Abstract
Cholangiocarcinoma (CCA) is a highly lethal disease because most patients are asymptomatic until they progress to advanced stages. Current CCA diagnosis relies on clinical imaging tests and tissue biopsy, while specific CCA biomarkers are still lacking. This study employed a translational proteomic approach for the discovery, validation, and development of a multiplex CCA biomarker assay. In the discovery phase, label-free proteomic quantitation was performed on nine pooled plasma specimens derived from nine CCA patients, nine disease controls (DC), and nine normal individuals. Seven proteins (S100A9, AACT, AFM, and TAOK3 from proteomic analysis, and NGAL, PSMA3, and AMBP from previous literature) were selected as the biomarker candidates. In the validation phase, enzyme-linked immunosorbent assays (ELISAs) were applied to measure the plasma levels of the seven candidate proteins from 63 participants: 26 CCA patients, 17 DC, and 20 normal individuals. Four proteins, S100A9, AACT, NGAL, and PSMA3, were significantly increased in the CCA group. To generate the multiplex biomarker assays, nine machine learning models were trained on the plasma dynamics of all seven candidates (All-7 panel) or the four significant markers (Sig-4 panel) from 45 of the 63 participants (70%). The best-performing models were tested on the unseen values from the remaining 18 (30%) of the 63 participants. Very strong predictive performances for CCA diagnosis were obtained from the All-7 panel using a support vector machine with linear classification (AUC = 0.96; 95% CI 0.88-1.00) and the Sig-4 panel using partial least square analysis (AUC = 0.94; 95% CI 0.82-1.00). This study supports the use of the composite plasma biomarkers measured by clinically compatible ELISAs coupled with machine learning models to identify individuals at risk of CCA. The All-7 and Sig-4 assays for CCA diagnosis should be further validated in an independent prospective blinded clinical study.