Does a frozen allosteric-site detector find allosteric sites? Week 44: 12 cases, 4 hits, and not one of the four is an allosteric site.

TL;DR (valid as of 26 October 2026). The frozen detector (hash 08ad337d, frozen 2026-09-07) stands at 12 hits in 27 prospective cases: 44.4%, 95% Wilson interval 27.6–62.7%, down from 53.3% last week. Twelve structures released in the 2026-10-19 to 2026-10-26 window were scored blind; four were hits. We then audited what each answer ligand chemically is. Exactly one of the twelve cases is a genuine allosteric site — 29CY, the SHP2 allosteric latch, targeted covalently on lysines — and the detector missed it with zero residue overlap, even though SHP2 is in its own 86-protein training set and all 14 ground-truth residues are distal. The other eleven answer ligands are active-site inhibitors, catalytic cofactors, a structural pigment and a membrane lipid. None of the four hits is an allosteric site: two are sulfamates on the catalytic zinc of carbonic anhydrases XII and XIII, one is chlorophyll a in photosystem II, one is a 1 kDa polyanion in cGAS. Two of those four also die under the routine's own mandatory checks: 31FD fails the distance-matched null (0.3405) and drops to rank 32 under anchor dilation; 22UK drops to rank 7 at R=6 A and stops being distal at R=8 A. The challenger stands at 5/28 against the incumbent's 12/28 — no promotion. All four quantitative readings of the scoreboard still contain our 32.6% retrospective rate. Zero forward hypotheses published. Zero QPU consumed.

The number, five ways

The scoreboard is reported under every admission rule we have on the table, because which cases are allowed in moves the number more than the detector does. Reading A is the rule actually in force. Reading E is new this week: it keeps only the cases whose answer ligand is, by chemical audit, an allosteric modulator rather than an active-site ligand, a cofactor or a lipid.

What each answer ligand actually is

A case counts as a hit when the top-1 predicted pocket touches at least one residue of the ligand's contact shell. That makes the chemistry of the ligand the whole game. Here it is, case by case.

The one real allosteric case, and the miss

29CY is human SHP2 (PTPN11) with a covalent fragment on lysines at the allosteric latch — the same region where SHP099 binds, the best-characterised allosteric site in the phosphatase and the basis of the allosteric inhibitors now in the clinic. The case is as favourable as a prospective case gets: the anchor has 9 annotated residues, all 14 ground-truth residues are distal (distality ratio 1.00), and the protein itself is in the detector's training set. The frozen detector predicted elsewhere, and its prediction was stable: rank 1 under anchor dilation at both R=6 A and R=8 A. It did not fail because the case was ambiguous. It failed with conviction.

What we know and what we do not know

What we know. Over 27 structures published after the model was frozen, the detector's top-1 pocket touches the answer ligand 12 times (44.4%, 27.6–62.7%). That interval contains our own retrospective leave-one-protein-out rate of 32.6%, so after eight weeks of prospective measurement we still cannot distinguish the two. We know the chemistry of every answer ligand in this week's 12 cases, and we know that only one of them is an allosteric modulator. We know the detector missed that one. We know the predicted pockets keep landing in the 5–19% distance percentile from the active site (median 12.6%) while real allosteric sites in our validated set live around the 28% percentile — the periphery bias we measured in August, now in its seventh consecutive week.

What we do not know. We do not know the detector's hit rate on allosteric sites proper, because our admission filter does not select for them: in 27 cases it has admitted exactly one audited allosteric modulator. We do not know whether the four weekly readings would separate from 32.6% at larger N, because n=27 is too small. We do not know whether the two carbonic-anhydrase hits mean anything: the same detector, on the same family, missed carbonic anhydrase II last week. We do not know whether a dilated catalytic anchor would fix the admission problem, because that change is proposed and not applied, and it needs a human decision before it touches the scoreboard.

Limit statement. These are prospective predictions from a frozen model against structures it never saw, and nothing here authorises anyone to discard a site. A miss by this detector is evidence about this detector, not evidence that a site is not allosteric. Our negatives do not license anyone to rule a site out.

Mechanics

The examinee is a set of frozen weights (10 de novo features, no annotation features) with hash 08ad337d; its seven pinned code files verified byte-exact against their recorded sha256 before any science ran. The primary metric is de novo without annotations. Every case carries a distance-matched null (±2 A band, 2000 permutations, seed 20260717), an anchor-dilation check at R=6 and R=8 A computed with the same frozen weights, and a druggability check by a declared geometric surrogate (fpocket still segfaults in qhull, so the surrogate is labelled as such and never reported as fpocket). The blind prediction timestamp precedes the unveiling timestamp in every case; both are in the seal. The challenger may retrain on anything; it is never published as our number. Promotion requires the challenger to win on paired cases with McNemar p<0.05, and this week the paired record is 12–5 for the incumbent with 9 discordant pairs (p=0.0391 in the incumbent's favour). Content hash of this run: sha256:5a15a963ac62072b881a3f9aeb433641b980c71d5e7fa8f8f58ac0daa8709ba3.

Sources

1. RCSB PDB 29CY — SHP2 allosteric latch, covalent lysine targeting
2. RCSB PDB 29SV — carbonic anhydrase XII with a sulfamate
3. RCSB PDB 29SY — carbonic anhydrase XIII with a sulfamate
4. RCSB PDB 31FD — photosystem II from Arabidopsis thaliana
5. RCSB PDB 22UK — human cGAS with NF023
6. RCSB PDB 12WW — FGFR3 kinase domain with inhibitor
7. RCSB PDB 9SXG — PDGFRA T674I with AZD7762
8. RCSB PDB 9SXL — ERK2 with a 194 Da fragment
9. RCSB PDB 9TLF — human TDO2, heme
10. RCSB PDB 9YN2 — mouse RyR1 T4706M
11. RCSB PDB 9YO2 / 9YO3 — integrin a4b1 headpiece with firategrast
12. RCSB Chemical Component Dictionary — ligand chemistry for all 12 answer ligands
13. UniProt — functional site annotation used as the anchor
14. Chen et al. 2016, Nature 535:148–152 — SHP099 and the SHP2 allosteric site
15. RCSB PDB Search API — release-date window query