Integration of latent factor analysis into multivariable Mendelian randomization.

Area of Science:

• Genetics
• Epidemiology
• Bioinformatics

Background:

• Mendelian randomization (MR) uses genetic variants to infer causality from observational data.
• Multivariable MR (MVMR) extends this to multiple exposures but struggles with highly correlated exposures, especially in high-dimensional multi-omics data.
• Conventional MVMR methods can face multicollinearity and reduced power with correlated exposures, limiting biological insights.

Purpose of the Study:

• To develop an enhanced MVMR framework that addresses challenges posed by highly correlated exposures in high-dimensional settings.
• To integrate latent factor analysis into MVMR for effective dimension reduction while preserving interpretability.
• To improve causal inference in multi-omics studies by accounting for shared latent factors or pathways.

Main Methods:

• Proposed an integration of latent factor analysis within the MVMR framework.
• Developed a method for dimension reduction in MVMR without compromising biological interpretability.
• Validated the approach through extensive simulation studies.

Main Results:

• The proposed method demonstrated a well-controlled false positive rate in simulations.
• Achieved superior sensitivity compared to conventional MVMR approaches for correlated exposures.
• Successfully applied the method to investigate causal links between DNA methylation and mitochondrial DNA copy number.

Conclusions:

• The novel latent factor-based MVMR method effectively handles highly correlated exposures, particularly in multi-omics data.
• This approach offers significant advantages for uncovering causal relationships driven by shared latent factors or pathways.
• The method provides new insights into molecular mechanisms underlying complex phenotypes by improving causal inference in high-dimensional genetic data.