Identifying and prioritizing potential human-infecting viruses from their genome sequences

Area of Science:

• Virology
• Genomics
• Machine Learning
• Epidemiology

Background:

• Identifying zoonotic viruses early is challenging, hindering preparedness for potential outbreaks.
• Genomic data is increasingly vital for virus discovery, but biological knowledge of new viruses is often limited.

Purpose of the Study:

• To develop machine learning models for identifying potential zoonotic viruses (zoonoses) using only genomic signatures of host range.
• To prioritize high-risk viruses for early investigation and enhance outbreak preparedness.

Main Methods:

• Trained machine learning models on a dataset of 861 viral species with known zoonotic status, using genomic host range signatures.
• Compared model performance against phylogenetic relatedness models.
• Applied models to a separate set of 645 animal-associated viruses and evaluated predictions for SARS-CoV-2.

Main Results:

• The machine learning approach achieved an AUC of 0.773, outperforming phylogenetic models.
• Identified high-risk viruses within families containing few human-infecting species and potential novel zoonoses.
• Predicted elevated zoonotic risk in viruses from nonhuman primates.
• Successfully identified SARS-CoV-2 as a high-risk coronavirus without prior knowledge of related zoonoses.

Conclusions:

• Genome-based analysis offers a rapid, cost-effective method for zoonotic risk assessment and virus surveillance.
• Machine learning models utilizing viral genome features can predict zoonotic potential, aiding in proactive public health strategies.
• This approach enhances the feasibility of downstream biological and ecological characterization of viruses with pandemic potential.