acceptodds
Under review as a conference paper at ICLR 2027

Hidden in the Fold: Cryptic Pocket Discovery with Local-Global Geometric Context

Abstract

Cryptic pockets are transient or weakly exposed binding sites that are often invisible in static protein structures, yet they play important roles in ligand recognition, allosteric regulation, and drug discovery. Unlike canonical surface pockets, cryptic pockets may only emerge through subtle conformational rearrangements, making their discovery a challenging problem that requires reasoning beyond apparent local concavity. In this work, we propose PocketMon, a context-conditioned equivariant residue field for cryptic pocket discovery. It treats cryptic pocket discovery as a local-global geometric inference problem, where visible residue geometry and latent residue-pair interactions jointly indicate hidden binding opportunities. To this end, PocketMon integrates equivariant geometric modeling with general pairwise residue priors, enabling whole-protein residue-level prediction. We further curate and standardize three cryptic-pocket discovery benchmarks to systematically evaluate this problem. Across these benchmarks, PocketMon consistently outperforms competitive baselines, with up to 14.83% absolute PR-AUC improvement on PocketMiner, and we further apply it to human proteome-scale cryptic pocket detection to provide a broad map of potential hidden binding sites.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
cryptic pocket discovery, geometric deep learning

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_qa1fOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes PocketMon, a cryptic pocket discovery framework that integrates local geometric features with global protein representations. It combines residue level pocket prediction with ligand ranking and ligand induced conformation modeling, demonstrating competitive performance across multiple benchmarks and potential for proteome scale cryptic pocket discovery.

Critical strengths
  1. Well-motivated architectural design. Integrating an SE(3) equivariant geometric encoder with AlphaFold2 pairwise representations effectively combines local geometric features with global protein context.
  2. Comprehensive evaluation and ablation studies. The method is evaluated on multiple datasets, including PocketMiner, CryptoBank, and AF2BIND, with extensive ablation studies examining key components.
  3. Superior performance over baselines. PocketMon consistently outperforms multiple baseline methods across different benchmarks, demonstrating its effectiveness in cryptic pocket prediction.
  4. Effective cryptic pocket identification in dynamic proteins. PocketMon identifies latent binding sites involving domain rearrangements and loop remodeling without costly conformational ensemble sampling.
Critical weaknesses
  1. Potential limitation of static geometric representations for cryptic pocket prediction. PocketMon constructs local geometric features and residue interaction graphs from a single input structure, whereas cryptic pocket formation may involve substantial conformational rearrangements that alter inter-residue distances, local neighborhoods, and backbone orientations. Consequently, residue pairs that become spatially proximate in the pocket forming state may not be directly connected in the input graph, and their target state geometry is not explicitly represented by the local geometric encoder.
  2. Limited methodological novelty. The proposed framework primarily combines established components, including SE(3) equivariant geometric encoding, residue pair representations from AlphaFold2, and GotenNet style interactions. Although their integration is reasonable, the paper does not sufficiently clarify the methodological advances beyond combining existing techniques. In particular, the conceptual formulation of cryptic pocket discovery as local global geometric inference does not, by itself, establish a substantially novel learning mechanism.
  3. The computational scalability of the global pairwise representation is insufficiently characterized. Incorporating residue pair features provides valuable long range context, but pairwise representations scale quadratically with protein sequence length. Given the proteome scale application, reporting inference time, memory consumption, and performance across proteins of different lengths would help establish the practical efficiency of the framework.
  4. Limited experimental validation of newly predicted pockets. The proposed workflow relies mainly on computational benchmarks, molecular ranking, and structure prediction. The large scale proteome analysis reports potential pockets and estimates their druggability using computational criteria, but these results do not establish that the predicted pockets are experimentally accessible or capable of binding ligands. Experimental validation or stronger independent evidence for selected novel predictions would substantially strengthen the biological claims.
Questions for authors
  1. Static Geometry. How does PocketMon account for substantial conformational rearrangements during cryptic pocket formation?
  2. Methodological Novelty. Beyond integrating established components such as SE(3) equivariant encoding, AlphaFold2 derived residue pair representations, and GotenNet style interactions, what are PocketMon's key methodological advances?
  3. Scalability. What are the inference time and memory costs for proteins of varying lengths?
5 comments

W1&Q1. How can a static apo graph identify cryptic residues when pocket formation changes local geometry and creates contacts absent from the input graph?

Thank you for highlighting this distinction. Predicting pocket-forming residues from an initially closed structure is precisely PocketMon’s objective. The model identifies where a cryptic site may form; it does not explicitly simulate the transition or require open-state contacts to exist in the input graph.

Our hypothesis is that static structures contain predictive cues about pocket-forming propensity, including residue composition, local packing, and backbone orientations. Supervised training learns associations between these features and annotated cryptic residues. The global pair representations additionally provide context beyond the local graph’s spatial cutoff, allowing information about non-neighboring residues to contribute to prediction (Sections 4.1–4.2). Our ablations and pair-only comparisons support the complementarity of these information sources (Figure 3; Appendix D.6.1, Table 11). Nevertheless, neither branch explicitly reconstructs future contacts, and substantial rearrangements with weak observable cues remain challenging. We will clarify this distinction between predicting cryptic-site propensity and modeling the conformational transition.

W2&Q2. What methodological advances go beyond combining established equivariant encoders, AF2 pair features, and GotenNet-style interactions?

Thank you for requesting clarification. We would like to clarify that our methodological contribution comprises three complementary advances tailored to cryptic-pocket discovery (Sections 4.1–4.3).

