Obtaining polygenic transcriptome risk scores (PTRS) directly from GWAS summary statistics

Abstract

Polygenic Transcriptome Risk Scores (PTRS) are variations of Polygenic Risk Scores (PRS) that use genetically predicted transcriptome as features for prediction instead of directly using genetic variants. We have shown that when PTRS is combined with PRS, they can yield improved prediction performance and portability across populations (Liang et al., 2022). Given the difficulty of training PTRS using large scale individual-level data (due to both computational burden and the lack of data access), we developed a user friendly software that infers PTRS using GWAS summary results and reference LD. We tested three summary statistics-based PTRS approaches: i) Clumping and thresholding (clump-PTRS), keeping trait associated genes while removing highly correlated ones; ii) Summary statistics-based elastic net PTRS (S-EN-PTRS), an extension of lassosum (Mak et al., 2017) to predicted transcriptome; iii) Naive-PTRS, the sum of predicted expressions of significantly associated genes weighted by PrediXcan-estimated effect sizes (Gamazon et al., 2015). Despite reports that individual-level trained elastic net PTRS outperformed clump-PTRS in (Liang et al., 2022), for most of the 11 traits used in the comparison, clump-PTRS outperformed S-EN-PTRS, which outperformed naive-PTRS.