Architecture of CoNSEPT. First the input enhancers are scanned with user-defined PWMs and passed to a pooling layer to extract the strongest matches. Next, TF profiles are integrated (multiplied) with the extracted scores and pair-wise features are built according to the user-defined TF–TF interactions. Each pair-wise feature is passed to a separate kernel capturing short-range regulatory interactions between pair of TFs. (Note: a TF pair may be homotypic.) The concatenated output of interaction convolutional kernels is passed to a sequence of activation functions (σ) and optional additional convolutional layers. The final activated output of the convolutional layers is passed to a dropout layer followed by a fully-connected layer and an activation function to predict the expression.
CoNSEPT is a tool to predict gene expression in various cis and trans contexts. Inputs to CoNSEPT are enhancer sequence, transcription factor levels in one or many trans conditions, TF motifs (PWMs), and any prior knowledge of TF-TF interactions.
