Loading…
Residue-level prediction of ligand-binding and cryptic pockets in one model, reading frozen ESM-2 embeddings with no 3D structure at inference.
No providers recorded yet. Browse all providers
In bovine carboxypeptidase A, Tyr248 sits solvent-exposed and pointing away from the active-site cleft until a ligand arrives, at which point it rotates roughly 120° into a closed conformation that hydrogen-bonds the ligand carboxylate. Read the apo coordinates and there is no pocket there to find. That is the cryptic-site problem, and it is why binding-site prediction has run as two separate toolchains: geometry-based finders such as fpocket and P2Rank for conventional ligand-binding sites, and CryptoSite or PocketMiner for cryptic ones, each with its own inputs, preprocessing and output format for two analyses a drug-discovery campaign performs together.
UniPocket, from Lening Zhao, Tianhua Zhai and Li Shen at the University of Pennsylvania, collapses that split by dropping structure from the input entirely. Per-residue embeddings from a frozen ESM-2 650M backbone feed a shared residual MLP that branches into two lightweight heads — one scoring ligand-contact residues, one scoring cryptic-pocket residues. The premise is that residues predisposed to open a pocket carry a co-evolutionary signature even when the apo backbone shows nothing.
The unification is structural rather than cosmetic. Training alternates ligand-labeled and cryptic-labeled mini-batches through the shared backbone, and an orthogonality regularizer penalizes the squared cosine similarity between the two heads' read-out weight vectors, forcing task-relevant structure into the backbone instead of into nearly parallel read-outs. The work follows CryptoBank and CryptoBench's pLM-NN baseline in treating pretrained sequence representations, rather than molecular dynamics, as the substrate for crypticity prediction, and was published at ACM-BCB 2026.
The backbone is esm2_t33_650M_UR50D, frozen and never fine-tuned, producing a 1,280-dimensional vector per residue. The trainable network compresses 1280→256, widens to 512, applies two residual blocks with LayerNorm and GELU, narrows to 256, and branches into two 256→128 heads. Ligand supervision comes from RCSB structures deposited between September 2021 and 2024 — 14,877 structures filtered to 4,051 carrying a non-polymer ligand, then to 2,823 chains after removing waters, ions, cryoprotectants and packing sugars and deduplicating at 50% identity with MMseqs2. A residue is positive if any heavy atom lies within 5.0 Å of the largest-contact ligand. Cryptic supervision is CryptoBench's per-residue labels used verbatim, 975 of 1,107 structures retained. A bidirectional MMseqs2 search between the two sets returned zero hits at both 50% and 40% identity. On held-out splits the unified model reaches macro ROC-AUC 0.816 on cryptic prediction, against 0.732 for the strongest geometry-augmented PocketMiner-embedding baseline, and 0.761 on ligand-binding, against 0.618 for P2Rank and 0.593 for fpocket. Multitask training lifts the ligand head from 0.744 while leaving the cryptic head statistically unchanged.
The natural use is early target prioritization, where thousands of proteins must be ranked before anyone commits to crystallography, fragment screening or enhanced-sampling simulation. Because the input is a sequence, targets with no experimental structure and no apo-holo pair are scored on the same footing as well-characterized ones. Converting residue scores into a docking box takes a post-hoc step: thresholding at 0.5 and clustering Cα coordinates with DBSCAN (ε = 8 Å, minimum 3 points) recovers the correct pocket as the top-ranked cluster for 192 of 193 CryptoBench test proteins, well above P2Rank at 59.1% and fpocket at 42.0%. That step needs 3D coordinates, so the sequence-only claim covers the residue scores, not the pocket-level workflow.
UniPocket's contribution is the demonstration that one frozen protein language model representation carries enough signal for both halves of pocket prediction, and that training them together helps the noisier ligand task rather than diluting either. The evidence is honest about its limits: the authors report PocketMiner scoring 0.499, essentially random, and explain that this reflects a per-residue protocol rather than CryptoBench's pocket-level overlap criterion, so the gap is not directly comparable. Retaining only the largest-contact ligand per protein underrepresents fragments and multi-site proteins, and both training sets skew toward structurally characterized families, limiting reach into intrinsically disordered targets. No code or trained weights have been released, so reproduction means rebuilding the pipeline from the published description.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.