Now

Amit asked me to stop “holding” and keep finding directions, borrowing from his two other sessions (virtual cell, virtual embryo). Their notes contributed a matched − shuffled control, a hop-test benchmark habit (we already had one), and a warning that label-free model selection mostly fails. Two corrections to my own assumptions: (1) “test labels are released in active learning” was only in my brief; the Synapse-confirmed spec says just 100 validation submissions (credits reset) + 2 test — and each validation submission returns a deterministic hit count, i.e. an oracle for real PGK2 binders; (2) the rules require a DEL-ML step (our fused buy satisfies it), and the test score includes full-library PR-/ROC-AUC. Results so far: the shuffled-reference control shows reference-similarity’s wins are target-specific (other-target references sit at or below chance), not generic vendor style; excluding DEL compounds outside the catalog’s property envelope starves the data (the DEL is much larger than the catalog: MW 536–598 vs 335) and pre-registered class-motif annotations did not meet the win criterion; our buys do not prefer big molecules (Spearman with MW −0.12; picks sit at the catalog median). The important one: a cross-paralog benchmark (BET→BRD4, p38→MAPK14, DNMT3B→DNMT3A) shows paralog references carry real but small signal — at PGK1’s identity band roughly 15% of native lift — so the 0.95 PGK1 anchor weight was too high (now scaled to 0.25) and true PGK2-native references are worth far more. Re-testing the fusion with paralog-quality references, equal-weight only wins when the paralog is strong; a supervised-heavy list is staged as a hedge. 🚩 The conflict: the analogs favor supervised-heavy, but PGK2’s own July history was 0/8 for DEL-primary supervised models vs 1 hit for the reference prior (tiny counts, but it’s PGK2-specific). Only the validation oracle can arbitrate on PGK2 itself — so the oracle protocol (simulated offline, ready for Amit’s token) now does double duty: find native binders and compare channels. Both candidate lists stay staged; neither is declared primary. Update (later the same day): the conflict is stronger than first stated — our faithful reproduction of the organizer’s LightGBM/ECFP4 baseline also scored 0 on PGK2 validation (the reported “~5 hits” was not reproduced). The shuffled control and a 14-fingerprint sweep confirm reference-similarity is target-specific and that FCFP6 stays. But a new same-target, independent-source test shows that most of the within-target reference win is same-series similarity (≈86% on p38, ≈100% on sEH/BRD4), and PGK2’s references sit at the floor of the chemical-overlap scale (median max Tanimoto to the catalog 0.20; only 172 of 184,632 compounds reach 0.40). The honest walk-back: the pivot’s claim that “real known binders transfer” holds only where references overlap chemically with the candidates — which PGK2’s PGK1-derived references do not. What remains: two weak channels, the validation oracle as the arbiter, and val-derived binders being scaffold-disjoint singletons (so they are worth only the weaker independent-source rung). Update (30 Sep, later): two more checks closed. Pocket retrieval (DrugCLIP) does not clear the bar (worse than chance on an unfamiliar pocket) and transcriptomic overlap is 0.7%, so both routes are closed. A size-confound audit of our own referee found that the decoys stop at ~34 heavy atoms, so on BELKA and KinDEL molecule size alone scores ROC 0.97–0.998. Re-scoring with size matched leaves every headline intact: reference-similarity stays target-specific, the paralog cliff stays, and the supervised channel stays above chance. What shrank: the BELKA top-50 ceiling was a size artifact, and weighting the supervised channel double gains only about +0.02 ROC. A second confound (DEL-vs-vendor chemistry style) is not removed by size matching. Nothing in the staged plan changes; the validation oracle remains the arbiter. Update (Oct 3–4): the organizers released the Blind Test labels (39 true PGK2 hits among 6,580 selected molecules). Scoring our channels against them: none of our three staged top-50 lists contains a hit; the supervised channel is above chance on PGK2 (median top 8.7% of the catalog), the paralog-reference channel is weaker (top 19%). Adding the released labels to the DEL-trained model, with each hit’s whole series held out, moves the median hit from the top 8.7% to the top 0.8% (13% of hits in the top 0.1%, 54% in the top 1%); a fake-hit control shows nothing, a stricter series split keeps most of it (top 2.3%), and the weighting, native-similarity fusion and richer features change little. It helps isolated hits most. But on analog benchmarks the same recipe gives only modest lift (median-level, top 1% mostly flat), even at PGK2’s negative ratio and with team-like selection, so the PGK2 gain is unconfirmed beyond PGK2 itself. Expected top-50 yield for new series is still under about one hit. Per Amit’s rule, no test submission yet; the direct test (validation credits, not test slots) awaits his decision. Update (Oct 5–6): new data types and design, all negative so far. A ChEMBL neighbor-knowledge model (sequence, pathway, family, structure facets, with unrelated-protein and shuffled controls) did not generalize across MAPK14, BRD4, sEH and HSA, and PGK2 has almost no neighbor ligand data (PGK1: 97 rows). A PGK2 interactome from expression, imaging, phenotype and evolutionary layers lists 884 candidates (66 measured, 790 inferred; best: GAPDHS, LDHC, PGAM2, ENO4, TPI1). Interface modeling (Boltz-2 on Explorer) found no confident PGK2–partner interface: partner confidence sits inside the random-control range, even for the known PGK1 paralog. De novo design with ShEPhERD and DiffSBDD against the 3-PG, nucleotide, terazosin and one fpocket site produced novel molecules far from catalog chemistry; similarity retrieval of the 39 released hits and DEL-model scoring of the designs are both at random level. Only the nucleotide site is druggable; no credible cryptic pocket. The labels-trained supervised model remains the strongest PGK2 evidence (median hit at the top ~1% across fold assignments). Test slots still on hold.

