PaperlayerView exact revision on arXiv

Source-linked AI summary

zPocket: Better Pocket Finding with Frozen Co-Folding Features

Tally Portnoi, Allison M Keys, Benjamin R DiFrancesco, Nate Gruver

arXiv:2610.17696v1cs.LGcs.AI

TL;DR

Public benchmarks do not adequately evaluate pocket prediction on sites beyond co-folding models’ training data. The paper builds PocketWatch and trains zPocket heads on frozen co-folding features; zPocket outperforms the tested baselines on per-residue prediction.

  • Problem

    Existing benchmarks predate AF3-era co-folding model training cutoffs, leaving no open benchmark to test pocket-prediction claims on unseen pockets.

    Source: §1 · p. 1

  • Method

    The authors construct PocketWatch, a temporal dataset stratified by sequence similarity, structural similarity, and crypticity, and train per-residue heads on frozen co-folding trunk features.

    Source: §1 · pp. 1–2, §3.2 · p. 4

  • Results

    0.745 per-target AUPRC for the best zPocket model, compared with 0.477–0.598 for the baselines; zPocket heads outperform every baseline on per-residue pocket prediction.

    Source: §4.1 · p. 5, §6 · p. 11

  • Takeaways & Limitations

    PocketWatch enables direct comparison of pocket finders on pockets that postdate co-folding model training data, and its training split lifts both retrained baselines.

    Source: §6 · p. 11

    The test set contains few truly novel pockets, aggregate performance is dominated by pockets with close structural precedents, and evaluation covers monomer pockets only.

    Source: §6 · p. 11

Abstract

from arXiv · show

Identifying protein-ligand binding sites (pockets) is important for understanding protein function and developing drugs. However, no public method has leveraged co-folding models for novel pocket prediction; furthermore, this task lacks a suitable benchmark as existing benchmarks are within the training distribution of co-folding models. We therefore developed PocketWatch,a curated dataset of biologically relevant ligand pockets. The PocketWatch test set consists of unseen pockets stratified by sequence similarity, structural similarity, and crypticity to enable evaluation of generalization to novel and challenging pockets. Using this dataset, we develop zPocket, a method that trains pocket-prediction heads on top of frozen features from co-folding models. We show that zPocket heads outperform baseline methods across all stratifications. Our best zPocket model achieves overall per-target AUPRC of 0.745 compared to 0.598 and 0.595 for the strongest baselines, VN-EGNN and UniSite-1D retrained on the PocketWatch split (0.559 and 0.531 when using the published weights); the widely used P2Rank reaches 0.477. We further show that zPocket can be used to outperform baseline methods on top-N+2 recall of predicted pocket centroids through clustering per-residue predictions or using zPocket to rescore centroids proposed by other models. Finally, we explore two applications of zPocket: crystallographic fragment screens and a proteome-scale scan for uncharacterized pockets on human proteins. Together these results suggest that zPocket heads built on co-folding trunk features can greatly advance ligand binding site prediction.

1 Introduction

Existing benchmarks do not adequately test pocket prediction on sites unseen by co-folding models. The paper introduces PocketWatch and zPocket, which reads per-residue pocket predictions from frozen co-folding features and outperforms baselines across test-set strata.

1 Introduction

  • PocketWatch provides a temporal test set of unseen pockets stratified by sequence similarity, structural similarity, and crypticity to assess generalization beyond memorization.Its per-UniProt consolidation includes biologically relevant pockets; cryptic pockets are included because they are particularly challenging for structure-based prediction.Source: §1 · p. 1
  • zPocket leads P2Rank, UniSite-1D, and VN-EGNN in every max_qcov bin of the PocketWatch test set.UniSite-1D and VN-EGNN are retrained on PocketWatch for this stratified comparison.Source: Figure 1 · p. 2, §1 · p. 2
  • zPocket trains per-residue prediction heads on frozen Protenix and ESMFold2 trunk features, with the main text reporting ESMFold2 heads.Protenix-v2 weights later moved to a proprietary license; the compared trunks have matched architectures in Table A5.Source: §1 · p. 2

2 Related work

Prior pocket-finding work ranges from geometric, conservation, and energetic scoring to learned cavity ranking and direct predictions from voxel grids or protein graphs. Existing benchmark datasets lack the temporal split needed to assess generalization to co-folding models’ training cutoffs.

