Incorporating tandem repeats improves gene expression prediction models

Tandem repeats (TRs), consisting of consecutive repeated units of 1+bp, are highly polymorphic features found throughout the genome that have been shown to have widespread impacts on gene expression and other traits. However, despite their prevalence, TRs are not included as features in polygenic prediction models, which typically focus on common single nucleotide polymorphisms (SNPs). Here, we develop a framework to incorporate TRs into gene expression models, which consider variants in a cis-window around each gene to predict expression levels. Simulation analyses demonstrate that including linear effects of TR copy number as features can result in substantially improved predictions, particularly when high impact causal TRs are poorly tagged by individual SNPs. We applied our framework to train and evaluate SNP+TR expression prediction models in lymphoblastoid cell lines (LCLs) using data from the Geuvadis cohort (n=359 European individuals) and compared these to SNP-only models. Overall, across 2,553 genes for which models trained successfully, we observed that SNP+TR models give significantly improved performance for 78 genes and a mean gain in adjusted R-squared 0.0037 across all genes, with most gains replicating in the independent Gene-Tissue Expression Project (GTEx) cohort. We additionally find that incorporating TRs results in simpler models and improves portability of models when testing on data for 89 Yoruban individuals from Geuvadis. Finally, we extended our models to consider nonlinear effects of TR copy number on expression. Under our nonlinear TR association testing, we identify 2,255 nonlinear significant genes not captured by linear association testing models, and identify exemplar genes, including PALD1, with highly polymorphic TRs for which incorporating their nonlinear effects results in substantially improved gene expression prediction. Overall, these results demonstrate the utility of including TRs in expression prediction models and show promise that they will have similar impact across other models such as polygenic risk scores for complex traits.