A deep learning model to predict silencers. (A) A schematic of the deep learning model used to predict cell type–specific silencers and enhancers. The number of kernels or neurons in a layer is listed in parentheses. The input of the model consists of 1-kb genomic sequences, and the output is a set of three probabilities of the input sequence being a silencer (ys), an enhancer (ye), or a nonfunctional sequence (yn). (B) ROCs and PRCs of the prediction models in three human cell types. The results in the additional three cell lines are presented in Supplemental Figure S1. (C) ROC and PRC classification accuracy in prediction of ReSE (Pang and Snyder 2020) and MPRA (Doni Jayavelu et al. 2020) experimentally characterized silencers. (D) Sharpr-MPRA MaxPos scores in K562 cells. EN and SL are the predicted enhancers and silencers, respectively. NonEN represents the DNase-seq peaks that overlap with H3K27ac ChIP-seq peaks but have not been predicted as enhancers in this study. NonSL represents the H3K27me3 ChIP-seq peaks not predicted as silencers, and the DHS column corresponds to all DNase-seq peaks. The count under a group label (x-axis) is the size of the corresponding sequence set. (E) The density of GWAS SNPs in the predicted silencers (SL), the sequences with H3K27me3 peaks but no H3K27ac peaks (H3K27me3/-H3K27ac), and the predicted enhancers (EN). The background consists of the genomic sequences having the GC content and repeat density matching to T cell SLs. (**) P < 10−10. White asterisks are the significance enrichments of GWAS SNPs compared with H3K27me3/-H3K27ac, and black asterisks are the significant enrichments in enhancers compared with silencers.
