Abstract
Pancreatic ductal adenocarcinoma (PDAC) presents significant diagnostic challenges at early stages due to the absence of specific symptoms and the rapid progression of the disease. Consequently, there is an urgent imperative to investigate the mechanisms underlying early detection and the biological processes that drive cancer progression. In response to this need, we conducted a paired design study employing differential protein analysis, Mendelian randomization (MR), and single-cell analysis to identify distinctive features associated with early-stage PDAC (E-PDAC). Our initial analysis in the RJ cohort identified 1,068 E-PDAC-related proteins from differential protein analysis. Subsequently, we employed a random forest approach to pinpoint 25 E-PDAC-specific proteins. These proteins informed the development of 13 machine learning models aimed at predicting E-PDAC risk, which demonstrated an area under the curve (AUC) of approximately 0.9 in the discovery cohort and approximately 0.8 in external validation. Furthermore, MR and single-cell analysis were utilized to explore causal relationships and the composition of the tumor microenvironment. Through MR, we identified STX7 as a risk factor (odds ratio = 1.26; confidence interval: 1.03-1.54; p = 0.02), with LUM exhibiting a dual role (pro-tumorigenic in proteomic analysis but anti-tumorigenic in MR analysis). Single-cell analysis revealed that LUM primarily aids in the generation of fibroblasts and T/B cells, serving as pro-tumorigenic and antitumorigenic agents, respectively. Our research offers valuable insights into protein biomarkers, cell types, and communication in E-PDAC, suggesting potential targets to improve screening efficiency.