Fig. 3: Cell type-specificity generated by TFBU.

a Illustration of measuring TFBS-context’s cell type-specificity with deep learning models. b Measured enhancer activity of TFBUs in different cell lines. The generated denotes the designed TFBS-context with the models; genome denotes TFBS-context selected from the genome sequence by the models; -SC denotes the core TFBSs were scrambled. The fold value is the mean activity fold change across cell types. The p-values are based on a paired one-tailed t-test, where ns represents not significant. c The fold change of the enhancer activity of TFBUs across cell types. The yellow bar denotes the TFBU’s core TFBSs were scrambled. The fold value denotes the mean effect on the fold change caused by scrambling the core TFBS. The p-values are based on a paired one-tailed t-test, where ns represents not significant. d The fold change of the enhancer activity across cell types before/after filtering out TFBUs with obvious TFBSs in their TFBS-context. The fold value was calculated with the mean value of each group. The box plots in b–d show the median (center line), interquartile range between Q1 and Q3 (box limits), minimum and maximum values within 1.5 times the IQR from Q1 and Q3 (whiskers), and outliers (hollow circles). Each point in b–d denotes the average of three biological replicates. Source data are provided as a Source Data file.