SUN-154 Machine Learning-driven Single-cell Genomics In Polycystic Ovary Syndrome: Advanced Clustering And Genetic Profiling

Abstract Disclosure: A. Chen: None. Polycystic ovary syndrome (PCOS) is a multifactorial and common endocrine condition characterized by ovarian-source hyperandrogenism, abnormal folliculogenesis, insulin resistance, and infertility. While extensive knowledge through bulk transcriptomics has identified central molecular alterations in ovarian theca and granulosa cells, the cellular resolution of these phenomena at the single-cell level, most critically, those that underlie androgen excess and metabolic aberrations, is not yet adequately described. A high-resolution map of ovarian cell populations in PCOS is urgently needed to uncover the mechanisms of its heterogeneity and to inform the creation of targeted therapies. In the current study, we utilized scGPT, a cutting-edge machine-learning-based large language model, to combine and dissect multiple single-cell RNA sequencing (scRNA-seq) datasets from ovarian tissue of PCOS women and age-matched healthy controls. Leveraging its zero-shot generalizability, scGPT was able to identify sparse, rare, and transitional cell subpopulations in heterogeneous sample types without needing to be retrained, thus offering an adaptive and scalable cross-sample comparison platform. Compared with traditional methods such as scVI and Harmony, scGPT offered improved batch effect correction with preservation of key biological variability, particularly in the theca, granulosa, stromal, and immune compartments. Differential gene expression analysis revealed transcriptional programs corresponding to the hyperandrogenic and metabolically deranged phenotype of PCOS. Ovarian tissue from metformin-treated subjects, in particular, exhibited partial reversal of gene expression patterns observed in the disease, demonstrating the drug's potential to modify the ovarian microenvironment and restore molecular homeostasis. Our findings reveal the paradigm-shifting promise of machine learning-augmented single-cell genomics to untangle PCOS cell-type-specific disease mechanisms. In addition to gaining an unparalleled understanding of ovarian dysfunction, this approach also provides a critical foundation for precision diagnosis, treatment stratification, and future multi-omic integration in reproductive endocrine disease research. Presentation: Sunday, July 13, 2025