On the illusion of a pre-treatment chemoresistance signal in HR+ breast cancer

Abstract

Small single cell RNA sequencing cohorts are increasingly used to search for pre-treatment cell states that predict chemotherapy resistance, but such studies are vulnerable to under-appreciated failure modes: cell-level pseudoreplication, patient batch effects masquerading as shared biology, and gene signatures that inadvertently capture proliferation rather than resistance. We reanalyzed a paired single cell RNA sequencing dataset of pre- and post-chemotherapy tumor biopsies from five HR+ breast cancer patients (three resistant, two sensitive) to test whether pre-treatment tumor cells resemble post-treatment surviving cells, then subjected the signal to four robustness checks. An initial per-cell analysis produced a seemingly positive result (one-sided Mann-Whitney U p = 0.069): resistant patients pre-treatment cells clustered with the post-treatment compartment more often than sensitive patients cells. This did not survive scrutiny. Harmony batch correction for patient identity showed the association was driven almost entirely by one patient contributing six pre-treatment tumor cells, whose post-treatment-compartment membership fell from 100% to 0% after correction (corrected p = 0.1665). A parallel signature derived from post-treatment upregulated genes scored higher, not lower, in pre-treatment sensitive-patient cells (p = 1.000); gene set enrichment analysis confirmed this signature was dominated by cell-cycle and mitotic pathways (Hallmark G2-M Checkpoint, adjusted p = 1.4x10^-9) rather than a resistance-specific program. A patient-balanced pseudobulk reanalysis found a consistent direction of effect across all three resistant patients but did not reach significance (paired Wilcoxon p = 0.125, n = 3). These results show that a superficially promising resistance signal in a small cohort can arise from patient-specific batch structure and proliferation confounding rather than reproducible biology, and that patient-level pseudobulk analysis, batch integration, and pathway enrichment testing are necessary, low-cost checks before such signals are reported as candidate biomarkers. We provide this as a case study and checklist for evaluating single cell resistance signatures in small cohorts.