Amit reopened autonomous work with three ideas (pull more DEL for larger finetunes; cell-state-matched med-chem refs; analogous-target validation sets → optimize → hop). Key correction to an earlier over-generalization: the campaign retired DEL-representation methods after they failed the hardened referee — but only on SPARSE-signal targets. Rich DEL-ML with real binding labels is a different regime, and it reopens the verdict decisively. Landed KinDEL (kinase/MAPK14 ATP pocket + real off-DNA K_d — the closest analog to PGK2's ATP site) and BELKA (BRD4/HSA/sEH) to Explorer. Bake-off: a supervised DEL model beats chance widely (KinDEL ROC 0.71, BELKA OOD 0.94) where the sparse suite had the whole family at chance. The load-bearing finding: which method wins depends on REFERENCE DENSITY, not target class — few real binders → reference-similarity wins (KinDEL 41 vs 12); many → supervised DEL wins even OOD (BELKA 40 vs 28). Hop to PGK2: PGK2 is few-reference, so reference-similarity stays PRIMARY (validates the current tool) — but PGK2's OWN DEL is rich (7.7M rows), so a supervised-DEL model on it is an ORTHOGONAL second channel (Spearman ≈0.17 — the partner earlier RRF fusion lacked). The fusion VALIDATED on the hardened referee in PGK2's exact few-reference regime: RRF beats the best single channel at n=3/5 refs (BELKA +3.4, KinDEL +2.2) whenever the supervised-DEL channel is above chance; it collapses only on sparse EAS-MS DELs (expected). PGK2 passed the GO-gate (within-DEL AUC 0.995) and its two channels are genuinely orthogonal (Spearman 0.062). Result: two evidence-backed buys are now staged for the active-learning phase — an EXPLOIT buy (reference-similarity, the validated primary) and an EXPLORE buy (RRF fusion, 30 new picks from PGK2's own rich DEL). Honest caveat: PGK2 supervised→drug-like transfer is unmeasured (inferred from the analogs), so reference-similarity stays primary. Separately, WS2 (cell-state-matched med-chem refs) closed as a clean negative: all 14 germ-cell/glycolytic candidates are off-fold and the pocket-gate correctly rejected them (same-pathway biology doesn't transfer without the same fold). WS4 (larger models / finetunes) also settled: a pretrained ChemBERTa-77M finetune on Explorer GPU does NOT transfer better than the simple LightGBM (KinDEL ROC 0.842 vs 0.892; BELKA 0.781 vs 0.945, decisively worse on the trustworthy OOD test) — so LightGBM/ECFP4 stays the supervised channel and a bigger model doesn't help. DIR-017 complete: the eventual PGK2 buy is stronger (orthogonal fusion channel) and the method choices are now empirically settled. DIR-018 (optimize-then-hop across 6 DEL benchmarks): the method GENERALIZES (a tuned config beats default on 5/6 held-out benchmarks with small gaps) — but the winning config is a SAFETY LEVER that only helps when the supervised channel is weak/below-chance; on the strong-supervised analogs (sEH, KinDEL) that match PGK2's regime it gives ~0 (±1 pick), so for PGK2 we keep equal-weight fusion defaults and adopt just the one zero-risk refinement (reference fingerprint ECFP4→FCFP6). Net: tuning to the benchmark average would overfit to the weak-supervised majority (the opposite of PGK2), and the honest read is that PGK2's own strong supervised channel is already near-optimal. DIR-019 tested two more ideas — both honest referee no-wins that sharpen the picture: (a) synthetic relabeling with molecular knowledge (QED/PAINS/Brenk) LOSES to the noisy DEL count label (KinDEL top-50 18.7→11.0) — the count itself IS the binding signal, and generic drug-likeness is orthogonal to it; pathway-molecules-as-labels collapse the channel (perm p=1.0), confirming same-pathway biology doesn't transfer. (b) A real PGK-family evolutionary/coevolution prior (596-seq MSA) doesn't lift top-50, but delivers a genuine selectivity design principle: the PGK1↔PGK2 divergence residues are the least-conserved pocket positions (so paralog references can't capture selectivity), and the one credible selectivity handle is Tyr242 at the hinge edge (not the plastic Ala255 tail) — a hypothesis whose clean test is a Boltz co-fold of a Tyr242 reacher. DIR-020 (co-folding pre-approved by Amit) closed both remaining structure angles: (a) docking as auxiliary FEATURES (not a standalone ranker) still doesn't lift the referee — as an RRF channel, a downweightable model feature, or an auxiliary target, all ≤ 0; and (b) that Boltz co-fold got run — 18/18 complexes on Explorer GPU, PGK2(Y242) vs PGK1(F242) — and refuted the Tyr242 selectivity hypothesis: ligands reach the 242 position but none H-bonds the Tyr hydroxyl and show no PGK2-favoring energy delta, so reaching ≠ a Tyr-specific contact. Honest null, no buy tie-break — and a clean example of co-fold doing what it's good for (killing a 2D hypothesis with 3D evidence). Net across DIR-017→020: reference-similarity (FCFP6) + orthogonal supervised-DEL fusion has now survived a long gauntlet (DEL-representation, DANN, docking ×3, FM finetune, config-tuning, knowledge-relabeling, evolutionary prior) — all referee no-wins against it — which is itself strong evidence it's the right method.

