Interpretable detection of novel human viruses from genome sequencing data

Jakub M Bartoszewicz; Anja Seidel; Bernhard Y Renard

(c) Spike protein gene, a small peak (positions 22,595-22,669, dashed line in Fig. 2b ) within the receptor-binding domain (predictied by CD-search). Binding to the receptor is crucial for entry to the host cell. Local host adaptation could help switch hosts between the animal reservoir and humans. describe how the filter reacts to presented data. It seems that the assumption of nucleotide independence -which is crucial for treating DeepLIFT as a method of estimating Shapley values for input nucleotides -is broken. Indeed, k-mer distribution profiles are frequently used features for modelling DNA sequences (as shown also by the dimershuffling method of generating reference sequences proposed by Shrikumar et al. (2019a) ). However, DeepLIFT's multiple successful applications in genomics indicate that the assumption probably holds approximately. We see information content and DeepLIFT's contribution values as two complementary channels that can be jointly visualized for better interpretability and explainability of CNNs in genomics.