  • Backbone-informed residue modeling combines directional fields, dihedral features, and local chemical context within an equivariant encoder to characterize residue environments even when an exposed cavity is absent.
  • Local–global integration combines these geometric descriptors with learned summaries of pretrained pair representations, bringing together local structural cues and broader residue-interaction context for cryptic-site prediction.
  • Probe-guided conformation prediction uses ProbeEBM to rank molecular probes for predicted patches, then couples the selected probes and site constraints with Boltz-2 to generate candidate open-pocket conformations.

Together, these contributions connect the representation of a closed structure, the localization of latent binding regions, and the generation of open-state structural hypotheses. We will clarify these task-specific advances and distinguish them from the established components on which they build.

W3&Q3 (1/2). Given quadratic pair-representation costs, report inference time, memory, and predictive performance across protein lengths.

Thanks for your important feedback on computational scalability. We have added per-protein timing, GPU memory measurements, and length-stratified performance evaluations across all three test sets. We report PocketMon's inference time and memory overhead separately across proteins of varying lengths, while distinguishing the AF2 feature-extraction cost from the PocketMon inference cost using cached features. The protocol and results are provided in Appendix D.3.1, Tables 5–7.

We performed all measurements on a single NVIDIA A100-SXM4-40GB GPU using the original fixed checkpoints, original test splits, batch size 1, and FP32 precision. Each protein was run at its actual length, warmed up 3 times, and timed synchronously 10 times. We first took the median per protein and then aggregated across predefined length bins. The inference time in Table 5 (reproduced below) includes graph construction, feature conversion and transfer to GPU, the geometric backbone, and the fusion readout. Memory is reported as the maximum PyTorch peak allocated memory within each bin, covering the model, inputs, and intermediate tensors.

DatasetResidue Length BinNumber of Timed ProteinsMedian Inference Time (ms)Max Peak Tensor Memory in Bin (GiB)
PocketMiner1–2001214.350.49
PocketMiner201–4001723.980.86
PocketMiner401–600433.111.34
PocketMiner601–800241.851.63
CryptoBank1–20032614.140.43
CryptoBank201–40029721.691.04
CryptoBank401–60011231.731.52
CryptoBank601–8002444.602.04
CryptoBank>8001158.063.16
AF2BIND1–200814.870.42
AF2BIND201–4004423.851.00
AF2BIND401–6001631.141.23

Full-feature timing covers 35 PocketMiner, 770 CryptoBank, and 68 AF2BIND test proteins. Taking the longest full-feature sample in CryptoBank (1,256 residues) as an example, inference takes 77.82 ms with a peak tensor memory of 3.16 GiB. These results support single-GPU per-protein processing across the tested length range. We further measured the extraction cost of residue-pair features separately on three real apo structures of different lengths using the same GPU model (Table 6, reproduced below).

Protein Length (Residues)First Prediction & Feature Retrieval Time incl. JIT (s)Peak Feature Extraction Process Memory (GiB)
12717.801.51
25717.741.51
66134.844.64

The timing includes the initial JIT compilation and excludes model initialization, input preparation, and file saving. Memory uses an on-demand allocation setting, sampling the total memory usage of the extraction process every 20 ms via NVML. These measurements clarify the one-time feature-extraction cost, while the generated features can be saved offline and reused, enabling subsequent PocketMon screening via per-protein loading and millisecond-level prediction.

W3&Q3 (2/2). Given quadratic pair-representation costs, report inference time, memory, and predictive performance across protein lengths.

We also re-evaluated all test proteins grouped by length under the original evaluation protocol. Taking the largest dataset, CryptoBank, as an example, the predictive performance across length bins is reported in Table 7 and reproduced below:

CryptoBank Residue Length BinNumber of ProteinsROC-AUC (%)PR-AUC (%)Recovery (%)
1–20032671.3022.3620.06
201–40029774.0915.9716.98
401–60011273.7310.3913.40
601–8002478.029.6514.36
>8001180.0817.3419.96

The longest bin in CryptoBank still achieves 80.08% ROC-AUC, and the three non-empty length bins in AF2BIND achieve ROC-AUCs of 92.51%, 94.16%, and 92.94%. These supplementary results show that PocketMon maintains millisecond-level inference speed and manageable memory requirements across the tested protein length range while preserving effective predictive performance on longer proteins, providing practical support for large-scale per-protein screening.

W4. Computational scores do not establish accessibility or ligand binding. Provide experimental validation or stronger independent evidence for newly predicted pockets.

Thank you for this important suggestion. We agree that computational predictions alone do not establish experimental accessibility or ligand binding. Our evaluation provides complementary evidence through residue-level benchmarks (Table 1), component ablations (Figure 3), site-level Top-K analysis (Appendix D.11.3, Table 13), and an end-to-end evaluation on 20 apo–holo targets (Appendix D.2.1, Table 4). Retrospective experimental structural evidence for a selected unmatched prediction is provided in Appendix D.11.2.

These results support predictive utility and selected structural hypotheses, but do not validate all newly predicted sites. We will therefore frame the proteome analysis as candidate prioritization and clearly distinguish retrospective evidence from prospective validation. We plan to pursue experimental assessment of selected predictions in future work to strengthen their biological interpretation.

Reviewer_KJ2yOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents PocketMon, a method for predicting cryptic binding sites from protein structures by combining local equivariant geometric features with global residue-pair representations. The authors evaluate it on PocketMiner, CryptoBank, and AF2BIND, and further combine predicted pockets with probe ranking and Boltz-2 structure prediction. They also apply the model to the human proteome to identify potential hidden binding sites.

