Fig. 6: Deep neural network (DNN) models validate regulatory elements in regulons.

a We train a DNN model on the ATAC-seq signals from cluster 1 and cluster 3 cells, respectively. We compare the predicted accessibility of either the wildtype sequence or the occluded sequence in the regulatory elements from the regulons inferred by Epiregulon. b Normalized changes in predicted accessibility if we occlude the sequences found in the regulatory elements of GATA6 and NKX2-1 regulons in either the DNN model trained on cluster 1 or cluster 3 cells. The number of regulatory elements for GATA6 is 402, and the number of regulatory elements for NKX2-1 is 134. Boxplots presented as median values ± 25%. Lower whisker is the smallest observation ≥25% quantile −1.5 × interquantile range (IQR). Upper whisker represents the largest observation ≤75%  + 1.5 × IQR. c Each cell was identified by the HTO tag corresponding to the well receiving virus encoding GATA6 or mNeonGreen, and this information served as the true cell labels. For each TF, a cell was classified into either expressing GATA6 or not. GATA6 ChIP-seq was obtained by merging ChIP-seq from ChIP-atlas and ENCODE. ChIP+ motif refers to ChIP-seq peaks that contain GATA6 motifs. We trained a DNN model on cells expressing GATA6 (cluster 1) using Basenji2 and predicted an importance score for each motif based on the difference between the original sequence and the motif occluded sequence. We filtered for those motifs with importance scores higher than the quartiles of scores.