Pillar D · State as of 2026-10-19

Week 43: 5 eligible structures, 2 hits. Prospective total 8/15 (53.3%, 95% CI 30.1-75.2%) — and both hits fail anchor dilation

As of 19 October 2026 the frozen allosteric-site detector stands at 8 hits in 15 prospective cases: 53.3%, 95% Wilson interval 30.1-75.2%, down from 60.0%. Five structures published in the 2026-10-12 to 2026-10-19 window were scored blind and two were hits. Neither hit survives the routine's own mandatory check: 29ZZ and 9O69 are the only two of the five cases that lose rank 1 when the functional anchor is dilated to 8 A, falling to rank 15 and rank 11, while all three misses stay at rank 1-3. None of the five answer ligands is an allosteric modulator: an Fe3S4 catalytic cluster, a vanadate transition-state model in the nucleotide pocket, a crosslinked substrate arm, nirmatrelvir on a catalytic cysteine, and a sulfonamide on the active-site zinc of carbonic anhydrase II. The run also pins the root cause of a failure mode logged since week 37: UniProt annotates catalytic residues, not catalytic sites, so in MYH11 the whole functional annotation is the ATP P-loop 178-185 and in cathepsin K and fabB it is a bare three-residue triad, leaving the rest of the catalytic machinery labelled DISTAL. All five readings of the scoreboard still contain our 32.6% retrospective rate. Zero forward hypotheses published. Zero QPU consumed.

→ Leer en español

State as of: 2026-10-19

Status as of: 19 October 2026. The frozen allosteric-site detector scored five structures published after it was frozen and got two right. The cumulative prospective rate falls from 60.0% to 8/15 (53.3%, 95% Wilson 30.1–75.2%). The number is not the story. None of this week’s five answer ligands is an allosteric modulator, and the two hits are the only two cases that lose their top rank when the anchor is dilated by 8 Å — the engine’s own mandatory check discards both. This run also identifies the root cause of a failure mode logged since week 37, and it is general: UniProt annotates catalytic residues, not catalytic sites.

What was measured?

124 entries were released in the seven-day window 2026-10-12 to 2026-10-19 at 3.0 Å or better with at least one non-polymeric small-molecule ligand. All 124 were inspected. 29 reached geometric verification, 10 after deduplicating by UniProt accession, and 5 were accepted. The detector predicted each pocket blind, from the apo structure only, with a timestamp written before the holo was opened; the reveal followed. Hits: 29ZZ (smooth muscle myosin, overlap 4) and 9O69 (cathepsin K, overlap 1). Misses: 28FU, 9MLX, 9SW8.

case	protein	answer ligand	what the ligand actually is	distality ratio	hit	matched null	cavity check	source
28FU 2.48 Å	hybO P69741	F3S 7 heavy / 295.8 Da	Fe3-S4 cluster: catalytic cofactor of hydrogenase-2	0.46 (6/13)	no	0.000 passes	fails	RCSB 28FU · UniProt P69741
29ZZ 2.30 Å	MYH11 P35749	VO4 5 heavy / 114.9 Da	vanadate ion: transition-state model for ATP hydrolysis at the catalytic site	0.46 (6/13)	yes	0.000 passes	passes	RCSB 29ZZ · UniProt P35749
9MLX 2.05 Å	fabB P0A953	MRJ 35 heavy / 523.6 Da	pantetheinyl-acyl arm of the FabB-ACP crosslink: the substrate	0.75 (24/32)	no	0.020 passes	passes	RCSB 9MLX · UniProt P0A953
9O69 1.90 Å	CTSK P43235	4WI 35 heavy / 501.5 Da	nirmatrelvir: covalent peptidomimetic active-site inhibitor	0.56 (10/18)	yes	0.033 passes	fails	RCSB 9O69 · UniProt P43235
9SW8 1.107 Å	CA2 P00918	A1JQ3 44 heavy / 647.7 Da	sulfonamide carbonic anhydrase II inhibitor: binds the active-site Zn	0.50 (10/20)	no	0.055 fails	fails	RCSB 9SW8 · UniProt P00918

Why neither hit survives the engine’s own checks

Rule 3 of this routine makes anchor dilation mandatory: re-run the prediction with the functional anchor grown to every residue within 6 and then 8 Å of an annotated catalytic residue, and check whether the predicted pocket keeps its rank. A pocket that is genuinely distal should not care. This week the result is unambiguous and it points the wrong way for the scoreboard: the only two cases that lose rank 1 under an 8 Å dilated anchor are the two hits — 29ZZ drops from rank 1 to rank 15, 9O69 from rank 1 to rank 11. The three misses stay at rank 1, 1 and 3. The hits are not robust to where the catalytic site is drawn; the misses are.