Critical strengths
  1. Predicting cryptic pockets from a single protein structure is a challenging and useful problem. I like the idea of combining local geometry with global residue-pair information, since cryptic pocket formation can involve structural changes that are not apparent from local surface geometry alone.

  2. PocketMon shows substantial improvements on PocketMiner, increasing PR-AUC from 66.39 to 81.22 compared with PocketMiner (Table 1). The results on CryptoBank are also encouraging, and the cross-dataset experiments provide additional evidence of generalization. The ablations help establish the contributions of local geometry and global pair information.

Critical weaknesses
  1. A major part of PocketMon's performance appears to come from the AlphaFold2 pair representations. Figure 3 shows that removing these features causes a substantial performance drop. However, the main comparisons on PocketMiner and CryptoBank are against methods that do not necessarily have access to the same pretrained information. I would like to see a pair-representation-only baseline trained on the same data, to better understand how much improvement comes from the proposed geometric architecture. On AF2BIND, where a strong pair-representation baseline is available, the improvement is relatively small (PR-AUC 68.63 versus 67.77). This makes it harder to judge the added value of the local-global integration.

  2. My main concern is whether the predicted sites are truly cryptic pockets rather than general binding sites. The AF2BIND benchmark includes conventional ligand-binding residues, not specifically cryptic sites. In the human proteome experiment, druggability is evaluated using Fpocket scores on the same static structures. This measures geometric pocket quality, but does not establish whether a predicted site is hidden in the apo state or can open through conformational changes. In fact, 78.9% of PocketMon's predicted sites overlap with Fpocket predictions (Appendix D.11). The remaining 10,212 unique sites are potentially more interesting, but their cryptic nature is not independently evaluated. I also find the downstream validation somewhat limited. ProbeEBM is evaluated using known ligand-contact patches, while the Boltz-2 experiments cover only three proteins with known ligands. These results are useful, but they do not demonstrate that the complete pipeline can discover a previously hidden pocket and generate a realistic open conformation. For practical drug discovery, I would be more interested in whether the top-ranked predicted sites correspond to experimentally observed cryptic pockets, particularly those missed by conventional pocket finders. A site-level top-K evaluation on held-out apo–holo examples would help establish this.

  3. I am concerned about possible overlap between training and evaluation data. Appendix D.5 checks identical PDB chain IDs across datasets, but does not assess sequence homology or structural similarity between different proteins. For cryptic pocket prediction, this is particularly important because homologous proteins, or even proteins with relatively low sequence similarity, may share very similar binding pockets. Excluding identical PDB IDs alone may therefore not be sufficient to establish independent evaluation. I would like to see a more systematic analysis of protein sequence identity, structural similarity, and especially local pocket similarity between training and test sets. The same concern applies to ProbeEBM, where similar ligand-binding pockets may appear across the PDBbind splits. This does not necessarily mean that information leakage occurred, but without these checks, it is difficult to assess how well the method generalizes to genuinely unseen binding sites.

Questions for authors
  1. Could the authors compare PocketMon with a pair-representation-only model trained on exactly the same PocketMiner and CryptoBank splits? How much improvement comes specifically from the equivariant geometric encoder?

  2. Among the 10,212 PocketMon sites not detected by Fpocket, is there evidence that these are genuine cryptic pockets? Could the authors evaluate top-ranked sites against experimentally characterized apo–holo pocket-opening events?

  3. Can the authors provide a more systematic evaluation of the complete pipeline, from predicted residues to probe selection and open-pocket structure prediction? In particular, how often does it recover the experimentally observed open conformation without using the known holo pocket or ligand as input?

  4. Have the authors evaluated sequence and structural similarity between training and test proteins across PocketMiner, CryptoBank, and AF2BIND? More importantly, have they checked local binding-pocket similarity, rather than only matching PDB chain IDs? Could the authors report performance after excluding highly similar pockets or using protein-family- and pocket-cluster-disjoint splits? Was a similar analysis performed for the ProbeEBM training and test sets?

7 comments

W1&Q1. Compare against pair-only models on identical PocketMiner/CryptoBank splits to establish the added value of the geometric encoder beyond pretrained AF2 features.

Thanks for your helpful question. We added pair-representation-only logistic regression and MLP baselines trained on PocketMiner and CryptoBank to examine the performance achievable by predictors using only pretrained residue-pair representations. The baseline definitions and results are provided in Appendix D.6.1 and Table 11.

Both baselines use the residue-pair representations extracted by AF2, keep these features fixed during training, and exclude PocketMon's equivariant geometric backbone. The specific setups are as follows.

  • Logistic regression. Following the linear classification approach of AF2BIND, a linear classification head is applied to the pair features corresponding to each residue, and residue probabilities are output via a sigmoid function.
  • MLP. We reuse PocketMon's pair-feature aggregation module to summarize the pair features of each residue into a 64-dimensional vector via learnable projection and attention weighting, followed by a 64→256→1 MLP with ReLU and a dropout of 0.1. This baseline allows joint training of the aggregation module and the nonlinear classification head, and thus tests whether increasing the readout capacity for pair features can approach the full model.
DatasetModelROC-AUCPR-AUCAccPrecRecallRecovery
PocketMinerPocketMon88.31±1.0781.22±1.2983.42±0.8475.65±1.3867.32±2.1339.22±1.92
PocketMinerpair-representation-only logistic regression59.34±0.7836.11±0.9665.49±1.127.95±0.231.24±0.1214.54±1.70
PocketMinerpair-representation-only MLP71.37±1.1557.04±1.8869.50±1.290.00±0.000.00±0.0016.68±1.64
CryptoBankPocketMon74.39±0.7116.78±1.1280.32±0.6711.78±1.5854.07±3.4717.73±2.36
CryptoBankpair-representation-only logistic regression64.10±0.937.68±1.2495.61±0.000.00±0.000.00±0.008.61±1.32
CryptoBankpair-representation-only MLP64.28±0.957.93±1.3795.61±0.000.00±0.000.00±0.009.11±1.45