Everything lines up. The campaign converged (5 method families on the honest referee → only known-binder-REFERENCE similarity wins; our refs are 84% PGK1-paralog, 0 PGK2-native). The active-learning phase opens ~today (Sep 16–Oct 15): test labels released, credits reset, 2 more test submissions. Released labels = confirmed PGK2 binders = the first PGK2-NATIVE references the whole campaign concluded we need — and our winning method is already built to consume them. What I need from Amit (time-sensitive): (1) confirm test-label release + provide the Synapse token (pasted live, never persisted, rotated after); (2) then it's deterministic — rerank_from_labels.py is built + validated: it folds released binders in as top-weight PGK2-native refs, re-ranks the 184,632 catalog by known-binder similarity, applies the quinazoline/leakage/selectivity gates + Murcko-diverse top-50, and I validate the re-rank on the hardened referee before spending a scarce credit. Plan for the 2 subs — both buys are now concretely staged (NOT submitted): (1) EXPLOIT = reference-similarity buy (the validated primary); (2) EXPLORE = the RRF-fused buy (reference-similarity + PGK2's own rich supervised-DEL channel, orthogonal at Spearman 0.062; 30 of 50 picks are new-from-fusion). When your real PGK2-native labels arrive, both regenerate automatically (reference channel via --new-refs; supervised-DEL channel already trained). Parallel unlock: the SGC/CoDA 5 PGK2 probes (real selective binders) via the SGC/UNC line.

