LINA: A Linearizing Neural Network Architecture for Accurate First-Order and Second-Order Interpretations

While neural networks can provide high predictive performance, it was a challenge to identify the salient features and important feature interactions used for their predictions. This represented a key hurdle for deploying neural networks in many biomedical applications that require interpretability, including predictive genomics. In this paper, linearizing neural network architecture (LINA) was developed here to provide both the first-order and the second-order interpretations on both the instance-wise and the model-wise levels. LINA combines the representational capacity of a deep inner attention neural network with a linearized intermediate representation for model interpretation. In comparison with DeepLIFT, LIME, Grad*Input and L2X, the first-order interpretation of LINA had better Spearman correlation with the ground-truth importance rankings of features in synthetic datasets. In comparison with NID and GEH, the second-order interpretation results from LINA achieved better precision for identification of the ground-truth feature interactions in synthetic datasets. These algorithms were further benchmarked using predictive genomics as a real-world application. LINA identified larger numbers of important single nucleotide polymorphisms (SNPs) and salient SNP interactions than the other algorithms at given false discovery rates. The results showed accurate and versatile model interpretation using LINA.