Non-invasive Transcriptomic Cell Profiling of the Human Endometrium with Generative Deep Learning

This study addresses the need for non-invasive methods to assess the human endometrium, which is critical for successful pregnancy and is implicated in unexplained infertility. The primary biological relevance lies in determining if extracellular vesicles in uterine fluid (UF-EVs) can serve as a proxy for the cellular state of the endometrial tissue. The intent is to validate a deep learning-based deconvolution approach that uses the transcriptomic profile of UF-EVs to accurately delineate the cellular composition of the endometrium throughout the menstrual cycle. Success in this area would enable detailed endometrial assessment without requiring invasive biopsies. The experimental workflow was designed to compare cellular profiles derived from UF-EVs against those from paired endometrial tissue. The primary dataset consisted of paired endometrial tissue and UF-EV transcriptomes collected from 19 fertile, healthy women across four menstrual cycle phases: proliferative (n=4), early-secretory (n=5), mid-secretory (n=5), and late-secretory (n=5). Two additional unpaired UF-EV samples were also included (phenotypes mapped in main_pheno.csv). For this cohort, small EVs were isolated using size-exclusion chromatography and RNA libraries were prepared with the TruSeq exome RNA library preparation kit. Raw sequencing data was processed with the nf-core/rnaseq pipeline (v3.14) and the RSEM outputs are provided. To map the spatial origin of EVs, the analysis also incorporated endometrial tissue Visium slides from two proliferative (n=2) and two early-secretory (n=2) phase samples (E-MTAB-9260). A deep learning deconvolution model, BulkTrajBlend, was trained on a human endometrial single-cell RNA-seq reference atlas. The trained model was then used to infer cell type proportions in the bulk transcriptomes and to generate pseudo-single-cell (pSC) data (weights in pseudosc_bio.pth and pseudosc_ev.pth), which was subsequently projected onto the spatial transcriptomic slides to visualize the potential tissue origins of the cells secreting EVs.