Ensemble Deep Learning Models on Raw DNA Sequences for Viral Genome Identification in Human Samples.

Area of Science:

• Virology
• Bioinformatics
• Computational Biology

Background:

• Detecting novel or divergent viruses is challenging for current diagnostic methods.
• Deep learning (DL) offers potential for analyzing 'viral dark matter' lacking known references.

Purpose of the Study:

• To develop a high-performance deep learning ensemble for identifying viral contigs in human metagenomic data.
• To improve viral detection capabilities for clinical diagnostics and pathogen surveillance.

Main Methods:

• An ensemble of deep convolutional neural networks (CNNs) was designed to process biological sequence data.
• The framework integrates complementary architectures to capture local and global genomic features.
• The model was evaluated on complex human metagenomic datasets.

Main Results:

• The DL ensemble achieved state-of-the-art performance with an AUROC of 0.939 on 300 bp contigs.
• It outperformed existing methods like transformer-based approaches, ViraMiner, and DeepVirFinder.
• The model demonstrated robustness to data degradation (10% nucleotide substitution) and generalized to unseen viral families.

Conclusions:

• The developed DL framework effectively identifies viral contigs in complex metagenomic datasets.
• It offers a robust and generalizable solution for detecting known and unknown viruses, crucial for emerging threat detection.
• Publicly available code and data promote reproducibility and further research in clinical sensing.