The hardened referee has now adjudicated FIVE method families on the identical property-matched-decoy task, and the verdict is consistent and decision-grade. Bake-off (DNMT3A, top-50 / ROC-AUC): 2D-ECFP→real-binder refs 22 / 0.886 · 3D-shape→refs 9 / 0.733 · receptor docking (smina) 2 / 0.524 · DANN 1 / 0.322 · DEL-similarity 1 / 0.317. Only known-binder-REFERENCE similarity beats chance and buys top-50; every DEL-representation method AND generic receptor docking sit at chance. Docking (Slurm job 10353643, DNMT3A 6BRR cofactor pocket) is PARKED — caveat owned that it's a weak docking case (cofactor not inhibitor pocket; no referee target has a clean enzyme active-site like PGK2's ATP pocket), so PGK2-ATP-pocket docking is retained only as an auxiliary consensus axis. 🚩 The convergent conclusion: combined with the reference audit (our 159 refs are 84% PGK1-paralog, 0 PGK2-native), the entire campaign now points to ONE lever — real PGK2-native, PGK2-over-PGK1-selective references (the SGC/CoDA 5 probes). This is the convergent verdict of the referee across 5 method families, not a single-method hunch.

The make-or-break structural-transfer test on the hardened referee (the same one that put every DEL method at chance). A 3D-shape/pharmacophore scorer (USRCAT) built on real known-binder references lifts the substantive-3 aggregate from ROC-AUC 0.275 → 0.737, buys top-50 > 0 on all three targets (DEL/DANN buy ≈ 0), and even lifts the grade-A WDR91 anchor (0.597 vs 0.414) where DANN did nothing. The decisive reading: it isn’t 3D-vs-2D — a 2D-to-the-same-references control also wins, so what transfers is having REAL known binders as references; DEL-enriched matter simply doesn’t resemble drug-like binders. (Honest caveat: on the enzyme targets the lift is partly DUD-E analog-bias — the clean, truly-independent signal is the WDR91 arm, where a distinctly-3D component is real but modest.) This cohered the entire campaign: our blind-test method — the DEL-independent canonical-binder prior (ROC 0.649) — WAS a known-binder-similarity method; the test confirms that’s the right family and explains the 0-hits: PGK2 is reference-STARVED (its only real references are ~one PGK1 chemotype). 🚩 The pivot, committed: the #1 bottleneck is now REAL, DIVERSE PGK2 binder references — marshal PGK1/parasite/literature binders, chase the SGC/CoDA 5 PGK2 probes (the highest-leverage data unlock), and add the receptor-pocket (docking) signal on Explorer where ligand refs are sparse. Next: receptor docking on Explorer.

The most important result of the campaign, and the reason the suite exists. After hardening the referee with realistic property-matched decoys (50/active, drawn from the real 184k-Enamine catalog, Tc<0.35 to actives — so top-50 = buying 50 of ~3000, the true task shape), the honest verdict landed: DEL→drug-like transfer is at CHANCE. The DEL-similarity baseline drops below chance (substantive-3 aggregate ROC-AUC 0.275; WDR91 0.61→0.41, DNMT3A 0.75→0.31, PLCG1 0.29→0.10) with top-50 ≈ 0 everywhere — the earlier “some transfer” was an artifact of easy wrong-target decoys. DANN (its large-decoy PARK precondition now met) reaches only chance, buys 0 top-50, and doesn’t help on WDR91 (the grade-A anchor) → KILL confirmed. Three DEL-representation methods have now failed the honest referee (similarity, Snorkel, DANN). Crucially the hardened suite reproduces our blind-test 0-hit reality — so this is the first time we can PROVE the 0-hits are a representation-content problem (DEL enrichment doesn’t encode which drug-like compound binds), not a tuning problem. 🚩 Strategic redirect (flagged for Amit): stop polishing DEL-enrichment representations; pivot the effort to ORTHOGONAL / STRUCTURAL positive signal (real binder structures, co-folding / pose scoring, pocket-similar-protein binders), and the Dir-015 foundation model must be structure/binding-aware, not DEL-count-only, or it inherits the same ceiling.

