A machine learning-assisted analytical approach for the prediction of non-small cell lung cancer using UHPLC-HILIC-FLD serum N-glycan profiles.

The lack of robust early-stage molecular biomarkers severely limits the timely diagnosis of non-small cell lung cancer (NSCLC). To address this analytical challenge, we developed an integrated workflow combining serum N-glycan profiling with machine learning (ML) algorithms to explore the discriminatory patterns of serum N-glycan profiles between NSCLC and control groups and across disease stages. N-Glycans were analyzed using ultra-high-performance liquid chromatography with hydrophilic interaction chromatography and fluorescence detection (UHPLC-HILIC-FLD). The resulting high-dimensional chromatographic data from patient and control cohorts were subjected to statistical and ML analyses. Serum glycan profiling revealed marked dysregulation of the glycome. Twenty of the 28 N-glycan peaks showed significant cancer-associated alterations, with selected features also exhibiting stage-specific differences. Notably, relative quantitation showed that the agalactosylated core-fucosylated glycan G1 (H3N4F1) increased 2.55-fold in the cancer cohort. To evaluate the diagnostic utility of these glycomic profiles, nine ML models were trained and evaluated using 5-fold cross-validation. A neural network (NN) model demonstrated strong classification performance, achieving an area under the curve (AUC) of 0.93 for binary classification (NSCLC vs. healthy controls) and 0.91 for three-class classification. Furthermore, SHapley Additive exPlanations (SHAP) analysis elucidated the model decision-making process, identifying glycans G7 (H5N4), G23 (H6N5S3), and G2 (H3N5F1 and H5N2) as key features contributing to classification. These findings indicate that integrating UHPLC-HILIC-FLD serum glycomic profiling with explainable machine learning provides an exploratory analytical framework for identifying candidate glycan features associated with NSCLC status and stage.