Fig. 2: Deep learning models predict accessibility and regulatory logic across cell types.

a, Schematic overview of the architecture of the deep learning (DL) models used to predict chromatin accessibility across cell types from the whole dataset. The models are trained for each cell type using cell-type-specific peak sets and process their one-hot encoded DNA sequences through a series of dilated convolutions with residual connections. The final output generates count predictions and profile log-likelihood via a dense layer or an additional convolutional layer, respectively. b, Coverage plots displaying observed (colored area) and predicted (black line) accessibility for cell-type-enriched genomic regions. Each peak is predicted using a different cell-type-specific model (rows). c, Bar plots reporting Pearson’s correlations between observed and predicted counts for each cell-type-specific model. Error bars report standard deviations calculated from the correlations obtained with fivefold cross-validations. On the right, the number of peaks used as input for each model is reported. d, Heat map highlighting the frequency (scaled by row) of top occurring motifs in sets of peaks called for each cell type. On the right, examples of transcription factors recognizing such sequences are reported. e, Sequence logos for representative DNA binding sites and their matching transcription factors. f, Nucleotide contribution scores calculated by each cell-type-specific model for representative genomic regions. Motifs identified carrying higher importance are highlighted (gray), and the transcription factors recognizing such sequences are reported under the colored area. The limits on the y axis are reported in brackets. g, Observed (colored area) and predicted (black line) accessibility profiles of the regions reported in f.