Interpretable detection of novel human viruses from genome sequencing data

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Rapid viral evolution necessitates advanced host prediction for biosecurity and biosafety.
• Current bioinformatics workflows struggle to detect novel human-infecting viruses.
• Accurate viral host prediction is crucial for pandemic preparedness.

Purpose of the Study:

• To develop and validate a deep learning approach for predicting viral human host tropism directly from next-generation sequencing (NGS) reads.
• To compare the performance of deep neural networks against traditional machine learning and homology-based methods.
• To introduce interpretability tools for understanding model decisions and identifying genomic features associated with host tropism.

Main Methods:

• Utilized deep neural network architectures for direct prediction of human infectivity from NGS data.
• Benchmarked deep learning models against shallow machine learning algorithms and homology-based approaches.
• Developed and applied novel convolutional filter visualization techniques for interpretability.
• Implemented methods as accessible software packages for broad usability.

Main Results:

• Deep neural networks significantly outperformed existing methods, reducing error rates by 50%.
• Models demonstrated strong generalization capabilities to novel viruses distant from training data.
• Interpretability tools successfully mapped genomic regions to the infectious phenotype, aiding in the identification of key viral features.
• Demonstrated utility in identifying regions of interest in known pathogens like SARS-CoV-2.

Conclusions:

• Deep learning offers a powerful and accurate solution for viral host prediction from NGS data.
• The developed interpretability tools enhance understanding of viral-host interactions and facilitate novel pathogen discovery.
• Accessible implementation of these methods empowers researchers to improve biosecurity and biosafety surveillance.
• This approach is vital for early detection and characterization of emerging infectious threats.