Preprint Uses FELINE Trial Profiles and Breast Cancer Organoids to Study Endocrine Insensitivity

Following Cells That Remain During Endocrine Treatment

A bioRxiv version-one preprint posted September 16, 2026, combines human tumor sampling, single-cell analysis and organoid experiments to investigate endocrine-insensitive breast cancer. FELINE is a human neoadjuvant trial in postmenopausal patients with estrogen-receptor-positive, HER2-negative disease. The study asks which cell populations persist during endocrine therapy and whether their expression profiles can nominate additional drug vulnerabilities.

The clinical analysis used paired day-zero and day-14 letrozole samples from nine patients: five responders and four nonresponders. The investigators classified populations by changes in abundance during treatment. Statistically depleted clusters were considered sensitive; enriched or unchanged clusters were considered insensitive. This definition describes behavior over the sampled treatment interval.

Insensitive Populations Were More Diverse

Every analyzed tumor contained both sensitive and insensitive populations. Estrogen-receptor expression or activity and conventional subtype assignments did not reliably explain their proportions. Fifteen clinical insensitive clusters included five patient-specific populations, whereas the sensitive states were more conserved across patients.

The investigators related cluster signatures to outcomes in 1,817 estrogen-receptor-positive tumors from the Molecular Taxonomy of Breast Cancer International Consortium (METABRIC) dataset. Some signatures were associated with poorer overall or relapse-free survival. These retrospective associations provide prognostic context for the states identified in the trial samples.

Longitudinal analyses incorporating day-zero, day-14 and day-180 samples produced a more qualified result. Cell-level mixed models identified response-by-time interactions for clusters 2 and 5, but the corresponding analysis of patient-averaged data was not statistically significant. The small, incomplete patient series therefore limits the strength of the longitudinal clinical inference.

Four Human Tumor Models Supplied Experimental Comparisons

The experimental arm used four patient-derived xenograft-derived organoid models, abbreviated PDxOs. The tumors were human, with an intervening passage through immunodeficient mice. HCI003 and HCI017 came from treatment-naive estrogen- and progesterone-receptor-positive tumors; UIC013 was estrogen-receptor-positive and progesterone-receptor-negative. UIC020 originated from a tumor initially classified as estrogen-receptor-negative that recurred as estrogen-receptor-positive.

Fourteen-day treatment with 1 μM fulvestrant or 4-hydroxytamoxifen produced partial growth responses that persisted after extended passaging. EdU nucleotide labeling showed reduced S-phase entry even in comparisons where organoid-area changes were not significant. Thus, the proliferation and size readouts captured different aspects of response.

Single-cell analysis of these models identified five relatively conserved sensitive clusters and nine more diverse insensitive clusters. Epithelial states were retained in differing proportions, while inflammatory and immune-associated signals were reduced in the xenograft-derived setting. This model composition matters when applying clinical tumor signatures to the organoid experiments.

Computational Priorities Were Tested Selectively

Drug Estimation from single-cell Expression Profiles, or DREEP, combined cluster expression signatures with the Cancer Therapeutics Response Portal version-two cell-line pharmacogenomic dataset. The analysis generated scores for 450 drugs; negative scores indicated predicted sensitivity. The subsequent experiments tested selected compounds using fourteen-day organoid-area measurements by image cytometry.

UIC013 showed greater sensitivity to dasatinib and pluripotin than the other models. Vandetanib activity was greater in UIC013 than in HCI017 and UIC020, whereas saracatinib produced only modest differences among models. AZD8055, an mTOR inhibitor nominated against shared populations, produced similar responses across the four PDxOs. Depending on the compound, area-under-the-curve comparisons used two or three independent biological experiments.

Separate MCF7 colony-formation and endocrine-resistant cell-line assays added pharmacological checks in simpler systems. Those experiments used different treatment formats and durations from the PDxO tests. Across computational comparisons, pathway activity and DREEP sensitivity were not consistently correlated for every compound.

A Route to Testing Heterogeneous Residual States

The preprint links a small clinical sampling study to measured drug responses in four human tumor-derived organoid models. It highlights the risk of losing rare insensitive populations in a bulk average and provides selected experimental support for model-specific drug hypotheses. The new candidates have not been evaluated as treatments in the FELINE patients. Broader donor sampling and prospective comparisons are needed to determine which cell-state signatures retain predictive value across tumors and treatment histories.