What the five answer ligands actually were

A hit only means something if the ligand that defines the answer is an allosteric modulator. We checked all five against the RCSB Chemical Component Dictionary and the primary literature. None of them is.

The vanadate in 29ZZ is a five-atom, 115 Da ion that models the transition state of ATP hydrolysis, bound in the nucleotide pocket coordinated to Mg (Smith & Rayment 1996). Its contact shell is switch I and switch II — catalytic machinery. The ligand in 9O69 is nirmatrelvir, a covalent peptidomimetic that targets the catalytic cysteine of cysteine proteases; the overlap that earns the hit is a single residue, A19, in the active-site cleft. 28FU’s F3S is a seven-atom Fe₃S₄ cluster, the catalytic cofactor of hydrogenase-2. 9MLX’s MRJ is the pantetheinyl-acyl arm of a deliberate ketosynthase-ACP crosslink: the substrate. 9SW8’s A1JQ3 is a sulfonamide that binds the active-site zinc of carbonic anhydrase II.

Cathepsin K is the case that teaches most, and it repeats last week’s lesson in a second protein. Cathepsin K has published allosteric sites: an ectosteric site in the surface loop spanning residues 84–100 (Law et al. 2017) and a second allosteric sector around Arg198/Tyr169/Lys119-Lys122-Arg123 (Novinec et al. 2014). The detector predicted residues A16–A21 plus A185–A186. That is neither of them. In a protein with two documented allosteric sites, the detector pointed somewhere else and was scored a hit anyway, because what it touched was one residue of an active-site inhibitor’s contact shell.

The admission filter proposed last week would have rejected all five. The ratio criterion — at least 80% of the ligand’s contact residues beyond 6 Å of the annotated functional site — fails for every case: 0.46, 0.46, 0.75, 0.56, 0.50. The size criterion additionally excludes F3S (7 heavy atoms) and VO4 (5 atoms, 115 Da). This is the second week in a row in which the week contributes zero cases under the proposed filter.

The root cause, with numbers

Why does a pocket sitting on catalytic machinery get labelled distal? Because the anchor is built from UniProt’s functional annotation, and for these proteins that annotation is a handful of residues, not a site. For MYH11 the entire functional annotation in UniProt is a single ATP binding site, residues 178–185. For cathepsin K and for fabB it is three active-site residues each — the bare catalytic triad. Everything else in the catalytic apparatus falls more than 6 Å from those residues and is therefore labelled DISTAL, which is to say eligible as an allosteric pocket.

protein	everything UniProt annotates as functional	anchor size	what is left labelled DISTAL	source
MYH11 (P35749)	one ATP Binding site, residues 178-185. Nothing else.	8 residues	switch I (A236-A241) and switch II (A459-A462) — the machinery that coordinates the γ-phosphate and the Mg	UniProt P35749
CTSK (P43235)	three Active site residues: 139, 276, 296 (Cys25/His162/Asn182 in the mature chain)	3 residues	the whole substrate cleft of the protease, including residue A19 that nirmatrelvir touches	UniProt P43235
fabB (P0A953)	three Active site residues: 163, 298, 333	3 residues	24 of the 32 contacts of the crosslinked substrate arm	UniProt P0A953

Three of this week’s five cases had anchors of 3 to 8 residues built this way, and two of them were flagged weak (fewer than 5 residues) by the routine’s own W38 rule. The fix follows directly and is conservative by construction: anchor on the dilated catalytic site — every residue within 6 to 8 Å of any annotated catalytic residue — rather than on the raw annotated residues. Growing the anchor can only subtract candidate territory; it can never manufacture a hit. It was not applied this week and nothing is reclassified retroactively. The record is append-only and the decision is Nicholas’s.

