Högskolan i Skövde

Glioma is one of the most aggressive central nervous system malignancies; yet diagnosis still depends on invasive neuroimaging and biopsy procedures. Plasma-based protein biomarkers offer a minimally invasive alternative, but identifying a reliable panel from high-dimensional proteomic data remains challenging. This study applied differential expression analysis, machine learning (ML)- based feature selection, and comparative classifier evaluation to identify a plasma protein panel capable of distinguishing glioma from other cancer types in a pan-cancer proteomics dataset (n = 1,477 cancer patients, 1,463 plasma proteins). Differential expression analysis identified 227 upregulated and 685 downregulated proteins in glioma. Unsupervised dimensionality reduction showed partial global separability, while supervised approaches revealed locally coherent proteomic structure. Gene ontology (GO) enrichment analysis identified catabolic and metabolic processes as the dominant enriched biological processes, reflecting tumor-driven metabolic reprogramming and extracellular matrix (ECM) catabolism. Consensus feature selection across Elastic Net (EN), Random Forest (RF), and Support Vector Machine (SVM), ranked GFAP (mean rank = 2.33), MMP12 (5.0), and ITIH3 (26.0) as the top discriminative proteins. Classification on an independent test set (n = 442) using the top 30 consensus proteins yielded AUC = 0.942 for linear discriminant analysis (LDA) (sensitivity = 0.860, specificity = 0.880), AUC = 0.937 for EN (sensitivity = 0.767, specificity = 0.940), and AUC = 0.925 for RF (sensitivity = 0.837, specificity = 0.960). Repeated cross-validation confirmed that most of the discriminative information was captured within 3-10 proteins and revealed systematic interactions between the feature selection method and downstream classifier performance. The three-protein panel (GFAP, MMP12, ITIH3) achieved AUC > 0.92. Formal statistical comparison through one-way ANOVA with Tukey post-hoc test confirmed that EN achieved the most stable classification performance, RF and LDA performed at equivalent levels, while k-nearest neighbours (kNN) was significantly inferior to all four. The results show that feature extraction method and classifier choice must be considered jointly in plasma proteomics-based classification pipelines.