2 Related work

  • Classical pocket finders score geometry, evolutionary conservation, or interaction energetics, while deep-learning methods rank cavities or predict pockets from voxel grids and protein graphs.P2Rank scores surface points and clusters them into pockets; DeepPocket reranks fpocket proposals, while VN-EGNN predicts from a protein graph.Source: §2 · p. 2
  • Lacuna detects pockets across a generated conformational ensemble rather than only on the input structure.Source: §2 · p. 3
  • UniSite and LIGYSIS curate pocket datasets, but neither provides a temporal split for evaluating AF3-era co-folding models.PocketWatch extends the RCSB ligand-binding record through the end of 2025 and stratifies by structural similarity and crypticity.Source: §2 · p. 3
  • AF2Bind reads binding-site information from AlphaFold2 trunk features using appended bait amino acids as small-molecule surrogates.Light heads trained on frozen trunk features are also used for other tasks, including affinity prediction in Boltz-2.Source: §2 · p. 3

3 Methods

The methods construct a temporally split, stratified pocket dataset, train prediction heads on frozen co-folding features, compare against existing pocket finders, and evaluate residue-level and centroid-level predictions.

3.1 PocketWatch dataset

  • PocketWatch aggregates ligand-contacting pocket instances by canonical UniProt sequence, clusters instances by ligand centroid into sites, and assigns chains to temporal splits using pocket release dates.Test chains have a site first appearing on or after 2024-01-01; test pockets are stratified by sequence identity, structural similarity, and crypticity.Source: §3.1 · pp. 3–4

3.2 zPocket: trunk-feature-based pocket prediction

  • zPocket trains heads on frozen Protenix-v2 or ESMFold2 features to predict per-residue pocket logits, selecting checkpoints by per-target validation AUPRC.Pair-based heads pool row means, column means, and self-pair diagonals; variants differ in pair mixing, while single heads read residue representations directly.Source: §3.2 · p. 4
  • Pocket centroids are decoded either by clustering high-scoring residues with weighted mean shift or by rescoring centroids proposed by other predictors.Hybrid decoders combine zPocket and baseline centroid proposals before reranking.Source: §3.2 · p. 4

3.3 Baselines

  • Comparisons use five pocket finders on ESMFold2-predicted folds, with VN-EGNN and UniSite-1D evaluated both using published weights and after PocketWatch retraining.The retraining uses the PocketWatch train/validation split and existing labels; the other baselines are published-only.Source: §3.3 · p. 4

3.4 Evaluation

  • Evaluation measures per-residue average precision and centroid detection within 4 Å among the top N + 2 predictions, alongside whole-list average precision and a detection ceiling.Seen, interface-only, and certain novel sites are masked; scored residues exclude masked pockets from negatives, and centroids landing on masked pockets are dropped.Source: §3.4 · p. 5

4 Results

zPocket’s frozen co-folding features support strong per-residue pocket prediction across baselines and trunks, while its predictions can also improve pocket-centroid ranking and proposals. Performance is lower on unfamiliar and multi-pocket cases, and the centroid methods have distinct precision and recall trade-offs.

