Identification of Glioma Phenotypic Subtypes From Multimodal MRI Data Using Hierarchical Multi-Kernel Learning

Area of Science:

• Neuro-oncology
• Medical imaging
• Computational biology

Background:

• Gliomas are primary brain tumors with significant variability.
• Non-invasive identification of glioma subtypes can enhance patient management.

Purpose of the Study:

• To develop a non-invasive method for identifying glioma phenotypic subtypes.
• To analyze distinct signaling pathways associated with these subtypes.
• To predict isocitrate dehydrogenase (IDH) genotype using machine learning.

Main Methods:

• Hierarchical multi-kernel learning applied to MRI data (T1CE, T2FLAIR).
• Validation using an independent TCGA/TCIA dataset.
• Analysis of pathway activity and prediction of IDH genotype using five machine learning models, including GA-KPLS.

Main Results:

• Identified two glioma phenotypic subtypes: high-risk and low-risk, with significant survival differences (p < 0.05).
• High-risk group showed activated JAK-STAT and TGF-β pathways; low-risk group showed activated Hypoxia and p53 pathways.
• GA-KPLS model achieved the highest IDH genotype prediction performance (AUC = 0.819).

Conclusions:

• The study presents a non-invasive approach for glioma subtype identification.
• Distinct signaling pathways are associated with identified glioma subtypes.
• This method can define therapeutically homogeneous subgroups for targeted glioma therapy.