Transmissibility of emerging viral zoonoses

Abstract

Effective public health research and preparedness requires an accurate understanding of which virus species possess or are at risk of developing human transmissibility. Unfortunately, our ability to identify these viruses is limited by gaps in disease surveillance and an incomplete understanding of the process of viral adaptation. By fitting boosted regression trees to data on 224 human viruses and their associated traits, we developed a model that predicts the human transmission ability of zoonotic viruses with over 84% accuracy. This model identifies several viruses that may have an undocumented capacity for transmission between humans. Viral traits that predicted human transmissibility included infection of nonhuman primates, the absence of a lipid envelope, and detection in the human nervous system and respiratory tract. This predictive model can be used to prioritize high-risk viruses for future research and surveillance, and could inform an integrated early warning system for emerging infectious diseases.

Fig 1. Relative influence of predictors, averaged across 40 models.

For each of the 40 boosted regression tree models in our primary ensemble, the normalized relative influence of each predictor variable was computed using Friedman’s algorithm ⁴⁵. This figure shows the average of these scores (mean relative influence) for each predictor variable in our dataset that was included in at least one model of the ensemble (mean relative influence > 0). Horizontal lines represent the interval formed by ± 1 standard deviations. Exact relative influence values are listed in S2 Table.

Fig 2. Predicted viral risk index.

This figure contrasts the observed transmission ability of all 224 viruses in our dataset (red = human-to-human transmission observed, blue = human-to-human transmission not observed) with their average model-predicted response probabilities, as assigned by the primary boosted regression tree models. This model ensemble accurately discriminates transmissible and non-transmissible viruses, as illustrated by the lack of “overlap” of the two groups in the rank-ordering. The highest ranked viruses that are not currently known to be transmissible between humans were Carnivore amdoparvovirus 1, Hendra virus, Cardiovirus A, Rosavirus A, Human T-lymphotropic viruses 3 & 4 (HTLV-3/4), and Simian Foamy virus. Crimean-Congo haemorrhagic fever virus was the lowest ranked species for human-to-human transmission has been documented.

Fig 3. Variable partial dependence plots.

Partial dependence plots show how the model-predicted probability that a virus is able to spread between humans is affected by individual viral traits when the effects all other predictors are controlled for. Dark lines represent the median predicted transmission probability across the 40 boosted regression tree models of the primary ensemble, while shaded regions represent the corresponding 95% confidence interval. Viral features are ordered by their mean relative influence within the primary boosted regression tree models from left to right, then top to bottom. Predictor variables with a mean relative influence score of 0 are not included in this figure. Trait definitions and exact relative influence scores are given in S1 and S2 Tables, respectively.