Integration of Multi-omics Data Based on Deep Learning for Subtyping of Low-Grade Glioma.

Low-grade gliomas (LGGs) represent a complex and aggressive category of brain tumors. Despite recent advancements in molecular subtyping and characterization, the necessity to identify additional molecular subtypes and biomarkers remains. To delineate survival subtypes in LGG, we propose a deep learning (DL)-based multi-omics SurvivalNet (MOST) model. By integrating histological RNA-seq, miRNA-seq, and DNA methylation data obtained from The Cancer Genome Atlas (TCGA), we applied the MOST model to analyze data from 497 LGG patients. We employed consensus clustering to reveal heterogeneous subtypes, validated our findings using an internal validation set through a supervised classification algorithm, and further evaluated the robustness of our model in an independent external cohort. The DL-based MOST model identified two optimal patient subtypes with significant differences in survival (P = 3.07E - 16) and demonstrated a robust model fit (C = 0.92 ± 0.02). This multi-omics model was validated using external Chinese Glioma Genome Atlas (CCGA) datasets, including RNA-Seq (N = 497, C = 0.85), miRNA array (N = 89, C = 0.80), and DNA methylation (N = 89, C = 0.61). High-risk subcategories exhibited increased expression of the homeobox (HOX) family genes, regulation of cholesterol homeostasis, glycolysis, epithelial-mesenchymal transition pathway enrichment, and a high density of M2 macrophages. Our study utilized deep learning to identify multi-omics features associated with differential survival outcomes in patients with LGG. This work is anticipated to significantly enhance prognosis prediction for LGG due to its robustness within the cohorts.