Single-cell RNA sequencing allows researchers to study gene activity in individual cells, making it possible to identify distinct cell types within complex tissues. However, analyzing these datasets is challenging because they contain thousands of genes, many missing values, and millions of possible relationships between cells. Developing computational methods that can accurately classify cell types while efficiently handling large datasets remains an active area of research.

Researchers at Nanjing University have developed a new computational framework called scGMB that combines graph convolutional networks with a deep learning architecture known as Mamba to improve cell classification from single-cell RNA sequencing data.

The architecture of scGMB.

The method begins by constructing a graph that represents relationships between cells. Instead of analyzing each cell independently, graph convolutional networks use these connections to learn how similar cells relate to one another. This allows the model to capture biological patterns that may not be obvious when examining gene expression alone.

The second component of the framework uses the Mamba architecture, a recently developed selective state space model designed to efficiently analyze large, high dimensional datasets. Because single-cell RNA sequencing data are often sparse, meaning many genes have little or no measurable expression in individual cells, Mamba helps extract meaningful biological features while requiring fewer computational resources than many conventional deep learning models.

The researchers integrated both components into a single end to end framework, allowing the model to simultaneously learn relationships between cells and patterns within gene expression data. This design improves computational efficiency while making the classification process easier to interpret.

To evaluate the method, the team tested scGMB on five widely used benchmark datasets, including Zheng68K, Zhengsorted, BaronHuman, BaronMouse, and AMB. Across all datasets, scGMB achieved classification accuracies ranging from 72.2% to 99.5%, consistently outperforming several existing cell classification methods.

Accurate cell type identification is essential for many areas of biomedical research, including cancer biology, developmental biology, and immunology. By improving classification performance, tools such as scGMB can help researchers better understand cellular diversity and discover previously unrecognized cell populations.

The authors also suggest that future versions of scGMB could be extended beyond single-cell RNA sequencing to analyze spatial transcriptomics and multiomics datasets. As artificial intelligence continues to advance, integrated deep learning methods like scGMB may become increasingly valuable for extracting biological insights from complex sequencing data.

Availability – The experimental and computational data that support the current study are available in a public repository (https://doi.org/10.5281/zenodo.3357167).