Multi-modal deep learning for single-cell mapping of somatic-to-germline reprogramming in infertility

Abstract

Infertility is a complex disease commonly linked with the disturbed germline development and oocyte maturation process. Here, we adopted a multimodal deep learning platform coupled with single-cell transcriptomic analysis to reveal the transcriptional variation and developmental differentiation trajectory of infertility. Public datasets, including three disease states with more than 25,000 cells and 3,000 expressed features, were harmonized using PCA and Harmony for batch effect correction, followed by UMAP visualization and Leiden clustering, which yielded seven transcriptionally distinct clusters. Diffusion pseudotime analysis inferred a developmental continuum between early somatic-like and late germline-enriched populations, identifying progressions in the absence of defined panels of canonical oocyte and germline markers. Differentially expressed top 10 marker genes per cluster indicated candidate control factors for reproductive failure. Functional enrichment analysis validated the remarkable participation of pathways in meiotic recombination, germline stem cell maintenance, and oocyte maturation. Predictive modeling with Logistic Regression and Random Forest classifiers fitted on integrated embeddings achieved a maximum AUROC of 0.998 and AUPRC of 0.996 using Random Forest and MLP models, confirming strong discriminative performance. Together, our findings show that the combination of deep learning and single-cell multi-omics offers a potent approach to map cellular heterogeneity in infertility, reconstruct developmental dynamics, and discover new biomarkers and therapeutic targets.