Extended Data Fig. 1: Characterizations of the person-specific models.

a, The consistency between the predictive contributions of regulators and their expression. Spearman’s correlation was used to evaluate the relation between DeepLIFT scores vs log2-TPMs in each tissue. A relation with the absolute correlation > 0.3 and FDR < 5% was defined as significant. b, The consistency of predictive contributions across tissues. We selected the regulators that showed the significant relationships between their DeepLIFT scores with log2-TPMs in multiple tissues. Then, the consistency of directions of the correlations was assessed and visualized as a histogram. c, The actual correlation profile between DeepLIFT scores and log2-TPMs. We selected 99 regulators that showed consistent relationships in more than four tissues. d, The pairwise similarity of per-gene prediction accuracy between tissues. Spearman’s correlation was used for this comparison. e, The association of per-gene prediction accuracy with the gene status. The gene status indicates whether genes are registered in multiple databases (Known) or only in the GENCODE database (Novel or Putative). f, Decomposition of variance in the per-gene prediction accuracy. To compare the variances in the per-gene prediction accuracy explained by the gene status, the number of promoter features, and RNA features, we used a variance decomposition method (lmg) implemented in relaimpo R package.