CoRE-ATAC: A deep learning model for the functional classification of regulatory elements from single cell and bulk ATAC-seq data
Fig 4
CoRE-ATAC predictions overlap with experimentally detected enhancers.
(A) Overlap of FANTOM enhancer annotationswith CoRE-ATAC (C) and ChromHMM (H) predictions in MCF7, A549, CD4+ T and PBMC samples. CoRE-ATAC predicted the majority of FANTOM enhancers as enhancers or promoters, recapitulating these experimentally identified enhancers. CoRE-ATAC annotations were similar toChromHMM annotations. (B) CoRE-ATAC predictions for active regulatory regions identified by STARR-seq in A549 cell line. The majority of active enhancers identified by STARR-seq were predicted as promoter or enhancer by CoRE-ATAC. (C) MIN6 MPRA log fold change values for genomic regions predicted as losing or gaining cis-RE function based on CoRE-ATAC probabilities for reference and alternative alleles. Significance for predicted loss and predicted gain categories was calculated using student’s t-test for MPRA log fold change values being less than or greater than 0 respectively. Significance comparing the predicted loss and predicted gain of MPRA fold change distributions was calculated using Mann-Whitney U test. We observed concordant direction of effect both for CoRE-ATAC predictions and MPRA activity levels when alternative and reference alleles are compared. (D) Genome browsers of 19 islet samples highlighting a loss of enhancer activity for rs11205653 (also highlighted in (C)) for the alternative allele (G). Values for enhancer and other represent the probability assigned to those classes of cis-REs by CoRE-ATAC. We observe that for 5 out of 7 individuals with the reference allele (TT) CoRE-ATAC predicted enhancer activity, reflecting ChromHMM reference annotations, while for the individuals with GT or GG alleles, we observed an enhancer activity loss for all but one individual based on CoRE-ATAC predictions.