A computational approach for predicting host specificity of adenoviruses

Adenoviruses are large viruses that consist of a complex protein cover enclosing its DNA sequence and core proteins. They belong to a large family called Adenoviridae with over hundred known members that can infect extensive range of vertebrates. The cross-species transmissions of adenoviruses between human, non-human primate, bat, cat, dog, pig, sheep and goat species are known. Understanding adenovirus host specificity and prediction of cross-species virus transmissions are important in the management of infections caused by adenoviruses. Nearly fifty of known adenoviruses infect human and can cause, sometimes serious, gastrointestinal, respiratory, urinary, and corneal infections. For the infection to start, adenoviruses establish a complex set of interactions with host cells, typically initiated by a binding between a host cell receptor and adenovirus fiber protein. Hence, understanding protein-protein interactions (PPIs) between receptor and fiber proteins, is an essential step towards developing a model to predict host specificity of a given adenovirus. In this work, our goal was to create a pipeline to predict host specificity of an adenovirus. For this purpose, we, first, conducted a literature review and identified all human receptors that are involved in the uptake of adenoviruses. Next, we computed human receptor orthologs from other adenovirus host species. Then, we computed a library of fiber proteins from various adenoviruses infecting different host species. Finally, we developed a meta-classifier predicting host specificity of adenoviruses. The data objects used in the classifier were the predictions computed for host receptor-fiber protein interactions by the existing PPI prediction methods, namely VHPPI, HOPITOR and DeNovo. For these data objects, four different class labels were computed based on candidate host and adenovirus's known host sharing the same taxa at taxonomic levels species, genus, family, and order. Several ensemble machine learning approaches were tested in meta-classification. The stacking ensemble classifiers gave the best accuracies. For stacking, we tested J48 decision tree, naïve bayes, multilayer perceptron, random forest, and support vector machine algorithms. The performance of classifiers were close the each other with J48 decision tree being slightly better (True positive rate:71.5%, true negative rate 71.5% and 71.0%). In this work, we demonstrated the potential and accuracy of using an ensemble PPI prediction methods in prediction of host specificity of adenoviruses. The diversity of Adenoviridae fiber and known host receptor proteins posed a challenge in the development of the pipeline. Although our results are quite promising, they also suggest the need for more accurate cross-species PPI predictions. In the future, we anticipate our work can be leveraged and improved for early detection of host shifts or in assessment of host specificity of adenoviruses. Furthermore, our approach has the potential to be extended to other virus families beyond Adenoviriade and this is particularly significant, given the rise in viral disease epidemics involving host shifts in recent years.