transferGWAS

transferGWAS performs genome-wide association studies directly on full medical imaging data by applying transfer learning with deep neural networks to extract semantically meaningful image representations for genetic association testing.

Key Features:

• Deep Learning Integration: A deep neural network is trained using transfer learning on independent but similar data to learn robust, semantically meaningful image representations.
• Direct Imaging GWAS: Performs association testing directly on full medical images rather than relying on predefined quantitative traits or biomarkers.
• Validation and Performance: Simulation studies with synthetic images show maintained type I error rates and demonstrate sufficient statistical power.
• Real-World Application: Applied to a GWAS of retinal fundus images from the UK Biobank, identifying 60 genomic regions including seven novel candidate loci linked to eye-related traits and diseases.
• Implementation: The method is implemented in Python.

Scientific Applications:

• Ophthalmology: Identify genetic loci associated with retinal fundus imaging phenotypes and eye-related traits.
• Oncology: Associate tumor imaging features with genetic variation captured in medical images.
• Neurology: Map genetic variants to brain imaging phenotypes represented in neuroimaging data.

Methodology:

Train a deep neural network using transfer learning to derive semantically meaningful image representations, then perform genome-wide association tests using those derived representations to identify associated genomic regions.