State-of-the-art sequencing techniques have significantly enhanced our understanding of human tissues at single-cell resolution. However, individual studies often involve a limited number of donors, raising concerns about the generalizability of findings. Discrepancies in cell type definitions across studies complicate this further. As a result, integrating multiple datasets has become a key strategy to address these challenges. In single-cell RNA sequencing, while many sample sequencing results are publicly available, the increasing data volume has rendered traditional integration methods inadequate for processing millions of cells due to high computational and time demands. Researchers are thus focusing on more efficient integration solutions. In this study, we utilize a Sketch strategy to construct a prostate cell atlas, integrating all available open-source single-cell data related to the prostate, while providing a valuable research tool for future investigations.

All upstream files for single-cell RNA sequencing were processed using STARsolo (version 2.7.11b) with the reference genome GRCh38-2024-A. The integration of multiple samples was performed using the Sketch strategy from Seurat V5, which alleviated computational burdens by sampling cells from each dataset and calculating their leverage scores, thereby preserving rare subpopulations. We employed scIB (single-cell integration benchmarking) to compare batch correction and biological conservation across strategies. A deep learning model using a fully connected neural network was developed for label transformation, facilitating reference mapping of quarry data.

We downloaded all available raw data for prostate samples from the GEO database, comprising 12 projects, 119 samples, and 720,000 cells, including prostate cancer samples (hormone-sensitive, castrate-resistant, neuroendocrine), normal prostate samples, and benign prostatic hyperplasia samples. In downstream processing, we used the Sketch strategy to sample 220,000 cells from the total and constructed a core map with cell annotation via rPCA integration. We compared results from Sketch-rPCA with those from PCA and Harmony applied to all 720,000 cells, finding good preservation of biological conservation within each cellular subpopulation despite sampling. Finally, we developed a prediction model for cell labels based on the neural network, enabling mapping of information from the core map to other prostate single-cell samples.

We assessed the feasibility of the Sketch-based integration strategy in large-sample single-cell projects, saving computational and time costs while yielding integration results comparable to de novo methods. We constructed the most extensive prostate single-cell atlas to date and mapped our integration results to quarry data using deep learning models, thus providing valuable tools for researchers in related fields.