Under the same representations, the MLP trained on PocketMiner improves ranking performance over logistic regression, increasing PR-AUC from 36.11% to 57.04%. This indicates that these pretrained features contain exploitable predictive signals. However, even when attention aggregation and nonlinear readout are trained jointly, it remains below PocketMon's 81.22%. Relative to this stronger pair-only baseline, PocketMon achieves gains of 16.94, 24.18, and 22.54 percentage points in ROC-AUC, PR-AUC, and Recovery, respectively, demonstrating the added value of our model on top of the same pretrained pair prior and showing that the improvement brought by the equivariant geometric encoder is substantial. These differences appear in both probability ranking and recovery metrics. These results show that relying solely on pretrained residue-pair representations is insufficient to explain PocketMon's performance gains. Precision and Recall values of 0.00±0.00 correspond to no positive predictions at the fixed 0.5 threshold. On CryptoBank, all predictions fall into the negative class below the 0.5 threshold, so Acc is determined by the roughly 95.61% negative proportion among evaluated residues, and similar classification outcomes across runs result in a standard deviation of 0.00. Therefore, a higher Acc does not indicate better cryptic-pocket detection.

W2&Q2. Are predictions genuinely cryptic rather than general binding sites? Support the 10,212 Fpocket-unmatched sites and evaluate site-level Top-K recovery of experimentally characterized cryptic regions, especially those missed by Fpocket.

We appreciate your careful comments and suggestion. We agree that ligand-binding-site prediction, druggability scoring of static cavities, and cryptic-pocket identification require different levels of evidence. The AF2BIND benchmark evaluates general binding-site prediction, and Fpocket scores measure the geometric druggability of matched cavities in static structures. To evaluate whether the predicted sites genuinely represent cryptic pockets, we added an independent experimental structure analysis of the original unmatched sites as well as a site-level top-K evaluation taking experimental apo structures as input, directly testing whether the predicted regions correspond to known ligand-contact regions and conformation-dependent binding sites.

(1) Study of unique predicted sites (Appendix D.11.2)

We retained the 10,212 unmatched sites and mapped them to experimental structural records in external structural resources. A representative example is IL-36γ (Q9NZH8, site 1), where three of the five residues predicted by PocketMon from the original static input belong to the A-552 contact region in the experimental holo structure 6P9E, and this prediction did not match any Fpocket cavity. After stable-core alignment, out of the 35 heavy atoms of A-552, 19 and 20 atoms clash within 2 Å with the unbound experimental structure 4IZE and the original AlphaFold input, respectively, compared to 0 in the experimental holo structure, with core RMSDs below 0.4 Å for both alignments. These results show that from a static input structure that cannot yet accommodate the ligand, PocketMon locates in advance a true conformation-dependent binding region missed by the original Fpocket matching pipeline. The experimental reference structure and the A-552 complex and original study provide independent experimental structural and functional evidence for this region.

(2) Site-level top-K evaluation (Appendix D.11.3, Table 13)

We conducted a site-level top-K evaluation on independent apo–holo test samples to verify whether top-ranked predicted sites correspond to experimentally observed cryptic pockets. Using the 24 positive proteins in the PocketMiner test set with their cryptic-site residue annotations as references, we evaluated the PocketMon model used in the original proteome scan. Predictions use only experimental apo inputs, while holo information and reference annotations are used solely for evaluation. The primary hit criterion requires at least 3 shared residues and Dice ≥0.30. Across all 24 targets, PocketMon achieves a Top-3 hit rate of 17/24 (70.8%). After further filtering based on the actual training and validation sequences to exclude targets with matches having ≥30% sequence identity and ≥80% coverage on either side against both task-model splits, 14 targets remain that pass both screens, on which PocketMon achieves a Top-1 hit rate of 8/14 (57.1%) and a Top-3 hit rate of 11/14 (78.6%). These results show that inspecting only the top three predicted sites per protein suffices to locate most annotated cryptic reference regions. Furthermore, focusing on cryptic sites missed by conventional pocket-detection methods, we selected proteins where none of Fpocket's output cavities met the predefined reference-site matching criterion. Taking 3ugkA as an example, the candidate site output by PocketMon achieves a Dice of 0.723 with the annotated cryptic reference region. In contrast, Fpocket outputs 25 cavities on the same input, yet the highest Dice with this reference region is only 0.231. This demonstrates that PocketMon can locate a known cryptic reference region from apo input alone that Fpocket fails to match under the same criterion.

On the one hand, the experimental analysis confirms the presence of conformation-dependent binding regions supported by independent experimental structures among the original unmatched predictions. On the other hand, the site-level top-K evaluation on experimental apo structures shows that PocketMon ranks multiple known cryptic reference regions near the top. In the human proteome scan, PocketMon can rapidly screen large-scale protein structures for cryptic pockets worthy of further investigation and provide high-confidence candidate sites characterized by druggability enrichment and amenable to further filtering with external experimental structure evidence for subsequent wet-lab experiments.

