Single-cell classification using graph neural network with cell and gene vectors

Abstract

Single-cell data holds significant importance in capturing subtle differences between individual cells and gaining an understanding of specific functional states. Recent studies using single-cell data have been conducted in various areas, such as cancer cell treatment, cell differentiation tracking, and the development of patient-specific drugs. In single-cell analysis, classification is an important process for downstream analysis as it identifies cell characteristics and marker genes. However, single-cell analysis still faces multiple challenges, such as sparsity and complex differential patterns of single- cell data. In this study, we introduce scGraphSor2vec, which classifies cell types using human and mouse single-cell RNA sequencing data. This study aims to accurately understand the characteristics of single-cell data and develop a model that functions even with unseen external data. The human dataset, comprising 10 tissues and 176,594 cells, underwent additional performance evaluations using an external dataset. Similarly, the mouse dataset, including 12 tissues and 112,844 cells, had its performance appraised using an external dataset. We incorporated the cor2vec methodology, capturing correlation information between cells or genes, into the existing benchmarked weighted graph sample and aggregation (GraphSAGE) model. This method used the relationships and interactions within the network, enhancing the predictive accuracy and biological relevance of the cell node classifications. We compared actual and predicted cell types using accuracy and F1 scores and verified clustering effectiveness using uniform manifold approximation and projection (UMAP) visualization. A slight but notable improvement in performance was observed in both internal and external datasets. UMAP clustering also confirmed effective clustering in both internal and external datasets.