The keystone Amit asked for: “construct ≥10 validation sets for different targets so we can be confident in our methods.” Until now WDR91 was our ONLY ground-truth yardstick (n=1), so every “the method transfers across the DEL→drug-like gap” claim was an anecdote. Built results/method_benchmark/<T>/ for 10 AIRCHECK targets (WDR91, WDR12, DNMT3A, PIAS1, PLCG1, ZER1, PLCZ1, DCAF7, NSP, PGK2) — each with normalized DEL counts, a synthon/BB-disjoint split, a truth set (external drug-like actives where they exist + a uniform DEL-internal holdout), and a graded card. One harness (run_method_suite.py) scores any method by top-50 recovery / EF@1% / BEDROC, self-testing WDR91 ≈ 9/19 (the ~11/19 ballpark). It immediately did its job: external drug-like transfer = mean ROC-AUC ≈0.51 (chance) for the DEL baselines (DNMT3A 0.75, WDR91 0.60, PLCG1 0.18) — the wall, quantified across targets. Honest verdict: a strong directional diagnostic + hard-negative stress test now (4 A/B-grade + a uniform n=10 internal panel), not yet a ≥6-target adjudicator; the fix is wiring per-target ECFP4 refs to grow the external panel 3→6.

A 12-idea cross-domain fleet (DANN, PU, optimal-transport, Snorkel, multi-task DEL-background-subtraction, conformal, bandits, generative bridge, hyperbolic, synthon-split, DCA prior, LLM-jury+TDA) was pressure-tested and converged into one flagship stack: DANN fixes the representation, weak-supervision fixes labels, multi-task grounds negatives, the synthon split is the common referee. It killed PU-via-competition-channel (PGK2-specific → unscorable on the suite; also confirmed by the schema: AIRCHECK validation files have that channel all-null). Snorkel PARKED honestly — on WDR91 it matched but never beat raw enrichment on top-50 (single-channel DEL → all labeling functions are enrichment-family). Now prototyping DANN (gradient-reversal) — the one idea attacking the DEL-vs-drug-like AUC≈1.0 separability head-on — measured on the suite external panel with kill-criteria (drop it if collapsing domain-AUC costs top-50: the over-regularization trap). Ranked list: crazy_ideas_ranked.md.

Amit’s steer: use Explorer for big data + train/finetune a DEL-ML foundation model. Plan (BRIEF_015): pretrain a molecular encoder across ALL AIRCHECK DEL targets with a DANN domain-adversarial head so it factors out library identity (crazy-ideas #5+#1 at the representation level), then finetune PGK2 with clean positives, scored on the 10-target suite. Explorer reconned and ready: A100 ×3–4 + H200 ×8, 1.5 PB scratch, /scratch/shenoy.am/pgk2_fm/ created. Gated on: (a) the suite as referee (done), (b) a torch env build on Explorer (existing venvs lack torch). Science-flagged: open-chem-FM init (MolFormer/ChemBERTa) + DANN vs from-scratch.

The current thrust (Amit's rational-design directive): our wall is top-50 CONVERSION, which is a clean-positives problem — DEL enrichment means “stuck to something,” not “binds PGK2.” A fleet is now: (A) denoising the DEL PGK2-specifically via the competition channel count_PGK2_with_inhibitor — a true site-binder drops under inhibitor competition, a matrix/tag/promiscuous false-positive doesn't (the cleverest FP filter, never made a core denoiser before); (B) augmenting positives from generative de-novo models (as a positive-chemistry source), co-fold predictions, and pocket-similar-protein binders — pocket-gated; (C) richer chemical representation + retrain + re-curate the test buy; (D) cross-domain positive-finding. Every lever validated on the WDR91 top-50 yardstick, not AUC. Running indefinitely; science decisions flagged for review.