4.1 zPocket beats all baselines on per-residue pocket prediction

  • 0.745 per-target AUPRC: ESMFold2 pairformer2 beats baselines at 0.477–0.598 on PocketWatch.Every z-based head on both trunks beats the published and retrained baselines; the reported comparison uses ESMFold2 folds for structure-reading baselines.Source: §4.1 · p. 5, Table 1 · p. 6
  • 0.803 per-pocket AUPRC in the most-covered bin falls to 0.327 below 20% coverage, though zPocket leads baselines in every similarity stratum.The aggregate is dominated by familiar pockets: 71.6% are in the top coverage bin. Performance also drops on chains with multiple pockets, with pairformer2 falling from 0.751 to 0.599.Source: §4.1 · p. 6

4.2 The pocket signal is not specific to one trunk

  • 0.733–0.757 per-target AUPRC across seven trunks shows the pocket signal is not specific to one co-folding model.Pairformer2 beats pooled by 0.08–0.12 on every trunk; averaging all seven pairformer2 heads reaches 0.777, though some ensemble gain may reflect averaging independent runs.Source: §4.2 · p. 7, Table 2 · p. 7, Table A5 · p. 7

4.3 Pocket centroids

  • 60% of zPocket clustering emissions land on real sites, but its 1.1 pockets per protein average limits recall and caps its ceiling at 0.524.It is competitive on DCCN+2, beating published baselines but falling short of PocketWatch-retrained VN-EGNN.Source: §4.3 · p. 8
  • 0.090 average DCCN+2 gain across six methods follows re-ranking with zPocket, except for PocketWatch-retrained VN-EGNN.The re-ranking comparison uses each method’s existing proposals; paired bootstrap results are described in Figure 4.Source: §4.3 · p. 8, Figure 4 · p. 8
  • 0.108 average DCCN+2 gain follows adding zPocket cluster proposals and re-ranking; retrained VN-EGNN plus these proposals reaches 0.675.The gain is reported across six methods and the union method; the passage notes that DCCN itself does not show the same improvement.Source: §4.3 · pp. 8–9, Figure 4 · p. 8

5 Applications

The fragment-screen evaluation corroborates zPocket’s pocket-prediction performance, while a human-proteome scan surfaces candidate pockets that illustrate both discovery potential and the tension between novelty and tractability.

5.1 In-Silico Fragment Screening

  • The evaluation retains 12 proteins with 35 sites after requiring support from at least three distinct fragments, limited crystal-contact dominance, and fold mapping on both trunks.Sites overlapping pockets in zPocket’s training split are excluded.Source: §5.1 · p. 9
  • 0.494 AUPRC for zPocket exceeds P2Rank’s 0.459, VN-EGNN’s 0.438, and UniSite-1D’s 0.385 on 35 fragment-soak sites, but no configurations separate statistically.Re-ranking fpocket cavities with zPocket reaches 0.391 DCCN+2, the best measured configuration; the 12-protein intervals are wide.Source: §5.1 · p. 9, Figure 5 · p. 10
  • Fragment-screen DCCN+2 reaches a 0.529 ceiling, versus 0.879 on PocketWatch; restricting proteins to their dominant site raises 0.391 only to 0.417.The fragment-screened proteins average 2.9 scored sites, compared with PocketWatch’s 1.23, limiting a decoder that emits about one pocket per protein.Source: §5.1 · pp. 9–10

5.2 Novel pocket hypotheses

  • A Protenix-v2 scan of about 20,000 reviewed human proteins produced about 22,000 predicted pockets, filtered for geometric plausibility and novelty.The scan was run while Protenix-v2 weights were openly licensed and has not been repeated on ESMFold2.Source: §5.2 · p. 10
  • After filtering, 359 pockets on 358 proteins remain when polymer interfaces are retained, or 155 pockets on 154 proteins when they are removed.Examples include a candidate degrader-recruiter site on DET1 and COP1’s peptide-binding interface, which zPocket re-identified despite no peptide-ligand training; the authors note novelty and tractability can conflict.Source: §5.2 · pp. 10–11

6 Discussion

The discussion presents co-folding features as a promising basis for blind binding-site prediction and PocketWatch as a resource for evaluating generalization. It also notes limits in test-set novelty, interface coverage, and the functional interpretation of predicted sites.

