A Transcriptomics-Based Meta-Analysis Combined With Machine Learning Identifies a Secretory Biomarker Panel for Diagnosis of Pancreatic Adenocarcinoma

Pancreatic ductal adenocarcinoma (PDAC) is generally incurable due to the late diagnosis and absence of markers that are concordant with expression in several sample sources (i.e., tissue, blood, plasma) and platforms (i.e., Microarray, sequencing). We optimized meta-analysis of 19 PDAC (tissue and blood) transcriptome studies from multiple platforms. The key biomarkers for PDAC diagnosis with secretory potential were identified and validated in different cohorts. Machine learning approach i.e., support vector machine supported by leave-one-out cross-validation was used to build and test the classifier. We identified a 9-gene panel (IFI27, ITGB5, CTSD, EFNA4, GGH, PLBD1, HTATIP2, IL1R2, CTSA) that achieved ∼0.92 average sensitivity and ∼0.90 average specificity in distinguishing PDAC from healthy samples in five training sets using cross-validation. These markers were also validated in proteomics and single-cell transcriptomics studies suggesting their prognostic role in the diagnosis of PDAC. Our 9-gene classifier can not only clearly discriminate between better and poor survivors but can also precisely discriminate PDAC from chronic pancreatitis (AUC = 0.95), early stages of progression [Stage I and II (AUC = 0.82), IPMA and IPMN (AUC = 1), and IPMC (AUC = 0.81)]. The 9-gene marker outperformed the previously known markers in blood studies particularly (AUC = 0.84). The discrimination of PDAC from early precursor lesions in non-malignant tissue (AUC > 0.81) and peripheral blood (AUC > 0.80) may assist in an early diagnosis of PDAC in blood samples and thus will also facilitate risk stratification upon validation in clinical trials.