Machine learning–driven decoding of maternal immune signatures in repeated pregnancy loss

Abstract

Repeated pregnancy loss (RPL) is a multifactorial condition in which the underlying immunological mechanisms, particularly the disruption of maternal-fetal tolerance, remain incompletely understood. Although immune tolerance is critical for pregnancy success, the specific immune dysregulations contributing to RPL, particularly in euploid pregnancies, have been difficult to characterize. To address this, we performed single-cell RNA sequencing of decidual tissues from RPL patients and first-trimester controls. Our analysis initially revealed elevated expression of a transcriptional module of immune activation genes in RPL decidual tissues. To dissect the cellular drivers of this complex landscape, we employed genotype-based origin analysis coupled with a supervised machine learning model and a transformer-based foundation model (scGPT). This hierarchical approach prioritized maternal T cells over other immune subsets as the population carrying the most distinct and generalizable RPL-associated signatures. Through the convergence of computational drug repurposing, network centrality analysis, and a rigorous origin-controlled expression filtering strategy, we identified CXCR4 and JUN as druggable molecular candidates strongly associated with this T-cell dysregulation. Collectively, our machine learning–driven approach characterizes the maternal immune landscape of euploid RPL in the context of immune tolerance breakdown, and nominates candidate targets for future functional investigation.