Transformer for one stop interpretable cell type annotation.

Area of Science:

• Computational Biology
• Genomics
• Bioinformatics

Background:

• Accurate cell type annotation is crucial for single-cell RNA sequencing (scRNA-seq) reproducibility.
• Existing deep learning methods often lack interpretability or struggle with data depth.
• Autoencoder-based tools offer speed but compromise on biological insight.

Purpose of the Study:

• To introduce TOSICA, a Transformer-based deep learning model for interpretable cell type annotation.
• To enable annotation using biologically relevant entities like pathways and regulons.
• To achieve fast, accurate, and batch-insensitive annotation for scRNA-seq data.

Main Methods:

• Developed TOSICA, a multi-head self-attention deep learning model utilizing the Transformer architecture.
• Enabled annotation transfer by integrating biologically meaningful entities (pathways, regulons).
• Applied TOSICA to scRNA-seq datasets, including tumor-infiltrating immune cells and COVID-19 monocytes.

Main Results:

• TOSICA demonstrated fast and accurate one-stop annotation capabilities.
• The model showed robustness to batch effects, ensuring reliable data integration.
• Biological interpretability was achieved, revealing insights into cellular behavior and disease dynamics.

Conclusions:

• TOSICA provides a powerful, interpretable deep learning approach for scRNA-seq annotation.
• It facilitates the discovery of rare cell types and heterogeneity in complex biological systems.
• The model aids in understanding disease progression and severity through dynamic trajectory analysis.