W2&Q3. Evaluate the complete predicted-patch → probe-selection → structure-generation pipeline: how often does it recover experimental open conformations without known holo pockets or ligands as inputs?

Thank you for your valuable suggestion to enrich our evaluation. The added end-to-end protocol and results are provided in Appendix D.2.1 and Table 4; the opening-recovery metric is defined in Eq. (41). To directly test the complete pipeline from predicted residues and probe selection to open-pocket structure prediction without using known holo pockets or ligands as input, we selected experimental apo–holo pairs from the CryptoBank test records. After residue mapping, checks for quantifiable opening-and-closing changes, deduplication, and sequence-homology exclusion against the PocketMon and ProbeEBM training and validation sets as well as existing development targets, we obtained 20 test target proteins. The newly added experiment starts from experimental apo structures, uses PocketMon to predict pockets, applies ProbeEBM to select ligand probes from a fixed candidate pool based on the predicted patches, and then feeds the selected ligands and predicted site constraints into Boltz-2 to generate structures. Reference holo structures are used only to define the evaluation criteria and compute scores, not to select generated outputs.

We predefine the pocket-opening recovery metric Ropen=1−E(pred)/E(apo)R_{\mathrm{open}}=1-E(\mathrm{pred})/E(\mathrm{apo}), where E(X)E(X) is the mean absolute error of Cα–Cα distances relative to the holo structure over opening-and-closing residue pairs predefined on the reference apo–holo pair. Successful recovery must simultaneously satisfy Ropen≥0.50R_{\mathrm{open}}\geq0.50 (reducing distance error by at least half relative to apo), improve pocket Cα RMSD relative to apo after alignment on a fixed stable core, and pass protein Cα geometry quality checks. These checks include sequence and coordinate completeness, adjacent Cα distances within 2.8–4.5 Å, and no Cα clashes below 2 Å between residues separated by more than 2 positions in sequence. Evaluated against the corresponding experimental holo structures under the same pocket-opening recovery criteria, the complete pipeline achieves a pocket-opening recovery rate of 45%, compared to 30% for the control using only the protein sequence and the same MSA. This represents an improvement of 15 percentage points, or 1.5 times the recovery rate of the control. The validation of the complete pipeline provides direct support for the end-to-end workflow starting from apo structures and proceeding through pocket prediction, probe screening, and constrained generation to recover open conformations.

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

What explicit criteria distinguish full success, partial success, and invalid submissions? Were these criteria fixed before inspecting results, and how are they reconciled with the allowance for arbitrary architectures? Please report raw functional scores and qualification outcomes separately.

Can you report the secondary metrics for every submission, particularly CMD uniqueness and novelty, ProteinMPNN perplexity, and RFdiffusion3 structural novelty and lookup fraction? For CMD, what anonymized task specification, data documentation, and evaluation artifacts can be released to permit independent verification?

What performance can a validated reference implementation achieve when trained under the same data, hardware, and time constraints? This would help separate limitations in agent implementation from limitations imposed by the training budget.

Please reconcile the CMD complexity scores of 25.7 and 51.9 and clarify whether complexity ratings consistently describe the restricted benchmark tasks or the complete original tools. Given only three evaluated tasks, how much of the complexity-based interpretation remains supported after this correction?

Critical strengths

A well-motivated combination of local geometry and pretrained structural context. Cryptic-site prediction from a closed structure is a reasonable setting for combining backbone orientation and neighborhood geometry with broader residue-interaction features. PocketMon provides a concrete implementation of this idea, and the component ablations suggest that both geometric processing and pair features contribute to performance.

Substantial improvements on relevant residue-level benchmarks. On PocketMiner, PR-AUC increases from 66.39% to 81.22%, while on CryptoBank it reaches 16.78%, compared with 11.00% for the reported PLM baseline. The paper reports multiple metrics and evaluates transfer between datasets with different label definitions. The paired-seed Boltz-2 comparison also provides an initial controlled test of whether predicted residue constraints can improve structural hypotheses, although its scope is limited.

Critical weaknesses

The contribution of the proposed geometric model is not fully separated from the pretrained features, and transfer leakage is insufficiently assessed. Removing AF2 pair fusion demonstrates that these features are useful, but does not establish how much the geometric architecture adds beyond an appropriately trained pair-feature-only predictor on PocketMiner and CryptoBank. The direct AF2BIND comparison addresses this question on the general binding-site benchmark, where the PR-AUC improvement is considerably smaller, from 67.77% to 68.63%. Matched-feature baselines on the cryptic-pocket datasets would better support the central architectural claim. Separately, Appendix D.5 explicitly states that cross-dataset overlap is checked using PDB chain identifiers without sequence-homology clustering. Removing the single identical chain is a useful sensitivity analysis, but cannot rule out homologous or alternative-structure overlap. The main-text statement that this analysis rules out training-set overlap is therefore too strong.

The downstream experiments do not yet establish the claimed scope of cryptic-pocket discovery. ProbeEBM retrieval uses observed ligand-contact patches rather than PocketMon-predicted patches, so it does not measure the effect of localization errors on probe selection. The Boltz-2 study covers only three selected proteins and does not clearly establish that its ligands were obtained through the proposed probe-selection workflow. These experiments support individual components, but not yet the full sequence from an apo structure to a newly selected probe and a recovered open pocket. The proteome analysis has a related limitation: druggability is assessed only for predictions matching pockets already detected by Fpocket on the same static structures. This measures enrichment among geometrically detectable cavities, rather than validating the unmatched predictions as cryptic sites. Moreover, PocketMon predicts substantially fewer sites, so higher average scores could partly reflect a different operating point. These findings should be framed as exploratory prioritization rather than evidence of validated proteome-wide cryptic-pocket discovery.

