27The genetic control of gene expression is a core component of human physiology. For the past 28 several years, transcriptome-wide association studies have leveraged large datasets of linked 29 genotype and RNA sequencing information to create a powerful gene-based test of association 30 that has been used in dozens of studies. While numerous discoveries have been made, the 31 populations in the training data are overwhelmingly of European descent, and little is known 32 about the generalizability of these models to other populations. Here, we test for cross-33 population generalizability of gene expression prediction models using a dataset of African 34American individuals with RNA-Seq data in whole blood. We find that the default models trained 35 in large datasets such as GTEx and DGN fare poorly in African Americans, with a notable reduction 36 in prediction accuracy when compared to European Americans. We replicate these limitations in 37 cross-population generalizability using the five populations in the GEUVADIS dataset. Via realistic 38 simulations of both populations and gene expression, we show that accurate cross-population 39 generalizability of transcriptome prediction only arises when eQTL architecture is substantially 40 shared across populations. In contrast, models with non-identical eQTLs showed patterns similar 41 to real-world data. Therefore, generating RNA-Seq data in diverse populations is a critical step 42 towards multi-ethnic utility of gene expression prediction. 43 44 Author summary 47 Advances in RNA sequencing technology have reduced the cost of measuring gene expression at 48 a genome-wide level. However, sequencing enough human RNA samples for adequately-49 powered disease association studies remains prohibitively costly. To this end, modern 50 transcriptome-wide association analysis tools leverage existing paired genotype-expression 51datasets by creating models to predict gene expression using genotypes. These predictive models 52 enable researchers to perform cost-effective association tests with gene expression in 53 independently genotyped samples. However, most of these models use European reference 54 data, and the extent to which gene expression prediction models work across populations is not 55 fully resolved. We observe that these models predict gene expression worse than expected in a 56 dataset of African-Americans when derived from European-descent individuals. Using 57 simulations, we show that gene expression predictive model performance depends on both the 58 amount of shared genotype predictors as well as the genetic relatedness between populations. 59Our findings suggest a need to carefully select reference populations for prediction and point to 60 a pressing need for more genetically diverse genotype-expression datasets. 61In the last decade, large-scale genome-wide genotyping projects have enabled a revolution in our 63 understanding of complex traits. [1][2][3][4] This explosion of genome sequencing data has spurred the 64 development of new methods that int...