Tumor-Made Androgen Marks Breast Cancer Subtype and Survival Odds

Researchers who measured hormones directly inside breast tumors report that one androgen, androstenedione, marks a distinct and largely favorable tumor state. In a study published in EMBO Molecular Medicine, a team led by Ramesh Narayanan at the University of Tennessee Health Science Center profiled steroid levels and gene expression in the same surgical specimens and found that tumors rich in intratumoral androstenedione carried gene activity tied to slower cell division, earlier disease stage and grade, and longer patient survival, while tumors with the lowest levels were enriched for the aggressive triple-negative subtype.

The work departs from decades of studies that relied on blood hormone measurements. Circulating androgen studies in breast cancer have produced confounding results, in part because they ignored local conversion of steroids by tumor enzymes and the fact that only hormone not bound by sex hormone binding globulin is biologically active. The authors argue that the gene activity a hormone correlates with inside the tumor offers a readout of that free, functional androgen action, sidestepping the difficulty of measuring the unbound fraction directly.

The clinical cohort comprised 73 women from the Memphis metropolitan area, split evenly between African American and Caucasian American patients, with more than half of tumors estrogen receptor positive and HER2 negative and roughly 70 percent androgen receptor positive. The team quantified dehydroepiandrosterone, androstenedione, testosterone and dihydrotestosterone in fresh-frozen tissue using gas chromatography mass spectrometry, normalizing concentrations to tissue weight. Dihydrotestosterone fell below the quantification limit in many specimens and was dropped. Androstenedione ranged from 0.05 to 23 nanograms per gram of tumor, dehydroepiandrosterone from 0.1 to 70 nanograms per gram, and testosterone from 0.06 to 17 nanograms per gram.

When the researchers correlated the expression of 13,458 protein-coding genes against tumor hormone levels in the 61 specimens with usable transcriptomic data, androstenedione stood out. At a false discovery rate threshold of 0.1, 1,388 genes significantly correlated with intratumoral androstenedione, whereas no genes reached significance for dehydroepiandrosterone and only two did for testosterone. Gene set enrichment analysis showed that the pathways most strongly negatively correlated with androstenedione were the E2F, Myc and G2M checkpoint programs, core drivers of cell cycle progression, with normalized enrichment scores near minus three and q values between 10−5 and 10−8. The mTORC1 pathway was also negatively correlated, and 98 of the androstenedione-associated genes overlapped curated androgen-related gene sets, including 11 genes from the hallmark androgen response set.

Unsupervised clustering of those 1,388 genes sorted the tumors into low, intermediate and high androstenedione groups along what the authors describe as a continuous transcriptional gradient. Clinical variables that played no role in gene selection or clustering differed sharply across the groups: the proliferation marker ki67, tumor grade and tumor stage all decreased from the low to the high cluster. Every tumor in the low cluster was grade 3, while the high cluster contained the fewest grade 3 tumors, a difference with a P value of 0.012 by Fisher’s exact test, and the low cluster had the fewest stage 1 tumors while the high cluster was mostly stage 1, with a P value below 10−5. The low cluster was enriched for triple-negative tumors and the high cluster for estrogen and progesterone receptor positive tumors.

The subtype context mattered. In estrogen receptor positive tumors, androstenedione levels inversely tracked with ki67, and the decline across clusters was statistically significant. In estrogen receptor negative tumors, androstenedione still rose across the clusters, but ki67 did not fall, echoing preclinical findings that androgen receptor signaling suppresses growth in hormone receptor positive disease but not in receptor negative disease. The authors note that only total hormone levels were measured, so the weak direct correlation between any single steroid and ki67 was expected, since the unbound fraction is what acts.

