GCAN: A data feature learning method for scRNA-seq data

Area of Science:

• Computational Biology
• Genomics
• Bioinformatics

Background:

• Single-cell RNA sequencing (scRNA-seq) data present challenges like sparsity, noise, and complex structures.
• Existing methods struggle with modeling intercellular relationships and stable feature selection.

Purpose of the Study:

• Develop a method for automatic identification of biologically relevant features in noisy scRNA-seq data.
• Model multi-scale cellular relationships for robust clustering.
• Generate interpretable representations for mechanistic insights.

Main Methods:

• Utilized adaptive Highly Variable Gene (HVG) selection with an improved Random Forest to address dropout artifacts.
• Employed a hybrid graph autoencoder, fusing Graph Attention Networks (local) and Graph Convolutional Networks (global).
• Incorporated biologically informed optimization using MMD regularization and Spearman-correlation feature filtering.

Main Results:

• GCAN demonstrated superior clustering performance compared to CellVGAE across 17 scRNA-seq datasets, evidenced by Silhouette Coefficient and Davies-Bouldin Index.
• Adaptive HVG selection effectively reduced technical noise while preserving biological signals.
• The method achieved accurate cell typing through adaptive feature screening and relational modeling.

Conclusions:

• GCAN offers a new paradigm for scRNA-seq analysis by integrating adaptive feature selection and context-aware relational modeling.
• The architecture facilitates robust biological interpretation and accurate cell typing from scRNA-seq data.