A 5-miRNA blood signature for gastric cancer diagnosis and therapeutic targeting derived from tumor pathways and validated by machine learning and single-cell RNA-seq

<title>Abstract</title> <p>Background Gastric cancer (GC) remains a leading cause of global cancer mortality, with late-stage diagnosis contributing significantly to poor patient outcomes. Circulating microRNAs (miRNAs) offer promise due to their stability in biofluids and established roles in carcinogenesis. However, existing miRNA biomarker candidates for GC suffer from inconsistent validation, limited specificity, and insufficient mechanistic links to gastric tumor biology. Establishing such mechanistic links for a miRNA signature would not only enhance its diagnostic value but also position it as a potential therapeutic target. Methods Differentially expressed genes (DEGs) and miRNAs (DEMs) were identified from tissue mRNA (GSE54129, GSE113255) and blood miRNA/mRNA datasets (GSE106817, GSE174302). We first identified shared dysregulated Reactome pathways between tissue DEGs and the target genes of blood-derived DEMs. miRNAs associated with these gastric cancer-specific pathways were selected as candidates. An ensemble of ten machine learning (ML) feature selection methods was then used to refine a minimal miRNA panel, which was subsequently validated using Random Forest and Naïve Bayes classifiers in independent dataset. To further validate them and explore their cellular origin, we performed single-cell RNA sequencing (scRNA-seq) on GC tissue. Additionally, a drug repurposing analysis was performed using the DGIdb database. Results A 5-miRNA panel (miR-124-3p, miR-23a-3p, miR-22-3p, miR-29b-3p, and miR-92a-3p) was identified. It demonstrated high diagnostic accuracy in the discovery set (AUC = 98.50%) and maintained robust performance upon external validation (AUC = 95.30%). scRNA-seq analysis confirmed the expression of target genes for a core 3-miRNA subset within key tumor microenvironment cell types (fibroblasts, endothelial, and epithelial cells). Furthermore, drug repurposing analysis identified four FDA-approved drugs (Olaparib, Doxorubicin Hydrochloride, Bithionol, and Daunorubicin Liposomal) as potential therapeutic candidates interacting with the signature's target genes. Conclusion This integrated multi-level transcriptomics and ML pipeline successfully discovered a high-performance, blood-based miRNA signature for GC. The signature is validated through scRNA-seq and ML, and linked to candidate therapeutics, offering a promising strategy for non-invasive diagnosis and therapeutic targeting.</p>