Integration of latent factor analysis into multivariable Mendelian randomization.

Mendelian randomization has emerged as a powerful tool for exploring causal relationships in observational studies by using genetic variants as instrumental variables. While multivariable Mendelian randomization extends this approach to simultaneously address multiple exposures, it faces significant challenges with highly correlated exposures, particularly in high-dimensional settings such as multi-omics data. Conventional MVMR methods, which are primarily based on linear regression models, may suffer from multicollinearity and reduced statistical power when analyzing correlated exposures. The increasing availability of high-dimensional multi-omics data has highlighted the limitations of conventional MVMR approaches in analyzing correlated exposures while maintaining biological interpretability. To address these challenges, we propose integrating latent factor analysis into the MVMR framework, enabling dimension reduction without compromising interpretability. Through extensive simulation studies, we demonstrate that our method maintains a well-controlled false positive rate and offers superior sensitivity compared to conventional MVMR approaches. We apply our method to investigate the causal relationship between DNA methylation and mitochondrial DNA copy number. Our method offers a significant advantage in scenarios with highly correlated exposures driven by common latent factors or shared pathways, especially when individual effects are sparse. By applying our method to correlated multi-omics data, we can uncover new insights into the molecular mechanisms underlying complex phenotypes.