Interpretable detection of novel human viruses from genome sequencing data

Jakub M Bartoszewicz; Anja Seidel; Bernhard Y Renard

Interpretability and explainability of deep learning models for genomics is crucial for their wide-spread adoption, as it is necessary for delivering trustworthy and actionable results. Convolutional filters can be visualized by forward-passing multiple sequences through the network and extracting the most-activating subsequences (Alipanahi et al., 2015) to create a position weight matrix (PWM) which can be visualized as a sequence logo (Schneider & Stephens, 1990; Crooks et al., 2004) . Direct optimization of input sequences is problematic, as it results in generating a dense matrix even though the input sequences are one-hot encoded (Lanchantin et al., 2016; . This problem can be alleviated with Integrated Gradients (Sundararajan et al., 2016; Jha et al., 2019) or DeepLIFT, which propagates activation differences relative to a selected reference back to the input, reducing the computational overhead of obtaining accurate gradients (Shrikumar et al., 2019a) . If a reference of all-zeros is used, the method is analogous to Layer-wise Relevance Propagation (Bach et al., 2015) . DeepLIFT is an additive feature attribution method, and may used to approximate Shapley values if the input features are independent (Lundberg & Lee, 2017) . TF-MoDISco (Shrikumar et al., 2019b) uses DeepLIFT to discover consolidated, biologically meaningful DNA motifs (transcription factor binding sites).