To test whether the signature generalizes, the team applied the gene set to The Cancer Genome Atlas breast cancer cohort of 973 tumors using a centroid-based prediction method, with 1,231 of the 1,388 genes overlapping the TCGA platform. Thirteen percent of samples landed in the low cluster and 22.9 percent in the high cluster. MKI67 gene expression, used as a proliferation surrogate because TCGA lacks ki67 staining, fell progressively from the low to the high cluster, a pattern seen in both estrogen receptor positive and negative tumors. Deceased patients were fewer in the high cluster, a trend confined to estrogen receptor positive disease, and five-year survival was significantly higher in the high cluster than the low cluster among estrogen receptor positive tumors, though the difference was not significant in receptor negative tumors. Estrogen receptor negative tumors made up 44 percent of the low cluster but only 19.3 percent of the high cluster, and the low cluster was enriched for African American patients, consistent with its triple-negative composition.

Because measuring intratumoral steroids routinely is impractical, the researchers distilled the signature into a 90-gene predictor using sparse partial least squares, selecting the gene count that minimized mean absolute error in fivefold cross-validation. The model’s predicted androstenedione values correlated strongly with measured levels, at a correlation of 0.77 and an R squared of 0.6, and 88 of the 90 genes overlapped the original 1,388. In the absence of an external cohort, leave-one-out cross-validation showed most of the 90 genes reselected across iterations, though the authors state plainly that independent external validation is needed before the classifier can be considered reliable. Applied to the clinical data, the 90-gene classifier reproduced the full signature’s assignments with only 4.9 percent discordance and no switches between the low and high extremes.

Multivariable Cox modeling in TCGA added an important nuance: the survival association of the androstenedione clusters was attenuated after adjustment for PAM50 subtype, age, stage and MKI67, but an interaction between cluster and subtype significantly improved model fit, indicating the prognostic signal is subtype dependent. Within luminal A tumors, the effect was strongest. After adjusting for age, stage and MKI67, patients with high-cluster tumors had significantly better overall survival than those with low-cluster tumors, with a hazard ratio of 0.13 and a 95 percent confidence interval of 0.02 to 0.77, P equal to 0.025. Single-nucleus RNA sequencing of three estrogen receptor positive specimens, two high and one low in androstenedione, supported the bulk findings: high specimens were enriched for luminal epithelial cells and T cells, while the low specimen was enriched for basal cells, fibroblasts and endothelial cells, and its cells spent more time in DNA synthesis phase of the cell cycle. The authors flag this three-specimen analysis as exploratory.

Preclinical experiments pointed toward mechanism. Androstenedione activated the androgen receptor in transactivation assays and dose-dependently inhibited proliferation of the estrogen receptor and androgen receptor positive cell lines ZR-75-1 and T47D, with roughly 40 percent growth inhibition over 14 days consistent with a cytostatic rather than cytotoxic effect. Silencing the androgen receptor in T47D cells blocked that growth effect, indicating the receptor mediates it. Treating ZR-75-1 cells with the steroid for 14 days altered 3,292 genes, and 2,083 of those overlapped the clinical dataset, clustering in the same two directions the patient tumors had shown. The authors caution that the cohort is small, that tumor purity, stromal content and intratumoral steroidogenic enzyme expression may influence the signature, and that the findings remain a prognostic association pending validation in larger independent cohorts.

Subject of Research: Intratumoral androstenedione levels and a 90-gene signature associated with breast cancer subtype and prognosis

Article Title: Intratumoral androstene-3,17-dione defines breast cancer subtype and prognosis

Article References: Asemota, S., Choi, H. Y., Effah, W., Holt, J., Young, K., Cripe, L., McNamara, K., Byun, D. J., Seo, H. J., Ponnusamy, S., Johnson, D., Lee, J.-Y., Khosrosereshki, Y., Grimes, B., Fleming, M. D., Pritchard, F. E., Hendrix, A., Berry, M., Fine, R., … Narayanan, R. (2026). Intratumoral androstene-3,17-dione defines breast cancer subtype and prognosis. EMBO Molecular Medicine. https://doi.org/10.1038/s44321-026-00530-5

Image Credits: AI Generated

DOI: 10.1038/s44321-026-00530-5

Keywords: breast cancer, androstenedione, androgen receptor, gene signature, prognosis, triple-negative breast cancer, tumor hormones, transcriptomics, The Cancer Genome Atlas, single-nucleus RNA sequencing, biomarker, Intratumoral