Mendelian randomization while jointly modeling cis genetics identifies causal relationships between gene expression and lipids

Abstract

Inference of causality between gene expression and complex traits using Mendelian randomization (MR) is confounded by pleiotropy and linkage disequilibrium (LD) of gene-expression quantitative trait loci (eQTL). Here, we propose an MR method, MR-link, that accounts for unobserved pleiotropy and LD by leveraging information from individual-level data, even when only one eQTL variant is present. In simulations, MR-link shows false-positive rates close to expectation (median 0.05) and high power (up to 0.89), outperforming all other tested MR methods and coloc. Application of MR-link to low-density lipoprotein cholesterol (LDL-C) measurements in 12,449 individuals with expression and protein QTL summary statistics from blood and liver identifies 25 genes causally linked to LDL-C. These include the known SORT1 and ApoE genes as well as PVRL2, located in the APOE locus, for which a causal role in liver was not known. Our results showcase the strength of MR-link for transcriptome-wide causal inferences.

Fig. 1. Graphical representation of the study.

The Biobank Integrative Omics Study (BIOS) cohort was used to identify expression quantitative trait loci (eQTLs) and characterize the genetic architecture of gene expression. Dashed outbox: Knowledge used in a simulation scheme that mimicked gene-expression traits, including linkage disequilibrium (LD) between eQTL single nucleotide polymorphism (SNPs). We used this simulation to assess the false positive rates and power for widely used Mendelian randomization (MR) methods. We applied our MR method, MR-link, to both the simulations and to individual-level data of low-density lipoprotein cholesterol (LDL-C) in 12,449 individuals (Lifelines) combined with BIOS and GTEx eQTL as well as protein quantitative trait loci (pQTL) summary statistics to identify gene-expression changes and protein level changes that are causally linked to LDL-C within or outside a genome-wide association study (GWAS) locus.

Fig. 2. Typical scenarios of pleiotropy in causal inference of gene expression changes as an exposure.

Typical scenarios to consider when performing causal inference in gene expression: a expression quantitative trait locus (eQTL) single nucleotide polymorphisms (SNPs) used as instrumental variables (IVs) for the same gene (exposure) are in linkage disequilibrium (LD) and pleiotropic effects are absent, b pleiotropy is present through LD between IVs for different exposures (pleiotropy through LD), and c pleiotropy is present through overlap of the IVs (pleiotropy through overlap). In each panel, the left image shows the genomic context while the right image is a schematic diagram of the corresponding causal effects. Please note that the unobserved exposure trait does not necessarily need to be a protein product: it could be any measured or unmeasured phenotype that is regulated by the genetic locus.

Fig. 3. Relative performance of different MR methods.

The figure shows performance of MR methods on simulations representing the pleiotropy through linkage disequilibrium (LD) scenario (depicted in Fig. 2b) when 1, 3, 5, or 10 causal expression quantitative trait locus (eQTL) single nucleotide polymorphisms (SNPs) were simulated (“Methods”). a False positive rates (at alpha = 0.05) in scenarios where no causal relationship is simulated. b Power to detect a small causal effect (at alpha = 0.05). c Power to detect a large causal effect (at alpha = 0.05). Note that MR-link is the only MR method that can adjust for pleiotropy when only one or two instrumental variables are available. MR methods that had fewer than 100 out of 1500 estimates in a scenario are not shown (“Methods”). Extended results, including those that are not shown in the figure, can be found in Supplementary Data 3.

Fig. 4. Biological interpretation of PVRL2.

Functional and statistical evidence for the causal effect of PVRL2 on low-density lipoprotein cholesterol (LDL-C) levels. The teal arrow indicates a positive causal relationship between PVRL2 expression in liver and LDL-C levels in plasma—this relationship was detected in our MR analysis. The red arrow indicates a negative causal relationship between PVRL2 expression and LDL-C uptake in hepatic cells—this relationship was detected in small interfering RNA (si) experiments described in Blattman et al..