Fig. 5: Comparison of transcriptome imputation model performance and TWAS results.

TWAS models were developed using data from GALA II and SAGE participants following the PredictDB pipeline and compared with TWAS models generated in GTEx v8 and MESA using the same approach. a,b, Violin plots showing the distribution of internal cross-validation r² values for genes with TWAS models available in each study. The box plots extend from the 25th to the 75th percentiles of the r² distributions, with median r² values indicated by squares. Predictive performance was compared between TWAS models trained in the pooled GALA II/SAGE population and MESA models trained in African American and Hispanic participants (MESA_AFHI) (a) and between TWAS models trained in African Americans (b). Statistical significance was determined by Wilcoxon test and the P values for differences in r² are two sided. c, Summary of TWAS associations with FDR < 0.05 across 28 traits in PAGE. BMI, body mass index; BP, blood pressure; CRP, C-reactive protein; eGFR, estimated glomerular filtration rate; HbA1c, hemoglobin A1c; LDL, low-density lipoprotein; WHR, waist-to-hip ratio. d,e, Correlation between TWAS z scores based on GALA II/SAGE pooled models and z scores using GTEx (d) or MESA_AFHI models (e) for the union of genes with FDR < 0.05 detected with either model. Associations between TWAS z scores from each model are visualized by linear regression lines with shaded 95% CIs. Genes highlighted in orange had FDR < 0.05 using GALA II/SAGE models but did not reach nominal significance (TWAS P value > 0.05) using GTEx or MESA models.