No novel TNBC repurposing candidate established: signature-reversal screening was non-specific and drug-response validation was not completed

Summary

This iteration built a stable TNBC-versus-luminal expression signature from TCGA-BRCA and queried it against L1000CDS2 for drugs that reverse it. The signature was sound: it recovered known basal biology and classified held-out tumors and breast cell lines with AUC about 0.94. The reversal screen was not specific, however. Luminal-flipped and random control signatures reached the same drug-hit frequencies (up to 0.99–1) as the real ones. Only 17 drugs separated from controls, and none was validated in drug-response data or checked against the literature. No drug can be called a TNBC repurposing candidate from this work. Only one of the three planned approaches (signature reversal) was run, and the dependency-matching and drug-response arms were not completed. All results are exploratory.

What was analyzed

• Expression data: 1111 TCGA-BRCA primary-tumor files were identified and 1106 downloaded (0 failures). They contain 19,938 protein-coding gene symbols, of which 16,298 were kept. Receptor status gave 191 TNBC, 646 luminal, 195 other and 63 unknown. The data were split into a discovery set of 586 tumors (134 basal TNBC) and a held-out set of 251 (57 basal).
• Signature construction: 300 bootstrap resamples of the discovery set, restricted to 9565 L1000 genes (947 landmark). Each resample used a moderated t-statistic and kept the top 50 up and top 50 down genes. Validation used 36 breast lines (22 TNBC/basal, 14 luminal) from a 1091-line expression matrix (9406 genes).
• Known-answer check: a separate run on 867 PAM50 primary tumors (basal vs LumA/LumB) with 5165 genes, plus L1000CDS2 queries on 40 bootstrap, 40 flipped and 40 random signatures.
• Reversal screen: 900 L1000CDS2 queries (300 real, 300 luminal-flipped, 300 random), 0 failed, 45,000 hit rows, 1119 distinct drugs. Approval was taken from PRISM "Launched" or ChEMBL max_phase 4 (the top 150 drugs by frequency were checked in ChEMBL); 168 drugs were approved.
• Not completed: PRISM primary screen and GDSC2 validation was in progress when time ran out, and the literature novelty ranking was not performed.
• Compute: 107 runs (90 succeeded), 3.5 CPU-hours, 4 work groups done, 1 dropped and 1 still in progress, 108 of 120 minutes used.

Key results

All of the following is exploratory.

• Signature biology (known-answer check, run adbc22f8): t(basal−luminal) was KRT5 +12.8, KRT14 +7.4, KRT17 +13.3, FOXC1 +41.6 and SOX10 +13.0. For GATA3 it was −29.1, ESR1 −42.8 and FOXA1 −36.0. All have the expected sign.
• Signature stability (run b881f68c):

The most stable up genes (frequency ≥0.99) included STAC, PSAT1, FAM171A1, YBX1, FZD9, UGT8, TTLL4, FOXC1, BCL11A, ART3, EN1, PDSS1 and CCNE1.

• Reversal screen (run 1bf89399):

Raw reversal frequency is therefore non-specific. The 17 control-filtered drugs are the only ones worth following up. Their names were not available in the inputs, and none has been validated.

• Small known-answer run (9b840649): top basal reversers with luminal and random frequency near 0 included palbociclib, erlotinib, trametinib, mitoxantrone, PD-0325901 and PHA-665752. Most of these are already studied in breast cancer, so they are not novel. Palbociclib, erlotinib and trametinib are approved drugs; mitoxantrone is an approved cytotoxic. CDK inhibitors scored as basal reversers (mean frequency 0.4–0.6), while HDAC inhibitors and some cytotoxics also hit the luminal and random controls.
• Rank test of a top-30 set in PRISM (run ff571fae): 30 top-ranked drugs were found in the 2034 PRISM drugs, 11 in basal and 8 in luminal lines. The mean z was −0.22 against a null of −0.011, with permutation p = 0.884. This gives no evidence of enrichment for TNBC-selective sensitivity.
• Claim f7e2f976 (specificity failure of reversal frequency): confirmation level candidate, novelty possibly_new, reviewer verdict inconclusive. The reviewer gave no verdict, so the claim is unreviewed.

No preregistered hypotheses were evaluated, and no evaluator leaderboard exists.

What did not work

• Raw L1000CDS2 reversal frequency does not discriminate TNBC signatures from luminal-flipped or random controls, so ranking by it would mostly return generic, broadly active compounds. HDAC inhibitors and cytotoxics hit controls at high frequency.
• The top-30 reverser set showed no enrichment for basal-line sensitivity in PRISM (permutation p = 0.88).
• The drugs most consistently reversing the signature are kinase and CDK inhibitors already studied in breast cancer, so the screen rediscovered known biology rather than new candidates.

Confidence and limitations

• Confidence is low-to-moderate that no strong candidate emerged. Nothing here is confirmed.
• The dependency-matching and drug-response arms, and any cross-approach consistency ranking, were not completed. The novelty check against literature and trials is also open.
• The known-answer run used only 5165 genes, which limits generality. The cell-line validation set is small (36 lines), and the TNBC label is a receptor-based approximation with PAM50-based variants.
• L1000 reversal reflects cell-line perturbations, mostly non-breast, so any candidate would need in vitro testing in TNBC lines and then in vivo work.

Next steps

1. Name and test the 17 control-filtered reversers in PRISM and GDSC2 breast lines, using a rank-based comparison of TNBC against luminal sensitivity that adjusts for proliferation.
2. Match TNBC-selective CRISPR/RNAi dependencies (DepMap) to targets of approved drugs, and intersect with the reversal and response results to rank by cross-approach consistency.
3. Check each surviving approved drug against the literature and ClinicalTrials.gov for existing TNBC testing.
4. Run dose-response assays of the top novel candidates in a panel of TNBC and luminal lines.