Abstract
The emergence of single-cell RNA sequencing (scRNA-seq) technology has transformed our understanding of cellular diversity, yet it presents notable challenges for cell type annotation due to data's high dimensionality and sparsity. To tackle these issues, we present scSorterDL, an innovative approach that combines penalized Linear Discriminant Analysis (pLDA), swarm learning, and deep neural networks (DNNs) to improve cell type classification. In scSorterDL, we generate numerous random subsets of the data and apply pLDA models to each subset to capture varied data aspects. The model outputs are then consolidated using a DNN that identifies complex relationships among the pLDA scores, enhancing classification accuracy by considering interactions that simpler methods might overlook. Utilizing GPU computing for both swarm learning and deep learning, scSorterDL adeptly manages large datasets and high-dimensional gene expression data. We tested scSorterDL on 13 real scRNA-seq datasets from diverse species, tissues, and platforms, as well as on 20 pairs of cross-platform datasets. Our method surpassed nine current cell annotation tools in both accuracy and robustness, indicating exceptional performance in both cross-validation and cross-platform contexts. These findings underscore the potential of scSorterDL as an effective and adaptable tool for automated cell type annotation in scRNA-seq research. The code is available on GitHub: https://github.com/kellen8hao/scSorterDL.