6 Discussion

  • zPocket heads outperform the tested baselines on per-residue prediction and improve rankings of pockets proposed by other methods.PocketWatch enables comparisons on pockets postdating co-folding-model training data, and retraining baselines on its training split improves their performance.Source: §6 · p. 11
  • The test set contains few truly novel pockets, and aggregate performance is dominated by pockets with close structural precedents.The evaluation covers monomer pockets only; extending it to interfaces is identified as future work.Source: §6 · p. 11
  • Predicted sites may not be functional, but the authors argue that non-orthosteric sites can be therapeutically actionable for degraders, molecular glues, and radioligand conjugates.The discussion frames the useful question as where a small molecule can bind, not only where a protein performs its chemistry.Source: §6 · p. 11

A.1 PocketWatch dataset

PocketWatch curates ligand-contacting residues into pocket sites on a per-UniProt basis, using explicit inclusion criteria and a common predicted-fold frame. Its evaluation also masks interface-only sites to support comparisons with single-chain baselines.

A.1 PocketWatch dataset

  • Pocket instances pair one receptor chain with one ligand occurrence; residues are defined as having a heavy atom within 5 Å of a ligand heavy atom.The dataset retains ligands with 10–50 resolved heavy atoms, chains with at least 10 contact residues and over 80% SIFTS coverage, and canonical UniProt sequences 40–1200 residues long.Source: §A.1 · p. 14
  • Pocket instances are aligned to a canonical UniProt Protenix-predicted fold, then joined into sites by single-linkage clustering when centroids lie within 5 Å.Site membership is shared across trunks; for ESMFold2 evaluation, centroids are re-projected onto that fold.Source: §A.1 · p. 14
  • Residues in any ligand-binding site receive positive labels, other residues receive negative labels, and masked residues are excluded from training loss and evaluation metrics.Per-residue labels and masks are indexed by canonical UniProt position.Source: §A.1 · pp. 14–15
  • The effective test set excludes four centroid-novel pockets whose residues lie inside an already-public pocket, then drops 11 pockets on four proteins unsupported by VN-EGNN.The exclusions leave 1,279 chains and 1,573 pockets after the first step.Source: Table A1 · p. 15
  • Interface-only sites, whose member instances all involve multiple contacting chains, are masked at evaluation to enable fair comparison with baselines run on single-chain folds.Whether such sites count as pockets is treated as a policy choice.Source: §A.1 · p. 15

A.1.1 Test set stratification

This section defines how PocketWatch measures sequence and pocket novelty and crypticity, then describes frozen-feature zPocket heads, baseline comparisons, evaluation metrics, and additional trunk comparisons.

A.1.1 Test set stratification

  • Only 16% of scored test proteins fall below 30% identity to anything seen, while 45% are at or above 70% and 32% at or above 95%.The sequence reference includes trunk pretraining and PocketWatch training and validation sequences; the authors caution that one average would overstate generalization.Source: §A.1.1 · p. 15
  • Pocket similarity uses max_qcov, the maximum structural coverage of a held-out site by pre-cutoff ligand sites, with no sequence-identity floor for hits.The search uses a pre-cutoff Foldseek database; despite conservative choices, max_qcov remains a lower bound on precedent, so the novel tail is an upper bound on novelty.Source: §A.1.1 · pp. 15–16
  • Of 1,562 scored test sites, 463 have crypticity labels: 40 cryptic and 423 not cryptic; the other sites remain unknown.Only 17 sites are both cryptic and not already open in the ESMFold2 fold, too few for a stable AUPRC; predicted-fold accuracy is analyzed separately.Source: §A.1.1 · p. 16

A.2 zPocket: trunk-feature-based pocket prediction

  • zPocket featurizes each protein once with Protenix-v2 or ESMFold2 and trains pocket-prediction heads on persisted trunk representations while keeping the trunks frozen.Stored features include pair and single representations, plus ESMFold2 language-model embeddings; reported models use plain unweighted BCE selected on validation.Source: §A.2 · pp. 16–17
  • The clustering decoder groups confident residues by position into pocket centroids, while the re-ranking decoder uses zPocket residue scores to reorder cavities proposed by other detectors.Clustering parameters are selected on validation; the hybrid decoder adds zPocket cluster centroids to the pooled detector proposals before re-ranking.Source: §A.2 · pp. 17–18

