transferGWAS: GWAS of images using deep transfer learning

Area of Science:

• Genomics
• Medical Imaging
• Computational Biology

Background:

• Medical images offer valuable insights into disease biology.
• Traditional genetic association studies struggle with complex imaging data.
• Novel methods are needed to link genetic variations with imaging phenotypes.

Purpose of the Study:

• To introduce transferGWAS, a novel method for genome-wide association studies (GWAS) directly on medical images.
• To enable the investigation of genetic associations with detailed imaging phenotypes.
• To identify genetic loci influencing traits observable in medical images.

Main Methods:

• Utilizing transfer learning to train deep neural networks on independent datasets for image representation learning.
• Developing semantically meaningful image representations.
• Performing genetic association tests on learned image representations using genome-wide data.

Main Results:

• Validated transferGWAS performance using simulation studies for type I error rates and power.
• Applied transferGWAS to UK Biobank retinal fundus images in a genome-wide association study.
• Discovered 60 genomic regions associated with retinal images, including 7 novel candidate loci for eye-related traits and diseases.

Conclusions:

• transferGWAS provides a powerful new approach for genetic analysis of medical imaging data.
• The method successfully identified novel genetic associations with ocular traits.
• This work opens new avenues for understanding the genetic basis of diseases through imaging.