Polygenic Prediction of Molecular Traits using Large-Scale Meta-analysis Summary Statistics

Abstract

Introduction: Transcriptome-wide association study (TWAS) integrates expression quantitative trait loci (eQTL) data with genome-wide association study (GWAS) results to infer differential expression. TWAS uses multi-variant models trained using individual-level genotype-expression datasets, but methodological development is required for TWAS to utilise larger eQTL summary statistics. Methods: TWAS models predicting gene expression were derived using blood-based eQTL summary statistics from eQTLGen, the Young Finns Study (YFS), and MetaBrain. Summary statistic polygenic scoring methods were used to derive TWAS models, evaluating their predictive utility in GTEx v8. We investigated gene inclusion criteria and omnibus tests for aggregating TWAS associations for a given gene. We performed a schizophrenia TWAS using summary statistic-based TWAS models, comparing results to existing resources and methods. Results: TWAS models derived using eQTL summary statistics performed comparably to models derived using individual-level data. Multi-variant TWAS models significantly improved prediction over single variant models for 8.6% of genes. TWAS models derived using eQTLGen summary statistics significantly improved prediction over models derived using a smaller individual-level dataset. The eQTLGen-based schizophrenia TWAS, using the ACAT omnibus test to aggregate associations for each gene, identified novel significant and colocalised associations compared to summary-based mendelian randomisation (SMR) and SMR-multi. Conclusions: Using multi-variant TWAS models and larger eQTL summary statistic datasets can improve power to detect differential expression associations. We provide TWAS models based on eQTLGen and MetaBrain summary statistics, and software to easily derive and apply summary statistic-based TWAS models based on eQTL and other molecular QTL datasets released in the future.