A.3 Baselines

  • VN-EGNN and UniSite-1D are evaluated with published and retrained variants, but their retrained versions do not reproduce the authors’ original training recipes.The VN-EGNN code paths differ, and the UniSite-1D retraining uses choices supplied by this study, so it is described as a lower bound for that architecture on PocketWatch.Source: §A.3 · pp. 18–19

A.4 Evaluation

  • Per-target AUPRC is the headline residue-level metric, complemented by AUROC, R-precision, and per-pocket metrics that isolate one pocket at a time.Centroid evaluation includes budgeted DCC recall, average precision over the full ranked list, pooled precision, mean list length, and a no-budget ceiling.Source: §A.4 · pp. 19–20

A.5 Additional co-folding trunks; B.1 Additional per-residue results

  • Protenix leads ESMFold2 on every matched zPocket head, but the margin is small relative to the choice of head architecture; the architecture ranking is shared across trunks.The additional-trunk comparison trains pooled and pairformer2 heads, while Table A6 extends per-target results to every trunk and head with paired AUPRC differences.Source: §A.5 · p. 20, Table A5 · p. 22

B.2 Additional results for additional co-folding trunks

Table A6 extends the per-target comparison across co-folding trunks and pocket-prediction heads.

B.2 Additional results for additional co-folding trunks

  • Table A6 gives the full comparison between different co-folding trunks.Source: §B.2 · p. 22

B.3 Additional pocket centroid results

The appendix extends centroid evaluation across proposal and ranking combinations, tests performance by pocket familiarity, and compares P2Rank on predicted folds with experimental apo structures.

B.3 Additional pocket centroid results

  • Table A7 compares centroid predictions across proposal and ranking combinations using DCC budgets, ceiling, average precision, precision, and emitted centroids per protein.It covers the same 1,562 pockets of 1,275 targets; a prediction succeeds within 4 Å, and each decoder is selected by a validation sweep.Source: §B.3 · p. 22, Table A7 · p. 24

B.1 Additional per-residue results

  • Per-target AUPRC by sequence-identity bin is roughly flat from novel through close and falls only in the near-identical bin, where the seen-pocket mask may explain the drop.Figure A1 reports bootstrap 95% intervals and bin sizes; zPocket outperforms all baselines.Source: Figure A1 · p. 23

B.4 How do predicted structures affect P2Rank?

  • 0.518 versus 0.505 per-target AUPRC compares P2Rank on experimental apo structures and ESMFold2 folds, respectively, with paired difference +0.013 and 95% CI [−0.003, +0.030].The authors interpret fold-based scores as a fair proxy for experimental-input performance; the estimated structure-source advantage is much smaller than zPocket’s margin over P2Rank.Source: §B.4 · p. 24
  • P2Rank’s apo-versus-predicted-fold gap comes almost entirely from 49 targets whose apo chain observes fewer than half of the scored residues.The comparison uses an evaluation set of 433 targets and 537 pockets, restricted to admissible apo chains and canonical positions observed experimentally.Source: §B.4 · p. 24

B.5 Crypticity versus predicted-fold accuracy

Fold accuracy tracks pocket-prediction performance more closely than crypticity labels, while the appendix also clarifies case-study calibration and reports trunk-dependent fragment-screen results.

