PredGCN: a Pruning-enabled Gene-Cell Net for automatic cell annotation of single cell transcriptome data.

Area of Science:

• Genomics
• Computational Biology
• Bioinformatics

Background:

• Accurate cell type annotation from single-cell transcriptomics (scRNA-seq) is vital for understanding cellular functions.
• Manual annotation is the gold standard but not scalable; automatic methods, especially deep learning, are crucial but depend on classifier architecture and training data quality.

Purpose of the Study:

• To develop an advanced deep learning framework, PredGCN, for robust and scalable cell type annotation in scRNA-seq data.
• To improve the accuracy and dynamic adaptability of automatic cell type identification methods.

Main Methods:

• Introduced Pruning-enabled Gene-Cell Net (PredGCN) with a Coupled Gene-Cell Net (CGCN) for representation learning.
• Integrated Gene Splicing Net (GSN) for feature extraction and Cell Stratification Net (CSN) for cell identification.
• Employed a multi-objective Pareto pruning operation (Pareto PrO) to optimize sub-network structures for annotation.

Main Results:

• PredGCN demonstrated superior performance over state-of-the-art methods on diverse scRNA-seq datasets across species.
• The method showed excellent scalability, including for cross-species datasets.
• PredGCN successfully identified unknown cell types and provided functional genomic insights by analyzing gene influence on cell clusters.

Conclusions:

• PredGCN offers a powerful and scalable solution for automated cell type annotation in scRNA-seq data.
• The framework provides novel perspectives for cell type identification and characterization.
• Availability of code and data facilitates further research and application in the field.