Screening, Validation, and Machine Learning-Based Evaluation of Serum Protein Biomarkers for Esophageal Squamous Cell Carcinoma Based on Single-Cell Subtype-Specific Genes

Abstract

Cellular heterogeneity of epithelial cells and fibroblasts is critical in esophageal squamous cell carcinoma development (ESCC). Identifying dysregulated subtype-specific genes in these cells is essential for early diagnosis and treatment. In this study, our pipeline integrated scRNA-seq, proteomics, and ELISA to screen biomarkers: scRNA-seq defined epithelial and fibroblast subtypes and their markers, while proteomics and secretory profiling identified dysregulated secretory proteins. Serum levels of five selected proteins were measured in 344 ESCC patients, 46 HGIN cases, and 390 normal controls. Machine learning was employed to construct diagnostic models. An interactive web tool was implemented in R Shiny. Six epithelial and four fibroblast subtypes, proportionally distinct between ESCC and normal tissues, were identified. Four validated dysregulated proteins were used to build diagnostic models; among 12 algorithms, the Support Vector Machine (SVM) achieved the best performance with AUCs of 0.829 and 0.767 in the training and validation sets, respectively (p > 0.05). The model effectively distinguished early- and late-stage ESCC and HGIN from normal controls. The web-based diagnostic tool is publicly available at https://zhangxz.shinyapps.io/P4_Pred/. The identified serum biomarkers may enhance early ESCC detection and diagnosis. Our pipeline, leveraging heterogeneity-related genes in fibroblasts and epithelial cells, is readily adaptable to other tumors.