A deep-research fleet (King → Queens → workers) on two method ideas from collaborator Sigge Hermann: (1) a physics-based “Drug Hunter” de novo designer for PGK2, and (2) “a BLAST-like algorithm for SMDs” (SMD ambiguous — Small-Molecule-Drugs vs Steered-MD; the fleet covers both and flags it). The task is landscape assessment, not implementation: map the SOTA of physics-based de novo design (Pocket2Mol, TargetDiff, DiffSBDD, FEP+, RGA, commercial) and of fast molecular/MD “BLAST-analog” search (LINGO/SMILES-k-mer, fingerprint-LSH, shape/USR, infiniSee/SmallWorld, dynamic-pharmacophore search), then assess where each idea could genuinely differentiate vs re-badge — and how they'd attack our residual top-50-conversion wall. Output: a positioning memo + questions for Sigge. Standing by for his algorithm details to sharpen it. Full memo: deepresearch/POSITIONING_MEMO.md.

Hopped to five new primitives (surface-similarity, ligand-surface reverse probe, fragment corpus, functional-genomics/KG, cryptic pocket). Honest asymmetric yield (the run was degraded by a model-availability wobble — 17/32 workers errored). No fold-independent surface cleared a clean STRUCTURAL gate as new drug-like chemistry, so no new anchor CLASS — but real leads surfaced: the ligand-surface reverse probe imported 12 novel catalog-buyable ATP-competitive kinase scaffolds (CDK/MAPK/CLK/Src-sourced, new at sim ≤0.33, one docking check from anchors); the fragment mine found 2 new on-target PGK co-crystals (P. vivax OTQ, PGK1 pyrazolopyrimidine CRT0063465/5NP8); and hidden-biology leads (TXNRD1, carbamate kinase from two independent routes). Next: structurally gate the 12 scaffolds by docking; pull CRT0063465 SAR. Don't re-buy on weight-bumps — wait for confirmed new chemistry.

G1 confirmed the hit lives in the 38 dual-method-consensus compounds. Rather than spend 6 credits identifying which one — that just chases val chemistry, which is scaffold-disjoint from test and would overfit — we exploit it as method validation. The 38's shared signature is a MECHANISM, not a scaffold: near-unanimously ATP-hinge-anchored (0.974 vs 0.05 in random val, ~19×) across 38 DISTINCT scaffolds. Running the same method variants on the 184K test set reproduces that mechanism in a 41-compound consensus (test hinge 0.83; 0 of 41 in the DEL top-500 — a total re-rank), staged with the test buy HELD. Forward step the analysis identified: the two variants are ~64% collinear, so a truly ORTHOGONAL structure/docking prior is what makes consensus a strong confidence signal — exactly what the leads→anchors docking arm can become.

The method's culmination, and it holds. Ground-truth check: an independent structure prior (smina) — with no knowledge of the similarity method — ranks the confirmed-enriched 38 above decoys at AUC 0.58, carried by hinge geometry (bidentate rate 0.58 vs 0.32; hinge-distance AUC 0.64, p=0.013) not affinity. So two independent methods re-derive the same ATP-hinge mechanism the confirmed val binder sits in. And they're genuinely orthogonal — structure×similarity Spearman 0.18 (vs the v3/v5 variants' ~64% collinearity) — so their consensus is a real confidence signal. Staged an orthogonal-consensus diverse-50 (90% new vs similarity-only; a 37-compound require-both high-confidence core) as the strongest, most method-validated TEST-buy candidate yet. Honest: the structure axis is weak (AUC 0.58) and docks into a mouse-PGK2 ortholog — value is as a consensus axis, not an oracle. Test buy HELD pending Amit's call.

Chased the field-jump leads: 5 became real pocket-gated anchors — CRT0063465 (the only new PGK-orthosteric chemotype with a real co-crystal, 5NP8, but weak at Kd ≈ 24 µM → low weight ~0.40) + 4 of the 12 kinase scaffolds (CLK1 diaminopyridoquinazolines etc.) that reinforce the existing hinge class. But the bigger win is the tool: the smina docking gate is validated (ATP −9.65 kcal/mol, bidentate hinge; decoys rejected) — the orthogonal structure prior the campaign needed. Honest caveat: rigid docking into a mouse-PGK2 ortholog (no human structure exists), so it's an orthogonal axis for consensus, not a standalone oracle. Also fixed a provenance error (5NP8 was mislabeled NG52; it's CRT0063465).

