proposal outputs and shapes flow (Steps 4–5)? Are the top-32 queries selected per window or shared across windows?

Thank you for pointing out these ambiguities step by step. To substantially improve clarity, we rewrote the CoPocket method exposition (Section 4.1–4.5, pp. 4–6, lines 194–280; Figure 2, p. 5): starting from Section 4.2, each inference subsection now focuses on clear algorithmic logic and maps one-to-one to a subpanel of Figure 2 (Section 4.2 →\to Figure 2a, Section 4.3 →\to Figure 2b, Section 4.5 →\to Figure 2c, with Section 4.4 dedicated to training matching), while Appendix C.1 (newly added Table S2, pp. 17–18, lines 878–932) tracks the exact notation and dimensionality changes of the core state tensors section by section (summarized in Table R1).

Specifically addressing your three pipeline questions:

  1. MSA runs before sliding-window segmentation (Section 4.2, Figure 2a; Appendix C.1): A single global MSA search is run first on the full-length protein sequence (L,) (capped at M≤1,024M \le 1{,}024 sequences to form the full-length alignment (M, L)), which is then sliced along the sequence axis into 384-residue windows (inference stride 192, right-padded to 384 when shorter, with valid length N≤384N \le 384) and fed window by window into the frozen trunk.
  2. Clarification of (N, N) and (N, 2) in the proposal module (Section 4.3–4.4; Appendix C.1): In the proposal head, the (N, N) pairwise mask prediction is used solely as auxiliary supervision during training (Section 4.4) and does not enter the decoder; meanwhile, the two per-residue scalar scores (objectness and quality, i.e., (N, 2) in your notes) are used only to rank the NN valid residues within the current window and select the top 32 seed positions.
  3. The 32 queries are strictly per-window (Section 4.3, Figure 2b): From the refined single-residue representation S(N,384)S(N, 384) of the current window, the top-32 residue projections within that window are added to 32 learnable embeddings to initialize window-specific queries Q(0)(32,384)Q^{(0)}(32, 384). Predicted masks from different windows are only mapped to full-sequence coordinates {1,…,L}\{1,\dots,L\}, grouped, and fused afterward in Section 4.5 (Figure 2c).

Table R1. Core state-tensor flow in CoPocket aligned one-to-one with main-text subsections and Figure 2 subpanels (Appendix C.1, Table S2). LL: full sequence length; M≤1,024M \le 1{,}024: MSA depth; N≤384N \le 384: valid residues per window; GG: fused protein-level pockets.

Main-Text Subsection (Figure 2)Stage OperationCore State Tensors & Shape Flow
Section 4.2 (Figure 2a)Full-length MSA, slicing, and frozen co-folding extractionFull MSA (M,L)→(M, L) \to window slice (M,N)→S(0)(N,384),  Z(0)(N,N,128)(M, N) \to S^{(0)}(N, 384),\; Z^{(0)}(N, N, 128)
Section 4.2 (Figure 2a)Joint refinement via 2 trainable Pairformer blocks(S(0),Z(0))→S(N,384),  Z(N,N,128)(S^{(0)}, Z^{(0)}) \to S(N, 384),\; Z(N, N, 128)
Section 4.3 (Figure 2b)Per-window top-32 query initializationS(N,384)→Q(0)(32,384)S(N, 384) \to Q^{(0)}(32, 384) (auxiliary (N,N)(N,N) masks stay in training supervision only)
Section 4.3 (Figure 2b)10-layer decoder and prediction heads(Q(0),S,Z)→Q(32,384)→P(32,N),  o,u,c(32)(Q^{(0)}, S, Z) \to Q(32, 384) \to P(32, N),\; o, u, c(32)
Section 4.5 (Figure 2c)Cross-window probability fusion & unified residue projection{P,c}windows→{Mg⊆{1,…,L}, c~g}g=1G→r(L)\{P, c\}_{\text{windows}} \to \{M_g \subseteq \{1,\dots,L\},\, \tilde c_g\}_{g=1}^G \to r(L)

Note that controlled ablations in Section 5.2 (Table 3, line 396) further confirm that the Pairformer blocks already fold pairwise geometry into SS, allowing the decoder's pair-derived attention bias to be safely removed as a simplified variant without changing the main tensor flow in Table R1.


Part B / Q3. Temporal splitting alone may not exclude homology from parallel discoveries across species; could the authors provide a quantitative analysis of sequence similarity (including binding-site regions) between training and validation/test sets?

