Interpretable detection of novel human viruses from genome sequencing data

Jakub M Bartoszewicz; Anja Seidel; Bernhard Y Renard

This protocol allowed us to test the impact of defining the negative class, while using the exactly same data as representatives of the positive class. We used three training and validation sets ("All", "Stratified", and "Chordata"), representing the fully open-view setting, a setting more balanced with regard to the host taxonomy, and a setting focused on cases most likely to be clinically relevant. In each setting, the validation set matched the composition of the training set. The evaluation was performed using all five test sets to gain a more detailed insight on the effects of negative class definition on the prediction performance.