Improving gene expression prediction models through the use of meta‐analysis summary statistics

Abstract

AbstractBackgroundBy focusing on a gene‐level rather than a single‐variant basis, transcriptome‐wide association studies (TWAS) assess the role of genetically regulated gene expression in disease risk and improve statistical power by reducing the multiple testing burden by over a 100‐fold. A key limitation of current TWAS approaches is the requirement for dense individual‐level datasets with both genotypes and gene expression measures to produce effective expression prediction models. We sought to improve the statistical power of AD TWAS by leveraging easily accessible summary statistics taken from large‐scale expression quantitative trait loci (eQTL) meta‐analyses.MethodWe used eQTLGen Whole Blood cis‐eQTL data (N = 31,684) as our source of meta‐analysis summary statistics. To perform gene‐based elastic‐net regressions using the summary statistics, we used the lassosum R package (Mak et al. 2017) with 1000 Genomes data as a genomic reference and GTEx v8 genotype, normalized Whole Blood gene expression and covariate data (N = 670) as a testing panel for model validation and dynamic selection of the optimal mixing/penalty parameters. The SNP‐weights generated by our models were subsequently applied to AD meta‐analysis summary statistics from Kunkle et al. 2019 (AD Cases = 21,982; Controls = 44,944) and Bellenguez et al. 2022 (AD/ ‘proxy’ AD Cases = 111,326; Controls = 677,663) via S‐PrediXcan to assess whether predicted gene expression is associated with AD risk.ResultWe found on average a 3.3% (SD: 0.05%) significant improvement in the predictive performance for 93.3% of genes models (N = 2,787) when using the same SNPs as the PredictDB model. Using models with the restricted SNP set, we identified four genome‐wide significant gene expression models (CTDNEP, p = 3.2e‐9; STK4, p = 1.6e‐13; RPS14, p = 8.8e‐38; POLR1H, p = 5.6e‐45) using Kunkle et al. 2019 summary statistics relative to only one using the PredictDB model (FBXO46, p = 6.6e‐15). When applying the Bellenguez et al. 2022 summary statistics, which contains a broader AD phenotype thru the inclusion of ‘proxy’ cases, we identified 12 genome‐wide significant gene expression models (Range: CHRNA2, p = 2.6e‐8; SLC24A4, p = 4.5e‐16) associated with AD.ConclusionOur findings show AD association studies based on TWAS approaches may be improved by meta‐analysis reference panels.