Sparse dimensionality reduction approaches in Mendelian randomisation with highly correlated exposures

Area of Science:

• Genetics
• Statistical Epidemiology

Background:

• Multivariable Mendelian randomisation (MVMR) extends Mendelian randomisation (MR) to multiple exposures but suffers from multicollinearity.
• Exposure correlation significantly impacts MVMR estimate bias and efficiency.

Purpose of the Study:

• To introduce sparse Principal Component Analysis (sPCA) for dimensionality reduction in MVMR.
• To improve the interpretability and reliability of MR estimates using sPCA-transformed exposures.

Main Methods:

• Applied sparse dimension reduction to transform variant-exposure summary statistics into principal components.
• Selected principal components using data-driven cutoffs and assessed instrument strength with adjusted F-statistics.
• Performed MVMR using the transformed exposures, validated via simulation and a lipid metabolite GWAS.

Main Results:

• Sparse PCA achieved a better balance of sparsity and biologically meaningful grouping of lipid traits compared to standard MVMR and MR GRAPPLE.
• Demonstrated the pipeline's effectiveness in simulation studies with highly correlated exposures.
• Successfully applied the method to 97 highly correlated lipid metabolites, assessing causal links to coronary heart disease (CHD).

Conclusions:

• Sparse PCA offers a robust approach to handle multicollinearity in MVMR.
• The proposed method yields more interpretable and reliable causal inference from genetic association studies.
• sPCA provides a superior alternative for analysing complex, correlated exposure data in MR studies.