We aligned all validation and test sequences against the final 15,444 training proteins using MMseqs2 (sensitivity 5.7, e <= 1e-3, query cov >= 0.5), quantitatively auditing both full-sequence identity and local binding-site conservation (Appendix B.3, p. 16; Appendix E.4, Figure S1 & Table S7, pp. 24–25). The audit shows that ChronoPocket is not dominated by close homologs: more than one-fifth of both validation and test proteins have no significant training hit (No hit), whereas high-identity targets (>90% sequence identity) account for only ~10%. CoPocket's test performance stratified by training-sequence similarity is shown in Table R2.

Table R2. Test-cohort counts and CoPocket performance on ChronoPocket test (2,890 proteins / 7,084 pockets) stratified by maximum training sequence identity (Appendix E.4, Table S7, which also reports per-bin binding-site local sequence identity and training-label overlap). Metrics: Proj. AP denotes pocket-projected per-residue macro AUPRC, and Nat. AP / ROC denote native full-sequence per-residue macro AUPRC and AUROC.

Sequence Identity BinProteins / PocketsAP0.3\mathrm{AP}_{0.3} / AP0.5\mathrm{AP}_{0.5}DCAn\mathrm{DCA}_nProj. APNat. AP / ROC
No hit681 / 1,4110.397 / 0.2490.6270.5080.581 / 0.890
(0, 20]%1 / 11.000 / 1.0001.0000.8000.985 / 0.999
(20, 30]%273 / 4980.563 / 0.4210.7430.6040.690 / 0.927
(30, 40]%416 / 9200.529 / 0.3910.7250.6220.701 / 0.932
(40, 50]%396 / 9470.639 / 0.5050.7800.6830.759 / 0.934
(50, 60]%277 / 7270.598 / 0.4950.7680.7000.774 / 0.939
(60, 70]%213 / 4870.644 / 0.5490.8090.7180.785 / 0.942
(70, 80]%161 / 4260.539 / 0.4380.7320.6760.749 / 0.928
(80, 90]%183 / 7010.620 / 0.5320.7860.7120.779 / 0.927
(90, 100]%289 / 9660.541 / 0.4580.7390.6930.771 / 0.935
  • Stability across similarity bins and strong zero-homology generalization (Table R2 and Figure S1, pp. 24–25): Across the (20, 30]% through (90, 100]% strata, CoPocket's performance remains stable without depending on high sequence identity; on the zero-homology No hit subset, CoPocket still substantially leads representative structure and sequence baselines (Figure S1), demonstrating cross-homology generalization rather than training-set memorization.
  • Strict homology filtering and bilateral de-overlap verification (Table 2, p. 7 and Appendix E.2, Table S5 Panel A, p. 23): After removing test proteins sharing ≥40%\ge 40\% identity over ≥90%\ge 90\% coverage with CoPocket's training set (excl. overlap, n=1,546n = 1{,}546 in Table R3), CoPocket maintains strong pocket precision and novel-site discovery, and preserves consistent advantages over all 20 baselines on the bilateral intersection excluding each baseline's own training homologs.

Part C / W2 / Q1. Benchmarks should compare against UniSite-1D, Seq2Pocket, AF2BIND, and "AlphaFold + DeepPocket / P2Rank / fpocket"; how do they compare?

We incorporated every method you listed into ChronoPocket and expanded our benchmark to 20 open-source baselines across all four categories (geometric cavity detectors, surface/volumetric/graph structure predictors, residue-centric classifiers, and language-model/set decoders; summarized in Table R3, corresponding to Section 5.1, lines 293–354, Table 2, p. 7, with all 21 methods and extended ranking depths in Appendix E.1, Table S4 and Appendix E.3, Table S6). To ensure fair comparison without experimental receptor structures, all structure-requiring baselines uniformly receive ligand-free AlphaFold 3 predicted receptor structures validated by 1-to-LL sequence alignment and SHA-256 checksums (2,890/2,890 passed; details in Appendix D.4). In addition, UniSite-1D is fully evaluated in Table 2 Panel A and Tables S4–S6, and excluded only on Table 2 Panel B (the external UniSite test split) because its public checkpoint uniste_1d_all.pth was trained on the entire UniSite-DS dataset including that split (Appendix D.3).

