What does a frozen allosteric detector score on structures released after it was frozen? 3 of 6

Week 41: one eligible structure, zero hits. The accumulated prospective scorecard of Rosetta Quantum's hash-frozen allosteric detector stands at 3 of 6 (50.0%, 95% Wilson interval 18.8-81.2%) under the eligibility rule currently in force, or 2 of 6 (33.3%, 9.7-70.0%) if last week's heme A is read as a catalytic cofactor; both intervals still contain the retrospective 32.6% the detector was built on, so neither reading distinguishes anything yet. The single case was 30KD, human SIRT2 with a myristoylated-lysine peptide, released 2026-09-30 and 1.45 A: the detector predicted pocket A99-A103/A106-A107/A134 blind at 12:49:23Z, the real myristoyl channel sits 3.8 A away, overlap zero. The miss is clean and the pocket also fails the distance-matched null (0.0805) and the cavity check. Valid as of October 5, 2026. This run could not be written to the public ledger database: the Cloudflare connector has been disconnected since October 1 and the record is pending insertion.

Status as of: October 5, 2026. Frozen detector: hash 08ad337d, frozen 2026-09-07. Run record: RQ-PROSP-2026-W41, content hash 0ba62d41.

What was measured, and on what?

The question this run answers every Monday is narrow on purpose: how often does a detector that has been frozen by hash, and cannot be retrained or adjusted, point at the allosteric site of a protein structure that did not exist when it was frozen. The detector was frozen on 2026-09-07; everything released after that date is, by construction, something it has never seen. This week the search window ran 2026-09-28 to 2026-10-05 and returned 192 released entries, all inspected. Thirty-nine passed the organism, method, resolution and small-molecule filters; deduplicated by UniProt accession they were 16; one survived the geometric distality check. The other 15 are recorded with their reason, because the rejections are data too.

The one case, number by number

Entry 30KD is the crystal structure of human SIRT2 in complex with a KMyrMe peptide, 1.45 A by X-ray diffraction, released 2026-09-30. SIRT2 maps to UniProt Q8IXJ6; the structure gives 290 modelled residues on chain A, of which 223 are distal to the 23 residues UniProt annotates as functional. The frozen weights scored the apo network alone and the prediction was written at 2026-10-05T12:49:23Z, before the holo was opened. The predicted pocket was A103 (top), A102, A101, A99, A106, A100, A107 and A134. The real answer — the residues in contact with the myristoyl group — is 14 residues, 7 of them distal. Overlap: zero. Closest approach: 3.8 A. The detector did not land next to the site; it landed one turn away and missed.

Two of those rows matter more than the miss itself. The distance-matched null is the primary one and it reads 0.0805, above 0.05: the predicted pocket is not distinguishable from a pocket picked at random at comparable distance from the active site. The contiguous null, reported alongside as the house rule requires, reads 0.0155 and would have said the opposite — the second week running where the cheaper null flatters the detector. And the identity figure puts this case in the harder-to-interpret bucket: SIRT1 (Q96EB6) is in the training set, and SIRT2 is its paralogue at 40.9 % identity, which is 0.009 above our own blindness threshold.

Why 3 of 6 still does not distinguish anything

Because six is not a number you can draw a conclusion from, and the interval says so. The 95 % Wilson interval around 3 of 6 runs from 18.8 % to 81.2 %. The retrospective rate the detector was built on — 32.6 % leave-one-protein-out on 86 proteins — sits inside it, and so does pure chance for most plausible definitions of chance. Under the alternative reading of last week's ambiguous case the scorecard is 2 of 6, 33.3 %, interval 9.7 % to 70.0 %, which also contains 32.6 %. The honest summary is that after five weeks of measurement the prospective scorecard still cannot tell us whether the detector is better, worse or the same as its retrospective number. It is published weekly anyway, which is the point: the number that gets published is the one that came out.

Split by sequence identity against the training set, the six prospective cases are 1 of 2 below 40 % identity and 1 of 3 at or above it; one case (36CV, last week) has no identity figure because it was not computed that week, so the split covers five of six.

The eligibility rule is now the fourth-week problem

The answer ligand this week was MYR — the myristoyl group of the substrate peptide itself, sitting in SIRT2's catalytic acyl tunnel. That is the same failure mode as ADP in week 37, AF3 and TDK in week 39, and heme A in week 40: UniProt annotates 23 functional residues for SIRT2, the acyl tunnel extends past 6 A of that anchor, and the geometric filter therefore calls it distal. Fourth consecutive week. MYR is not on the cofactor exclusion list, so the case was rendered under the rules in force and is not reclassified retroactively; the register is append-only.

What is new is the direction of the bias. Last week the pending rule change would have lowered the published number; this week the same change would raise it, from 3 of 6 (50 %) to 3 of 5 (60 %), because this time the questionable case counted as a miss. A rule whose effect flips sign depending on the week is a rule that has to be decided before the cases arrive, not after. The proposal is already written and waiting on a human decision; this report does not adopt it.

