scMRA: a robust deep learning method to annotate scRNA-seq data with multiple reference datasets.

Area of Science:

• Computational Biology
• Genomics
• Bioinformatics

Background:

• Single-cell RNA sequencing (scRNA-seq) is crucial for understanding cellular heterogeneity.
• Integrating multiple scRNA-seq datasets is essential for comprehensive cell-type annotation.
• Existing methods struggle with batch effects and the lack of high-quality reference data.

Purpose of the Study:

• To develop an effective and efficient cell-type identification method for scRNA-seq data integration.
• To address the challenge of limited and diverse reference datasets in cell annotation.
• To overcome batch effects inherent in multi-dataset scRNA-seq analysis.

Main Methods:

• Introduced a deep learning-based tool named single-cell Multiple Reference Annotator (scMRA).
• Constructed a knowledge graph to represent cell-type characteristics across datasets.
• Employed a graph convolutional network as a discriminator to maintain cell-type relationships.

Main Results:

• scMRA effectively transfers knowledge from multiple reference datasets to unlabeled target data.
• The method preserves intra-cell-type similarity and inter-cell-type relative positions.
• scMRA demonstrates superior performance over state-of-the-art methods in multi-reference experiments and removes batch effects.

Conclusions:

• scMRA is the first method to utilize multiple insufficient reference datasets for target data annotation.
• This approach offers a robust solution for cell-type identification in complex scRNA-seq data.
• scMRA provides a significant advancement in analyzing and annotating multiple scRNA-seq datasets.