Ranking Cryptic Pockets from Closed Structures

A pocket that is closed in an apo structure can become accessible in other conformations, including transiently without a ligand. We asked how well cryptic-pocket detectors recover such a site from the closed, experimental apo structure alone, on two targets where the answer is now known.

On SMURF1, the alpha and surface detectors ranked a pocket at the inhibitor cavity 2nd and 3rd. Their pocket centres were 0.79 angstroms apart, with four shared contact residues. Static fpocket first reached the site at rank 18. On hnRNPA1/UP1, all four methods placed a qualifying pocket in their top five. The two detectors' best pockets were 16.8 angstroms apart, however, with no contact residue in common.

Under our permissive hit rule, a top-five rank says a detector touched part of the site. Residue-level agreement helped narrow the candidate region in SMURF1. These are different detector modes within Lacuna, with shared pipeline components; statistical independence has not been established. Two selected targets cannot turn that observation into a general rule.

Added targets

2

SMURF1 shared contact residues

4

Alpha and surface best hits

Pockets scored

115

Rothman et al. (Cell, 2025) found SMURF1 inhibitors that bind a cryptic cavity in the N-lobe of the HECT domain, away from the catalytic cysteine. In the unbound enzyme, Ile631 and its neighbours fill that cavity; the inhibitor displaces helix H10 and extends it across a conserved glycine hinge (Gly634). We used their apo crystal structure, 9FSK (2.75 angstroms), as detector input and the compound 8 complex, 9FSJ (ligand TJQ, 2.05 angstroms), as the answer.

Dunnett et al. (JBC, 2025) combined cryptic-pocket mapping with an X-ray fragment screen on UP1, the RNA-binding core of hnRNPA1. The input here is the deposited representative model of the NMR structure 2LYV; the answer is fragment K3A in 9F53 (1.80 angstroms).

We checked whether the bound ligand pose fits the apo input. After fitting the holo protein onto the apo input, 20 of 27 TJQ heavy atoms land within 2.2 angstroms of an apo protein atom. That fit uses the SMURF1 N-lobe (251 residues, RMSD 3.34 angstroms). The whole-protein fit is distorted by the large domain motion (RMSD 12.55 angstroms) and puts 26 of 27 atoms in clash, so the N-lobe count is the conservative one. For UP1 all 10 K3A heavy atoms clash after a global fit (175 residues, RMSD 2.98 angstroms), and every one of the 20 deposited NMR models still blocks the site. These clashes support occlusion in the fitted apo structures; they do not establish that opening requires ligand binding.

The detectors come from Lacuna, an open-source cryptic-pocket tool by Clayton W. Moore (Moore, bioRxiv 2026). Lacuna builds a conformational ensemble from one structure, finds pockets in every conformer and ranks the persistent sites with fitted models. We thank the author for releasing the code, weights and benchmark folds under an MIT licence, which made the exposure audit below possible. We ran version 1.1.1 on the Crane Protein Intelligence Platform: normal-mode ensembles of 20 conformers plus the input, the top 15 pockets, and three detectors (alpha, surface and surface fusion). Static fpocket ran on byte-identical inputs with default settings.

A pocket counts as a hit when its lining includes at least three of the residues within 4.5 angstroms of the experimental ligand. The endpoint is the rank of the first such pocket. Two detectors "agree" only if their best hits sit within 8 angstroms of each other and share at least three contact residues. The targets, inputs, answer keys and scoring code were fixed and recorded before any detector ran, so a later change to any of them would be obvious. The per-target verdicts were written before this interpretation, and nothing was retuned afterwards. The full frozen protocol and raw detector reports are not provided on this page.

This exploratory hit rule is not equivalent to Lacuna's benchmark recovery. Lacuna's v1.1.1 benchmark requires a top-five pocket with Jaccard overlap of at least 0.25, or a pocket centre within 4 angstroms of the site centroid. Jaccard divides the shared residues by all distinct residues in the pocket and reference site, penalising oversized pockets. Our three-contact rule does not.

Rank of the first pocket with at least three contact residues; top five means rank 5 or better. MALT1 comes from an earlier cycle and was not rerun. Its structure is in Lacuna's evaluation set, so it is a comparison, not evidence of generalisation.

Rank does not measure hit quality. SMURF1 alpha shares seven contact residues, while fpocket at rank 18 shares exactly three. MALT1 alpha at rank 1 and fpocket at rank 5 also meet only the three-residue minimum.

Distance is measured from the predicted pocket centre to the mean C-alpha position of the scored contact residues in the apo input, not to the ligand centre. Every distance listed exceeds 4 angstroms; UP1 surface, fusion and fpocket nevertheless exceed 0.25 Jaccard. These are the first hits under our three-contact rule, not necessarily the best pockets under either localisation metric. The values use our contact keys and are not a reproduction of Lacuna's benchmark.

