HostNet: improved sequence representation in deep neural networks for virus-host prediction

Area of Science:

• Virology
• Bioinformatics
• Computational Biology

Background:

• Emerging viruses necessitate rapid host identification for public health and animal welfare.
• Traditional virus host range determination is labor-intensive and time-consuming.
• Accurate computational tools are crucial for predicting hosts of novel viruses and controlling infectious diseases.

Purpose of the Study:

• To develop a highly accurate computational tool for predicting virus-host associations.
• To address challenges in virus-host prediction, including data deficiency and imbalance.

Main Methods:

• Introduced HostNet, a deep learning framework employing a Transformer-CNN-BiGRU architecture.
• Implemented two enhanced sequence representation modules: k-mer to vector and an adaptive sliding window.
• Pre-trained k-mer representations to mitigate data deficiency and used adaptive sliding windows to handle varying sequence lengths and data imbalance.

Main Results:

• HostNet demonstrated superior performance over state-of-the-art methods on benchmark and in-house datasets.
• Achieved higher host-prediction accuracies and F1 scores.
• Enhanced sequence representation modules improved training generalization, stability, and performance on challenging viral classes.

Conclusions:

• HostNet is a robust framework for predicting virus hosts from genomic sequences, effectively handling sparse and variable-length data.
• The framework shows significant potential as a valuable tool in virology and public health.
• Deep neural network-based virus-host prediction offers an accurate and efficient approach for disease prevention and vaccine development.