TransferGWAS: GWAS of images using deep transfer learning

Abstract

Medical images can provide rich information about diseases and their biology. However, investigating their association with genetic variation requires non-standard methods. We propose transferGWAS, a novel approach to perform genome-wide association studies directly on full medical images. First, we learn semantically meaningful representations of the images based on a transfer learning task, during which a deep neural network is trained on independent but similar data. Then, we perform genetic association tests with these representations. We validate the type I error rates and power of transferGWAS in simulation studies of synthetic images. Then we apply transferGWAS in a genome-wide association study of retinal fundus images from the UK Biobank. This first-of-a-kind GWAS of full imaging data yielded 60 genomic regions associated with retinal fundus images, of which 7 are novel candidate loci for eye-related traits and diseases.

ContIG: Self-supervised Multimodal Contrastive Learning for Medical Imaging with Genetics

“…Besides pigmentation, the GWAS catalog records associations with an array of cardiovascular traits (such as BMI, pulse pressure, large artery stroke, and blood biomarkers), as well as eyespecific associations (cataract and astigmatism). Similar results were found by [59], albeit with a larger sample size.…”

“…A GWAS is a statistical analysis that correlates individual genetic markers sampled along the full genome with a trait of interest, such as a specific disease. GWASs usually require a low-dimensional, well-defined trait for association analysis; there is only little work yet on leveraging full medical imaging data in a GWAS setting [7,59]. Here, we follow the transferGWAS [59] framework to evaluate the embeddings learned by ContIG.…”

“…GWASs usually require a low-dimensional, well-defined trait for association analysis; there is only little work yet on leveraging full medical imaging data in a GWAS setting [7,59]. Here, we follow the transferGWAS [59] framework to evaluate the embeddings learned by ContIG. In this framework, images are projected onto their latent space embeddings and then the dimensionality is further reduced with a Principal Component Analysis.…”

“…In addition to its successful applications to medical imaging [74], deep learning also found success in applications on genomics [33,60,121,138,139]. There is a growing number of recent works that utilize deep neural networks to jointly learn from both modalities, such as [7,10,21,28,37,42,59,81,117,135]. However, these methods are all either highly application specific or fully supervised.…”

“…We again correct for the same confounders in the linear model. Finally, we aggregate the summary statistics of the ten individual features into a single p-value for each SNP by using a Bonferroni-correction of the factor 10, following [59].…”