Short k-mer abundance profiles yield robust machine learning features and accurate classifiers for RNA viruses

Area of Science:

• Bioinformatics
• Computational Biology
• Virology

Background:

• High-throughput sequencing generates vast biological data requiring automated annotation.
• Current methods often overlook RNA viruses due to their genomic complexity.
• Machine learning models exist but are limited to DNA genomes.

Purpose of the Study:

• Develop a robust method for distinguishing viral RNA from host transcripts.
• Address the gap in automated annotation for RNA viruses.
• Improve machine learning classification for viral sequence data.

Main Methods:

• A novel short k-mer based sequence scoring method was developed.
• 18 machine learning classifiers were trained to differentiate viral RNA from human transcripts.
• Performance was evaluated against BLASTn, BLASTx, and HMMER3 using curated and raw RNA-Seq data.

Main Results:

• The developed models achieved near-perfect accuracy on clean sequence data, outperforming existing methods.
• On de novo assemblies of Ebola virus-infected cells, the model's ROC AUC ranged from 0.6 to 0.86.
• The classifier successfully re-classified false positives from BLAST and HMMER3 searches.

Conclusions:

• The novel k-mer based scoring method provides robust sequence information for machine learning.
• This approach significantly enhances automated annotation capabilities for viral RNA sequences.
• The findings establish a foundation for advanced machine learning in sequence data annotation.