Predictive Radiomics for Evaluation of Cancer Immune SignaturE in Glioblastoma: the PRECISE-GBM study

Technical Summary: Predictive Radiomics for Evaluation of Cancer Immune SignaturE in Glioblastoma (PRECISE-GBM)

Problem Statement
The immune tumor microenvironment (TME) in IDH-wildtype (IDH-wt) glioblastoma is characterized by significant heterogeneity and an immunosuppressive state, which poses challenges for the efficacy of immunotherapy. Current strategies for identifying immune biomarkers rely on invasive tissue biopsies, histopathology, and immunohistochemistry. These methods provide detailed molecular insights but are limited by spatial heterogeneity, the inability to sample deep-seated or inoperable lesions, and the lack of non-invasive tools for pre-operative patient stratification. The authors hypothesize that pre-operative MRI images contain radiological biomarkers capable of predicting the tumor's immune status, specifically immune cell signatures derived from transcriptomic data, thereby offering a non-invasive avenue for clinical trial enrichment.

Methodology
The PRECISE-GBM study is a retrospective, multicenter biomarker development and analytical validation study utilizing the largest known open-access dataset of matched pre-operative MRI and transcriptomic data for IDH-wt glioblastoma.

• Data Curation and Harmonization: The study integrated data from five public repositories: TCGA-GBM, CPTAC, IvyGAP, REMBRANDT, and CGGA, comprising 176 patients. To address missing FLAIR sequences in the CGGA dataset, synthetic FLAIR images were generated using a pix2pix-based Generative Adversarial Network (GAN); these synthetic images were restricted to training data to ensure holdout test integrity. Transcriptomic data (RNA-seq and microarray) were harmonized using ComBat to remove batch effects.
• Immune Label Construction: Digital cytometry (CIBERSORTx) was applied to bulk transcriptomic data using two reference signature matrices: a pan-cancer matrix and a glioblastoma-specific matrix. This process estimated scores for 17 immune cell categories (including T cell subsets, macrophage subtypes M0/M1/M2, microglia, NK cells, and dendritic cells). These continuous scores were binarized into "high" (1) or "low" (0) classes using a Gaussian Mixture Model (GMM).
• Radiomic Feature Extraction: Deep learning-based auto-segmentation (DeepBraTumIA) delineated three tumor regions: necrotic core, enhancing tumor, and edema. Radiomic features (shape, first-order, and higher-order texture features from GLCM, GLSZM, GLDM, NGTDM, and GLRLM) were extracted from T1, T1CE, T2, and FLAIR sequences.
• Model Architecture and Validation:
  • Preprocessing: Outliers were removed via Principal Component Analysis (PCA) and Hotelling's T2 ellipse. Radiomic features were harmonized using neuroCombat and normalized via Z-score.
  • Feature Selection: Least Absolute Shrinkage and Selection Operator (LASSO) with nested 5-fold cross-validation was used to select discriminative features for each immune label.
  • Classification: Two model types were trained: Support Vector Machines (SVM) and an Ensemble model (SVM weighted 2:1:1 with Random Forest and Histogram-based Gradient Boosting).
  • Validation Strategy: Three independent cross-cohort experiments were conducted. In each, models were trained on a combination of datasets and validated on a strictly held-out external dataset (IvyGAP, TCGA, or CPTAC) to assess generalizability. Performance was evaluated using precision, balanced accuracy, and Matthews Correlation Coefficient (MCC). Bootstrapped resampling (10,000 iterations) was used to compare model performance statistically.

Key Results

• Feature Selection: After selection, 2069 distinct radiomic features were identified across the 17 immune categories. Features were distributed relatively evenly across tumor regions (necrotic core, enhancing, edema) and MRI sequences.
• Model Performance:
  • Macrophage M0 Signature: The ensemble model using the pan-cancer signature matrix (ENS_pan) demonstrated the most stable performance for predicting the macrophage M0 subtype. Across the three independent holdout datasets, it achieved a mean balanced accuracy of 0.67 and a mean precision of 0.89. The Matthews Correlation Coefficient (MCC) was moderately positive (mean 0.22).
  • T Cell Signatures: SVM models showed promise for T cell-related categories (e.g., CD4, CD8, overall T cells) in specific holdout sets, achieving high balanced accuracy in some instances (up to 1.00), but these results were less stable across all three experiments compared to the M0 model.
  • Other Categories: Many other immune categories (e.g., Tregs, M1/M2 macrophages) resulted in "non-evaluable" or unstable metrics due to extreme class imbalance in the binarized holdout test sets, where one class was represented by a single case or absent entirely.
• Explainability: SHAP (Shapley Additive exPlanations) analysis for the M0 model indicated that predictions were driven by higher-order texture features (GLCM, GLSZM, GLDM) reflecting spatial heterogeneity and structural organization, rather than single dominant features.
• Prognostic Correlation: Patients predicted to have a "high" M0 signature demonstrated significantly longer overall survival compared to those with a "low" signature (log-rank p = 0.04).

Key Contributions

1. Dataset Curation: The study created and harmonized the largest open-access radiogenomic dataset for IDH-wt glioblastoma, integrating matched MRI and transcriptomic data from five major repositories.
2. Methodological Framework: The authors developed a rigorous pipeline for radiogenomic analysis that includes synthetic image generation for missing sequences, batch correction for both imaging and genomic data, and a robust cross-cohort validation strategy using three independent external holdout sets.
3. Biomarker Identification: The study identified the macrophage M0 immune signature as the only immune-related feature reliably predictable from pre-operative MRI using the developed pipeline.
4. Analytical Validation: Unlike previous studies that often rely on internal cross-validation, this study provided analytical validation across three distinct external cohorts, demonstrating the stability of the M0 predictive model.

Significance and Claims
The authors claim that the PRECISE-GBM study successfully developed and analytically validated a non-invasive MRI-based pipeline to interrogate glioblastoma immunobiology, specifically for the macrophage M0 subtype. The significance of this work lies in its potential to stratify patients for immunotherapy trials prior to biopsy, addressing the need for trial enrichment in a disease with high heterogeneity. The authors position these biomarkers as a complementary tool to tissue-based analysis, particularly for inoperable patients or deep-seated lesions.

The paper modestly frames these findings as a "proof-of-concept" and a "plausible pathway" for future clinical application. It explicitly states that the models require prospective clinical validation in "shadow mode" within neo-adjuvant immunotherapy trials before clinical deployment. The study emphasizes that while the M0 signature showed stability, the broader application to other immune cell types was limited by data distribution and class imbalance, necessitating larger, more balanced datasets for future refinement. The pipeline architecture and evaluation methods are presented as a template for future radiogenomic studies.