Integrated analysis of single-cell and bulk RNA-seq establishes a novel signature for prediction in gastric cancer

BACKGROUND: Single-cell sequencing technology provides the capability to analyze changes in specific cell types during the progression of disease. However, previous single-cell sequencing studies on gastric cancer (GC) have largely focused on immune cells and stromal cells, and further elucidation is required regarding the alterations that occur in gastric epithelial cells during the development of GC. AIM: To create a GC prediction model based on single-cell and bulk RNA sequencing (bulk RNA-seq) data. METHODS: In this study, we conducted a comprehensive analysis by integrating three single-cell RNA sequencing (scRNA-seq) datasets and ten bulk RNA-seq datasets. Our analysis mainly focused on determining cell proportions and identifying differentially expressed genes (DEGs). Specifically, we performed differential expression analysis among epithelial cells in GC tissues and normal gastric tissues (NAGs) and utilized both single-cell and bulk RNA-seq data to establish a prediction model for GC. We further validated the accuracy of the GC prediction model in bulk RNA-seq data. We also used Kaplan–Meier plots to verify the correlation between genes in the prediction model and the prognosis of GC. RESULTS: By analyzing scRNA-seq data from a total of 70707 cells from GC tissue, NAG, and chronic gastric tissue, 10 cell types were identified, and DEGs in GC and normal epithelial cells were screened. After determining the DEGs in GC and normal gastric samples identified by bulk RNA-seq data, a GC predictive classifier was constructed using the Least absolute shrinkage and selection operator (LASSO) and random forest methods. The LASSO classifier showed good performance in both validation and model verification using The Cancer Genome Atlas and Genotype-Tissue Expression (GTEx) datasets [area under the curve (AUC)_min = 0.988, AUC_1se = 0.994], and the random forest model also achieved good results with the validation set (AUC = 0.92). Genes TIMP1, PLOD3, CKS2, TYMP, TNFRSF10B, CPNE1, GDF15, BCAP31, and CLDN7 were identified to have high importance values in multiple GC predictive models, and KM-PLOTTER analysis showed their relevance to GC prognosis, suggesting their potential for use in GC diagnosis and treatment. CONCLUSION: A predictive classifier was established based on the analysis of RNA-seq data, and the genes in it are expected to serve as auxiliary markers in the clinical diagnosis of GC.