Enhanced Integrated Gradients: improving interpretability of deep learning models using splicing codes as a case study

摘要

The high accuracy of deep neural networks (DNN) in areas such as computer vision, natural language processing, and robotics has led to the fast adaptation of DNN in biomedical research. In genomics, deep learning models have outperformed previous state-of-the-art methods on tasks such as predicting protein binding sites [1] or mRNA alternative splicing from genomic sequence features [2]. However, the interpretation of these complex models remains a challenge [3, 4]. Approaches to model interpretation include approximation with simpler models [5], identifying the most influential samples [6], or finding the most relevant features for a specific sample or a task by a variety of metrics [7]. Here, we focus on the last approach, which is naturally appealing for biomedical tasks. In this context, interpretability is defined as attributing the prediction of a DNN to its input features. We focus on the recently developed method called Integrated Gradients (IG) [8]. Both IG and DeepLIFT, which was recently used for protein DNA binding sites [9], identify features associated with a model’s prediction for a sample with respect to a baseline. The usage of a baseline is appealing as it serves as the model’s proxy to human counterfactual intuition. This implies that humans assign blame for difference in two entities on attributes that are present in one entity but absent in the other. IG computes feature attribution by aggregating gradients along a linear path between the sample and the baseline. Compared to other interpretation methods, IG offers two desirable theoretical guarantees motivating its usage. The first is sensitivity, which states that for every input and baseline that differ in one feature but have different predictions, the method will give a non-zero attribution for that differing feature. The second is implementation invariance, which states that regardless of network architecture, if two models are functionally equivalent (same output given any input), then their feature attributions will also be equivalent (see more details in [8]).