Radiogenomics and machine learning predict oncogenic signaling pathways in glioblastoma

Area of Science:

• Neuro-oncology
• Radiology
• Computational Biology

Background:

• Glioblastoma (GBM) is an aggressive brain tumor with poor prognosis, necessitating novel therapies.
• Standard treatments are insufficient, and identifying genetic targets often requires invasive methods.
• Radiogenomics offers a non-invasive approach to link imaging data with genetic information.

Purpose of the Study:

• To explore the utility of radiogenomics and machine learning (ML) for non-invasively predicting oncogenic signaling pathways in GBM.
• To assess the association between MRI-derived radiomic features and key signaling pathways (RTK-RAS, PI3K, TP53, NOTCH).
• To inform personalized therapeutic strategies for GBM patients.

Main Methods:

• Utilized MRI scans (T1w, T1c, FLAIR, T2w) from the BRATS-19 dataset linked with TCGA/CPTAC genetic data.
• Extracted radiomic features using PyRadiomics and applied dimensionality reduction.
• Trained five ML models to predict signaling pathways, optimizing with Grid Search and 5-fold cross-validation.

Main Results:

• Demonstrated a positive association between most signaling pathways and radiomic features.
• Achieved high AUC scores for predicting RTK-RAS (0.7), PI3K (0.8), and TP53 (0.75) pathways.
• Indicated the potential of ML models to accurately predict oncogenic pathways from radiomic data.

Conclusions:

• Presented a novel radiogenomic approach for non-invasive prediction of signaling pathway deregulation in GBM.
• Highlighted the integration of radiomics and genomic data for better understanding GBM behavior and treatment response.
• Advanced precision medicine in GBM by enabling informed, personalized therapeutic decisions.