Tissue Tropism and Transmission Ecology Predict Virulence of Human RNA Viruses

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

Therefore, we also included an additional 'viraemia' category in this variable to indicate only 3 5 0 blood presence was known. Binary variables were also constructed denoting whether viruses were ever known to utilise a) more than one transmission route/tissue tropism, and b) each reproductive, sensory, skin, muscular and endocrine tropism). Human-to-human transmissibility was specified using infectivity/transmissibility levels, based Binary variables were also sourced as to whether infection was known within a) humans only, The copyright holder for this preprint (which was not peer-reviewed) is the author/funder.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 pertained to natural or unintentional artificially-acquired human infection only and data from taxonomy was included in analyses by specifying both genome type and taxonomic family as Comparative risk factor analyses were firstly carried out by constructing a classification tree learning models that aim to optimally classify data points into their correct category of the tree was pruned back to the optimal branching size, taken as most common consensus 3 8 9 size over 1000 repeats of 10-fold cross-validation. To validate the predictive power of the 3 9 0 classification tree, predictions of virulence rating were generated when applied to the test set. Tree accuracy was then calculated comparing the proportion of correct predictions compared 3 9 2 to literature-assigned ratings (assuming these to be 100% accurate as the 'gold standard' or 3 9 3 'ground truth'). As virulence ratings were imbalanced (i.e. only a minority of viruses cause 3 9 4 severe disease, so correct nonsevere classifications are likely to be achieved by chance), predicted 'nonsevere' for all viruses. Additional diagnostics of interest (sensitivity, specificity, 3 9 7 negative predictive value, and True Skill Statistic [60]) were also obtained. Due to their high structuring, random forest models cannot give a simple parametric predictor 4 1 6 effect size and direction (e.g., an odds ratio). Instead, potential virulence risk factors were 4 1 7 evaluated using two metrics: variable importance and partial dependence. Variable importance is calculated as the mean decrease in Gini impurity following tree splits on the predictor and can be considered as how informative the risk factor was towards correctly 4 2 0 predicting virulence. Partial dependence is calculated as the mean relative change in log- reflect any complex risk factor interactions. Therefore, to test hypotheses regarding virulence 4 2 5 risk factors, we present both random forest partial dependences and the less robust but more and Feifei Zhang for assistance in data collection. 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 15. 38. Allen T, Murray KA, Zambrana-Torrelio C, Morse SS, Rondinini C, Marco MD, et al.