CryptoBank

Many proteins that matter therapeutically have no obvious pocket to drug. Some of them, however, have a cryptic one: a cavity that is closed in the ligand-free structure and opens only in a higher-energy conformation that a binder can capture and stabilize. Cryptic sites are the main route to expanding the druggable proteome beyond enzyme active sites and known allosteric pockets, but progress has been throttled by data scarcity — before this work, the largest ensemble-aware collection of cryptic sites held on the order of 1,500 structures.

CryptoBank, from the Gervasio lab at the University of Geneva with University College London, addresses that shortage and then builds on it. The first contribution is a database: a supervised scoring function applied to more than 6 million structural alignments of unbound (apo) and bound (holo) chains from the Protein Data Bank, yielding roughly 574,000 apo-holo-ligand combinations that carry a crypticity signal. The second is a predictor: a protein language model fine-tuned on that database to score crypticity per residue directly from sequence, with no structure required at inference time.

Both are served from a public web interface. The work was posted as a bioRxiv preprint in April 2025 and published in Science Advances in April 2026.

#Key Features

• Sequence-only cryptic site prediction: The fine-tuned model assigns a per-residue crypticity probability from amino acid sequence alone, so it can be run on targets with no apo structure.
• Ensemble-based crypticity scoring: A site's score is averaged over every apo-holo-ligand comparison that maps to it, so transient fluctuations are separated from pockets that genuinely stay closed in the unbound state.
• Database scale: 5,151 sites across 3,643 distinct 95%-identity clusters are classified as cryptic — 18.4% of all clusters analysed, and roughly two orders of magnitude more examples than prior collections.
• Simulation-validated predictions: Predictions for four proteins absent from training were followed up with enhanced-sampling molecular dynamics, which opened pockets in the predicted regions.
• Fragment library design: 9,400 ligands found exclusively at cryptic sites are clustered into 60 chemically diverse groups, giving a starting library for cryptic-pocket screening campaigns.

#Technical Details

The structural pipeline filters PDB X-ray and cryo-EM entries at 2.5 Å or better, splits them into chains, assigns apo/holo status after excluding ions, solvents, and sub-60 Da compounds, and supplements sparse apo coverage with one AlphaFold Protein Structure Database model per UniProt accession. This yields ~81,000 apo and ~163,000 holo chains across 19,781 clusters and 56,279 distinct binding sites. A supervised classifier — trained on 199 curated apo-holo pairs (71 cryptic, 128 non-cryptic) with five-fold cross-validation, reaching 89% accuracy — scores each alignment by counting protein atoms in concentric shells around ligand segments, converting steric clashes with the aligned apo state into a crypticity probability in [0, 1].

The predictor fine-tunes ProtTrans's Prot-T5-XL-UniRef50 with low-rank adaptation on the attention layers plus a 1D convolutional head (512 channels, kernel size 3) for residue-level binary classification, trained 50 epochs at a learning rate of 3 × 10⁻⁴ on a single RTX 4090. Splits were made with MMseqs2 at 20% identity by connected component, guaranteeing that no test sequence exceeds 20% identity with training data: 6,345 training, 793 validation, and 793 test sequences. ROC AUC is 0.97 / 0.92 / 0.70 and PR AUC 0.79 / 0.65 / 0.11 across train, validation, and test. That test-set collapse is the model's honest boundary — precision on truly novel sequences is roughly twice random rather than an order of magnitude better — and the authors say so directly.

#Applications

The database gives medicinal chemists experimentally grounded structural evidence for pockets that are invisible in apo structures, including in targets from families conventionally labelled undruggable: 26.3% of single-pocket druggable targets cross-referenced against Open Targets carry a cryptic site, and 33.8% of disease-associated clusters harbour one. The predictor is best used as a triage step, pointing expensive enhanced-sampling simulations or fragment screens at the regions most likely to open — the role it played for TPP1, the VWF A3 domain, and the ROR2 cysteine-rich domain.

#Impact

CryptoBank's main contribution is supply: it turns cryptic-site discovery from a serendipity-driven exercise into one with a training corpus large enough to fit models on. The accompanying language model shows that crypticity leaves a detectable signature in sequence, which is a non-obvious result given that crypticity is defined structurally. Its limits are equally clear. The database inherits the PDB's biases toward well-studied families and conditions; the predictor degrades sharply below 20% identity; and the authors themselves suggest structural information will be needed to close that gap. Code, train/validation/test splits, ligand collections, and the database snapshot are archived on Zenodo under CC BY 4.0, with the GitHub repository as a mirror covering the structural pipeline; the GitHub tree itself carries no license file.