To test the predictive accuracy of the CMC-derived DLPFC models, and to benchmark this against existing GTEx-derived prediction models, GREX was calculated in an independent DLPFC RNA-sequencing dataset (the Religious Orders Study Memory and Ageing Project, ROSMAP19,20). We compared predicted GREX to measured ROSMAP gene expression for each gene (Replication R2, or RR2) for the CMC-derived DLPFC models and twelve GTEx-derived brain tissue models15,21 (Figure 1, Supplementary Figure 2b). CMC-derived DLPFC models had higher average RR2 values (Mean RR2 = 0.056), more genes with RR2 > 0.01, and significantly higher overall distributions of RR2 values than any of the twelve GTEx models (ks-test, p < 2.2 × 10−16 across all analyses; Figure 1). Median RR2 values were significantly correlated with sample size of the original tissue set (rho = 0.92, p = 7.2 × 10−6), the number of genes in the prediction model (rho = 0.9, p = 2.6 × 10−5), and the number of significant ‘eGenes’ in each tissue type (rho = 0.95, p = 5.5 × 10−7; Figure 1c). Notably, these correlations persist after removing obvious outliers (Figure 1c).
