A new Mendelian Randomization method to estimate causal effects of multivariable brain imaging exposures

Abstract

The advent of simultaneously collected imaging-genetics data in large study cohorts provides an unprecedented opportunity to assess the causal effect of brain imaging traits on externally measured experimental results (e.g., cognitive tests) by treating genetic variants as instrumental variables. However, classic Mendelian Randomization methods are limited when handling high-throughput imaging traits as exposures to identify causal effects. We propose a new Mendelian Randomization framework to jointly select instrumental variables and imaging exposures, and then estimate the causal effect of multivariable imaging data on the outcome variable. We validate the proposed method with extensive data analyses and compare it with existing methods. We further apply our method to evaluate the causal effect of white matter microstructure integrity (WM) on cognitive function. The findings suggest that our method achieved better performance regarding sensitivity, bias, and false discovery rate compared to individually assessing the causal effect of a single exposure and jointly assessing the causal effect of multiple exposures without dimension reduction. Our application results indicated that WM measures across different tracts have a joint causal effect significantly impact on the cognitive function among the participants from the UK Biobank.

Causal Inference on Neuroimaging Data with Mendelian Randomisation

“…Unfortunately, due to the much weaker SNP–IDP association strengths (which are partly due to comparatively small sample size of the imaging GWAS), most causal effect estimates involving IDPs, either as exposure or as outcome, are small and often inconclusive. Multi-variable MR methods, possibly combined with a dimensionality-reduction and orthogonalisation step for highly-correlated imaging exposures as recently proposed (Mo et al, 2021), are one direction of ongoing methodological development to address some of these issues.…”

“…Neuroimaging MR analyses in particular are thus likely to require careful consideration of potential pleiotropic effects. Recent studies have relied on multi-variable MR approaches to account for (known) pleiotropy between multiple IDPs (Mo et al 2021). Additionally, high-level phenotypes can be assumed to be more likely to exhibit non-linear associations with SNPs.…”

Integrating multi-omics summary data using a Mendelian randomization framework

“…(i) Transformation-based tests require a transformation of the measurement of exposures into a single new synthetic exposure through principal component analysis (PCA) or other dimensionality reduction techniques [ 12 , 13 ]. Note that as opposed to the other two summary data-based approaches, individual-level data access is necessary to enable the transformation.…”