VirusPredictor: XGBoost-based software to predict virus-related sequences in human data

Area of Science:

• Genomics
• Bioinformatics
• Machine Learning

Background:

• Identifying novel viruses without reference genomes is challenging due to unmappable sequences in high-throughput data.
• Existing software lacks specialized capabilities for accurate viral sequence prediction in human samples.

Purpose of the Study:

• To develop and validate a machine learning method for predicting viral sequences, including uncharacterized viruses and endogenous retroviruses (ERVs), from human data.
• To create a user-friendly software tool, VirusPredictor, for accurate classification of unmappable sequences.

Main Methods:

• Developed a two-step XGBoost classification model utilizing an in-house viral genome database.
• The first step classifies sequences into infectious virus, ERV, or non-ERV human categories.
• The second step further classifies infectious viral sequences into six taxonomic subgroups.

Main Results:

• Prediction accuracy increased with sequence length, reaching 0.98 for sequences >2000 bp.
• Classification accuracy for infectious viruses ranged from 0.92 to >0.98 based on sequence length.
• VirusPredictor demonstrated high accuracy when applied to real genomic and metagenomic datasets.
• This study is the first to classify ERVs within infectious viral sequence prediction and combine virus subgroup predictions.

Conclusions:

• VirusPredictor accurately predicts the origin of unmappable sequences in human data, including novel viruses and ERVs.
• Longer sequences (ideally >850 bp) improve prediction accuracy; de novo assembly of short reads is recommended.
• VirusPredictor is a valuable open-source tool for advancing viral discovery and diagnostics.