,
11 tweets,
2 min read
Read on Twitter
Announcing another robust method for Mendelian randomization: biorxiv.org/content/10.110…. I hear you cry - do we really need another new robust method for Mendelian randomization? Well this one is a bit different. Tweetorial follows!
Most methods for Mendelian randomization assume that all genetic variants target the same causal parameter. However, as the number of variants associated with a given trait is expands into the 100s and 1000s, this assumption is increasingly unlikely.
We present a different paradigm: we search for genetic variants with similar estimates that might represent a particular causal mechanism. The "contamination mixture method" finds clusters of genetic variants with similar causal estimates in a likelihood-based framework.
We use the method to investigate the causal relationship between HDL-cholesterol and coronary heart disease (CHD) risk. Here, the likelihood function is bimodal, suggesting the presence of at least two groups of variants with mutually similar but distinct causal estimates.
The method identified variants in 9 gene regions associated with increased levels of HDL-cholesterol, decreased triglycerides, and decreased CHD risk. Variants were also associated with haemoglobin concentration, platelet distribution width, and red cell distribution width.
Directions of associations with these traits were concordant for all the variants, and in seven of the nine gene regions, evidence of colocalization was observed between HDL-cholesterol, CHD risk, and at least one of the above blood cell traits.
This suggests a possible mechanism linking lipids to CHD risk via platelet aggregation.
The contamination mixture method can also be used as a conventional robust method for MR. In a set of realistic simulated scenarios, the method had the best performance amongst robust methods (including MR-PRESSO, median, mode-based and MR-Egger) in terms of mean squared error.
The method gave estimates with low bias and low Type 1 error rate inflation when up to 40% of variants were invalid instruments. It has linear computational time in the number of genetic variants, and can perform analysis with hundreds of variants in a fraction of a second.
Software code is included in the paper, and the method will be included in the next update of the MendelianRandomization package (hopefully within a week!).
I commend the method to the community! Thanks to Chris Foley, James Staley, Elias Allara and Joanna Howson for collaboration on this project! /END
