Attention-Enhanced Kolmogorov–Arnold Network for Interpretable Subtype Classification and Biomarker Discovery in Lower-Grade Glioma

Abstract

Precision molecular classification and valid biomedical marker identification are necessary for the optimization of personalized therapy in lower-grade glioma (LGG). In this study, we present a transparent deep learning model built on the Kolmogorov–Arnold Network (KAN) for effective LGG subtype classification and biomarker discovery. The method integrates multiple-omics (mRNA expression, DNA methylation and miRNA profiles) data from the TCGA-LGG sample set for comprehensive modelling of tumour heterogeneity. Before training the model, we performed variance-based feature screening, retaining molecular features and excluding class labels