Does a detector frozen three weeks ago find the allosteric site in structures that did not exist when it was frozen?

TL;DR (2026-09-28 · valid as of this date): Week 2026-W40: of 190 PDB entries released between 2026-09-21 and 2026-09-28, exactly one survived to be rendered — 36CV, cyanate-bound bovine cytochrome c oxidase. The frozen detector (hash 08ad337d, 86 training proteins, 10 annotation-free features) predicted blind at 12:22:00Z and picked a pocket around residue A439. That entry carries three ligands, and the answer key is not unique: the pocket overlaps the heme A site by 4 residues (hit), and misses the Cu-A centre and the cyanate site entirely (miss, miss). Under the rule currently in force — score against the ligand with the most distal residues — the case is a hit and the prospective scoreboard reads 3 of 5, 60%, Wilson 95% 23–88%. Read heme A as what it is, the catalytic prosthetic group of a respiratory enzyme, and the case is a miss: 2 of 5, 40%, 12–77%. Under the stricter rule proposed last week the case would not have been eligible at all and the board would stay at 2 of 4. All three intervals contain the same detector's retrospective rate of 32.6% (N=86), so the ambiguity changes no conclusion. It should change the rule, and that decision is open. Independently of the ligand question, the predicted pocket fails the mandatory distance-matched null (frac_above 0.4145) and moves under anchor dilation at both 6 Å and 8 Å: even the reading that scores it a hit describes a pocket indistinguishable from a random one at comparable distance from the active site. The student model, which residualises the examinee's score against distance, hit the cyanate site where the examinee missed it and missed the heme where the examinee hit it; McNemar has 3 discordant pairs and no power, so no promotion is proposed. No forward hypothesis was published. The blind-retrospective line holds at 0 of 1.

What we know

One structure out of 190 released this week was eligible, and it was rendered blind: the prediction is timestamped 2026-09-28T12:22:00Z, before the holo was opened. The frozen detector's hash recomputes to 08ad337d and all seven pinned code files reproduce their SHA-256 byte for byte, so the model that predicted this week is the same object that predicted in week 37. The predicted pocket sits at the 8.5th percentile of distance from the active site inside its own protein, which is where this detector has always fired and is not where real allosteric sites live (28th percentile). The pocket does not pass the distance-matched null and does not survive anchor dilation.

What we do not know

Whether this case is a hit. Entry 36CV carries heme A, a Cu-A centre and cyanate; the detector's pocket touches the first and neither of the other two. The rule in force picks the ligand with the most distal residues, which is heme A — the catalytic prosthetic group of the enzyme. Nothing in the frozen detector decides this; it is a question about how a case is defined, and it is the third week running that incomplete functional annotation has made a catalytic cofactor look distal. We also do not know whether the detector is better or worse than 32.6%: with five prospective cases every interval on this page contains that rate. And we do not know whether the student's inverted result — it found the cyanate site, the examinee found the heme — is signal or noise, because three discordant pairs cannot tell us.

Method, unchanged

Primary metric is de-novo prediction without annotations: ten features, no UniProt annotation columns, frozen weights, no refitting. The null is matched by distance to the active site (±2.0 Å band); the contiguous null is reported beside it, never in its place. Anchor sensitivity is tested by geometric dilation at 6 Å and 8 Å. Druggability uses an explicit geometric substitute — burial by CA neighbours within 10 Å plus hydrophobic fraction — and is labelled as such, because fpocket still segfaults in its qhull call in this environment. The quantum arm did not run: it only runs on a promotion, and no promotion was proposed. No QPU time was consumed.

Limit statement

Our negatives do not authorise anyone to rule out a site. This scoreboard measures whether one frozen detector points at the right pocket in structures published after it was frozen. A miss is information about the detector, not about the biology. No result on this page should be used to conclude that a protein lacks an allosteric site, and no result on this page is a discovery: every predicted pocket is a dated hypothesis.

Sources

1. RCSB PDB — entry 36CV, cyanate-bound bovine cytochrome c oxidase, released in window 2026-09-21 to 2026-09-28 (queried 2026-09-28)
2. RCSB PDB — entry 35VQ, engineered retinol-binding protein 2 variant with ZFJ (rejected: no annotated active site)
3. RCSB PDB — entry 23JV, human cytosolic NADP-dependent malic enzyme with 2OP (rejected: site not distal)
4. RCSB PDB — search API, release-date window 2026-09-21 to 2026-09-28, 190 entries, all inspected (queried 2026-09-28)
5. UniProt — functional-site annotations used as anchor: P00396 (MT-CO1, 12 annotated residues), P48163 (ME1), P50120 (RBP2)
6. Rosetta Q run record RQ-PROSP-2026-W40, sealed 2026-09-28, content_hash sha256:0e36c8a1…, convention rosettaq-archive/v2
7. Rosetta Q run record RQ-PROSP-2026-W39, prospective scoreboard 2 of 4, content_hash sha256:82ff28f3…
8. Rosetta Q — 2026-W37 frozen detector, recorded hash 08ad337d (86 training proteins, 10 annotation-free features), retrospective LOPO rate 32.6%
9. Wilson, E. B. — Probable inference, the law of succession, and statistical inference, J. Am. Stat. Assoc. 22, 209 (1927)
10. Yoshikawa & Shimada — Reaction mechanism of cytochrome c oxidase, Chem. Rev. 115, 1936 (2015) — heme a/a3 and Cu-A as catalytic prosthetic groups