Genetic InfoMax: Exploring Mutual Information Maximization in High-Dimensional Imaging Genetics Studies

Area of Science:

• Neuroimaging
• Genetics
• Machine Learning

Background:

• Genome-wide association studies (GWAS) link genetic variations to traits, but applying them to high-dimensional medical imaging data requires effective trait representation.
• Existing representation learning methods face challenges unique to imaging genetics and GWAS, limiting their utility.
• Integrating medical imaging and genetic data for trait discovery remains an underexplored area.

Approach:

• Introduced Genetic InfoMax (GIM), a trans-modal learning framework addressing GWAS challenges from a mutual information (MI) perspective.
• Developed a regularized MI estimator and a novel genetics-informed transformer within the GIM framework.
• Established standardized evaluation protocols for comparing GIM against existing approaches using 3D MRI brain data.

Key Points:

• GIM effectively extracts lower-dimensional, informative representations from high-dimensional medical imaging data for GWAS.
• The framework incorporates a novel genetics-informed transformer and a regularized MI estimator.
• Demonstrated significantly improved GWAS performance on 3D human brain MRI data.

Conclusions:

• GIM offers a powerful new approach for representation learning in imaging genetics.
• The developed framework enhances the ability to discover genetic associations with traits from medical imaging.
• This work paves the way for more effective utilization of medical imaging in genetic studies.