Where the scoreboard stands, under five readings

reading	what it changes	cases	hits	rate	95% Wilson	contains 32.6%?
A — rule in force	what the code does today: n_distal ≥ 5 and nothing else	15	8	53.3%	30.1–75.2%	yes
B — W40 heme as cofactor	W40 heme A read as a catalytic prosthetic group, so 36CV becomes a miss	15	7	46.7%	24.8–69.9%	yes
C — extended cofactor list	COFACTORES extended with Fe-S clusters and phosphate mimics, so 28FU and 29ZZ do not enter	12	6	50.0%	25.4–74.6%	yes
D — proposed admission filter	A plus the proposed admission filter (distality ratio ≥ 0.80 and ligand ≥ 12 heavy atoms and ≥ 150 Da)	6	3	50.0%	18.8–81.2%	yes
D+B — both filters	both filters at once	6	2	33.3%	9.7–70.0%	yes

All five readings still contain the 32.6% retrospective leave-one-protein-out rate measured on the 86 training proteins. After seven weeks of prospective measurement the scoreboard still cannot distinguish the detector from its own retrospective performance, and the move from 60.0% to 53.3% sits well inside the noise of its own interval.

entries	reason for rejection
64	no small-molecule ligand
31	non-model organism or no UniProt mapping
2	catalytic cofactor (ADP in DHX9, ANP in MCM2)
3	not distal (PKLR/oxalate 1/10, SF4 3/11, BEF 4/10)
19	duplicate of the same UniProt accession

The challenger

Challenger v0 — the frozen examinee plus a residualisation of the score by distance decile, which corrects the measured periphery bias without touching the frozen weights — missed all five. Paired totals: incumbent 8/16, challenger 2/16. McNemar on the paired cases gives 8 discordant pairs, 7-1 in the incumbent’s favour, p = 0.070: no power, since fewer than 10 discordant pairs cannot reach p < 0.05 even in the best case. No promotion is proposed. The trained challenger was not run: with 16 paired cases, rebuilding the 86 training proteins buys no power. The quantum arm did not run — it only runs during a promotion. Zero QPU consumed.

What we know / what we do not know

What we know. The frozen model is byte-identical to the one frozen on 2026-09-07: all seven pinned code files reproduce their sha256 and the weights block validates by two independent paths (canonical hash 08ad337d, declared block sha256 b0bf214f). Five structures published after the freeze were scored blind, with prediction timestamps preceding the reveal. Two were hits. Both hits lose rank 1 under an 8 Å dilated anchor. None of the five answer ligands is an allosteric modulator. The cumulative rate is 8/15 under the rule in force and the 95% Wilson interval is 30.1–75.2%.

What we do not know. Whether the detector has any real skill at finding allosteric sites: with n = 15 and every reading containing our retrospective rate, the prospective scoreboard has not yet separated skill from chance. Whether the dilated anchor would change the historical cases — we did not recompute them, and we will not reclassify retroactively. Whether the two pockets the detector predicted in myosin and cathepsin K are allosteric sites at all: they were scored against ligands that are not allosteric, so this week’s data cannot say. Whether the three misses would have been hits against a real modulator in the same proteins. We also do not know whether the ligands rejected this week by the cofactor rule would have been hits, because we did not score them.

Limit statement. Our negatives do not authorise anyone to rule out a site. This week our positives do not authorise anyone to assert one either: both hits are against ligands that are not allosteric modulators, and neither survives anchor dilation. Myosin and cathepsin K both have published allosteric sites and the detector predicted none of them. Nothing here is a finding; every predicted pocket is a dated hypothesis. This run publishes zero forward hypotheses: the only druggable pocket far from its answer was 9MLX/fabB (A172–A179, 15.5 Å away), and its anchor is the bare three-residue catalytic triad, so by this week’s own lesson its distal label is not trustworthy.

Reproducibility

Examinee hash 08ad337db711d7e77a4d66a6208aa88956fb7f71507683f2db47b61649ef0dd0, frozen 2026-09-07, 86 training proteins, 10 features, no annotations. Primary metric is de novo without annotations; the matched-by-distance null is primary and the contiguous null is reported beside it, never instead of it. Cavity screening used an explicit geometric surrogate, not fpocket, which still segfaults in qhull in this sandbox — labelled as a surrogate wherever it appears. Run sealed as RQ-PROSP-2026-W43, content hash sha256:1e7f642fbc75af5aff56d250c3c1a54e6dbb09d4a36e526eedc0d6d80aae4813.