Interpretable detection of novel human viruses from genome sequencing data

Jakub M Bartoszewicz; Anja Seidel; Bernhard Y Renard

While defining a human-infecting class is relatively straightforward, the reference negative class may be conceptualized in a variety of ways. The broadest definition takes all non-human viruses into account, including bacteriophages (bacterial viruses). This is especially important, as most of known bacteriophages are DNA viruses, while many important human (and animal) viruses are RNA viruses. One could expect that the multitude of available bacteriophage genomes dominating the negative class could lower the prediction performance on viruses similar to those infecting humans. This offers an open-view approach covering a wider part of the sequence space, but may lead to misclassification of potentially dangerous mammalian or avian viruses. As they are often involved in clinically relevant host-switching events, a stricter approach must also be considered. In this case, the negative class comprises only viruses infecting Chordata (a group containing vertebrates and closely related taxa). Two intermediate approaches consider all eukaryotic viruses (including plant and fungi viruses), or only animal-infecting viruses. This amounts to four nested host sets: "All" (8,187 non-human viruses), "Eukaryota" (5,114 viruses), "Metazoa" (2,942 viruses) and "Chordata" (2,078 viruses). Auxiliary sets containing only non-eukaryotic viruses ("non-Eukaryota"), non-animal eukaryotic viruses ("non-Metazoa Eukaryota") etc. can be easily constructed by set subtraction.