The selectivity-aware buy (val 9774699) scored N_hits=1, N_clusters=1 — same as v3, so the re-ranking didn't increase the count (though it improved every honest offline metric: PGK1-sep 0.861, selectivity index 1.86). The valuable takeaway isn't the count — it's that v5 and v3's p-values are byte-identical, which revealed the group-testing lever above. v3 and v5 share ~38/50 picks, so the hit is most likely the same single binder; group testing would pin exactly which compound.

Reasoned about PGK2 by ANALOGY (its identity as a testis-specific germ-cell isozyme / male-contraceptive target, its reaction mechanism, analogous pathways, isozyme-selectivity design) — a Queen per analogy. Verdict: even the richest analogy yields ZERO new pocket-justified anchors — every germ-cell/metabolic isozyme (LDHC, GAPDHS, PKM2, HK…) is a different fold from PGK's phosphotransferase cleft, so the pocket-gate correctly rejected them all (no FBP1-style mirage). The data ceiling holds across analogy as well as similarity. The real prize: a crystal-verified PGK2-over-PGK1 selectivity rule (exploit Tyr242, avoid Ala255; allosteric-site selectivity ruled out — PGK is monomeric) — now feeding the v5 buy above.

Built + dry-run-validated (Team_PGK2_val_v4_pgk1_margin.txt). Vs the v3 buy: PGK1-human anchors 38→44, T. brucei 7→2 (only the 2 independently-strong ones kept), unanchored 1→0; terazosin quartet retained; margin-weighting shipped (d 0.224→0.254). Jaccard 0.72 vs v3 (8 picks swapped); all gates clean (50 series, 0 PAINS/quinazoline/leakage). Honest: it's a re-weight of the same 159-anchor corpus, not new chemistry — it rationally raises the EXPECTED hit rate by concentrating the budget on the highest-confidence anchors, but at N=50 with a borderline single-hit baseline, 1→2+ is stochastic. Ready for Amit's token.

The pocket-gated v3 prior is built (159 anchors: PGK1-human + terazosin allosteric quartet + low-weight T. brucei ATP). The pocket-gate is validated: the off-fold-inclusion control collapses PGK1 recovery to AUC 0.465 — chance. A sharp catch: a naïve merge that keeps the good core beside off-fold noise inflates the aggregate to 0.887 (specificity rots under a rescued headline) — so discipline must be enforced at anchor-admission, not read off a post-hoc AUC. v3's own headline dips (0.92→0.81) but that's a diversity artifact of broadening, not lost signal. The deepest conclusion of the campaign: the ceiling is DATA, not method — the positive corpus is one dominant chemotype and no diverse pocket-validated PGK set exists anywhere. Further mining has diminishing returns.

Mined PGK2-related proteins three ways (deep-hop pathway/HPA/STRING, sequence, and Foldseek structural) and spawned a Queen per protein to scrape each exhaustively. Verdict, with pocket validation + permutation nulls: every off-fold / deep-hop candidate FAILS (HK1, FBP1, NAE1/UBA3, G6PD, GCKR) — FBP1's 2D-anchor z=16 collapses to pocket-weight 0.07, the false-transfer trap made concrete. Exactly one new pocket-validated protein — Schistosoma PGK (all 3 sub-sites, RMSD 0.23 Å) — but it has no binders of its own. Net: ~+96 real PGK1 anchors (a 1 nM quinoxaline series) + the terazosin allosteric quartet, NOT the ~3,000 a naive union would wrongly admit. The durable win is the pocket-gate discipline — now being applied in the v3 rebuild above.

The biology-first Meta-King completed clean (32/32 agents). It asked what PGK2 is evolved to bind and delivered a DEL-independent 4-component prior (ATP-anchor ⊕ known-binder similarity ⊕ proximal QSAR ⊕ PGK1 kNN) that scores both 429K splits to a 50/50-distinct-series, 0-PAINS top-50 — and proved that folding the DEL in as primary swaps 54-64% of picks and lowers mechanistic quality (the inversion, quantified). It found drug-like nanomolar PGK binders inside the Enamine catalog (CHR-6494, Z57346765), confirmed PGK1 is the only real homology proxy (29 SD above null; 8/8 anchor residues conserved), and imported SAIR + BindingDB + PDBbind + BioLiP + Papyrus (all offloaded to O:). Honest negative: the 2D anchor can't yet separate PGK-fold from generic ATP-kinase chemistry — needs a pose-based gate.

