scKanFormer: A Transformer-KAN framework with biologically informed attention for cell type annotation in large-scale...

Area of Science:

• Genomics
• Computational Biology
• Bioinformatics

Background:

• Accurate cell type annotation is crucial for single-cell RNA sequencing (scRNA-seq) data analysis.
• Existing deep learning methods struggle with batch effects, interpretability, and handling large datasets.

Purpose of the Study:

• To develop a supervised framework, scKanFormer, for precise and robust cell type annotation in large-scale scRNA-seq data.
• To address limitations of interpretability and batch effect sensitivity in current annotation methods.

Main Methods:

• Utilized a Transformer architecture integrated with Kolmogorov-Arnold Networks (KAN) and Convolutional Neural Networks (CNN).
• Incorporated biological information, local, and global attention mechanisms.
• Avoided dimensionality reduction to maintain feature traceability.

Main Results:

• scKanFormer demonstrated precise, robust, and transferable high-resolution cell type annotations across benchmark datasets.
• The model showed insensitivity to batch effects and enhanced biological interpretability.
• Outperformed nine state-of-the-art methods in various cell type annotation scenarios.

Conclusions:

• scKanFormer offers a significant advancement in scRNA-seq data analysis, providing interpretable and batch-effect-resilient cell type annotations.
• The framework's design enhances the identification and interpretation of cellular features, paving the way for more reliable biological insights.