Discovering and interpreting transcriptomic drivers of imaging traits using neural networks

Area of Science:

• Oncology
• Bioinformatics
• Medical Imaging

Background:

• Cancer exhibits heterogeneity across biological levels, complicating clinical outcomes.
• Current radiogenomic studies often employ limited models, hindering a deep understanding of imaging-trait molecular underpinnings.
• A need exists for advanced methods to connect organ-level imaging data with cellular-level genomic information.

Purpose of the Study:

• To develop and validate a neural network approach for non-linear mapping between high-dimensional gene expression and imaging traits.
• To interpret these radiogenomic models to identify specific transcriptomic drivers of imaging phenotypes.
• To assess the clinical utility of derived radiogenomic traits for predicting patient outcomes.

Main Methods:

• Utilized a neural network to map gene expression data to imaging traits in glioblastoma patients.
• Employed gene masking and saliency techniques for model interpretability and identification of key genes.
• Validated interpretation methods by identifying known gene-molecular subtype relationships.

Main Results:

• The neural network models demonstrated superior performance over traditional classifiers for radiogenomic mapping.
• Identified specific transcription patterns associated with imaging traits like edema and cellular invasion.
• Discovered 10 radiogenomic traits significantly predictive of patient survival.
• Demonstrated that neural networks can effectively model transcriptomic heterogeneity reflected in imaging data.

Conclusions:

• Neural networks offer a powerful tool for deep radiogenomic analysis, moving beyond linear models.
• The developed interpretation methods facilitate understanding of molecular drivers behind imaging traits.
• Radiogenomic traits derived from this approach hold significant clinical value for predicting cancer patient survival.