Fig. 2

Structure and sequence predictions of forward and backward neural networks. a Distance-normalized Hi-C contacts predicted from the forward-model CNN in a 5 Mbp region of the test data set that was left out of the fitting procedure (chrom 3R 15–20 Mbp). A correlation of 0.68 between the original and predicted counts was obtained. b Weights of the sigmoid-activated convolutional filter applied to the chromatin factor sequences in order to generate the 1D sequence profile. c Histogram of the values of the 1D sequence obtained after the convolutional layer for all sites, transcriptionally active sites and transcriptionally inactive sites. d An independent DNN (backward model) was built to predict the 1D sequence from Hi-C data. From bottom to top, multiple chromatin factors were converted to this 1D sequence by running them through the convolutional filter of the forward model. Then, the backward model was used to predict those 1D sequences from Hi-C contacts. The predicted 1D sequence (red) and the original 1D sequence (below) showed a correlation of 0.73. (Top) A Gaussian-smoothed version of the original sequence showed a correlation of 0.93 with the predicted sequence from the backward model. The genomic region shown in (d) is the same 5 Mbp region of the test set shown in (a)