B.5 Crypticity versus predicted-fold accuracy

  • Crypticity labels describe experimental structures, not whether the predicted apo fold reproduces the holo pocket; some labeled cryptic sites are within 2 Å in predicted folds.The authors therefore stratify by predicted-fold accuracy as well as crypticity.Source: §B.5 · pp. 24–25
  • Fold accuracy predicts binding-site performance better than crypticity: zPocket and P2Rank score higher when the predicted fold is within 2 Å, and every method degrades otherwise.zPocket still outperforms baselines across the reported stratifications, but three of five buckets contain only 33, 23, and 17 pockets.Source: §B.5 · p. 25, Figure A3 · p. 26

B.6 Calibration and the main-text case-study figure

  • The plain-BCE head scores higher on validation than the weighted-focal case-study head, and ranks S2 residues higher despite making the site less visually prominent.The plain-BCE model has per-pocket AUPRC 0.403 on S2 versus 0.247 for the weighted-focal variant; Figure A4 shows the reported model with its own 0–1 color ramp.Source: §B.6 · p. 25, Figure A4 · p. 26

B.7 Fragment screens on the Protenix-v2 trunk

  • On Protenix-v2, zPocket clusters lift fpocket’s ceiling 0.438 to 0.542 and arm 0.273 to 0.382, and Lacuna’s ceiling 0.306 to 0.410 and arm 0.247 to 0.352.Pooling zPocket clusters into those detectors adds nothing on ESMFold2; P2Rank and VN-EGNN show different trunk-dependent patterns, while zPocket clustering remains on its own ceiling.Source: §B.7 · p. 27, Figure A5 · p. 27

C Sourcing and curation of the fragment-soak test set

The fragment-screen test set combines leakage-screened Fragalysis campaigns with eligible group depositions, then applies uniform structure-confidence and site-curation rules. The novelty-filter appendix motivates resolving known sites through homologues because direct accession annotations leave most scanned proteins without evidence.

C Sourcing and curation of the fragment-soak test set

  • Of 61 Fragalysis campaigns, 11 survived head- and trunk-leakage screens; selection excluded training-set accessions or close relatives and trunk-saturated campaigns.The closest retained head-training relative was 44% identical; campaigns with more than 15 pre-cutoff PDB entities at ≥70% identity were excluded.Source: C Sourcing and curation of the… · p. 28
  • The curation added six new group-deposition campaigns after screening 132 campaigns; nine duplicates and FabF were excluded.FabF failed the head-axis leakage screen because its closest training relative was 66% identical.Source: C Sourcing and curation of the… · p. 28
  • A mean pLDDT threshold of 85 excluded ORF9b and MYLIP; the remaining constructs scored between 85.5 and 97.5.ORF9b reached 41.7 and MYLIP 69.5.Source: C Sourcing and curation of the… · p. 28
  • Sites were counted from soaked-compound coordinates, clustered at 5 Å, and screened for sparse fragments, crystal packing, mapping, interface, and fold-transfer issues.For A71EV2A, only 113 fragment-screen events were retained, not the 812 later lead-like compounds.Source: C Sourcing and curation of the… · pp. 28–29

D Pocket novelty filters

  • Table A8 groups the pocket-filtering funnel by evidence-propagation route, while Table A9 removes interface evidence and reports a strict superset of survivors.The funnel captions state that total survivors are independent of filter order, although per-filter removal counts are not.Source: Table A8 · p. 29, Table A9 · p. 30

C Sourcing and curation of the fragment-soak test set

  • 35 sites across 12 proteins were scored—29 from Fragalysis and 6 from group depositions—with one Fragalysis campaign retaining none.Across both sources, 155 sites were assembled and 117 masked under uniform rules.Source: C Sourcing and curation of the… · p. 29

D Pocket novelty filters

  • The appendix proteome scan used Protenix-v2 heads and folds and has not been repeated on ESMFold2.The scan was run while Protenix-v2 weights were openly licensed.Source: D Pocket novelty filters · p. 29

D.1 Why known sites are resolved by homology

  • Direct accession-based annotation reached only 2,444 of 20,092 scanned proteins, leaving 87% of pockets without evidence and motivating homology-based resolution.An earlier shortlist based on that direct gate contained four leads later found to be known functional sites.Source: §D.1 · p. 30

D.2 The ligand-site annotation

The appendix specifies how ligand-site evidence is annotated, transferred, and filtered to assess candidate pockets in the human proteome. Homology and motif evidence expand coverage, but many surviving pockets remain untested and alternative proposals do not rescue sites the head missed.

