Improving interpretability of deep learning models: splicing codes as a case study

Despite the success and fast adaptation of deep learning models in a wide range of fields, lack of interpretability remains an issue, especially in biomedical domains. A recent promising method to address this limitation is Integrated Gradients (IG), which identifies features associated with a prediction by traversing a linear path from a baseline to a sample. We extend IG with nonlinear paths, embedding in latent space, alternative baselines, and a framework to identify important features which make it suitable for interpretation of deep models for genomics.