An iterative approach to detect pleiotropy and perform Mendelian Randomization analysis using GWAS summary statistics

Abstract

Motivation: The overall association evidence of a genetic variant with multiple traits can be evaluated by cross-phenotype association analysis using summary statistics from genome-wide association studies. Further dissecting the association pathways from a variant to multiple traits is important to understand the biological causal relationships among complex traits.

Results: Here, we introduce a flexible and computationally efficient Iterative Mendelian Randomization and Pleiotropy (IMRP) approach to simultaneously search for horizontal pleiotropic variants and estimate causal effect. Extensive simulations and real data applications suggest that IMRP has similar or better performance than existing Mendelian Randomization methods for both causal effect estimation and pleiotropic variant detection. The developed pleiotropy test is further extended to detect colocalization for multiple variants at a locus. IMRP will greatly facilitate our understanding of causal relationships underlying complex traits, in particular, when a large number of genetic instrumental variables are used for evaluating multiple traits.

Availability and implementation: The software IMRP is available at https://github.com/XiaofengZhuCase/IMRP. The simulation codes can be downloaded at http://hal.case.edu/∼xxz10/zhu-web/ under the link: MR Simulations software.

Supplementary information: Supplementary data are available at Bioinformatics online.

Fig. 1.

The association paths of genetic IVs, exposure (Y₁) and outcome (Y₂). $G_1,\mathrm{ }{G}_2,\mathrm{ }\dots ,\mathrm{ }{G}_{n_1}\mathrm{ }$represent mediation variants, which are valid IVs in MR analysis. $G_1^{\mathrm{\text{'}}},\mathrm{ }{G}_2^{\mathrm{\text{'}}},\mathrm{ }\dots ,\mathrm{ }{G}_{n_2}^{\mathrm{\text{'}}}\mathrm{ }$represent pleiotropic variants, which are invalid IVs in MR analysis. Each γ represents a direct contribution of a genetic variant. β represents the causal effect from exposure Y₁ to outcome Y₂. U represents confounding factors

Fig. 2.

Type I error and power of $T_{\mathrm{pleio}}$ when testing pleiotropy/colocalization against mediation using two variants at a locus. Type I error rate and power were evaluated at 0.05 significance level based on 1000 replicates. (A–C) Two traits have complete overlapped samples and the LD between two variants are r²=0.7, 0.5 and 0.3, respectively; (D–E) similar to (A–C) but with 50% overlapped samples; and (G–I) similar to (A–C) but with 0% overlapped samples (two-sample model). ρ represents trait correlation

Fig. 3.

Type I error and power of $S_{\mathrm{pleio}}$ when testing pleiotropy/colocalization against mediation using two variants at a locus. Type I error rate and power were evaluated at 0.05 significance level based on 1000 replications. (A–C) Two traits have complete overlapped samples and the LD between two variants are r²=0.7, 0.5 and 0.3, respectively; (D–E) similar to (A–C) but with 50% overlapped samples; and (G–I) similar to (A–C) but with 0% overlapped samples (two-sample model). ρ represents trait correlation