Tissue Tropism and Transmission Ecology Predict Virulence of Human RNA Viruses

Liam Brierley; Amy B. Pedersen; Mark E. J. Woolhouse

To quantify the effects of the most informative risk factors, partial dependences were Table) . Averaging across other predictors, presented the lowest risk. An increased probability of severe virulence was also observed for 1 4 5 viruses transmitted by direct contact or respiratory routes, and those with known but limited 1 4 6 human-to-human transmissibility. Although the single classification tree model predicted the training set well, it did not appear 1 5 0 generalisable to novel data within the test set. The single tree correctly predicted virulence gave better predictive accuracy, correctly predicting virulence ratings for 28 of 31 test set accuracy (exact binomial one-tailed test, p = 0.025). The random forest also achieved 1 5 7 superior performance when considering sensitivity, specificity, True Skill Statistic, and the 'severe'-rated viruses ( Table 1 ). The random forest also outperformed the classification tree in All misclassifications from the random forest occurred within the genus Flavivirus (S2 Table) . Within the test set, there were two flaviruses rated as severe from literature protocols that The copyright holder for this preprint (which was not peer-reviewed) is the author/funder. It . https://doi.org/10.1101/581512 doi: bioRxiv preprint Tropism and Transmission Ecology Predict Viral Virulence -Brierley et al. 9 were predicted to be nonsevere (Rio Bravo virus, Yellow fever virus), and one nonsevere 1 6 4 flavivirus predicted to be severe (Usutu virus). The observed predictor importances and risk factor directions were robust to constructing 1 6 7 random forest models for subsets of viruses, removing those with low-certainty data or data 1 6 8 from serological evidence only (S1 Fig, S2 Fig) , and similar performance diagnostics were obtained (S5 Table) . Redefining our virulence measure to integrate information on known 1 7 0 fatalities and differences with subspecies or strains in an ordinal ranking system (S5 Table) 1 7 1 did not improve predictive performance (S6 Table) . Using alternative virulence 1 7 2 measurements, the most informative variables and virus traits predicting severity showed 'severe' virulence were widened, hepatic tropism became an informative predictor towards 1 7 5 disease severity. We present the first comparative analysis of virulence across all known human RNA virus 1 7 8 species to our knowledge. We find that disease severity is non-randomly distributed across virus families and that beyond taxonomy, severe disease is predicted by risk factors of tissue Primary tissue tropism was the most informative non-taxonomic risk factor (Fig 4) and the first 1 8 9 split criteria in the classification tree (Fig 2) , with specific neural tropism and generalised predicted how tissue tropism should influence virulence. The identified risk factor tropisms 1 9 2 could be explainable as a simple function of pathology occurring in multiple or sensitive 1 9 3 tissues respectively, increasing intensity of clinical disease. However, it has been suggested 1 9 4 that an excessive, non-adapted virulence may result if infections occur within non-target tissues that do not contribute to transmission [28] . Furthermore, the evolutionary determinants of tissue tropism themselves are not well understood [29] . Tissue tropism should be a key 1 9 7 consideration for future comparative and evolutionary modelling efforts. The copyright holder for this preprint (which was not peer-reviewed) is the author/funder. It . https