| Background:Transcriptome-wide association study(TWAS)aims to integrate genomewide association study(GWAS)and expression quantitative trait loci(eQTL)study to explore the associations between gene expression and complex diseases.With a great number of summary data publicly available,summary data-based TWAS methods were becoming more popular.However,in summary data-based TWAS analysis,the linkage disequilibrium matrix(LD matrix)of the GWAS population would be estimated from an external reference panel.The statistical noise introduced by the discrepancy between the reference LD matrix and the actual LD structure of GWAS population,small sample size of the LD reference panel,the insufficient estimation accuracy,and the non-invertibility of LD matrix,etc.,have brought great challenges to summary data-based TWAS methods.Currently,in TWAS analysis,LD matrix would be firstly estimated from LD reference panel data and then undergo further regularization or approximation to ensure its numerical stability.There have not been studies to systematically evaluate and compare the performance of different LD matrix estimation methods in TWAS analysis.Objectives:Using 1000 Genomes Project as reference panel,applying LD matrix directly estimated from 1000 Genomes Project and estimated by different LD matrix regularization methods to summary data-based TWAS methods,respectively,to systematically evaluate their impacts on TWAS statistical inference,thus to further provide the guidance for the choice of LD matrix estimation methods and to ensure the rationality of TWAS analysis.Methods:Based on five commonly used summary data-based TWAS methods,including PrediXcan,TWAS,DPR,LDA MR-Egger and PMR-Egger,using four different LD matrix estimation methods:①empirical LD matrix Σref drived from 1000 Genomes Project;②Σa=λΣref+(1-λ)I;③ΣB=Σref+λI;④(?)(?);in which λ is the shrinkage parameter,and I is the identity matrix,we systematically evaluated the impacts of different LD matrix estimation methods on the performance of TWAS analysis through both statistical simulations and real data applications.All data used in this study were publicly available and were restricted to European ancestry,among which eQTL data was from GEUVADIS study,GWAS data from UK Biobank,and LD reference panel data from 1000 Genomes Project.1.We randomly selected 10,000 genes,obtained corresponding cis-SNPs of these genes,and used actual genotype data to generate gene expression and phenotype data.Following the standard analysis pipeline of TWAS,LD matrices estimated from 1000 Genomes Project and estimated by three different methods were applied to five summary data-based TWAS methods(PrediXcan,TWAS,DPR,LDA MR-Egger and PMR-Egger),respectively.Besides,we conducted comphrehensive simulations under different scenarios,such as different gene expression heritability,different gene expression genetic architecture(sparse or polygenic assumption),in the absence or presence of horizontal pleiotropy,in the absence or presence of causal effect,etc,to evaluate the type I error control and statistical power for testing causal effect and horizontal pleiotropy of these summary data-based TWAS methods.Then,compared with the performance of TWAS analysis using gold standard LD matrix which was estimated from GWAS individual-level genotype data,we systematically evaluated the impacts of different LD matrix estimation methods on TWAS statistical inference.2.We performed real data applications using GWAS summary data of four lipid traits from UK Biobank,including total cholesterol(TC),high density lipoprotein cholesterol(HDL),low density lipoprotein cholesterol(LDL),and triglyceride(TG).Of note,the performance in capturing the underlying gene expression genetic architecture would be different for TWAS methods under different modeling assumptions.Therefore,for each lipid trait,using the significant genes identified by five summary data-based TWAS methods using different LD matrix estimation methods,we performed GO and KEGG enrichment analysis by Metascape to identify underlying biological pathways and test the robustness of the results.Results:1.In simulations,TWAS methods based on traditional two-stage inference strategy,including PrediXcan,TWAS,and DPR,showed better performance under no horizontal pleiotropy and polygenic assumption scenarios.For these three methods,the performance of type Ⅰ error control and statistical power for testing causal effect when usingΣref,ΣA,ΣB,ΣC was comparable to the performance when using gold standard LD matrix.Specifically,in type Ⅰ error simulations,inflated p-values were produced by these three