A hybrid 1DCNN-GRU deep learning framework for classifying caprine granulosa cell fertility potential using single-cell transcriptomics.

Abstract

Background and aim: Granulosa cells (GCs) are crucial mediators of follicular development and oocyte competence in goats, with their gene expression profiles serving as potential biomarkers of fertility. However, the lack of a standardized, quantifiable method to assess GC quality using transcriptomic data has limited the translation of such findings into reproductive applications. This study aimed to develop a hybrid deep learning model integrating one-dimensional convolutional neural networks (1DCNNs) and gated recurrent units (GRUs) to classify GCs as fertility-supporting (FS) or non-fertility-supporting (NFS) using single-cell RNA sequencing (scRNA-seq) data.

Materials and methods: We analyzed publicly available scRNA-seq datasets from monotocous and polytocous goats. A set of 44 differentially expressed genes (DEGs) (False discovery rate ≤0.01, log2 fold change ≥1.5) was identified and used to distinguish FS-GCs and NFS-GCs through Leiden clustering. The expression profiles of these DEGs served as input to train a hybrid 1DCNN-GRU classifier. Model performance was evaluated using accuracy, precision, recall, and F1 score.

Results: The optimized hybrid model achieved high classification performance (accuracy = 98.89%, precision = 100%, recall = 97.83%, and F1 score = 98.84%). When applied to scRNA-seq datasets, it identified a significantly higher proportion of FS-GCs in the polytocous sample (87%) compared to the monotocous sample (10.17%). DEG overlap across samples further confirmed the model's biological consistency and generalizability.

Conclusion: This study presents the first application of deep learning-based classification of goat GCs using scRNA-seq data. The hybrid 1DCNN-GRU model offers a robust and quantifiable method for evaluating GC fertility, holding promise for improving reproductive selection in livestock breeding programs. Future validation in larger datasets and across species could establish this model as a scalable molecular tool for precision livestock management.