Questions for authors

Can you compare PocketMon against pair-feature-only logistic regression and MLP baselines on PocketMiner and CryptoBank using identical features, splits, and model-selection procedures? Please also specify the AF2 feature-extraction settings, including template use and how the input backbone is incorporated.

Do the cross-dataset transfer gains persist after excluding sequence homologues and structurally related proteins, rather than only identical PDB chain identifiers?

Were the ligands in the three Boltz-2 case studies selected by ProbeEBM from PocketMon-predicted patches, or were they known ligands? Can you evaluate the complete workflow on a predefined held-out set and report pocket-opening recovery separately from ligand-pose accuracy?

Does the proteome-wide druggability advantage persist when methods are compared at matched numbers of predicted sites? What independent evidence supports the crypticity of predictions that do not overlap with Fpocket cavities?

4 comments

W1&Q1. Separate architectural gains from AF2 features using pair-only LR/MLP baselines with identical features, splits, and model selection; clarify AF2 extraction, template use, and backbone input.

Thank you for your insightful question. To address this, we added pair-feature-only logistic regression and MLP baselines on PocketMiner and CryptoBank to examine what performance a predictor relying solely on pretrained residue-pair representations can achieve. The baseline definitions, extraction settings, and results are also provided in Appendix D.6.1 and Table 11.

(1) Pair-feature-only logistic regression and MLP baselines

Both baselines use the residue-pair representations extracted by AF2, keep these features fixed during training, and exclude PocketMon's equivariant geometric backbone. The specific setups are as follows.

  • Logistic regression. Following the linear classification approach of AF2BIND, a linear classification head is applied to the pair features corresponding to each residue, and residue probabilities are output via a sigmoid function.
  • MLP. We reuse PocketMon's pair-feature aggregation module to summarize the pair features of each residue into a 64-dimensional vector via learnable projection and attention weighting, followed by a 64→256→1 MLP with ReLU and a dropout of 0.1. This baseline allows joint training of the aggregation module and the nonlinear classification head, and thus tests whether increasing the readout capacity for pair features can approach the full model.
DatasetModelROC-AUCPR-AUCAccPrecRecallRecovery
PocketMinerPocketMon88.31±1.0781.22±1.2983.42±0.8475.65±1.3867.32±2.1339.22±1.92
PocketMinerpair-feature-only logistic regression59.34±0.7836.11±0.9665.49±1.127.95±0.231.24±0.1214.54±1.70
PocketMinerpair-feature-only MLP71.37±1.1557.04±1.8869.50±1.290.00±0.000.00±0.0016.68±1.64
CryptoBankPocketMon74.39±0.7116.78±1.1280.32±0.6711.78±1.5854.07±3.4717.73±2.36
CryptoBankpair-feature-only logistic regression64.10±0.937.68±1.2495.61±0.000.00±0.000.00±0.008.61±1.32
CryptoBankpair-feature-only MLP64.28±0.957.93±1.3795.61±0.000.00±0.000.00±0.009.11±1.45

Under the same representations, the MLP trained on PocketMiner improves ranking performance over logistic regression, increasing PR-AUC from 36.11% to 57.04%. This indicates that these pretrained features contain exploitable predictive signals. However, even when attention aggregation and nonlinear readout are trained jointly, it remains below PocketMon's 81.22%. Relative to this stronger pair-only baseline, PocketMon achieves gains of 16.94, 24.18, and 22.54 percentage points in ROC-AUC, PR-AUC, and Recovery, respectively. These differences appear in both probability ranking and recovery metrics, supporting the added value of our model on top of the same pretrained pair prior. Precision and Recall values of 0.00±0.00 correspond to no positive predictions at the fixed 0.5 threshold. On CryptoBank, all predictions fall into the negative class below the 0.5 threshold, so Acc is determined by the roughly 95.61% negative proportion among evaluated residues, and similar classification outcomes across runs result in a standard deviation of 0.00. Therefore, a higher Acc does not indicate better cryptic-pocket detection.

(2) AF2 feature-extraction settings

We extract residue-pair representations using the pretrained AlphaFold2 model. The inputs are the experimental or apo structures corresponding to the task labels, and template settings follow each benchmark. The target protein in the input PDB is used as the sole structural template, without introducing any external homologous structures. PocketMon additionally uses the N, Cα, C, and O backbone coordinates from the same input structure to build the residue graph, explicitly encoding local distances, orientations, and backbones. The precomputed residue-pair representations serve as global context, which is projected and fused with the geometric representations. Thus, the model utilizes both local 3D geometric information from the input backbone and global residue-coupling information from AlphaFold2. This also enables the newly added pair-feature-only baselines to directly test whether relying solely on pretrained residue-pair representations is sufficient to explain PocketMon's performance gains.

W1&Q2. Chain-ID checks do not exclude homologous proteins. Do transfer gains persist after sequence and structural exclusion? Please qualify the claim that training-set overlap is ruled out.

Thanks for raising this critical concern. First, to clarify, the cross-dataset evaluation in Table 2 uses PocketMon throughout, comparing its performance on the same 35-protein PocketMiner test set after training on PocketMiner, CryptoBank, and AF2BIND, respectively. In this evaluation, distinguishing identical-chain overlap from sequence and structural similarity is indeed important. We have now extended the analysis to sequence and whole-chain structural similarity, and re-evaluated the retained test proteins using the original checkpoints. Transfer advantages on several metrics still hold, especially the ROC-AUC and Recovery advantages of PocketMon trained on AF2BIND.