Alpha and surface agree under the frozen rule: 0.79 angstroms apart, sharing Ile506, Gly508, Phe510 and Tyr515. The alpha pocket lines 7 of 22 contact residues, including Ile631 (the residue that fills the cavity in the apo state); 7 of the 20 residues lining it are contacts. That is partial localisation, not a reconstruction of the open pocket.

Those counts give the alpha pocket a Jaccard overlap of 0.20 (7 / 35), below Lacuna's 0.25 threshold. Its archived distance to our apo contact-residue C-alpha centroid is 7.50 angstroms. The 0.79-angstrom separation above measures agreement between the two predictions, not their distance to the reference site. These metrics use our contact key; they are not a rerun of Lacuna's benchmark protocol.

Surface fusion, which pools both detectors' candidates, dropped the best hit from rank 2 to rank 8 and lost top-five status in this single fixed-seed run. Run-to-run variance is unmeasured, so this does not establish that fusion generally ranks the site worse. The result sits alongside a finding in Lacuna's own benchmarking (Moore, bioRxiv 2026): detector consensus gave no measurable gain at a budget of five candidates, though it does gain at larger budgets. That finding pools candidates across separate tools, while surface fusion pools Lacuna's own two detectors, so the two are adjacent rather than the same test. Static fpocket did propose a qualifying pocket, only at rank 18. Because detector and ranking differ as well as sampling, the gap to fpocket cannot be credited to the ensemble alone.

How much did conformational sampling contribute? The 12.55-angstrom whole-protein RMSD compares the experimental apo and holo structures, not the generated ensemble. It cannot tell us whether the detected region was already accessible near the apo state or explain the fusion rank change. A controlled follow-up would compare apo-only and ensemble Lacuna runs with detector and ranker held fixed, and run fpocket over the same 21 conformers. We have not run those controls.

Every method put a qualifying pocket in its top five (ranks 1, 4, 3 and 4). The alpha pocket reaches Glu93, Asp94, Ala100 and Lys179; the surface pocket reaches Arg88 to Arg92 and His101. Those centres are 16.8 angstroms apart. A pocket needs only three residues in common with the ligand site to count, so two top-five hits can sit on opposite sides of it.

Static fpocket also passed at rank 4, so UP1 offers no case for an ensemble-only advantage. The UP1 study itself mapped cryptic pockets with fpocket over molecular dynamics snapshots started from the same 2LYV structure, cross-checked against PocketMiner's graph neural network predictions.

A recent PDB release does not prove a model never saw the outcome, so we audited Lacuna's public training and evaluation data directly: all 1,107 apo entries and 5,493 apo/holo answer associations in CryptoBench, plus sequence-homolog searches and Lacuna's saved evaluation outputs. The three fitted model files and the checked runtime source match the pinned release and source commit byte for byte. These runs also embedded each sequence with ESM-2, and the image does not report which ESM-2 checkpoint revision it loaded, so that component is recorded but unverified.

UP1 had no matching protein or bounded homolog. SMURF1 had no same-protein or same-site match, but its relative WWP1 (52 percent identical, sharing 11 of the 22 contact-key positions) sits in the test fold, annotated at a different site that does not overlap the TJQ cavity. SMURF1 is therefore not a family-held-out test. Lacuna's documentation also states that the test fold was measured repeatedly during development, and private development history is not public. The supported description is two targets with no identified same-protein or same-site overlap in the audited records; this is not a certified blind holdout.

This workflow fits a project with an apo structure, a suspected allosteric or cryptic site, and a choice about where to spend fragment screening or mutagenesis effort. It generates hypotheses for experimental follow-up. Docking still requires a suitable open receptor conformation; a ranked region alone does not supply one. For one target, the output is a ranked pocket list from several detectors, residue-level agreement between them, an occlusion check against any holo structure you hold, and an exposure audit of the models used. Pocket prediction needs only one apo structure. Scoring against known contacts and checking occlusion require a holo reference. Compute per target runs in minutes, so the work that takes real time is building the answer key and the exposure audit.

Two targets, chosen for mechanism and geometry rather than sampled at random, so no pooled accuracy figure is justified. Each call ran once with fixed seeds, which leaves run-to-run variance unmeasured, and no compound went near a bench. The next step that would change the picture is a larger, prespecified panel with replicate ensembles.

• Moore, C. W. (2026). Lacuna: Cryptic Binding Pocket Discovery via Conformational Ensemble Analysis. bioRxiv. Code: github.com/mooreneural/lacuna.
• Moore, C. W. (2026). Cryptic binding sites are detected but not ranked. bioRxiv.
• Rothman, A. M. K. et al. (2025). Therapeutic potential of allosteric HECT E3 ligase inhibition. Cell.
• Dunnett, L. et al. (2025). Enhanced identification of small molecules binding to hnRNPA1 via cryptic pockets mapping coupled with X-ray fragment screening. Journal of Biological Chemistry.
• Škrhák, V. et al. (2025). CryptoBench: cryptic protein-ligand binding sites dataset and benchmark. Bioinformatics 41(1), btae745; dataset download.