Discovering interpretable models of scientific image data with deep learning

In this study, we demonstrate the possibility of finding interpretable, domain-appropriate models of biological images, and propose that such a strategy can be used to derive scientific insight in domains involving raw data. This is achieved by the novel, concerted application of existing methods, namely, disentangled representation learning, sparse deep neural network training and symbolic regression. We demonstrate their relevance to the field of bioimaging using a well-studied test problem of classifying cell states in microscopy data. We find that such methods can produce highly parsimonious models that achieve ~ 98% of the accuracy of black-box benchmark models, with a tiny fraction of the complexity, and greater domain-appropriateness, as tested by adversarial attacks. As such, we provide proof of concept that interpretable, high-performing models can be used to produce scientific explanations of some underlying biological phenomenon.