YuelPocket

Locating where a small molecule binds a protein precedes almost everything else in structure-based drug discovery: docking, virtual screening, and lead optimization all presuppose the pocket is known, and screening failures often trace to the wrong site rather than the scoring function. YuelPocket, from Jian Wang and Nikolay Dokholyan at the University of Virginia School of Medicine, recasts pocket detection as a joint protein-small molecule problem: it scores the compatibility between a local protein environment and a specific query ligand, so one protein can return different sites for different molecules.

The central design choice targets a scaling problem. Representing every residue-atom interaction directly needs edges proportional to the product of protein and ligand size, so earlier graph models confine message passing to local neighborhoods and lose long-range context. YuelPocket instead adds a virtual joint node, a global hub wired to every residue and every ligand atom, so any residue reaches any atom in two hops — a global receptive field at an edge count linear in system size.

Posted to bioRxiv in September 2025 and published in the Proceedings of the National Academy of Sciences in March 2026, the released model trains on PLINDER, superseding the smaller MOAD-trained version in the preprint. The authors frame it against a gap left by structure prediction: systems such as AlphaFold 3 transformed how structures are obtained without solving the distinct problem of locating functional sites.

#Key Features

• Virtual joint node: One hub links all residues to all ligand atoms, capturing long-range interactions in two message-passing steps rather than through dense all-to-all edges, at linear cost.
• Ligand-conditioned scoring: Pocket probability comes from a residue embedding interacting with the query ligand's embedding, making predictions molecule-specific rather than generic cavity detection.
• Two output modes: A residue-level mode ranks contact residues; a coordinate-level mode ranks pocket centers sampled on the solvent-accessible surface, returning 3D coordinates for a docking box.
• Coarse-grained residues: Each residue becomes a backbone node at the Cα plus a side-chain node at the side-chain centroid, keeping side-chain orientation without an all-atom graph.
• Robust to predicted structures: Accuracy on AlphaFold models of the test proteins stays close to that on experimental structures, extending use to never-crystallized targets.

#Technical Details

The network runs 16 message-passing layers at hidden dimension 128 over a heterogeneous graph. Intraprotein edges connect nodes within 8.0 Å, intraligand edges are chemical bonds, and global aggregation edges attach every node to its virtual node. Nodes carry one-hot residue or atom types plus a 3D structural mask; edges carry Euclidean distance and four binary type indicators. Residue-level predictions pass the element-wise product of a residue embedding and the ligand virtual-node embedding through an MLP and sigmoid. Coordinate-level predictions seed Shrake-Rupley surface probes, score a local interaction graph around each, and cluster high scorers into sites.

Training used PLINDER: 309,140 systems for training, 832 for validation, and a curated 1,036-complex test split, separated at 30% protein sequence and protein-ligand interaction similarity and 50% ligand Tanimoto similarity. On that test set at a strict 4 Å threshold, residue-level Top-1 success is roughly 48%, above 70% for Top-3 and 90% for Top-10. Across 340 Holo4k complexes, Top-1, Top-3, and Top-10 reach roughly 55%, 75%, and 88%. At the coordinate level, Top-1 Distance to Closest Atom success is about 62% at 4 Å versus roughly 55% for P2Rank, and Center-to-Center about 45% versus 40%. On AlphaFold-predicted structures Top-3 still exceeds 70%, and Top-1 DCA runs about 10 percentage points above P2Rank.

#Applications

The immediate use is defining a search box for docking or virtual screening: a coordinate-level prediction goes straight to a docking engine, while a residue-level prediction checks the site is biochemically sensible. Because scoring is conditioned on the query molecule, the model also addresses selectivity, indicating which candidate site a given scaffold prefers. Its stability on predicted structures suits annotating proteins with models but no bound-ligand crystal structure. Predicted pockets feed pocket-conditioned generative tools such as PocketXMol and complement blind docking methods such as CWFBind.

#Impact

YuelPocket's contribution is architectural: one inexpensive hub node recovers the long-range context local graph models discard, without the quadratic edge budget that motivated discarding it. Peer review in PNAS, a deduplicated PLINDER split, and gains over P2Rank on two geometric metrics make test-set leakage an unlikely explanation, and the AlphaFold experiments address a question pocket predictors are rarely evaluated on. Caveats bound it: the authors report residual sensitivity to side-chain placement as model RMSD grows; the model requires a query ligand, so it does not replace ligand-agnostic cavity detection; evaluation compares mainly against P2Rank rather than the wider deep-pocket-predictor field; and all results are computational. Source code and both checkpoints are on GitHub and benchmark data on Zenodo under CC-BY-4.0, but the repository carries no license file, leaving reuse terms unsettled.