RADIOMAP-IvyGAP

Subcompartment-Level Radiomic Features Associate with Regional Transcriptomic Programs in Glioblastoma: An Exploratory Analysis of the Ivy Glioblastoma Atlas Project

Daniele Piccolo, Marco Vindigni

Overview

This repository contains the analysis code for an exploratory study linking MRI-derived radiomic features from the IVYGAP-RADIOMICS dataset (Pati et al., 2020) with zone-specific transcriptomic pathway enrichment scores from the IvyGAP RNA-seq atlas (Puchalski et al., 2018).

The study tests whether radiomic features extracted from BraTS-style MRI subcompartments (enhancing tumor, non-enhancing tumor, peritumoral edema) associate with transcriptomic programs computed via single-sample Gene Set Enrichment Analysis (ssGSEA) using 24 gene sets across 28 matched patients.

Repository Structure

RADIOMAP-IvyGAP/
├── R/
│   ├── 00_download_and_audit.R        # Data download, patient matching, GO/NO-GO gate
│   ├── 01_ssgsea_computation.R        # 24 ssGSEA pathway enrichment scores
│   ├── 02_prepare_analysis_data.R     # Zone-to-subcompartment aggregation, feature merge
│   ├── 03_statistical_analysis.R      # Mixed-effects models, Elastic Net, permutation tests
│   └── 03b_nested_cv_analysis.R       # Nested LOPO-CV (primary predictive analysis)
├── results/                           # Pre-computed output tables and model objects
├── README.md
├── .gitignore
└── LICENSE

Requirements

R (>= 4.3)

Packages are installed automatically by the scripts if missing:

• Bioconductor: GSVA, ivygapSE, SummarizedExperiment, msigdbr, org.Hs.eg.db, AnnotationDbi
• CRAN: lme4, lmerTest, performance, glmnet, pmsampsize, caret, dplyr, tidyr, broom.mixed, ggplot2, patchwork, readxl

Reproduction

Scripts must be run sequentially from the repository root:

cd RADIOMAP-IvyGAP

# Step 1: Download data and match patients (~5 min, downloads ~500 MB)
Rscript R/00_download_and_audit.R

# Step 2: Compute ssGSEA enrichment scores (~10 min)
Rscript R/01_ssgsea_computation.R

# Step 3: Zone-to-subcompartment aggregation and feature merge (~1 min)
Rscript R/02_prepare_analysis_data.R

# Step 4: Mixed-effects models, Elastic Net, permutation tests (~30 min)
Rscript R/03_statistical_analysis.R

# Step 5: Nested cross-validation with internal feature selection (~3-4 hours)
#         Includes 1000-permutation test for each significant pathway
Rscript R/03b_nested_cv_analysis.R

Notes:

• Script 00 downloads both datasets automatically (IvyGAP via Bioconductor, IVYGAP-RADIOMICS from TCIA). If TCIA download fails, manual download instructions are provided.
• Script 03 saves a session file (results/03_analysis_session.RData, ~100 MB) that is required by script 03b.
• Script 03b runs 1000 nested permutations per significant pathway; total runtime is approximately 3-4 hours.
• The results/ directory contains pre-computed outputs for reference. Re-running the full pipeline will regenerate these files.

Data Sources

• IvyGAP RNA-seq: Allen Institute for Brain Science — https://glioblastoma.alleninstitute.org/
• IVYGAP-RADIOMICS: The Cancer Imaging Archive — https://doi.org/10.7937/9j41-7d44

Both datasets are publicly available. No IRB approval was required.

Key Results

Inflammatory Response was the only pathway supported by both analytical frameworks: nested LOPO-CV (R²_cv = 0.185, 95% CI [0.071, 0.355], permutation p = 0.008, BH-FDR = 0.096 across 24 pathways) and the exploratory linear mixed-effects model (FDR = 0.024, ΔR²_m = 0.214 beyond subcompartment effects). T2-derived texture features were selected in 100% of LOPO folds.

Angiogenesis (R²_cv = 0.209, 95% CI [0.028, 0.353], permutation p = 0.006, nested FDR = 0.096) reached significance in the nested CV but was not corroborated by the LMM (FDR = 0.445). It is reported as a tentative single-framework signal requiring independent validation.

The IvyGAP CTpan module (R²_cv = 0.133) is derived from the same transcriptomic atlas that provides the outcome data, did not survive the 24-pathway FDR correction (FDR = 0.104), and is retained only as an internal-consistency check — not as discovery.

21 of 24 pathways showed no predictive signal (R²_cv ≤ 0). At N = 28 (vs. N ≈ 240 required by Riley 2020 criteria for P = 5, R² = 0.20, shrinkage ≥ 0.90), the absence of signal does not constitute evidence of biological inaccessibility.

Citation

Piccolo D, Vindigni M. Radiomic Features of MRI Subcompartments Associate with Angiogenic and Inflammatory Transcriptomic Programs in Glioblastoma: An IvyGAP Exploratory Analysis. Cancers (2026), accepted for publication. Full bibliographic reference will be added once the article is assigned a DOI.

License

This project is licensed under the MIT License. See LICENSE for details.