ViraMiner: Deep learning on raw DNA sequences for identifying viral genomes in human samples

Area of Science:

• Virology
• Bioinformatics
• Machine Learning

Background:

• Detecting novel or divergent viruses in human samples is challenging.
• Conventional sequencing methods often misclassify viral sequences as 'unknown' due to low similarity to known genomes.

Purpose of the Study:

• To develop and validate ViraMiner, a deep learning-based method for identifying viruses in human biospecimens.
• To improve the detection of highly divergent and unknown viral sequences.

Main Methods:

• ViraMiner utilizes a dual-branch Convolutional Neural Network architecture.
• The model analyzes raw metagenomic contigs for patterns and pattern frequencies.
• Training involved sequences from 19 metagenomic experiments, labeled using BLAST.

Main Results:

• ViraMiner demonstrates significantly improved accuracy in viral genome classification compared to other machine learning methods.
• Achieved an area under the ROC curve of 0.923 using 300 bp contigs.
• This represents the first machine learning approach for detecting viral sequences in raw metagenomic contigs from diverse human samples.

Conclusions:

• ViraMiner effectively identifies viral sequences within raw metagenomic data.
• The model can serve as a recommendation system for investigating sequences flagged as 'unknown' by traditional methods.
• Enhanced detection of divergent viruses can advance the understanding of infectious disease causes.