Sparse latent factor regression models for genome-wide and epigenome-wide association studies

Area of Science:

• Genomics and Epigenomics
• Statistical Genetics
• Bioinformatics

Background:

• Analyzing associations between phenotypes and genomic/epigenomic data is challenged by unobserved confounding factors like ancestry and cell-type composition.
• Penalized latent factor regression models address high-dimensional data, but the benefits of sparsity penalties require further evaluation.
• Existing methods may struggle to accurately capture relevant associations in the presence of complex confounding variables.

Purpose of the Study:

• To develop and evaluate least-squares algorithms for sparse latent factor regression models that jointly estimate effect sizes and confounding factors.
• To compare the statistical performance of sparse latent factor regression models against other sparse and non-sparse methods using simulated and empirical data.
• To apply these models to real-world genome-wide and epigenome-wide association studies.

Main Methods:

• Development of least-squares algorithms for joint estimation of effect sizes and confounding factors within sparse latent factor regression frameworks.
• Performance evaluation using simulated datasets (generative and empirical) and comparison with least absolute shrinkage and selection operator (LASSO) and Bayesian sparse linear mixed models.
• Application to a genome-wide association study (GWAS) in *Arabidopsis thaliana* for a flowering trait and an epigenome-wide association study (EWAS) of smoking status in pregnant women.

Main Results:

• Sparse latent factor regression models demonstrated superior statistical performance compared to other sparse methods in simulated data.
• These models showed robustness to model departures in empirical data simulations, outperforming non-sparse approaches.
• Applications in GWAS and EWAS successfully estimated non-null effect sizes, addressed multiple testing issues, and identified novel relevant genes.

Conclusions:

• Sparse latent factor regression models provide a robust and effective statistical approach for analyzing high-dimensional genomic and epigenomic data with unobserved confounding.
• The developed algorithms facilitate accurate estimation of genetic and epigenetic associations, enhancing biological discovery.
• These models offer a valuable tool for identifying significant genetic and epigenetic associations in complex human and plant studies.