CoRE-ATAC: A deep learning model for the functional classification of regulatory elements from single cell and bulk ATAC-seq data
Fig 3
CoRE-ATAC can predict REs across cell-types.
CoRE-ATAC was evaluated in 7 cell types using 40 samples that are not used in model training. (A) Average precision scores for predicting cis-REs. Micro-average precision was used to calculate class average scores. CoRE-ATAC is predictive across cell types and different functional classes with an exception of insulators in islets, which is due to CTCF ChIP-seq quality in islets. (B) De novo motif enrichment results for regions predicted as insulators by CoRE-ATAC but were not annotated as insulators by ChromHMM. Note that these regions are significantly enriched for the CTCF motif (0.983 similarity), suggesting that CoRE-ATAC insulator predictions are functionally relevant.(C) Distribution of CoRE-ATAC predictions. Prediction distributions are similar to those observed by ChromHMM state annotations. (D) Comparison of CoRE-ATAC to baseline/naive predictions based on thresholds for distance to TSS, MACS2 FDR, and number of CTCF motifs. CoRE-ATAC improves upon baseline performances. (E) CoRE-ATAC performances for i) predictions overlapping regions used in model training (O), and ii) predictions within regions that are on held-out test chromosomes (E). Note the performance similarity between these two prediction categories across all classes. (F) CoRE-ATAC model performances (top) and the average number of promoters and enhancers observed (bottom) by cell-type-specificity. We observed that CoRE-ATAC was more effective in predicting common promoters and cell-type-specific enhancers, for which we had more examples represented in the data. CoRE-ATAC’s ability to predict cell-type-specific enhancers demonstrates its usefulness for interrogating individual and cell-type-specific enhancers.