TransferGWAS of T1-weighted brain MRI data from UK Biobank

Area of Science:

• Genetics
• Neuroimaging
• Artificial Intelligence

Background:

• Genome-wide association studies (GWAS) traditionally focus on single traits, limiting discovery in complex datasets.
• Large cohorts like UK Biobank offer rich imaging data suitable for advanced genetic analysis.
• Current GWAS methods for imaging data often rely on predefined features, potentially missing novel patterns.

Purpose of the Study:

• To apply TransferGWAS, a deep neural network (DNN) approach, to brain MRI data for novel genetic discovery.
• To identify genetic loci associated with brain imaging features and explore their links to various phenotypes.
• To assess the utility of DNN-derived features for polygenic risk prediction and genetic correlation analyses.

Main Methods:

• Utilized TransferGWAS with DNNs trained on Alzheimer's Disease Neuroimaging Initiative and ImageNet datasets to encode UK Biobank brain MRI scans.
• Performed multivariate genome-wide association testing on the low-dimensional imaging features.
• Constructed polygenic scores from deep features and computed genetic correlations with other phenotypes.

Main Results:

• Identified 289 independent genetic loci associated with traits including bone density, brain structure, and cardiovascular factors.
• Discovered 11 novel brain regions with no previously reported genetic associations.
• Demonstrated improved prediction of bone mineral density and other traits using multi-polygenic scores of deep features.
• Revealed potential genetic links between diffusion MRI traits and type 2 diabetes.

Conclusions:

• Deep learning-based feature extraction in GWAS (TransferGWAS) can identify novel heritable brain variability beyond predefined measures.
• This approach links brain imaging genetics to a broader spectrum of non-brain phenotypes.
• TransferGWAS enhances the discovery potential of large-scale neuroimaging datasets in genetic association studies.