Table R3. CoPocket versus key representative baselines on ChronoPocket test (2,890 proteins) and ChronoPocket-D (308 proteins; CoPocket denotes the released checkpoint; full 21-method results in Table S4). †\dagger marks methods with runtime failures scored as empty predictions (AF2BIND f=3f = 3); Proj. / Nat. AUPRC denote pocket-projected and native full-sequence per-residue macro AUPRC; bold marks the best value per column.

MethodInputAP0.3\mathrm{AP}_{0.3}AP0.5\mathrm{AP}_{0.5}DCAn\mathrm{DCA}_nProj. / Nat. AUPRCDisc.@5 New / Novel-low (%)
fpocketPred. struct.0.0830.0310.4040.282 / —6.1 / 5.6
P2RankPred. struct.0.2270.0740.4910.386 / 0.49810.2 / 10.7
DeepPocketPred. struct.0.1980.0920.4370.375 / —11.2 / 10.7
AF2BIND$^\dagger$Seq. + pred. struct.0.0670.0100.2870.221 / 0.4300.6 / 0.3
Seq2PocketSeq. + pred. struct.0.1170.0170.5700.379 / —3.4 / 3.8
UniSite-1DSequence0.2120.1040.5650.409 / —4.7 / 2.6
UniSite-3DSeq. + pred. struct.0.2280.1280.5280.422 / —6.3 / 4.3
CoPocketSeq. + MSA0.5470.4290.7330.632 / 0.70820.7 / 19.9
CoPocket (excl. overlap, n=1,546n = 1{,}546)Seq. + MSA0.5020.3570.7050.577 / 0.65521.7 / 24.3

Table R3 shows that classical and recent 3D baselines using ligand-free predicted structures often excel at only one granularity (e.g., Seq2Pocket achieves 0.570 center DCAn\mathrm{DCA}_n yet only 0.017 strict AP0.5\mathrm{AP}_{0.5}; UniSite-3D reaches 0.128 AP0.5\mathrm{AP}_{0.5}). From sequence and MSA alone, CoPocket leads across pocket boundaries (0.429), center localization (0.733), residue ranking (0.708), and low-overlap novel-site discovery (19.9%), and maintains this margin after removing training homologs (Section 5.1, lines 336–354; Table S5).

In interpreting Table R3 (and Table S4), we would also like to note an objective contextual observation: because different baselines were originally trained on varying datasets and pocket-definition conventions while ChronoPocket evaluates all methods under a unified ligand-free protocol and consolidated pocket annotations, several methods perform poorly on pocket-level detection metrics (even under the looser AP0.3\mathrm{AP}_{0.3} threshold) while remaining acceptable on per-residue classification metrics (Nat. AUPRC). We find that this happens because methods trained with per-residue classification objectives generally rank binding residues near the top, yet lack the ability to organize predicted binding residues into independent pocket instances (Section 5.1, lines 336–354). This motivated our use of UniSite's set-prediction training formulation; however, our results also indicate that current strategies for converting point/residue predictions into set predictions are still suboptimal, and we hope to identify stronger set-prediction formulations in future work.


Q2. Could the authors provide a deeper analysis of working and non-working cases and their underlying physical mechanisms?

We added a dedicated Section 5.4 (pp. 9–10, lines 458–510, Figure 3) and Appendix F.4 (pp. 28–29, Table S12) comparing ligand-free predicted structures against holo conformations across representative targets where CoPocket succeeds and where all 21 methods fail:

