Final Report — Lasques Mission (Iteration 3): TNBC Drug-Selectivity Screen

Summary

The central question of this iteration — whether the seven iteration-2 "TNBC-selective" GDSCv2 drug hits are real — was answered negatively on two independent legs. First, re-running the 9-vs-9 within-breast TNBC screen with exact permutation p-values (2,000 label shuffles per drug) instead of Welch p-values, as the iteration-2 mechanism study recommended, restores FDR control and zero of 172 drugs survive at q < 0.10; the top hit, mk-1775/adavosertib, keeps the largest effect (mean AUC delta −0.137, permutation p = 0.002) but its q-value is 0.17. Second, the DepMap 25Q2 CRISPR triage (the item that never started in iteration 1) shows none of the 15 hit-target genes is a TNBC-selective dependency after pan-essential filtering (best BH q = 0.52, PSMB5 — and that trend runs in the opposite direction). The goal of discovering a new pattern, formula, or effect that iteration 2 did not establish was not met: the exploratory agent's n-scaling extension of the sign-flip law (n_TNBC = 5–26) failed on a code defect that was fixed too late to execute, so no new conjecture was computed, tested, or checked for novelty this iteration.

Key results

• Permutation re-screen (runs de5a1701 and b7635665, exploratory). 172 drugs with ≥7 lines per group, 2,000 permutations/drug, BH-FDR: 0 drugs at q < 0.10, 0 at q < 0.05. mk-1775/adavosertib: mean AUC delta −0.137 (TNBC more sensitive), permutation p = 0.002, q = 0.17. Robustness: even at rank 1 of 172, its q would be ≈0.34 under the most conservative ranking, so the zero-hits conclusion does not depend on ranking details.
• Basal-lineage confound (subtype table, exploratory). In AUC ~ TNBC + basal regressions (7/9 TNBC lines are basal; BL1/BL2/Basal-like labels), mk-1775's TNBC coefficient remains nominally significant (p = 0.0196, pre-FDR): the effect is not purely a basal-lineage artifact, but 0% of hits lose their TNBC effect after basal correction and nothing is FDR-confirmed.
• DepMap triage (run 3ed5e97b, exploratory). 372 MB DepMap 25Q2 Chronos matrix streamed in chunks; 53 breast lines (22 TNBC, 31 non-TNBC). Of 15 hit-target genes (WEE1, CHEK1/2, PSMB1/4/5, MET, BCL2-family, IGF1R/INSR, PCNA/POLR2A/RPA3/PLK1), none is a TNBC-selective dependency; best BH q = 0.52 (PSMB5, permutation p = 0.034), and the proteasome dependency delta is +0.46 — TNBC lines are less proteasome-dependent, opposite to the drug-screen direction. WEE1 is not a TNBC-selective dependency.
• Sensitivity check (run b7635665, exploratory). With the stricter all-18-lines requirement (65 drugs), mk-1775's permutation p = 0.001 and q = 0.065 — the only configuration in which one hit passes q < 0.10. The zero-hits headline rests on the larger 172-drug set.
• New-pattern discovery (goal item): not achieved. The jackknife sign-flip n-scaling analysis (logit(P_flip) = a(n) + b(n)|t| for n_TNBC = 5–26, Gaussian benchmark, held-out validation) never produced numbers; no conjecture was stated, tested, or novelty-checked.
• Published knowledge: two negative results (cc4580e3, 72f52dff). No claims, no preregistrations, no leaderboard, no novelty checks — all results below are exploratory.

What did not work

• The iteration-2 hit list did not survive statistical correction. All 7 Welch+BH "hits" are consistent with the anti-conservativeness mechanism: exact-permutation FDR eliminates all of them, including mk-1775 (q = 0.17). This contradicts and supersedes the three iteration-2 hit-list knowledge items.
• Orthogonal genetic-dependency support failed. No hit target is a TNBC-selective DepMap dependency; the proteasome signal points in the wrong direction, and WEE1 dependency is not TNBC-selective (consistent with adavosertib sensitivity tracking TP53 loss rather than TNBC/basal status).
• Wet-lab/docking triage: the docking portion of the DepMap triage did not run, and no independent PRISM/CTRP replication cohort existed (zero drug overlap on this mirror), so mk-1775 remains externally unreplicated.

Confidence and limitations

• All results are exploratory (single runs, no preregistration, no independent reproduction or novelty check this mission). Agent 1's screen result: confidence 0.9; agent 3's DepMap triage: confidence 0.8; agent 2's untested sign-flip scaling: confidence 0.4 — no empirical claim from that approach should be relied on.
• The n_TNBC = 5–26 sign-flip scaling configurations stayed untested because the compute runs failed on a scalar-conversion bug whose verified fix came after the turn limit.
• The mk-1775 nominal effect (permutation p = 0.002, survives basal correction at p = 0.0196) is a real candidate signal but needs wet-lab validation across the 9+9 line panel or an independent replication cohort before any claim.
• The 65-drug strict-panel run gives q = 0.065 for mk-1775 — borderline; a preregistered confirmatory run with fixed line assignments would settle it.

Next steps

1. Wet-lab validation of adavosertib (WEE1 inhibition) across the 9 TNBC vs 9 non-TNBC breast lines — the decisive test of the top nominal hit.
2. Preregister the mk-1775 selectivity claim with fixed line assignments and permutation statistics, ideally alongside a PRISM/CTRP subset with actual drug overlap, so it can be formally confirmed or rejected (nutlin-3a shows replication is detectable when overlap exists).
3. Re-run the jackknife sign-flip n-scaling experiment from the verified snapshot (streaming fix already in place, fits in ~950 MiB): fit b(n) on fit-drugs, test the closed-form conjecture on held-out n ∈ {7, 9, …, 23}, and compare against the Gaussian benchmark — framed within the FDR-valid permutation regime established this iteration.
4. Complete the docking triage if a target ever survives both the screen and the dependency filter.