Drug Repurposing Candidates for Triple-Negative Breast Cancer

Builds on 'Own the end-to-end pipeline and final ranked table' (stopped before finishing in iteration 1): Actually execute the full screen. First a tiny-sample validation (columns, encoding, presence of MDA-MB-231/MDA-MB-468/HCC1937/BT-549 labels, Database==GDSCv2 filter), then Range-request streaming in 10-25 MB chunks with float32 and incremental accumulation to avoid the previous OOM, then per-drug TNBC-vs-other Welch t-tests with BH-FDR. Deliver the first actual ranked table with q-values and the reproducibility floor (per-drug GDSCv2|PRISM Spearman) attached to every hit. Deliverable is data + table, no claim beyond what survives the floor.

Compute many exact small cases and look for formulas: on small drug subsets, compute leave-one-TNBC-line-out jackknife statistics for every drug's t-statistic, fit the empirical sign-flip/instability probability as a function of t and effect size (H2), and test the fitted law on held-out drugs. Separately, quantify the basal-confound conjecture (H1): compute per-drug selectivity with and without basal non-TNBC lines in the comparator group and derive the 'basal-explained fraction' formula; test whether this fraction is constant across drug classes. State both laws as precise conjectures with exact small-case evidence.

Builds on the abandoned 'Attack the emerging shortlist' role: act as skeptic and counter-screener. (1) Permutation-test the BH-FDR imbalanced-design conjecture (H3) — simulate null t-tests with 4-vs-N group sizes on shuffled data and measure actual FDR versus nominal BH q. (2) Counter-screen the lead's top hits: pan-essential-target filter, basal-lineage confound check, prior-art/literature flag for every top drug (novelty verification). (3) Try to break the pattern hunter's stability law with counterexample drugs and probe the H4 non-oncology enrichment claim with a hypergeometric test. Nothing is claimed until it survives these attacks.