  1. Working cases (local side-chain or loop occlusion on a well-positioned backbone; Figure 3a,b and Table S12 upper panel): When the backbone of a ligand-free predicted structure (AF3) is already properly positioned, yet individual side chains or transmembrane helices locally occlude the binding cavity or the site borders a broad groove, geometry-based cavity baselines easily fail; by adapting frozen co-folding residue and pair representations through trainable Pairformer blocks, CoPocket can recover partially cryptic sites and distinguish ordinary grooves from true pockets (aided as well by our strategy of consolidating PDB pocket residue annotations onto a unified reference structure):
    • POLRMT (8U8V, homology-filtered target): Despite low local Cα\mathrm{C}_\alpha RMSD (1.27 Å), unbound AF3 side chains clash with 28/31 ligand heavy atoms. The runner-up baseline Seq2Pocket spills into an adjacent groove (33 predicted residues including 17 false positives, IoU=0.390\mathrm{IoU} = 0.390), and UniSite-3D reaches 0.298; CoPocket recovers 15/24 true contact residues at Rank 2 (15/15 precision, IoU=0.625\mathrm{IoU} = 0.625), while none of the 20 baselines enter the Top-5.
    • CHRNA7 (9IIR, homology-filtered transmembrane allosteric site): Local Cα\mathrm{C}_\alpha RMSD is 0.67 Å, yet a transmembrane helix locally closes the internal cavity (21/48 ligand atoms occluded). All 20 baselines miss the site (best UniSite-3D IoU=0.143\mathrm{IoU} = 0.143), whereas CoPocket recovers 9/11 contact residues at Rank 2 (9/11 precision, IoU=0.692\mathrm{IoU} = 0.692). Similar working cases include DTX1 (0.733), TPCN2 (0.533), and TRPC3 (0.500).
  2. Non-working cases (pockets physically absent in the unbound ground state; Figure 3c,d, Section 6 lines 512–525, and Table S12 lower panel): When pocket formation requires somatic mutations or large inter-domain rearrangements, evaluating on a single wild-type sequence without ligand priors keeps both predicted structures and ligand-free co-folding trunks in the closed ground state, causing all 21 methods to fail:
    • Mutation-gated pocket — PI3Kα\text{PI3K}\alpha (8V8H, H1047R-selective inhibitor YO4): The wild-type activation loop (up to 12.0 Å shift) and C-terminal tail (up to 41.7 Å shift) physically occupy the binding pocket (27/33 ligand atom clashes); CoPocket's Top-5 IoU is 0.000 (best candidate at Rank 53: 0.229), and all baselines remain ≤0.242\le 0.242 in Top-5.
    • Large domain-rotation interfaces — WRN (8PFL) and EGFR (9P9U): In WRN, the HRO761 pocket only forms after a 48.8 Å rotation of domain D2 relative to D1 into the inhibited state (in the unbound state, D1/D2 contact residues are separated and clash with 39/48 ligand atoms; CoPocket Top-5 0.136, best Rank 35: 0.149; baselines ≤0.270\le 0.270). In EGFR, a 15.1 Å local interface rearrangement leaves CoPocket at 0.000 in Top-5 (Rank 38: 0.107; all baselines ≤0.071\le 0.071).

Overall, model capability still requires substantial improvement on cryptic pockets driven by large-scale domain rearrangements and mutation-gated binding sites.


Your line-by-line review of our notation, workflow, and baseline comparisons, together with your suggestion on case studies, has directly helped us substantially improve the clarity and experimental completeness of the manuscript. We have worked carefully to refine both the narrative and the case analyses. If the rewritten Section 4, tensor-flow summary (Table S2), MMseqs2 sequence and binding-site similarity audit, 20-baseline comparison, and working/non-working case analyses begin to address your earlier concerns, we kindly ask that you re-evaluate the revised manuscript and consider adjusting your score. We welcome any further feedback and discussion. Thank you!

Comments on the paper

1 comment

Dear Area Chair and Reviewers,

We sincerely thank all reviewers for your time and constructive suggestions. In uploading our revised manuscript (with all edits and additions highlighted in blue), we would like to provide a brief roadmap explaining the scope of the revision:

  1. Core Framework, Benchmarks, and Main Conclusions Remain Unchanged: The CoPocket architecture, the ChronoPocket / ChronoPocket-D benchmark formulation, and our core conclusions remain intact. The only architectural simplification is formally adopting the pair-bias-free decoder after our controlled ablations confirmed Reviewer Wq85 and Reviewer 5o3y's observation that the additive decoder pair bias is redundant (Section 5.2, Table 3).
  2. Narrative Re-organization for Clarity: To directly address Reviewer nZ8o's feedback on readability without altering the underlying method, we rewrote Section 4 (lines 194–280) so each inference subsection maps one-to-one to a subpanel of Figure 2, and consolidated the core tensor-shape flow into Appendix C.1 (Table S2).
  3. Expansion Driven Strictly by Requested Experiments: The increase in total pages (from 20 to 29 pages) mainly reflects new results added to answer specific reviewer requests: expanding comparisons to open-source baselines (Section 5.1, Table 2, Appendix E.1–E.3), analyzing sequence and binding-site similarity (Appendix B.3 & E.4), unifying controlled ablations under matched conditions (Section 5.2, Table 3, Appendix F.1–F.2), testing alternative co-folding/protein-folding trunks (Section 5.3, Table 4, Appendix F.3), and adding structural case studies (Section 5.4, Figure 3, Appendix F.4).

We hope this summary makes it easy to navigate the revised PDF alongside our point-by-point responses below.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.