methods when using ΣA and ΣC,and the tendency of inflation gradually became obvious as the degree of shrinkage increased;in contrast,p-values became more conservative as the parameter λincreased when using EB.However,no obvious trend was observed for the power to test the causal effect with the change of parameter λ.Besides,in simulation scenarios with large horizontal pleiotropy effect,for these three methods,the type Ⅰ error for testing causal effect became inflated and the power was low when using gold standard LD matrix or four different LD matrix estimation methods.Compared with simulation results when using gold standard LD matrix,both LDA MR-Egger and PMR-Egger produced inflated p-values under various null simulation scenarios for testing causal effect and horizontal pleiotropy when using Σref,ΣA,ΣB,ΣC.For LDA MR-Egger,the tendency of type Ⅰ error inflation for testing causal effect became more obvious when using ΣC than using ΣA and ΣB;in contrast,the tendency of typeⅠ error inflation for testing horizontal pleiotropy became more obvious when using ΣA and ΣB than using ΣC.The power of LDA MR-Egger for testing both causal effect and horizontal pleiotropy was low,and no obvious trend was observed for the power with the change of parameter λ.For PMR-Egger,compared with simulation results using Σref,the type Ⅰ error inflation tendency was alleviated when using ΣC.Besides,when using ΣA,ΣB,ΣC,the tendency of type Ⅰ error inflation gradually became obvious as the degree of shrinkage increased.PMR-Egger showed relatively stable power for testing causal effect and horizontal pleiotropy when using Σc.2.In real data applications,different summary data-based TWAS methods using different LD estimation mentods identified a total of 320,486,266,and 419 genes which were significantly associated with TC,HDL,LDL,and TG,respectively.GO and KEGG pathway enrichment analysis showed that these genes were significantly enriched in the pathways related to blood lipid metabolism,such as cholesterol transport and cholesterol metabolism.In addition,TG-related genes were significantly enriched in neurodegenerative disease pathways,which was consistent with the findings in previous studies,suggesting the robustness of the results from different TWAS methods using different LD matrix estimation methods.Conclusion:The impacts of using LD matrix estimated from 1000 Genomes Project and different LD estimation methods on the performance of different summary data-based TWAS differed,which was closely related to the principles of different TWAS analysis methods,the characteristics of different LD matrix estimation methods,and different simulation scenarios.1.For TWAS methods based on traditional two-stage design,including PrediXcan,TWAS,and DPR,the LD matrix is used to construct the test statistics to test the associations between genes and complex traits in the second stage,which does not involve matrix inversion.Therefore,the influence on the statistical inference of summary data-based TWAS analysis was small when using different LD matrix estimation methods,and these TWAS methods were well-performed when directly using the LD matrix derived from 1000 Genomes Project.However,under the scenarios with horizontal pleiotropy or with the violation of polygenic assumption,it should be cautious in controlling the false positive findings of these three methods.2.Both LDA MR-Egger and PMR-Egger could account for widespread horizontal pleiotropy,which increases the complexity of statistical inference when using summary data and determines that the LD matrix would be applied in multiple inference procedures.Since the small sample size of the LD reference panel would result in insufficient estimation accuracy and matrix inversion might aggravate the numerical instability,potential discrepancies between the LD matrix from the LD reference panel and the actual LD structure of the GWAS sample might also be amplified,which would have much influence on TWAS statistical inference.For the methods that could simultaneously estimate and test causal effect and horizontal pleiotropy effect,PMR-Egger showed better performance,but more attention should be paid in controlling false positive findings.3.Based on TWAS analysis for four blood lipids,we found that the significant genes related to specific lipid trait identified by different summary data-based TWAS methods using different LD matrix were significantly enriched in lipid metabolism related biological pathways,suggesting that TWAS methods under different modeling assumptions could be complementary to each other and provide more biological information. |
