Abstract
Identifying novel biomarkers for early detection of lung cancer is crucial. Non-invasively available saliva is an ideal biofluid for biomarker exploration; however, the rationale underlying biomarker detection from organs distal to the oral cavity in saliva requires clarification. Therefore, we analyzed metabolomic profiles of cancer tissues compared with those of adjacent non-cancerous tissues, as well as plasma and saliva samples collected from patients with lung cancer (n = 109 pairs). Additionally, we analyzed plasma and saliva samples collected from control participants (n = 83 and 71, respectively). Capillary electrophoresis-mass spectrometry and liquid chromatography-mass spectrometry were performed to comprehensively quantify hydrophilic metabolites. Paired tissues were compared, revealing 53 significantly different metabolites. Plasma and saliva showed 44 and 40 significantly different metabolites, respectively, between patients and controls. Of these, 12 metabolites exhibited significant differences in all three comparisons and primarily belonged to the polyamine and amino acid pathways; N(1)-acetylspermidine exhibited the highest discrimination ability. A combination of 12 salivary metabolites was evaluated using a machine learning method to differentiate patients with lung cancer from controls. Salivary data were randomly split into training and validation datasets. Areas under the receiver operating characteristic curve were 0.744 for cross-validation using training data and 0.792 for validation data. This model exhibited a higher discrimination ability for N(1)-acetylspermidine than that for other metabolites. The probability of lung cancer calculated using this model was independent of most patient characteristics. These results suggest that consistently different salivary biomarkers in both plasma and lung tissues might facilitate non-invasive lung cancer screening.