(1) Excluding sequence-homologous proteins (Appendix D.5.1, Table 8)

We use MMseqs2 and treat pairs with sequence identity ≥30%, coverage ≥80% on at least one side, and E-value ≤10⁻³ as hits, while also excluding identical sequences and identical PDB chains. After removing test proteins that hit either reference set under these criteria, 11 proteins remain. Please refer to Table 8 for detailed results.

On this common subset, the AF2BIND-trained model still achieves a higher ROC-AUC than the PocketMiner-trained model (92.08 vs. 89.99), and both transferred models achieve higher Recovery.

(2) Excluding structurally related proteins (Appendix D.5.2, Table 9)

We use Foldseek and TM-align scores, defining each match by the larger of the query-normalized and target-normalized TM-scores. In the TM≥0.70 analysis, any match reaching the threshold to a reference protein is excluded. In the table, AF2BIND-filtered denotes the 13 test proteins retained after excluding proteins that meet the above sequence or structural matching thresholds against the AF2BIND training and validation sets. Joint-filtered denotes the 5 test proteins retained using both the CryptoBank and AF2BIND training and validation sets as references under the same exclusion criteria.

Test SubsetN (Pos/Neg)PocketMon Training SourceROC-AUCPR-AUCAccPrecRecallRecovery
AF2BIND-filtered13 (8/5)PocketMiner90.7783.5186.0571.5172.3245.32
AF2BIND-filtered13 (8/5)AF2BIND94.1780.7685.5089.2546.8955.94
Joint-filtered5 (3/2)PocketMiner85.3873.6382.6669.8451.7635.32
Joint-filtered5 (3/2)CryptoBank86.6165.7085.5576.9258.8246.49
Joint-filtered5 (3/2)AF2BIND95.3174.9582.0879.4936.4756.52

After structural filtering, PocketMon trained on AF2BIND still retains advantages in ROC-AUC, Precision, and Recovery over the model trained on PocketMiner. The model trained on CryptoBank still achieves higher values on most metrics in the cross-dataset evaluation, except for a lower PR-AUC. These supplementary experiments directly extend the original chain ID check and show that under the described sequence and structural exclusion criteria, transfer advantages persist.

We would also like to clarify the purpose of this cross-dataset experiment. Table 2 is an extended analysis of transferability across different supervision sources, designed to test whether knowledge learned from general ligand-contact labels and conformational-change-related labels can identify cryptic-pocket-related residues defined by MD in PocketMiner. AF2BIND, CryptoBank, and PocketMiner construct their labels differently, so homology between proteins does not automatically imply identical residue-level supervision. Even if two proteins are homologous, their positive labels may not point to the same residues, binding sites, or conformational events. Nevertheless, we agree that differences in task and label definitions alone cannot rule out transfer benefits from related proteins, so we further conducted sequence and whole-chain structural exclusions. At the same time, we restrict our conclusions to the metrics and filtering conditions where the advantages hold, allowing a more rigorous assessment of transferability across supervision sources.

W2&Q3. Were the original Boltz-2 ligands known or retrieved from predicted patches? Evaluate the complete workflow on a predefined held-out set, reporting pocket-opening recovery separately from ligand-pose accuracy.

Thank you for this constructive suggestion. The three Boltz-2 case studies in the original manuscript used known ligands rather than ligands retrieved by ProbeEBM. The added end-to-end protocol and recovery results are reported in Appendix D.2.1 and Table 4, with the opening-recovery metric defined in Eq. (41). To directly test the complete pipeline, we selected experimental apo–holo pairs from the CryptoBank test records. After residue mapping, checks for quantifiable opening-and-closing changes, deduplication, and sequence-homology exclusion against the PocketMon and ProbeEBM training and validation sets as well as existing development targets, we obtained 20 test target proteins. Since the main goal of this work is to identify cryptic sites from apo structures and recover open-pocket conformations, we use pocket-opening recovery rate as the primary evaluation metric for the end-to-end pipeline. The newly added experiment starts from experimental apo structures, uses PocketMon to predict pockets, applies ProbeEBM to select ligands from a fixed candidate pool based on the predicted patches, and then feeds the selected ligands and predicted site constraints into Boltz-2 to generate structures.

We predefine the pocket-opening recovery metric Ropen=1−E(pred)/E(apo)R_{\mathrm{open}}=1-E(\mathrm{pred})/E(\mathrm{apo}), where E(X)E(X) is the mean absolute error of Cα–Cα distances relative to the holo structure over opening-and-closing residue pairs predefined on the reference apo–holo pair. Successful recovery must simultaneously satisfy Ropen≥0.50R_{\mathrm{open}}\geq0.50 (reducing distance error by at least half relative to apo), improve pocket Cα RMSD relative to apo after alignment on a fixed stable core, and pass protein Cα geometry quality checks. These checks include sequence and coordinate completeness, adjacent Cα distances within 2.8–4.5 Å, and no Cα clashes below 2 Å between residues separated by more than 2 positions in sequence. Evaluated against the corresponding experimental holo structures under the same pocket-opening recovery criteria, the complete pipeline achieves a pocket-opening recovery rate of 45%, compared to 30% for the control using only the protein sequence and the same MSA. This represents an improvement of 15 percentage points, or 1.5 times the recovery rate of the control. In this pipeline, the ligand probe primarily serves as a conditioning input for generating open-pocket conformations, and accurately recovering ligand binding poses is not the primary evaluation objective of this work. We focus on whether the generated structures capture the key geometric changes of pocket opening and form candidate conformations close to the experimental open state, thereby providing a structural starting point for downstream complex modeling and experimental validation. The validation of the complete pipeline provides direct support for the end-to-end workflow starting from apo structures and proceeding through pocket prediction, probe screening, and constrained generation to recover open conformations.

