SUMMIT: An integrative approach for better transcriptomic data imputation improves causal gene identification

Abstract

Genes with moderate to low expression heritability may explain a large proportion of complex trait etiology, but such genes cannot be sufficiently captured in conventional transcriptome-wide association studies (TWASs), partly due to the relatively small available reference datasets for developing expression genetic prediction models to capture the moderate to low genetically regulated components of gene expression. Here, we introduce a method, the Summary-level Unified Method for Modeling Integrated Transcriptome (SUMMIT), to improve the expression prediction model accuracy and the power of TWAS by using a large expression quantitative trait loci (eQTL) summary-level dataset. We apply SUMMIT to the eQTL summary-level data provided by the eQTLGen consortium. Through simulation studies and analyses of genome-wide association study summary statistics for 24 complex traits, we show that SUMMIT improves the accuracy of expression prediction in blood, successfully builds expression prediction models for genes with low expression heritability, and achieves higher statistical power than several benchmark methods. Finally, we conduct a case study of COVID-19 severity with SUMMIT and identify 11 likely causal genes associated with COVID-19 severity.

SUMMIT-FA: A new resource for improved transcriptome imputation using functional annotations

“…The 2, 177 lower heritability genes produce 642 significant gene-trait associations, a relatively similar ratio to that of genes with R 2 ≥ 0.01: 12, 132 genes and 4, 049 significant genetrait associations. This gives further credence to the notion that lower heritability genes have notably larger causal effect sizes on complex phenotypes [6, 22].…”

“…The power of TWAS is determined by the accuracy of the expression prediction model in Step 1 and the sample size of GWAS in Step 2 [6, 13]. As the sample size of GWAS continues to increase, thanks to the extensive consortium efforts, the prediction accuracy of the expression prediction model remains a limiting factor.…”

“…As the sample size of GWAS continues to increase, thanks to the extensive consortium efforts, the prediction accuracy of the expression prediction model remains a limiting factor. To address this challenge, several methods have been proposed to improve the prediction model accuracy by leveraging auxiliary information from other tissues [10, 12], functional annotations or atlases of regulatory elements [14, 15], and summary-level expression reference panels (SUMMIT) [6]. As an example, by using the summary-level expression reference panel with a much larger sample size, SUMMIT outperforms many benchmark methods in terms of expression prediction model accuracy and subsequent TWAS power [6].…”

“…To address this challenge, several methods have been proposed to improve the prediction model accuracy by leveraging auxiliary information from other tissues [10, 12], functional annotations or atlases of regulatory elements [14, 15], and summary-level expression reference panels (SUMMIT) [6]. As an example, by using the summary-level expression reference panel with a much larger sample size, SUMMIT outperforms many benchmark methods in terms of expression prediction model accuracy and subsequent TWAS power [6]. However, SUMMIT only relies on the summary-level expression reference panel and ignores comprehensive functional annotations that may be useful for improving expression prediction model accuracy.…”

“…To test our hypothesis, we develop SUMMIT-FA (Summary-level Unified Method for Modeling Integrated Transcriptome using Functional Annotations), an extension of the SUMMIT [6], that improves the accuracy of gene expression prediction models by leveraging annotation resources from the Multi-dimensional Annotation-Class Integrative Estimation (MACIE) database [18] to prioritize functional variants. MACIE synthesizes multiple categories of annotations, such as evolutionary conservation annotations and epigenetic annotations, and consistently provides the state-of-the-art performance in discriminating between functional and non-functional variants [18].…”