ML-MAGES: A machine learning framework for multivariate genetic association analyses with genes and effect size...

Area of Science:

• Genetics and Bioinformatics
• Statistical Genomics
• Machine Learning in Biology

Background:

• Genome-wide association (GWA) studies aim to link genetic variants to traits.
• Key challenges include inflated effect estimates and analyzing multiple traits concurrently.

Purpose of the Study:

• To introduce ML-MAGES, a computationally efficient machine learning method for GWA studies.
• To address effect size inflation and enable simultaneous multi-trait analysis.

Main Methods:

• Utilizes neural networks for shrinkage of GWA effect sizes, mitigating inflation from variant non-independence.
• Employs variational inference for clustering variant associations across multiple traits.
• Compares neural network shrinkage to regularized regression and fine-mapping.

Main Results:

• Neural network shrinkage outperforms existing methods in approximating true effect sizes in simulations.
• The infinite mixture clustering approach effectively distinguishes trait-specific, shared, and spurious associations.
• ML-MAGES demonstrates high precision and recall in identifying gene-level associations.

Conclusions:

• ML-MAGES provides a flexible, data-driven approach for multi-trait genetic association analysis.
• Application in UK Biobank data identified trait-specific and shared genetic variants.
• The method suggests potential shared genetic architecture underlying complex traits.