Interpretable detection of novel human viruses from genome sequencing data

Jakub M Bartoszewicz; Anja Seidel; Bernhard Y Renard

We ran a similar analysis of S. aureus using the built-in DeePaC models (Bartoszewicz et al., 2019) and our interpretation workflow. While a viral genome contains usually . CC-BY-ND 4.0 International license author/funder. It is made available under a The copyright holder for this preprint (which was not peer-reviewed) is the . https://doi.org/10.1101/2020.01.29.925354 doi: bioRxiv preprint only a handful of genes, by compiling a ranking of 870 annotated genes of the analyzed S. aureus strain we could test if the high-ranking regions are indeed associated with pathogenicity. Indeed, out of three top-ranking genes, sarR and sspB are directly engaged in virulence, while hupB regulates expression of virulence-involved genes in many pathogens (Stojkova et al., 2019) . Fig. 2b presents a GWPA plot for the whole genome of the SARS-CoV-2 coronavirus. We highlighted the score peaks aligning with the gene encoding the spike protein, which plays a significant role in host entry (Li, 2016) , as well as the E and N genes, which were scored the highest (apart from an unconfirmed ORF10 of just 38aa downstream of N) by the CNN and the LSTM, respectively. Fig. 2c shows the nucleotide-level contributions in a small peak within the receptor-binding domain of the S protein, crucial for recognizing the host cell. The domain location was predicted with CD-search (Marchler-Bauer et al., 2017) . While this could suggest a host adaptation, more research is needed.