Identifying viruses from metagenomic data using deep learning.

Area of Science:

• Bioinformatics
• Computational Biology
• Virology

Background:

• Metagenomic sequencing enables large-scale analysis of microbial and viral genomes without culturing.
• Current methods struggle to efficiently identify novel or short viral sequences in metagenomic datasets.

Purpose of the Study:

• To develop a reference-free and alignment-free machine learning approach for viral sequence identification in metagenomic data.
• To improve the accuracy and efficiency of viral discovery from complex biological samples.

Main Methods:

• Developed DeepVirFinder, a deep learning-based method for identifying viral sequences.
• Utilized a reference-free and alignment-free strategy for sequence analysis.
• Trained and evaluated the model on viral RefSeq datasets and metavirome samples.

Main Results:

• DeepVirFinder outperformed existing methods like VirFinder across various sequence lengths, achieving high Area Under the Receiver Operating Characteristic Curve (AUROC) scores.
• Incorporating additional metavirome data improved accuracy for under-represented viral groups.
• Analysis of human gut metagenomic samples identified numerous viral sequences, with ten bins associated with colorectal carcinoma (CRC) status.

Conclusions:

• DeepVirFinder significantly enhances the accuracy of viral identification in the era of metagenomics.
• The tool aids in studying the role of viruses, potentially in diseases like colorectal carcinoma.
• Deep learning applied to metagenomic data represents a powerful approach for viral discovery.