W2&Q4. Does druggability enrichment persist at matched site budgets? Provide independent evidence for Fpocket-unmatched predictions and distinguish exploratory prioritization from validated proteome-wide cryptic-pocket discovery.

We appreciate your thoughtful feedback on this analysis. We first clarify that this analysis follows the human proteome scanning and druggability evaluation framework adopted by the AF2BIND study published in Nature Methods, which identifies candidate binding sites on predicted protein structures and uses structural scoring to assess their drug-discovery potential. This analysis is designed to evaluate druggability enrichment and experimental prioritization of candidate sites with a clear methodological and scientific basis, whereas whether the predicted sites are genuinely cryptic requires separate wet-lab validation.

(1) Analysis of predicted site counts and druggability advantages (Appendix D.11.1, Table 12)

Regarding the difference in the number of predicted sites across methods, we wish to clarify that PocketMon outputs fewer candidate sites while achieving a higher proportion of high-scoring sites within the Fpocket-matched subset, which helps focus downstream validation on more promising regions. To further control for the effect of prediction count, we added a comparison under equal site budgets.

Across the same 20,548 proteins, we rank predictions by each method's own score and compare multiple equal site budgets. At the largest common budget of 48,396, PocketMon retains all its predictions, while AF2BIND and P2Rank each retain their top 48,396 predictions. The proportions of high-scoring sites within the Fpocket-matched subset are 17.70%, 16.42%, and 16.88%, respectively. PocketMon improves over the two baselines by 1.28 and 0.82 percentage points, showing that under this common budget, the conditional druggability enrichment advantage persists after controlling for the number of predictions.

(2) Validation of cryptic pocket properties (Appendix D.11.2)

To verify the cryptic nature of unmatched sites, we further searched independent experimental structures, with IL-36γ (Q9NZH8, site 1) as a representative example. Three of the five residues predicted by PocketMon from the original static input belong to the A-552 contact region in the experimental holo structure 6P9E, and this prediction did not match any Fpocket cavity. After stable-core alignment, out of the 35 heavy atoms of A-552, 19 and 20 atoms clash within 2 Å with the unbound experimental structure 4IZE and the original AlphaFold input, respectively, compared to 0 in the holo structure, with core RMSDs below 0.4 Å for both alignments. These results show that from a static input structure that cannot yet accommodate the ligand, PocketMon locates in advance a true conformation-dependent binding region missed by the original Fpocket matching pipeline. The experimental reference structure and the A-552 complex and original study provide independent experimental structural and functional evidence for this region.

In summary, the equal-budget analysis confirms that PocketMon's enrichment for highly druggable sites persists after controlling for prediction count. The independent experimental structure analysis, exemplified by IL-36γ, further shows that some predicted sites missed by conventional pocket finders indeed correspond to potential cryptic pockets with conformational-change characteristics and ligand-binding evidence. The core contribution of the human proteome scan is to use PocketMon to rapidly screen large-scale protein structures for cryptic pockets worthy of further study, providing high-confidence candidate sites characterized by druggability enrichment and amenable to further filtering with external experimental structure evidence for subsequent wet-lab experiments.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

Dear Area Chair and Reviewers,

We sincerely thank you for your careful assessments and constructive suggestions. We are encouraged that reviewers recognize that (i) the importance of predicting cryptic pockets from a single structure (Reviewer KJ2y), (ii) the well-motivated combination of local geometry and global residue-pair context (all reviewers), and (iii) the strong benchmark performance (all reviewers).

The shared concerns center on performance attribution and evaluation independence. We summarize our responses below, followed by detailed replies to each reviewer. All section and table references below refer to the revised manuscript.

1. Performance attribution and methodological contribution.

We added pair-only logistic regression and MLP baselines using the same pretrained representations. PocketMon outperforms both on the cryptic-pocket benchmarks, supporting the added value of the complete local–global model (Appendix D.6.1, Table 11). We further clarify three complementary contributions: backbone-informed residue modeling, local–global integration, and probe-guided open-conformation prediction through ProbeEBM and constrained Boltz-2 generation (Sections 4.1–4.3).

2. Evaluation independence and transferability.

We extended the chain-ID check to sequence, whole-chain structural, and local-pocket similarity analyses (Appendices D, Tables 8–10). Transfer advantages persist after filtering, particularly AF2BIND-trained ROC-AUC and Recovery, although PR-AUC gains are not consistent across subsets.

3. End-to-end evaluation and evidence for cryptic sites.

We added an end-to-end evaluation on 20 apo–holo targets, connecting predicted patches, probe selection, and open-pocket structure generation (Appendix D.2.1, Table 4). Equal-budget comparisons, retrospective experimental structural evidence, and site-level Top-K evaluation further assess the proteome predictions and site prioritization (Appendices D, Tables 12–13). These results support candidate prioritization and structural hypotheses; they do not replace prospective experimental validation.

We believe these additions strengthen the empirical support and clarify the scope of our contribution. We appreciate the reviewers' guidance and look forward to further discussion.

Best regards,

Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.