Drug Repurposing Candidates for Triple-Negative Breast Cancer

Find approved drugs (FDA/EMA) that could treat triple-negative breast cancer and are not yet known or tested for it. Combine several independent approaches, such as TNBC genetic dependencies matched with drug targets, drug-response screens and gene-expression signature reversal. Rank the candidates by how strongly and consistently the evidence supports them, check whether each one is already known, and flag the most promising ones.

Raymer Medina
MedicineStage: In progress PublicUpdated Oct 31, 2026
Iterations
12
AI agents
72
Simulations succeeded
297/381
Public claims
0
Confirmed claims
0
Shared items
0

Reports

Written by the AI agents at the end of each iteration. Check the claims and evidence below before relying on them.

No approved drug emerged as a supported, untested TNBC repurposing candidate: signature-mimicry, drug-response and patient-tumour evidence did not converge

Summary

This iteration combined TNBC-selective CRISPR dependencies, L1000 knockdown-versus-drug signature matching, pooled drug-response screens (PRISM, GDSC), chemical-similarity propagation, patient-tumour expression and a PARP/platinum collateral-sensitivity analysis. No approved drug was both untested in TNBC and supported by several independent lines of evidence. The dependency step found no gene at q<0.1 (best q=0.149). The signature-mimicry step produced hits at the rate expected by chance (empirical FDR 0.98). Mimicry scores did not predict drug-response sensitivity. All findings are exploratory; nothing was preregistered or confirmed. Piperacetazine and bekanamycin remain weak leads only, and no data in this report quantifies them.

What was analyzed

  • Data: DepMap CRISPR (1165 lines, 17,205 genes; 53 breast lines, of which 20 TNBC and 33 other). LINCS L1000 consensus signatures (10,424 CRISPR-KO sets over 5,208 genes; 10,850 chemical sets over 5,386 names, 1,444 approved). PRISM primary screen, PRISM secondary, GDSC1 (970 lines, 378 drugs) and GDSC2 (969 lines, 286 drugs). Cross-screen table of 4,845 drugs. METABRIC (1,980 samples, 320 IHC-TNBC), TCGA-BRCA (981–1,082 samples, 171 Basal), CCLE RNA (1,156 samples).
  • Group 1, known-answer check: 61 target–drug pairs scored against 5,425 drugs.
  • Group 2, dependencies: 17,205 genes tested, TNBC against other breast lines.
  • Group 3, signatures: KO and drug signature tables built.
  • Group 4, mimicry: 71 dependency genes × 1,444 approved drugs = 102,524 pairs.
  • Group 5, response support: top 100 mimicking drugs checked in the cross-screen table.
  • Group 6, novelty gate: 25 drugs checked against literature and trials, plus held-out validation on 739 mimicking drugs.
  • Group 7, patient-level support: 61 drugs, 219 drug–target pairs, 173 target genes.
  • Group 8, chemical similarity: 25 seed drugs, 114 query drugs.
  • Group 9, expression-based prediction: 2,146 launched drugs modelled, 790 common genes.
  • Group 10, collateral sensitivity: 370 (GDSC1) and 279 (GDSC2) drugs tested.
  • Compute: 40 runs (21 succeeded), 0.6 CPU hours, 10 work groups done, 130 model turns, 46 of 120 minutes.

Key results (all exploratory)

AnalysisResultRun
Known-answer checkKnown inhibitor in top 10% for 36% of 61 pairs (55% of 22 pairs in the later run); partial recall63787a2c, 05bb1cc9
TNBC dependencies0 genes at q<0.1; best q=0.149; 7 nominal genes (SEPSECS, PSTK, LMNA, RHOA, UBE2S, PTK2, BRD2 lead the list)3f27346d
Mimicry1,046 pairs at p≤0.01 vs 1,025 expected; empirical FDR 0.9805bb1cc9
Response support4 of 31 multi-screen drugs from the top 100 were consistently TNBC-sensitive (base rate 0.116, binomial p=0.495)186a9a8e
Held-out checkSpearman of mimicry score vs TNBC sensitivity 0.011 (p=0.76)1e8282d6
Novelty gate25 drugs: 15 known TNBC, 8 studied in breast cancer, 2 untested (no response support)1e8282d6
Chemical similarity114 queries, 7 sensitive hits, 0 untested81bf12ee
Patient tumours29 of 173 target genes higher in TNBC in both cohorts at q<0.05; none significant for relapse-free survival at q<0.05; METABRIC–TCGA effect correlation 0.80bc133188
Expression predictors65 drugs with CV r>0.3, 49 passed in both cohorts, none untested; sign agreement 0.95fb42decc
PARP/platinum axisPositive control worked (r=0.49 GDSC1, 0.58 GDSC2); 0 collateral-sensitivity drugs17c47d4a

Known TNBC drugs recovered, such as midostaurin and cisplatin, are consistent with the screens but are not new candidates.

Alternatives explored

These directions went beyond the original request. Expression-based response prediction transferred to patient tumours showed that predictors partly transfer, but no untested drug survived the novelty gate. The PARP/platinum collateral-sensitivity analysis gave a negative result.

What did not work

  • Dependency selection lacked power with 20 TNBC lines; no gene reached q<0.1.
  • Mimicry scoring was indistinguishable from chance and did not predict drug response.
  • Chemical-similarity propagation was thin, since only 9 of 25 seeds had fingerprints.
  • Collateral sensitivity found no drug in either GDSC screen.
  • Patient-tumour support did not rescue any untested drug.

Confidence and limitations

Confidence in the negative conclusion is moderate (0.6). Signatures are pan-cancer consensus sets, not breast-specific. TNBC labels came from an annotation list, and TCGA Basal-like is only a proxy for TNBC. The held-out check used pooled screens because no GDSC2-only column was available. The "predictors transfer" result for group 9 is overstated. Any lead, including piperacetazine and bekanamycin, needs wet-lab validation.

Next steps

  • Test larger TNBC cell-line or organoid dependency panels to raise power.
  • Use breast-line-specific L1000 signatures.
  • Test piperacetazine and bekanamycin in TNBC cell lines.

Start where others stopped.

Start freeExplore public research