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

Repeated pregnancy loss (RPL) is a multifactorial condition with incompletely understood immunological mechanisms, particularly the immune disruptions contributing to RPL independent of fetal aneuploidy. To dissect this immune dysregulation, we performed single-cell RNA sequencing of decidual tissues from RPL patients and controls. Our analysis revealed that in RPL, fetal immune cells increased while fetal trophoblasts were reduced, and notably, RPL immune cells displayed transcriptional signatures resembling acute transplant rejection. Using machine learning and foundation models, we identified broad T-cell–derived transcriptomic signatures that distinguish RPL immune cells. To prioritize these candidates, we integrated network analysis and assessed their potential for therapeutic reversal using drug response data. After this rigorous filtering, we then controlled for biological confounding factors, a process which robustly identified CXCR4 and JUN in maternal T cells as the key RPL-associated signatures. The drug response analysis also highlighted three specific compounds as candidates for repurposing. Together, our approach identifies critical maternal immune signatures in RPL and links them to potential therapeutic opportunities, thereby providing both mechanistic insights and new avenues for treatment. Overall design: First-trimester decidual tissues were collected via dilation and curettage from women undergoing elective termination of uncomplicated pregnancies (Norm) and from patients with repeated pregnancy loss (RPL). Decidual samples were dissociated into single-cell suspensions and subjected to droplet-based scRNA-seq on the 10x Genomics Chromium platform with sequencing on Illumina HiSeq-X and NovaSeq 6000.