The IDG-DREAM relational (compound×target) pipeline now runs fully (OpenProtein/SW kernel ⊕ compound kernel ⊕ the validated CGKronRLS solver → a PGK2 ranking → a staged diverse-50, not submitted). But the honest test deflates its premise: on leave-PGK1-out (the one edge we can ground-truth), the target-similarity axis adds ≈nothing over compound-similarity alone — as built it's a compound-similarity model in a Kronecker coat. And OpenProtein's kernel lost to Smith-Waterman on compound ranking (AUC-PR 0.61 vs 0.05) — its panel-separation geometry win didn't translate. Not broken (sanity checks pass; terazosin ranks 87th pct), but its domain-gap value is unproven without a PGK-matched yardstick. Owned, logged.

OpenDDE ruled out (a Boltz-2 wrapper that disclaims affinity ranking). Physics MM-GBSA keeps hitting real runtime failures on Explorer: the OpenMM/CUDA env builds and smina docking gives sane poses (−7.7 to −7.9 kcal/mol), but the first validation silently failed on an AmberTools PATH bug (fixed) and the re-run now errors inside OpenMM (traceback being diagnosed). Still no valid AUC vs the 0.45 null — the physics arm is proving operationally stubborn, consistent with the honest read that pose fidelity, not the scoring form, is the bottleneck. Lower priority than the pocket-gated prior above.

The first autonomous 3-tier swarm picked 6 frontiers, each transferring a specific prior-challenge winner (board below); it ran ~1.5M tokens across 32 agents. Two produced real validated results — the IDG-DREAM relational kernel is built end-to-end, and the CACHE #3 novelty paradox is quantified on our own data. Synthesized manually after the Workflow's synthesis step kept dying on session limits.

Decomposes each compound into BRICS/synthon fragments and transfers per-fragment binding evidence (PGK2-DEL per-building-block enrichment ⊕ PGK1 actives, max-pooled). Recovers 3/3 hits·series guarded on WDR91 where whole-molecule similarity gets 0 — genuine fragment transfer, not leakage. PGK1 actives enrich exactly the ATP-site/hinge motifs; 20,070 of 40,000 sampled test compounds carry fragment evidence. Fold-back wired into the re-rank + pseudo-labels. P_frag.

Asks whether a compound's neighbourhood in ChEMBL / PubChem carries measured binding evidence. Mechanically validated: its on-target positive control recovers 49–50/50 same-class actives (AUC 0.958, guarded). Scores 0 on WDR91 only because WDR91 is off-target for an ATP-site prior. Pulled 3,053 on-target glycolysis/ATP-site ChEMBL actives for PGK2. P_struct.

Runs anchor-similarity target prediction over the test set. Genuinely orthogonal to the DEL model (corr −0.02) but thin with the anchors that exist (1/19 on WDR91), because PGK2 has near-zero direct ChEMBL annotation. Belongs as a tie-break / pseudo-positive nominator; the strong PCM/PIDGIN version needs Explorer GPU. P_mech.

The feared ~TB JUMP download was wrong — the assembled Cell-Painting profiles are a single 2.7–11.5 GB parquet, and 1,253 test compounds already have a real 737-d morphology vector in JUMP-CP (zero training, domain-gap-immune). Can't be validated on WDR91 (a WD40 protein with no metabolism), so it needs a glycolysis-specific check. P_bio.

Built the reusable WDR91 top-50 harness every prior plugs into — it reproduces the validated model⊕Boltz 11/19 and demonstrates the leakage guard biting (a self-referential prior: 19/19 fiction unguarded → 13/19 guarded). Defines the two fold-back modes (re-rank vs transductive augmentation) and a 5-clause pre-registered promotion rule.

The autonomous AI-scientist loop keeps running on cloud crons — one picks a research frontier and writes a directive each tick; the other mines hackathons for transferable insight (87 catalogued). Both commit to the repo.