scIMGCN: an Automatic Single-Cell Type Annotation Method Based on Interpretable Graph Convolutional Network.

Area of Science:

• Computational biology
• Genomics
• Bioinformatics

Background:

• Single-cell RNA sequencing (scRNA-seq) enables detailed study of cellular heterogeneity.
• Accurate cell type annotation from large scRNA-seq datasets remains a significant challenge.
• Graph convolutional networks (GCNs) show promise but suffer from limited interpretability.

Purpose of the Study:

• To develop an innovative method, scIMGCN, for automated and interpretable cell type annotation in scRNA-seq data.
• To address the limitations of traditional GCNs in handling complex single-cell data structures and interpretability.

Main Methods:

• scIMGCN integrates network augmentation for enhanced graph structure representation.
• An enhanced Transformer module dynamically models global relationships, mitigating long-range dependencies and noise.
• A Kolmogorov-Arnold network (KAN)-based GCN variant improves feature representation and nonlinearity.
• An interpretability masking mechanism enhances decision transparency.

Main Results:

• scIMGCN achieved annotation accuracy between 94.8% and 100% across ten real datasets.
• The method demonstrated improvements of 4.7%, 7.1%, and 5.6% in accuracy through its distinct components.
• scIMGCN outperformed traditional methods by over 15% and state-of-the-art graph-based methods by 4.8%.

Conclusions:

• scIMGCN significantly enhances cell type annotation accuracy and scalability for scRNA-seq data.
• The method effectively models complex intercellular relationships and improves model interpretability.
• scIMGCN offers a robust and generalizable solution for automated cell type classification.