Every worked example, including the misses

Kinase, 298 aa — canonical ATP-competitive site.

Clean rank-1 recovery: the top-ranked pocket cluster contains all 22 of 22 real ATP-site contact residues from a real co-crystal structure (PDB 1HCK) — 100% — including the hinge residues and the catalytic lysine. Whole-chain CA RMSD vs. 1HCK is 1.87 Å, 0.82 Å on the core domain excluding the activation loop (known to be flexible without cyclin bound).

Epigenetic reader, 1362 aa full-length — the acetyl-lysine pocket behind the BET-inhibitor chemical series.

Clean rank-1 recovery: 86% of the real JQ1 contact residues (PDB 3MXF), including the signature Asn140 hydrogen bond. CA RMSD vs. 3MXF is 0.89 Å on the folded BD1 domain. Full-length BRD4 has two bromodomains and large disordered linker regions; the ranker correctly ignored the disorder and landed on the folded BD1 pocket — the viewer’s confidence coloring shows the disordered regions as low-confidence extended chain rather than folded structure, which is what AlphaFold actually predicts for them.

Metalloenzyme, 260 aa — classic fragment-screening validation target.

The strongest structural match in the batch: rank-1 recovery of 8 of 11 real pocket-wall residues (PDB 2CBA) around the catalytic zinc, including the full His94/His96/His119 coordinating triad.

PPI target, 491 aa — the p53-binding groove nutlin-class inhibitors occupy.

Rank-1 recovery of the real p53-binding interface (PDB 1YCR, the p53-peptide co-crystal) — 8 of 22 real contact residues, landing precisely on the druggable PPI cleft rather than the disordered regions elsewhere in this 491-residue protein.

GTPase, 189 aa — the switch-II cryptic pocket that made this target druggable.

Rank-1 recovery of the real switch-II/sotorasib pocket (PDB 6OIM) — 6 of 21 real ligand-contact residues, spanning the P-loop, switch I, and switch II. Found unprompted, on the single most demo-relevant oncology target in this batch.

Chaperone, 732 aa — one of the most fragment-screened ATP pockets in structural biology.

The honest limitation, stated plainly: the real geldanamycin pocket (PDB 1YET) is present in the report, extremely well-resolved — 95% real contact-residue overlap — but ranked #2, not #1. Naive whole-domain RMSD is a poor 12.8 Å, almost entirely driven by the well-documented N-domain "lid" that closes on ligand binding; the core domain matches at 0.70 Å. Real induced-fit flexibility, plausibly the same reason the ranker didn’t put it first.

Immunophilin, 108 aa — the textbook NMR/fragment-screening pocket.

The honest limitation, stated plainly: the real FK506/rapamycin pocket (PDB 1FKJ) is present in the report, correctly identified — but ranked #2, not #1. This was the batch’s own designated sanity-anchor target, chosen specifically because a miss here would call the rest of the batch into question. The pocket wasn’t missed; it was demoted.

GPCR, 413 aa — known-difficult target class, no membrane context in the model.

Deliberately picked as a known-difficult case: this AlphaFold DB model has no lipid-bilayer context for a 7-transmembrane receptor. The real orthosteric pocket (PDB 2RH1, carazolol-bound) is nonetheless fully recovered — all 20 of 20 real contact residues — but at rank 2, not rank 1. A geometry win and a ranking miss, reported as both.

Enzyme, 286 aa — an ensemble-vs-static demonstration, not a ligand-recovery benchmark.

A different kind of finding: the rank-1 pocket only appears in later ANM ensemble frames, absent from the static AlphaFold structure. It lands on a real, independently-published allosteric site (the H3–H4 loop, Ambler numbering 99–114) — not the specific Bowman-lab cryptic pocket this target was originally chosen to test against. We checked the published literature directly rather than assume the original hypothesis held, and report the corrected finding, not the original claim.