Deep learning identified glioblastoma subtypes based on internal genomic expression ranks

Area of Science:

• Computational biology
• Genomics
• Machine learning in oncology

Background:

• Glioblastoma (GBM) exhibits distinct genomic subtypes (Proneural, Neural, Classical, Mesenchymal).
• Standardizing diverse genomic expression profiles for GBM subtype classification is challenging.
• Manual classification of GBM samples into subtypes is difficult.

Purpose of the Study:

• To develop an algorithm for unifying heterogeneous GBM genomic profiles.
• To train and evaluate deep neural network (DNN) and convolutional DNN (CDNN) models using standardized data.
• To assess the accuracy and generalization capacity of CDNN models for GBM subtyping.

Main Methods:

• Developed a novel algorithm to transform GBM genomic profiles into a standardized normal distribution (SND) based on gene expression ranks.
• Trained DNN and CDNN models on both original and SND GBM datasets.
• Utilized expanded TCGA datasets to enhance CDNN model robustness and generalization.

Main Results:

• SND transformation preserved data distribution and internal gene expression ranks.
• CDNN models trained on SND data significantly outperformed models trained on primary data.
• CDNN models demonstrated high accuracy across various GBM datasets, including independent ones.
• The Neural (NE) subtype was consistently classified with lower accuracy, supporting its potential exclusion.

Conclusions:

• Unified SND data enables the training of highly accurate and generalizable CDNN models for GBM classification.
• The developed CDNN models offer a robust tool for GBM subtyping.
• Findings suggest the Neural (NE) subtype may be incompatible with the established 4-subtype classification system for glioblastoma.