Detection of Human Cholangiocarcinoma Markers in Serum Using Infrared Spectroscopy

SIMPLE SUMMARY: Cholangiocarcinoma is a form of liver cancer that is found, predominantly, in Thailand. Due to the non-specific symptoms and laboratory investigation, it is difficult to rule out cholangiocarcinoma from other liver conditions. Here, we demonstrate the development of a diagnostic tool for cholangiocarcinoma, based on the ATR-FTIR analyses of sera, coupled with multivariate analyses and machine learning tools to obtain a better specificity. The innovative approach that shows highly promising results for this otherwise difficult to diagnose cancer. ABSTRACT: Cholangiocarcinoma (CCA) is a malignancy of the bile duct epithelium. Opisthorchis viverrini infection is a known high-risk factor for CCA and in found, predominantly, in Northeast Thailand. The silent disease development and ineffective diagnosis have led to late-stage detection and reduction in the survival rate. Attenuated total reflectance-Fourier transform infrared spectroscopy (ATR-FTIR) is currently being explored as a diagnostic tool in medicine. In this study, we apply ATR-FTIR to discriminate CCA sera from hepatocellular carcinoma (HCC), biliary disease (BD) and healthy donors using a multivariate analysis. Spectral markers differing from healthy ones are observed in the collagen band at 1284, 1339 and 1035 cm(−1), the phosphate band ([Formula: see text]) at 1073 cm(−1), the polysaccharides band at 1152 cm(−1) and 1747 cm(−1) of lipid ester carbonyl. A Principal Component Analysis (PCA) shows discrimination between CCA and healthy sera using the 1400–1000 cm(−1) region and the combined 1800—1700 + 1400–1000 cm(−1) region. Partial Least Square-Discriminant Analysis (PLS-DA) scores plots in four of five regions investigated, namely, the 1400–1000 cm(−1), 1800–1000 cm(−1), 3000–2800 + 1800–1000 cm(−1) and 1800–1700 + 1400–1000 cm(−1) regions, show discrimination between sera from CCA and healthy volunteers. It was not possible to separate CCA from HCC and BD by PCA and PLS-DA. CCA spectral modelling is established using the PLS-DA, Support Vector Machine (SVM), Random Forest (RF) and Neural Network (NN). The best model is the NN, which achieved a sensitivity of 80–100% and a specificity between 83 and 100% for CCA, depending on the spectral window used to model the spectra. This study demonstrates the potential of ATR-FTIR spectroscopy and spectral modelling as an additional tool to discriminate CCA from other conditions.