Interpretable detection of novel human viruses from genome sequencing data

Jakub M Bartoszewicz; Anja Seidel; Bernhard Y Renard

Several computational, genome-based methods exist that allow to predict the host-range of a bacteriophage (a bacteriainfecting virus). A selection of composition-based and alignment-based approaches has been presented in an extensive review by Edwards et al. (2016) . Prediction of eukariotic host tropism (including humans) based on known protein sequences was shown for the influenza A virus (Eng et al., 2014) . Two recent studies employ k-mer based, k-NN classifiers and deep learning (Mock et al., 2019) to predict host range for a small set of three well-studied species directly from viral sequences. While those approaches are limited to those particular species and do not scale to viral host-range prediction in general, the Host Taxon Predictor (HTP) (Gałan et al., 2019) uses logistic regression and support vector machines to predict if a novel virus infects bacteria, plants, vertebrates or arthropods. Yet, the authors argue that it is not possible to use HTP in a read-based manner; it requires long sequences of at least 3,000 nucleotides. This is incompatible with modern metagenomic next-generation sequencing workflows, where the DNA reads obtained are at least 10-20 times shorter. Another study used gradient boosting machines to predict reservoir hosts and transmission via arthropod vectors for known human-infecting viruses (Babayan et al., 2018) . Zhang et al. (2019) designed several classifiers explicitly predicting whether a new virus can potentially infect humans. Their best model, a k-NN classifier, uses k-mer frequencies as features representing the query sequence and can yield predictions for sequences as short as 500 base pairs (bp). It worked also with 150bp-long reads from real DNA sequencing runs, although in this case the reads originated also from the viruses present in the training set (and were therefore not "novel").