Understanding molecular features that facilitate aggressive phenotypes in glioblastoma multiforme (GBM) remains a major clinical challenge. Accurate diagnosis of GBM subtypes, namely classical, proneural, and mesenchymal, and identification of specific molecular features are crucial for clinicians for systematic treatment. We develop a biologically interpretable and highly efficient deep learning framework based on a convolutional neural network for subtype identification. The classifiers were generated from high-throughput data of different molecular levels, i.e., transcriptome and methylome. Furthermore, an integrated subsystem of transcriptome and methylome data was also used to build the biologically relevant model. Our results show that deep learning model outperforms the traditional machine learning algorithms. Furthermore, to evaluate the biological and clinical applicability of the classification, we performed weighted gene correlation network analysis, gene set enrichment, and survival analysis of the feature genes. We identified the genotype–phenotype relationship of GBM subtypes and the subtype-specific predictive biomarkers for potential diagnosis and treatment.

Classifying lower-grade gliomas (LGGs) is a crucial step for accurate therapeutic intervention. The histopathological classification of various subtypes of LGG, including astrocytoma, oligodendroglioma and oligoastrocytoma, suffers from intraobserver and interobserver variability leading to inaccurate classification and greater risk to patient health. We designed an efficient machine learning-based classification framework to diagnose LGG subtypes and grades using transcriptome data. First, we developed an integrated feature selection method based on correlation and support vector machine (SVM) recursive feature elimination. Then, implementation of the SVM classifier achieved superior accuracy compared with other machine learning frameworks. Most importantly, we found that the accuracy of subtype classification is always high (>90%) in a specific grade rather than in mixed grade (~80%) cancer. Differential co-expression analysis revealed higher heterogeneity in mixed grade cancer, resulting in reduced prediction accuracy. Our findings suggest that it is necessary to identify cancer grades and subtypes to attain a higher classification accuracy. Our six-class classification model efficiently predicts the grades and subtypes with an average accuracy of 91% (±0.02). Furthermore, we identify several predictive biomarkers using co-expression, gene set enrichment and survival analysis, indicating our framework is biologically interpretable and can potentially support the clinician.

No abstract