Fully Interpretable Deep Learning Model of Transcriptional Control

Abstract The universal expressibility assumption of Deep Neural Networks (DNNs) is the key motivation behind recent work in the system biology community to employ DNNs to solve important problems in functional genomics and molecular genetics. Because of the black box nature of DNNs, such assumptions, while useful in practice, are unsatisfactory for scientific analysis. In this paper, we give an example of a DNN in which every layer is interpretable. Moreover, this DNN is biologically validated and predictive. We derive our DNN from a systems biology model that was not previously recognized as having a DNN structure. This DNN is concerned with a key unsolved biological problem, which is to understand the DNA regulatory code which controls how genes in multicellular organisms are turned on and off. Although we apply our DNN to data from the early embryo of the fruit fly Drosophila, this system serves as a testbed for analysis of much larger data sets obtained by systems biology studies on a genomic scale.