Serum Metabolic Profiling of OSCC
Collected 819 serum samples consisting of 241 HC and 578 OSCC were divided into a development cohort and a validation cohort for serum marker discovery and confirmation. There were 367 ions selected to characterize the global metabolic profiles of HC and OSCC. The average mass spectra of OSCC and HC are shown in Fig. 1A . From the average mass spectra, it can be clearly observed that the peaks atm/z 135.0029 (lactic acid, [M+2Na-H]+), 203.0527 (glucose, [M+Na]+), 304.2611 (oleamide, [M+Na]+) were elevated in the OSCC group compared to the HC group. More discriminative fingerprint peaks had to be found by statistical tests, which are described later.
To visualize the difference between HC and OSCC metabolite patterns, an unsupervised machine learning method, t-stochastic neighbor embedding (t-SNE), was introduced to reduce the high-dimensional metabolite ions information into a three-dimensional (3D) feature space. In the constructed 3D feature space, serum cases from the same group were well clustered whereas those cases from different groups can be separated (Fig. 1B ). This result demonstrated there exists a substantial difference in serum metabolic profiles that can be used for OSCC and HC prediction.