Evaluation of computational phage detection tools for metagenomic datasets

Area of Science:

• Bioinformatics
• Computational Biology
• Virology

Background:

• Rapid development of computational tools for phage detection in metagenomes necessitates systematic benchmarking.
• Existing tools vary widely in their approaches and performance, leading to inconsistent results.

Purpose of the Study:

• To systematically benchmark 19 phage detection tools using diverse datasets.
• To assess tool performance based on fragment length, viral content, taxonomy, contamination, and resource usage.
• To compare phage community predictions in real human gut metagenomes and viromes.

Main Methods:

• Assessed 9 scalable tools on benchmark challenges including fragmented genomes and simulated metagenomes.
• Evaluated robustness to eukaryotic contamination and computational resource usage.
• Analyzed real human gut metagenomes and viromes to compare predicted phage communities.

Main Results:

• Tools showed strikingly different results, with homology-based methods (e.g., VirSorter2) having low false positives and high robustness.
• Sequence composition-based tools (e.g., VirFinder) exhibited higher sensitivity, detecting phages with limited reference data.
• Significant discrepancies in predicted phage communities were observed in human gut samples, with low tool overlap.

Conclusions:

• No single tool consistently outperformed others across all metrics.
• Developed benchmark datasets are publicly available for future tool evaluation and comparability.
• The study highlights the need for standardized evaluation of phage detection tools in metagenomics.