A multi-task CNN learning model for taxonomic assignment of human viruses

Area of Science:

• Bioinformatics
• Genomics
• Virology

Background:

• Accurate taxonomic assignment of human viral pathogens from sequencing reads is crucial for public health.
• Existing alignment and k-mer based tools may struggle with significantly divergent viral sequences.
• Current methods often lack the incorporation of genomic coverage for robust taxonomic assignment.

Purpose of the Study:

• To develop an advanced computational pipeline for the identification and ranking of human viruses from sequencing data.
• To improve the accuracy of viral taxonomic assignment, especially for divergent sequences.
• To integrate genomic coverage into the likelihood estimation for viral identification.

Main Methods:

• Development of a multi-task learning model based on convolutional neural network (MT-CNN) for read taxonomic assignment.
• Utilized a Bayesian ranking approach to determine the most likely human virus.
• Incorporated genomic region assignment from MT-CNN to estimate genomic coverage.

Main Results:

• The MT-CNN model demonstrated superior performance over Kraken 2, Centrifuge, and Bowtie 2 for divergent HIV-1 genomes.
• MT-CNN showed higher sensitivity in identifying SARS-CoV-2 in RNA sequencing datasets.
• The pipeline effectively integrated read counts and genomic coverage for ranking viral likelihoods.

Conclusions:

• A novel pipeline combining MT-CNN for divergent virus identification and genomic region assignment with Bayesian ranking has been developed.
• The pipeline accounts for both read counts and genomic coverage, enhancing taxonomic assignment accuracy.
• The developed pipeline is publicly available on GitHub for research use.