Flu-CNN: predicting host tropism of influenza A viruses via character-level convolutional networks

Abstract

Throughout history, Influenza A viruses (IAVs) have caused significant harm and catastrophic pandemics. The presence of host barriers results in viral host tropism, where infected hosts are subject to strict restrictions due to the hindered spread of viruses across hosts. Therefore, the identification of host tropism of IAVs, particularly in humans, is crucial to preventing the cross-host transmission of avian viruses and their outbreaks in humans. Nevertheless, efficiently and effectively identifying host tropism, especially for early host susceptibility warnings based on viral genome sequences during outbreak onset, remains challenging. To address this challenge, we propose Flu-CNN, a deep neural network model based on classical character-level convolutional networks. By analyzing the genomic segments of IAVs, Flu-CNN can accurately identify the host tropism, with a particular focus on avian influenza viruses that may infect humans. According to our experimental evaluations, Flu-CNN achieved an accuracy of 99% in identifying virus hosts via only a single genomic segment, even for subtypes with a relatively small number of viral strains such as H5N1, H7N9, and H9N2. The superiority of Flu-CNN demonstrates its effectiveness in screening for critical amino acid mutations, which is important to host adaptation, and zoonotic risk prediction of viral strains. Flu-CNN is a valuable tool for identifying evolutionary characterization, monitoring potential outbreaks, and preventing epidemical spreads of IAVs, which contribute to the effective surveillance of influenza A viruses.