Predicting Influenza A Tropism with End-to-End Learning of Deep Networks

Area of Science:

• Virology
• Bioinformatics
• Machine Learning

Background:

• Host specificity (tropism) is crucial for understanding virus infectivity, disease diagnosis, and epidemic control.
• Predicting virus phenotype from genomic sequences is challenging but vital for public health.
• While machine learning has been used, deep learning methods like CNNs have not been applied to host tropism prediction.

Purpose of the Study:

• To design and evaluate deep convolutional neural network (CNN) models for predicting host tropism in influenza A viruses.
• To assess the performance of CNNs compared to existing machine learning approaches for this task.
• To explore the utility of CNNs in visualizing viral strain similarity.

Main Methods:

• Development of deep CNN models utilizing viral protein sequences.
• Training and testing models for binary prediction of host tropism (human vs. avian influenza A viruses).
• Integration of CNN models with principal component analysis for similarity analysis.

Main Results:

• Deep CNN models achieved high accuracy (99% mean accuracy) in predicting host tropism, comparable to existing methods.
• The CNN approach enables end-to-end learning, eliminating the need for handcrafted features.
• CNNs combined with principal component analysis effectively quantified and visualized viral strain similarity.

Conclusions:

• Deep CNNs are a powerful and effective tool for predicting virus host tropism from genomic data.
• This method provides an automated and efficient approach for disease diagnosis and epidemic response.
• The findings highlight the potential of deep learning in advancing viral bioinformatics and public health strategies.