Phage-bacterial contig association prediction with a convolutional neural network.

Area of Science:

• Microbiology
• Bioinformatics
• Genomics

Background:

• Phage-host associations are crucial in microbial ecosystems, but identifying hosts from metagenomic data is challenging due to fragmented genomes.
• Existing computational tools often rely on whole genomes and struggle with short DNA contigs common in metagenomic studies.
• Accurate phage-host prediction is vital for understanding microbial community dynamics and viral functions.

Purpose of the Study:

• To develop a novel computational model, ContigNet, for predicting phage hosts from short DNA contigs.
• To evaluate ContigNet's performance against existing methods like VirHostMatcher (VHM) and WIsH.
• To assess ContigNet's applicability to large-scale metagenomic datasets and its potential for predicting plasmid-host associations.

Main Methods:

• Development of ContigNet, a convolutional neural network-based model.
• Training and validation using phage and host DNA sequences, including short contigs.
• Comparison of ContigNet's predictive accuracy (AUROC scores) with VHM and WIsH on validation sets and the Metagenomic Gut Virus (MGV) catalogue.

Main Results:

• ContigNet significantly outperformed VHM and WIsH on short contigs (200 bp–50 kbps), achieving 72-85% AUROC compared to a maximum of 68%.
• Applied to the MGV catalogue, ContigNet achieved 60-70% AUROC, outperforming VHM and WIsH (52%).
• ContigNet demonstrated high accuracy in predicting plasmid-host contig associations, suggesting its utility for diverse mobile genetic elements.

Conclusions:

• ContigNet is an effective tool for predicting phage-host associations from short contigs in metagenomic data.
• The model offers improved accuracy over existing methods, facilitating a deeper understanding of viral roles in microbial communities.
• ContigNet's capability extends to predicting plasmid-host interactions, highlighting its broad applicability in mobile genetic element research.