Learning distance-dependent motif interactions: an interpretable CNN model of genomic events

In most biological studies, prediction is used primarily to validate the model; the real quest is to understand the underlying phenomenon. Therefore, interpretable deep models for biological studies are required. Here, we propose (the -parameter eplainable Motif framework), a new architecture that learns biological motifs and their distance-dependent context through explicitly interpretable parameters that are immediately understood by a biologist. This makes more than a decision-support tool; it is also a hypothesis-generating tool designed to advance knowledge in the field. We demonstrate the utility of our model by learning distance-dependent motif interactions for two biological problems: transcription initiation and RNA splicing.