Learning Composition Rules for Mammalian Circuits with Neural Attention

Abstract

The expression of each gene in mammalian cells is controlled by regulatory sequences called enhancers. Regulatory logic encoded at enhancers is interpreted by transcription factors (TFs), which recognize individual ‘words’ in the genome. Here we describe a neural network with an attention mechanism that learns to discriminate enhancers from random genomic sequences. We distill the parameters learned by our model to identify combinations of TF target sequences that signal activation of a gene in response to specific cellular stimuli. Download paper here