A self-training interpretable cell type annotation framework using specific marker gene.

Area of Science:

• Computational Biology
• Genomics
• Bioinformatics

Background:

• Single-cell RNA sequencing (scRNA-seq) enables high-resolution biological process studies.
• Accurate cell type annotation is crucial for scRNA-seq analysis, typically relying on marker genes.
• Existing methods often separate clustering and assignment, limiting marker information guidance and failing to capture complex cell dependencies.

Purpose of the Study:

• To develop a novel marker-based cell type annotation method for scRNA-seq data.
• To improve the accuracy and interpretability of cell type identification.
• To address limitations of existing two-stage annotation methods.

Main Methods:

• Introduced Interpretable Cell Type Annotation based on self-training (sICTA), a marker-based method.
• Integrated self-training with pseudo-labeling and Transformer networks for nonlinear association capture.
• Incorporated biological prior knowledge (genes, pathways) via an attention mechanism for transparency.

Main Results:

• sICTA demonstrated superior performance compared to state-of-the-art methods across 11 public scRNA-seq datasets.
• Ablation studies confirmed the synergistic benefits of self-training and dependency capture for model performance.
• The method achieved robust prediction accuracy across diverse cell types and datasets, with interpretable attention matrices.

Conclusions:

• sICTA offers a powerful and interpretable approach for cell type annotation in scRNA-seq data.
• The combination of self-training and nonlinear modeling significantly enhances annotation accuracy.
• The method provides valuable insights into cell type identification and biological relationships.