ContigNet

ContigNet predicts associations between phage and plasmid contigs and their bacterial hosts from metagenomic sequence data using a convolutional neural network, enabling host assignment for short contigs (200 base pairs–50 kilobases).

Key Features:

• Convolutional Neural Network Architecture: Employs a CNN model tailored to extract sequence features from contigs ranging from 200 base pairs (bps) to 50 kilobases (kbps) for host prediction.
• Prediction Accuracy (AUROC): Demonstrated AUROC scores between 72% and 85% on a validation set, outperforming VirHostMatcher (VHM) and WIsH, which reported a maximum of 68% for similar contig lengths.
• Application to the MGV catalogue: Applied to the Metagenomic Gut Virus (MGV) catalogue of draft genomes from metagenomic samples, achieving AUROC scores of 60%–70% and outperforming VHM and WISH (52%).
• Versatility for Mobile Genetic Elements: Capable of predicting plasmid-host contig associations with high accuracy, supporting analysis of genetic exchanges between mobile elements and bacterial hosts.

Scientific Applications:

• Microbial ecology and evolution: Enables identification of phage-host interactions within complex microbial communities without requiring complete genomes.
• Metagenomic studies with short contigs: Facilitates host assignment from fragmented contigs typical of metagenomic sequencing (200 bp–50 kb).
• Horizontal gene transfer and plasmid analysis: Supports investigation of plasmid-host relationships and mechanisms of genetic exchange.

Methodology:

ContigNet trains a convolutional neural network on known phage-host and plasmid-host pairs to learn sequence patterns indicative of associations, then predicts potential host organisms for new contigs based on their sequence-derived features.