Background: Hepatocellular carcinoma (HCC) pathogenesis involves the alteration of multiple liver-specific metabolic pathways. We systematically profiled cancer- and liver-related classes of metabolites in HCC and adjacent liver tissues and applied supervised machine learning to compare their potential yield for HCC biomarkers.
Methods: Tumor and corresponding liver tissue samples were profiled as follows: Bile acids by ultra-performance liquid chromatography (LC) coupled to tandem mass spectrometry (MS), phospholipids by LC-MS/MS, and other small molecules including free fatty acids by gas chromatography-time of flight MS. The overall classification performance of metabolomic signatures derived by support vector machine (SVM) and random forests machine learning algorithms was then compared across classes of metabolite.
Results: For each metabolite class, there was a plateau in classification performance with signatures of 10 metabolites. Phospholipid signatures consistently showed the highest discrimination for HCC followed by signatures derived from small molecules, free fatty acids, and bile acids with area under the receiver operating characteristic curve (AUC) values of 0.963, 0.934, 0.895, 0.695, respectively, for SVM-generated signatures comprised of 10 metabolites. Similar classification performance patterns were observed with signatures derived by random forests.
Conclusion: Membrane phospholipids are a promising source of tissue biomarkers for discriminating between HCC tumor and liver tissue.
Keywords: diagnosis; hepatocellular carcinoma; machine learning; metabolomics; molecular imaging; phospholipids; positron emission tomography.