Fig. 1
Overview of Bichrom’s neural network architecture and approach. a Bichrom’s bimodal sequence and preexisting chromatin network consists of two sub-networks: the sequence sub-network (BichromSEQ) which uses one-hot encoded DNA sequence as input; and the chromatin sub-network (BichromCHR) which uses binned normalized tag counts from chromatin experiments such as ATAC-seq and histone modification ChIP-seq. The sequence and chromatin sub-network activations embed the training data into a lower-dimensional plane, which is then used by a sigmoid-activated node for TF binding label classification (i.e., bound/unbound). b Overview of Bichrom’s training strategy. BichromSEQ is trained using training batches within which positive and negative training samples are matched in their prior accessibility status. The weights of the convolutional and LSTM layers of BichromSEQ are fixed, and Bichrom is trained using both sequence and preexisting chromatin data. c Overview of the Ascl1 data: Ascl1 expression is induced in mouse embryoid bodies (mEBs) using a Dox-inducible promoter and Ascl1 binding is measured 12 h post induction. Bichrom training uses 12 prior chromatin datasets from mEB and mES cell types