Translational adaptation of human viruses to the tissues they infect

Xavier Hernandez-Alias; Martin Schaefer; Luis Serrano

Next, from the perspective of the translational selection hypothesis, we would expect that viral proteins are translationally adapted to their target tissues. In consequence, we tried to test our hypothesis using a completely blind and unbiased random forest classifier, which applies machine learning in order to predict the tropism of each viral protein based on the SDA to different tissues (see Methods). The resulting performance of the models, based on the Area Under the Curve (AUC) of their Receiver Operating Characteristic (ROC) curves, ranges between 0.74-0.91 ( Fig. 2A) , clearly higher than the no-skill model of 0.5. Similar results are also obtained from complementary prediction performance metrics such as Precision-Recall curves ( Fig. 2A) . These results indicate that our machine learning model is able to predict the tropism of a viral protein based on its SDA to tissues with high accuracy. In concordance, a linear discriminant analysis of the average SDA of each virus across tissues can similarly separate different clusters of viral tropism based on their translational efficiencies (Extended Data Fig. 1 ).