VirusPredictor

VirusPredictor predicts virus-related sequences in human genomic and metagenomic datasets using XGBoost machine learning to identify sequences unmappable to known human or pathogen genomes and classify infectious viruses and endogenous retroviruses.

Key Features:

• Two-Step Classification Model: Uses a two-step XGBoost-based classifier where step 1 assigns sequences to infectious virus, endogenous retrovirus (ERV), or non-ERV human.
• Taxonomic Subgroup Classification: Sequences classified as infectious viruses are further assigned to one of six virus taxonomic subgroups.
• Analysis of Unmappable Sequences: Targets sequences unmappable to known human or pathogen genomes to facilitate detection of uncharacterized viruses.
• Performance Metrics: Prediction accuracy increases with sequence length: 0.76 for 150–350 bp (Illumina short reads), 0.93 for 850–950 bp (Sanger), and up to 0.98 for 2000–5000 bp; taxonomic subgroup classification accuracy rises from ~0.92 to >0.98 beyond 850 bp.
• Assembly Recommendation: Recommends de novo assembly of Illumina short reads into contigs of approximately 1000 bp or longer prior to prediction.

Scientific Applications:

• Pathogen Discovery: Enables detection of disease-causing pathogens in datasets lacking reference genomes.
• Novel Virus Identification: Facilitates identification of potential new viruses from human genomic and metagenomic data.
• Endogenous Retrovirus Analysis: Distinguishes and classifies endogenous retroviruses within human sequences.
• Viral Taxonomy Insight: Provides taxonomic subgroup predictions to inform studies of viral origins and roles in human disease.

Methodology:

Two-step XGBoost models trained on an extensive in-house viral genome database: step 1 classifies sequences as infectious virus, endogenous retrovirus (ERV), or non-ERV human, and step 2 classifies infectious-virus sequences into one of six taxonomic subgroups.