scMapNet: Marker-based cell type annotation of scRNA-seq data via vision transfer learning with tabular-to-image...

Area of Science:

• Computational Biology
• Genomics
• Bioinformatics

Background:

• Single-cell RNA sequencing (scRNA-seq) is crucial for understanding cellular heterogeneity.
• Current cell-type annotation methods face challenges with efficiency, consistency, and leveraging marker knowledge from unlabeled data.
• Supervised methods show promise but struggle to integrate existing biological knowledge.

Purpose of the Study:

• Introduce scMapNet, a novel deep learning method for scRNA-seq data analysis.
• Enable effective utilization of cellular marker knowledge and unlabeled data for cell-type annotation.
• Enhance the accuracy and consistency of cell-type identification.

Main Methods:

• scMapNet employs a self-supervised deep learning framework utilizing masked autoencoders (MAE) and vision transformers (ViT).
• Treemap transformations are integrated to leverage cell marker information effectively.
• The model is pretrained on large volumes of unlabeled scRNA-seq data.

Main Results:

• scMapNet demonstrated superior performance compared to six existing methods across multiple datasets.
• The method achieved high accuracy and robustness against batch effects.
• scMapNet effectively extracts attention gene information, aiding in biological interpretation and cell type identification.

Conclusions:

• scMapNet offers significant improvements in cell-type annotation accuracy and batch insensitivity.
• The model provides good interpretability, offering valuable biological insights for researchers.
• The scMapNet models and code are publicly available for broader research use.