Adversarial training improves model interpretability in single-cell RNA-seq analysis

Abstract

For predictive computational models to be considered reliable in crucial areas such as biology and medicine, it is essential for them to be accurate, robust, and interpretable. A sufficiently robust model should not have its output affected significantly by a slight change in the input. Also, these models should be able to explain how a decision is made. Efforts have been made to improve the robustness and interpretability of these models as independent challenges, however, the effect of robustness and interpretability on each other is poorly understood. Here, we show that predicting cell type based on single-cell RNA-seq data is more robust by adversarially training a deep learning model. Surprisingly, we find this also leads to improved model interpretability, as measured by identifying genes important for classification. We believe that adversarial training will be generally useful to improve deep learning robustness and interpretability, thereby facilitating biological discovery.