Does a frozen allosteric-site detector get better or worse as the prospective sample grows?

TL;DR (valid as of 2 November 2026). The frozen detector (hash 08ad337d, frozen 2026-09-07) stands at 16 hits in 42 prospective cases: 38.1%, 95% Wilson interval 25.0-53.2%, down from 44.4% last week. Fifteen structures released between 2026-10-26 and 2026-11-02 were scored blind, each predicted before the ligand was uncovered. Four were hits. None of the four is on an allosteric site. Exactly one of the fifteen answer ligands is a verified allosteric modulator, and the detector missed it. Ninth week of prospective measurement; every quantitative reading still contains the 32.6% retrospective rate.

The harvest

305 entries were released in the seven-day window and all 305 were inspected: 130 had no non-polymeric small-molecule ligand, 103 were neither human nor a model organism or carried no UniProt mapping, leaving 72 candidates. Deduplicating by accession removed 35, and the standing exclusion rule for accessions already scored in an earlier week removed 1. Of the 36 unique candidates, 18 — exactly half — failed on the anchor: UniProt annotates fewer than three mappable functional residues, so the protocol cannot define what counts as distal. Three more were dropped because the real site had fewer than five distal residues. Fifteen were scored.

That half-of-the-harvest anchor failure is the dominant failure mode of the week, and it is an admission failure, not a detector failure. Those eighteen cases are not evidence about the model in either direction.

The fifteen cases

The chemical audit of the answer ligands

This is the measurement that answers the project question, and it is the second consecutive week in which it contradicts the headline rate. Of the fifteen answer ligands: one is a verified allosteric modulator (24PT, SB225002, occupying the colchicine and vinblastine regions of tubulin per Liu et al., J Biol Chem 2026); one is a probable but unverified cryptic-cavity binder (9ZRW); five are active-site or competitive ligands; six are catalytic nucleotides, substrates or an orthosteric agonist; and two are crystallisation artefacts. Zero of the four hits is an allosteric site.

The single verified allosteric case was the one the detector missed, and it missed it with conviction: 24PT has a 17-residue anchor, distality ratio 0.85, matched-null 0.008 and rank-1 invariance at both R=6 and R=8. That is the same failure shape as 29CY (SHP2) last week. Across 42 prospective cases the admission filter has let through two verified allosteric modulators, and the frozen detector has missed both: reading E is 0/2, 95% interval 0 to 65.8%.

Two admission bugs, both measured rather than anecdotal

One. The answer-ligand selector picks whichever candidate ligand has the most distal contact residues. In 9XB0 that chose free glycine over the actual inhibitor 6IC: selecting by residue name picks up every scattered glycine in the crystal, producing a 71-residue pseudo-site with 52 distal residues, at which point a hit is nearly automatic. One of this week four hits is that. Two. P4K, a polyethylene glycol, passed the common-additive filter in 9SYW; that filter excludes PEG, PG4, PGE, P6G, 1PE, 2PE, PE4 and PE8 but not P4K.

Neither case is removed from this week numbers. 9XB0 is a hit and 9SYW is a miss, so dropping them lowers the figure rather than raising it; we publish that as a separate reading and fix the rule for next week, not for this one. The rule has to be decided before scoring, not case by case.

fpocket, for the first time in this series

Druggability has been gated by an explicitly declared geometric substitute since week 37 because fpocket was unavailable. This week it was built from source and run for real. The two criteria disagree in an informative way: fpocket says the predicted pockets are genuine cavities (13 of 15 have volume at or above 100 cubic angstroms, median 228) while the geometric substitute passed only 6 of 15. But fpocket own druggability score is 0.005 or below in 14 of the 15; the only pocket it scores as druggable is 9XB0 at 0.62, which is the glycine artefact case. The series continues to report the substitute, because switching criteria mid-series without saying so is exactly the error this engine exists to avoid. Both numbers are in the seal.

Where they agree is the thing that matters: four cases clear the matched null at or below 0.05 and rank-1 invariance at both R=6 and R=8 (9SYP, 9SZ0, 9XBT, 24PT), and all four die on druggability, with fpocket scores of 0.000, 0.001, 0.002 and 0.000. Zero forward hypotheses this week.

The periphery bias, eighth consecutive week

The fifteen predicted pockets sit at the 3.9 to 38.3 percentile of distance to the annotated active site, median 13.2%. The real sites of these same fifteen cases sit at a median of 29.5% — which is, to within a point, the 28% of the validated training set. The detector fires too close to the active site, the measurement is stable across eight weeks, and it is sufficient on its own to predict that the prospective rate will stay near the retrospective one.

Challenger

Challenger v0 (the examinee score residualised by decile of distance to the anchor) scored 1 of 15 this week against the incumbent 4 of 15. Cumulatively the incumbent is 16/43 and the challenger 6/43 on paired cases, with 12 discordant pairs, 11 of them in the incumbent favour, exact McNemar p = 0.0064. No promotion. The promotion rule only fires when the challenger wins. The quantum arm did not run; it only runs inside a promotion. Zero QPU consumed.

All readings, published together

All six quantitative readings contain the retrospective leave-one-protein-out rate of 32.6% (28/86). Nine weeks of prospective measurement and the prospective scoreboard still cannot be distinguished from the retrospective one. Two of this week hits, 20ZO (VCP) and 9XB0 (KRAS), are proteins that are in the frozen training set: identity 1.00. Identity for 24PT and 9SYP could not be measured because UniProt returned no sequence for them through either route; they are declared unmeasured rather than imputed.

What we know and what we do not know

What we know. The detector hash is unchanged and verified byte-exact, so all 42 cases were scored by the same model. Every prediction carries a timestamp earlier than the timestamp at which the answer was uncovered. The cumulative rate fell from 44.4% to 38.1% as the sample grew by 15. None of the four hits this week is an allosteric site; the one verified allosteric modulator was missed. Four pockets survive the matched null and both dilation radii and none survives druggability.

What we do not know. Whether the rate settles above or below the retrospective 32.6%: at n=42 the interval is still 28 points wide. Whether the two crystallisation artefacts are the only ones in the series; weeks 37 to 44 were not re-audited for that specific failure. Whether 9ZRW is genuinely an allosteric case, which is why it is not counted in reading E. Whether fpocket near-zero druggability scores reflect the pockets or a mismatch between fpocket training distribution and single-chain cryo-EM inputs. And whether the anchor failure that removed half this week harvest is biased with respect to the answer, which would make the admitted set unrepresentative in a direction we have not measured.

Limit declaration

Every predicted site is a hypothesis with a date, labelled FORWARD. It is not a finding and not a discovery. And our negatives authorise nobody to dismiss a site: when this detector misses a pocket, that is a fact about the detector, not evidence against the pocket. A site we rank low may be real and well characterised; four of this week fifteen answers are sites we ranked low and that are in the literature.

Reproducibility

Frozen examinee: hash 08ad337db711d7e77a4d66a6208aa88956fb7f71507683f2db47b61649ef0dd0, frozen 2026-09-07, 86 training proteins, seven code files pinned by sha256 and verified byte-exact this run. Primary metric is de novo without annotations, ten features, no UniProt columns. Matched-distance null is mandatory, 2000 permutations, seed 20260717. Run seal RQ-PROSP-2026-W45, content hash sha256:7249a9bf, sealed with rosettaq_seal.py 3.1.0.