D.2 The ligand-site annotation

  • Ligand contacts use a fixed 5 Å heavy-atom policy and nonpolymer ligands; the 10-residue minimum censors smaller pockets, and a length cap limits direct annotation coverage.The reference annotation’s length cap explains why the direct gate reaches only 2,444 of 20,092 scanned proteins.Source: §D.2 · p. 30

D.3 Donor construction

  • Donors include structural ligand contacts, UniProt site annotations, and assembly-derived interfaces; transfer ranks same-family donors before clan-only donors and retains the 12 highest-identity alignments.Clan-level site transfers require at least two independent donors, while interface transfers remain restricted to family level.Source: §D.3 · p. 31, §D.4 · p. 31

D.5 Geometric thresholds

  • Pocket thresholds use a 39.9-atom burial cutoff and 30-residue upper size bound, while a 9-residue floor is paired with a 0.70 confidence floor.The size floor and explicit confidence floor replace a lower-bound choice unsupported by the censored reference distribution; missing per-residue pLDDT values were backfilled.Source: §D.5 · p. 31

D.6 Motif-propagated evidence

  • Motif matches to ligand-contact residues in 233,054 PDB structures can discard a pocket, but the same-family or EC-class sub-ångström rule fired only once and is under-powered.The ligand-contact motif rule carries the filter; most matched target chains lack annotation for comparison.Source: §D.6 · pp. 31–32

D.7 Evidence coverage

  • 89 of 155 surviving pockets had no evidence from any gate; homology reduced the share with no evidence from 87% to 57%, but did not eliminate it.A protein with no annotated member in its Pfam family remains unreachable by homology transfer.Source: §D.7 · p. 32

D.8 Alternative pocket proposals

  • None of 735 qualifying proposals from fpocket, P2Rank, or Lacuna on 346 proteins rescued a site the head had not already proposed.The proposals came from 1,797,462 candidates; no protein with no head-predicted pocket was rescued.Source: §D.8 · p. 32

E Candidate Pockets

  • Two candidate pockets were chosen after filtering because mechanistic hypotheses existed, not because they scored highest.Each is reported alongside evidence against it and agreement from conventional cavity detectors.Source: E Candidate Pockets · p. 32

E.1 Agreement with geometric cavity detectors

The cavity-detector comparisons show that agreement varies by site: DET1 is proposed by all three tools but ranked poorly by fpocket, whereas COP1 is not proposed by geometric detectors. The two candidate sites also differ in their biological interpretation and practical caveats.

E.1 Agreement with geometric cavity detectors

  • For DET1, P2Rank ranked the site fourth of 17, fpocket 39th of 40, and Lacuna first; for COP1, no geometric detector proposed it, while Lacuna ranked a comparable site 24th of 52.P2Rank’s four COP1 proposals included a top proposal sharing only three residues with the predicted pocket; the other three shared none.Source: §E.1 · pp. 32–33, Table A10 · p. 32

E.2 DET1

  • DET1’s predicted 17-residue pocket lies across five sequence-distant segments, but it is not pre-formed in apo structures and is not thereby shown to be unligandable.The lining is solvent-exposed across 16 experimental copies; the apo-reference comparison matters because deleting ligands from holo structures also makes validated sites appear enclosed.Source: §E.2 · p. 33

E.3 COP1

  • COP1’s 15-residue predicted pocket includes the TRIB1 degron hot-spot subsite, but occupancy would inhibit substrate recognition and its direction of effect is substrate-dependent.No small molecule is known at the site; the experimental literature describes the VP pocket as shallow.Source: §E.3 · p. 33
Loading 2610.17696v1…