Locoregional Radiogenomic Models Capture Gene Expression Heterogeneity in Glioblastoma

Adrien Depeursinge, Tünde Szilágyi, Yan Liu, Kázmér Kovács, Reena Thomas, Kristen W. Yeom, Nancy J. Fischbein, Daniel L. Rubin, Michael, Olivier GevaertbioRxiv 2018 · DOI

TL;DR

This paper introduces a novel radiogenomic approach to map gene expression at a locoregional level within glioblastoma tumors, capturing their inherent heterogeneity. By combining multi-omics data with patch-based MRI analysis, they create detailed "heatmaps" that visualize where specific gene modules are active, paving the way for more personalized cancer treatment and monitoring.

What Problem Does This Paper Address?

Glioblastoma (GBM), a highly aggressive brain cancer, is notoriously heterogeneous. This means that even within a single tumor, different regions can have distinct molecular profiles. Traditional radiogenomics often treats tumors as a single entity, averaging out this crucial local variation. However, understanding this intratumoral heterogeneity is key to developing effective treatments, as different parts of the tumor might respond differently to therapies. The challenge is to non-invasively capture this detailed, spatially resolved molecular information from medical images.

Methods

The researchers developed a three-step pipeline:

  1. Multi-Omics Data Fusion: They used a tool called AMARETTO to integrate gene expression, DNA methylation, and DNA copy number data from GBM tumors. This process generated 100 "gene modules" – groups of genes that are co-expressed and likely involved in specific biological pathways.
  2. Locoregional Image Analysis: Instead of looking at the whole tumor, they divided contrast-enhanced T1-weighted MRI scans into small, overlapping circular "patches." For each patch, they extracted intensity features. These patches were categorized as either "core" (deep within the tumor) or "boundary" (at the tumor's edge). Generalized Linear Models (GLMs) were then used to link the expression levels of the gene modules to the imaging features within these local patches.
  3. Expression Heatmaps: The GLM's predictions were used to create "heatmaps" overlaid on the MRI scans. These heatmaps visually highlight regions within the tumor where specific gene modules are predicted to be over- or under-expressed, effectively visualizing the tumor's molecular landscape.

Datasets

The study utilized two publicly available datasets:

  • The Cancer Genome Atlas (TCGA) glioblastoma dataset: Used for developing and training the radiogenomic models.
  • REpository of Molecular BRAin Neoplasia DaTa (REMBRANDT) dataset: Used for validating the developed models.

Key Insights and Ideas

  • Local Matters: The study strongly suggests that focusing on locoregional gene expression, rather than a global average, is crucial for understanding GBM and developing targeted therapies.
  • Visualizing Molecular Heterogeneity: The developed heatmaps offer a powerful, intuitive way to visualize where specific molecular pathways are active within a tumor, directly on the patient's MRI.
  • Intensity Features are Key: While texture features were explored, the study found that simple intensity-based features from MRI patches were more effective for predicting gene module expression.
  • Potential for Precision Medicine: By linking imaging patterns to specific gene modules, this approach could guide treatment decisions, predict treatment response, and monitor tumor evolution.

Key Results

  • The models achieved Area Under the Receiver Operating Characteristic Curve (AUC) values above 0.65 for several gene modules across different validation strategies, indicating good predictive performance.
  • Modules related to key GBM pathways like neural differentiation, cell cycle, and receptor tyrosine kinase signaling showed good predictive performance.
  • For example, Module 32 (neural differentiation) and Module 45 (ion transport) showed distinct heterogeneity patterns on heatmaps, with Module 45's under-expression linked to specific MRI signal characteristics.
  • Module 84, linked to the gene BCAN (involved in tumor growth and motility), was best predicted using patches from the tumor boundaries, suggesting its importance in tumor invasion.
  • Two potential drug targets, TNFRSF1A and ERBB3, were identified as regulators of modules that could be predicted from imaging.

Limitations and Future Work

  • Texture Features: The study found that texture features did not effectively capture local heterogeneity, which they attribute to the small size of the image patches used. Previous studies using larger regions or different texture analysis methods might explain discrepancies.
  • Model Generalizability: While validation was performed on an external dataset, further validation across larger and more diverse cohorts would be beneficial.
  • Clinical Translation: The authors envision future work focusing on using these radiogenomic maps for treatment allocation, monitoring treatment response, and guiding biopsies for more precise molecular profiling. The heatmaps could also guide future sampling points for validation.

Summary updated

Abstract

Radiogenomics mapping noninvasively determines important relationships between the molecular genotype and imaging phenotype of various tumors, allowing advances in both clinical care and cancer research. While early work has shown its technical feasibility, here we extend radiogenomic mapping to a locoregional level that can account for the molecular heterogeneity of tumors. To achieve this, our data processing pipeline relies on three main steps: 1) the use of multi-omics data fusion to generate a set of 100 interpretable gene modules, 2) the use of patch-based image analysis (specifically of contrast-enhanced T1-weighted weighted MR images) combined with Generalized Linear Models (GLM) to establish potential links between module expressions and local MR signal, and 3) the use of expression heatmaps based on GLMs decision values to explore visualization of tumor molecular heterogeneity. The performance of the proposed approach was evaluated using a leave-one-patient-out crossvalidation method as well as a separate validation data set. The top performing models were based on a small set of 20 features and yielded Area Under the receiver operating characteristic Curve (AUC) above 0.65 on the validation cohort for eight modules. Next, we demonstrate the clinical and biological interpretation of four modules using molecular expression heatmaps superimposed on clinical radiographic images, showing the potential for assessing tumor molecular heterogeneity and the utility of this method for precision treatment in clinical decision making and imaging surveillance.
BibTeX
@misc{depeursinge2018locoregional,
  author = {Adrien Depeursinge and Tünde Szilágyi and Yan Liu and Kázmér Kovács and Reena Thomas and Kristen W. Yeom and Nancy J. Fischbein and Daniel L. Rubin and Michael and Olivier Gevaert},
  title = {{Locoregional Radiogenomic Models Capture Gene Expression Heterogeneity in Glioblastoma}},
  year = {2018},
  doi = {10.1101/304105},
  howpublished = {bioRxiv},
}
Figure