No novel TNBC repurposing candidate established: three exploratory arms ran, none converged

Summary

This iteration did not find a new, supported repurposing candidate for triple-negative breast cancer (TNBC). The signature-reversal arm and the TNBC-vs-luminal analysis produced no result. The three arms that ran (network proximity, genotype-conditioned RNAi dependencies, immune-axis target correlation) are exploratory and gave no drug that passed two independent approaches. No preregistered hypothesis was run, so nothing here is confirmed. The one observation that may be new (network-proximity enrichment of ER agonists and GABA-A drugs) is probably an artifact.

Key results (all exploratory)

• Network proximity. A 300-gene basal-vs-other TCGA module (171 basal, 810 other tumours; 224 genes in STRING) was scored against 1,345 launched non-oncology drugs.
  • 200–206 drugs had z < -2, depending on the random draw.
  • Top hits were ER agonists (e.g. mestranol/quinestrol, z ≈ -10) and GABA-A drugs (diazepam, flumazenil, clonazepam, propofol; z ≈ -6 to -7).
  • Reruns reproduced the classes but not the exact drug order.
  • The ER hits are circular because ESR1 is in the module.
  • The GABA-A hits plausibly reflect large target sets and hub-heavy modules.
  • Novelty label: possibly_new (no close prior work found). Confirmation level: reproduced. This reflects computational reproducibility, not biological support.
• Genotype-conditioned RNAi (DEMETER2, 34 breast lines). 3,525 mutant-vs-wild-type tests per gene and genotype gave 61 with p<0.01, 4 with q<0.05 (minimum q = 0.027) and 30 hits at p<0.01. Seven replicated in 467 non-breast lines. All seven corresponding drugs already have trials or papers, so none is novel.
  • Group sizes were TP53 23, PTEN 4, BRCA1 3 and RB1 1 (skipped), so the signal is mainly TP53-driven and not TNBC-specific.
  • Novelty label: similar.
• Immune axis. Among basal-like/ER-HER2- tumours (TCGA n=171, METABRIC n=309), 217 of 1,187 targets of launched non-oncology drugs correlated with a marker-gene immune score. The correlation was adjusted for stromal and epithelial scores, replicated in METABRIC, and had the same direction. Cross-cohort rho was 0.60. Reproduced exactly across runs. Novelty label: similar (the closest prior work is TNBC immune/ICAM1 studies). The drug mapping is uncurated and no pair has been checked.
• Pharmacoepidemiology (FAERS). The breast-cancer reporting odds ratio was uninformative. The median ROR was ~0.56 in both periods, versus 1.96 for the headache negative control. Early/late agreement was weak (Spearman 0.32–0.37), and finasteride had an extreme ROR of 66 from tiny counts. This signal is a reporting artifact and the data carry no TNBC subtype.
• Literature check. Statins, beta blockers, metformin and aspirin are already studied in breast cancer. The RNR inhibitor class is known.

What did not work

• Disproportionality analysis of spontaneous reports does not separate drug effect from reporting background, and it cannot resolve TNBC.
• Proximity ranking is dominated by target-set size and hub effects (ER-agonist circularity, GABA-A targets), so it gives no usable candidates without controls.
• Genotype-conditioned dependencies in breast lines are underpowered for RB1, PTEN and BRCA1, and every replicated hit's drug already has prior records.
• No candidate passed two independent approaches, so the cross-arm convergence ranking is empty.

Confidence and limitations

• Confidence is low.
• No preregistration, evaluator score or formal proof exists.
• Untested: the TNBC-vs-luminal signature with METABRIC validation and perturbation reversal, genome-wide CRISPR dependencies matched to non-oncology targets, PRISM AUC with CTRP/GDSC replication, the basal-like vs mesenchymal check, the lipophilicity test matched on target class, and the six metabolic hits.
• The proximity module used no rewiring or degree-preserving controls and no external validation.
• Drug-action direction was not curated.
• Any candidate would need in vitro validation in TNBC lines.

Next steps

1. Derive the TNBC-vs-luminal signature on TCGA-BRCA, validate it on METABRIC, and score approved drugs by reversal.
2. Compute genome-wide TNBC-selective CRISPR dependencies and match them to targets of approved non-oncology drugs.
3. Redo the PRISM analysis with AUC, replicate in CTRP/GDSC, and compare basal-like with mesenchymal lines.
4. Re-run the proximity analysis excluding ESR1-linked targets, with degree-matched and rewired controls and METABRIC validation.
5. Curate action direction for the 217 immune-axis genes and check each drug-gene pair against the literature and trial registries.
6. Rank drugs by agreement across arms, drop those already studied, and preregister the top few for testing.