The challenger

The challenger runs for the record, never as a published Rosetta figure. Challenger v0 — the frozen detector plus a residualisation of its score by distance decile, which exists to correct the measured periphery bias — also missed this week, predicting A58. Across the seven paired cases accumulated so far the incumbent is 3 of 7 and the challenger 0 of 7. McNemar on the paired cases gives 3 discordant pairs, all in the incumbent's favour, p = 0.25: with fewer than ten discordant pairs the test cannot reach p < 0.05 even in the best case, so it is recorded as having no power and no promotion is proposed. The trained challenger, which would need the full 86-protein rebuild, was not run for the same reason. The quantum arm did not run: it runs only on a promotion. QPU consumed: zero.

Step 0: the frozen detector, verified two ways without the snapshot

The detector verified intact: recomputed hash 08ad337d, all seven pinned code files byte-exact, code unchanged. This week it got there by a different road, and that is worth recording. The content-addressed snapshot transport was not available in this run, so the seven files were materialised another way and checked against the same pinned sha256 values: two came byte-exact from a clone of the public evidence repository; conservation.py was recovered by deleting two whitespace bytes at offsets 2914 and 2943, the defect found in week 40 and reproduced here; engine.py by deleting one byte at offset 20377, the defect found in week 39 and now reproduced for the third consecutive week; three arrived intact. The frozen weight block itself was validated twice over — its own hash recomputes 08ad337d, and the sha256 of its canonical serialisation recomputes b0bf214f, the value the snapshot declares for that block. The examinee is demonstrably the same object by two routes that do not depend on the snapshot. In every case the arbiter was the hash: nothing was repaired into agreement.

What we know and what we do not know

What we know. The frozen detector, applied without retraining, missed the one eligible structure released this week, and the miss is unambiguous: zero residue overlap, 3.8 A from the real site. The predicted pocket does not pass the distance-matched null and does not pass the cavity check. The accumulated prospective scorecard is 3 of 6 under the rule in force and 2 of 6 under the alternative reading, and both intervals contain the retrospective rate. The detector's measured periphery bias reappeared: it fired at the 14.8 % distance percentile while the real site sits at 20.6 %.

What we do not know. We do not know whether the prospective rate differs from the retrospective 32.6 %; six cases cannot answer that, and we will not pretend otherwise by re-running. We do not know how much of this week's miss is the detector and how much is the eligibility rule, because the answer ligand is a substrate acyl chain in a catalytic tunnel — the fourth week of that confound. We do not know whether the 40.9 % identity to a training protein helped or hurt, and with one case per bucket we cannot find out. We do not know what fpocket would say about this cavity: fpocket still segfaults in its qhull call, so the cavity check is a declared geometric substitute, not fpocket. And we do not know whether this week's pocket would have survived a different anchor: it is invariant at 6 A and moves at 8 A, so the verdict depends on where the boundary is drawn.

Limit declaration

This detector's negatives do not authorise anyone to rule out a site. A miss here means our frozen model, with ten de novo features and no annotations, did not rank that pocket first — nothing more. It is not evidence that the pocket is undruggable, not evidence that no allosteric site exists there, and not a reason for any laboratory to stop looking. The predicted pocket of a missed case is likewise not a hypothesis: this week's pocket failed both the matched null and the cavity check and is published as a dated prediction that was wrong, labelled FORWARD and nothing else. We publish hypotheses with dates, never findings.

Sources

• RCSB PDB — entry 30KD, human SIRT2 in complex with KMyrMe peptide, 1.45 A, released 2026-09-30
• RCSB PDB search API — entries released 2026-09-28 to 2026-10-05, X-ray or cryo-EM, resolution ≤ 3.0 A (192 entries, all inspected)
• UniProt Q8IXJ6 — human SIRT2 functional annotations (23 residues, the anchor for this case)
• UniProt Q96EB6 — human SIRT1, the training-set protein the case is 40.9 % identical to
• Teng et al. — Efficient demyristoylase activity of SIRT2 revealed by kinetic and structural studies (Sci Rep 5:8529, 2015). Retrieved via PubMed.
• Wang et al. — Deacylation mechanism by SIRT2 revealed in the 1'-SH-2'-O-myristoyl intermediate structure (Cell Chem Biol 24:339-345, 2017). Retrieved via PubMed.
• Wilson — Probable inference, the law of succession, and statistical inference (J. Am. Stat. Assoc. 22:209, 1927)
• Run record RQ-PROSP-2026-W41 — this run, content hash sha256:0ba62d41370f4dc5, rosettaq-archive/v2, verify() = True
• Run record RQ-PROSP-2026-W40 — previous run, content hash sha256:0e36c8a1, the ambiguous heme A case that reading B concerns
• Frozen detector RQ-PROSP-2026-W37 — hash 08ad337d, frozen 2026-09-07, 86 training proteins, 10 de novo features, no annotations
• Retrospective reference — 32.6 % leave-one-protein-out on N = 86, de novo without annotations; reported alongside, never pooled with the prospective scorecard

Rosetta Quantum publishes verdicts with raw, reproducible data. This is a prospective run-report of our own frozen detector, published whether it hits or misses. Educational content, not a product claim.