Virtifier: a deep learning-based identifier for viral sequences from metagenomes.

Area of Science:

• Virology
• Bioinformatics
• Computational Biology

Background:

• Viruses are Earth's most abundant biological entities and significant human pathogens.
• Accurate identification of viral sequences from metagenomes is crucial for understanding viral communities and disease.
• Existing methods using one-hot vectors for nucleotide sequences are often ineffective for viral identification in metagenomic data.

Purpose of the Study:

• To propose Virtifier, a novel deep learning-based method for identifying viral sequences in metagenomic data.
• To develop an effective nucleotide sequence encoding method (Seq2Vec) and a deep learning model (LSTM with attention) for viral prediction.

Main Methods:

• Developed Seq2Vec, a nucleotide sequence encoding method utilizing a trained embedding matrix to capture codon relationships.
• Employed an attention-based long short-term memory (LSTM) neural network to analyze codon relationships and extract key features.
• Evaluated Virtifier on three diverse metagenomic datasets.

Main Results:

• Virtifier accurately identifies short viral sequences (<500 bp) from metagenomes.
• Virtifier surpasses the performance of widely used methods like VirFinder, DeepVirFinder, and PPR-Meta.
• Comparable performance was achieved by Virtifier for longer viral sequences (>5000 bp).

Conclusions:

• Virtifier offers a significant advancement in viral sequence identification from metagenomic data.
• The Seq2Vec encoding and attention-based LSTM model are effective for viral prediction.
• Virtifier provides a robust and accurate tool for viral analysis in complex biological samples.