RaFAH

RaFAH predicts virus-host associations for viruses infecting Bacteria and Archaea using random forest machine-learning models.

Key Features:

• Machine Learning Approach: Uses random forest models implemented with the Ranger package in R to learn associations between viral genes and host taxonomy.
• High Predictive Accuracy: Demonstrated an F1-score of 0.97 at the phylum level in comparative analyses.
• Large-scale Dataset: Trained and validated on nearly 200,000 viral genomes.
• Broad Biome Coverage: Applied to datasets from eight distinct biomes with medical, biotechnological, and environmental significance.
• Discovery of Archaeal Viruses: Identified 537 genomic sequences from previously unknown archaeal viruses, including novel auxiliary metabolic genes.
• Inclusion of Hypothetical Proteins: Incorporates hypothetical proteins as genomic features to improve predictive capability.

Scientific Applications:

• Host prediction for uncultured viruses: Assigns putative bacterial and archaeal hosts to viral genomes derived from metagenomes and isolates.
• Viral ecology and evolution: Enables exploration of viral community composition and virus–host interactions across diverse biomes.
• Discovery of novel viral lineages and AMGs: Facilitates identification of new archaeal viral lineages and associated auxiliary metabolic genes.
• Taxonomic host assignment: Provides high-accuracy host assignments at taxonomic ranks such as phylum.

Methodology:

Trains random forest models using the Ranger package in R on genomic features extracted from viral genomes—including hypothetical proteins—using a dataset of ~200,000 viral genomes.