Benchmarking and integration of viral prediction tools to uncover viral dark matter across divergent low abundance metagenomes

Abstract

While existing benchmark studies have focused on viral detection tools using highly-abundant virus-enriched, their performance in low-abundance metagenomic environments remains unexplored. We aimed to benchmark eleven viral prediction tools and assess combinations of the best-performing-tools in low-abundance metagenomes. We generated an artificial dataset simulating the low-abundant metagenomic samples including human,bacterial and viral sequences, using complete-Refseq genomes. Viral sequences were fragmented into varying lengths and subjected to increased levels of in silico mutations to evaluate recall and F1 score of eleven tools tested individually. We further evaluated complementary tool combinations to improve sensitivity and specificity, and assessed the best-performing multi-tools formulations on low-abundance viromes from human lower respiratory tract and air metagenomic samples. DeepVirfinder(DVF) and DIAMOND achieved the highest recall (median of 67% and 58%, respectively) on the RefSeq artificial contigs dataset along increased levels of induced mutations. On the artificially mutated RefSeq dataset, the highest sensitivity and F1 scores were achieved by 4 two-tools combinations : DVF+DeepMicroclass,DVF+ViraLM,DVF+geNomad and DVF+DIAMOND, combined with Kraken2X-filtering to remove false-positive contigs classified as Eukaryota,Bacteria,or Archaea. Using Kraken2X-filtering, DVF+ViraLM and DVF+DeepMicroclass achieved the best sensitivity across real human airway and air datasets. These combinations expanded the previously identified viral repertoire using gold-standard methods, increasing the number of viral contigs with functionally annotated CDSs and revealing previously undetected viral diversity. Benchmarking of the eleven most used tools for viral prediction showed that the complementary combination approach, improves the detection of viral sequences while limiting false positives. This provides a robust strategy for uncovering divergent viral dark matter.