scDeepSort: a pre-trained cell-type annotation method for single-cell transcriptomics using deep learning with a...

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Single-cell RNA sequencing (scRNA-seq) enables transcriptome profiling of individual cells.
• Accurate cell-type annotation is crucial but challenging for scRNA-seq data.
• Current methods often rely on marker genes or RNA-seq profiles, limiting extrapolation.

Purpose of the Study:

• Introduce scDeepSort, a novel tool for automated cell-type annotation of scRNA-seq data.
• Address the limitations of existing cell-type annotation methods.
• Provide a robust and accurate annotation solution for large-scale transcriptomic datasets.

Main Methods:

• Developed scDeepSort, a deep learning tool utilizing a weighted graph neural network (GNN).
• Employed a pre-trained GNN model for cell-type annotation.
• Validated scDeepSort on extensive human and mouse scRNA-seq datasets from various tissues.

Main Results:

• scDeepSort demonstrated high performance and robustness in labeling 764,741 cells across 56 human and 32 mouse tissues.
• Outperformed existing methods on 76 external test datasets, achieving 83.79% accuracy.
• Showcased universality across challenging datasets and different scRNA-seq technologies.

Conclusions:

• scDeepSort is the first tool to use a pre-trained GNN for scRNA-seq cell-type annotation.
• Enables accurate cell-type identification without requiring additional references like marker genes or RNA-seq profiles.
• Offers a significant advancement for analyzing complex single-cell transcriptomic data.