VIBRANT: automated recovery, annotation and curation of microbial viruses, and evaluation of viral community function from genomic sequences

Abstract

Background: Viruses are central to microbial community structure in all environments. The ability to generate large metagenomic assemblies of mixed microbial and viral sequences provides the opportunity to tease apart complex microbiome dynamics, but these analyses are currently limited by the tools available for analyses of viral genomes and assessing their metabolic impacts on microbiomes. Design: Here we present VIBRANT, the first method to utilize a hybrid machine learning and protein similarity approach that is not reliant on sequence features for automated recovery and annotation of viruses, determination of genome quality and completeness, and characterization of viral community function from metagenomic assemblies. VIBRANT uses neural networks of protein signatures and a newly developed v-score metric that circumvents traditional boundaries to maximize identification of lytic viral genomes and integrated proviruses, including highly diverse viruses. VIBRANT highlights viral auxiliary metabolic genes and metabolic pathways, thereby serving as a user-friendly platform for evaluating viral community function. VIBRANT was trained and validated on reference virus datasets as well as microbiome and virome data. Results: VIBRANT showed superior performance in recovering higher quality viruses and concurrently reduced the false identification of non-viral genome fragments in comparison to other virus identification programs, specifically VirSorter, VirFinder, and MARVEL. When applied to 120,834 metagenome-derived viral sequences representing several human and natural environments, VIBRANT recovered an average of 94% of the viruses, whereas VirFinder, VirSorter, and MARVEL achieved less powerful performance, averaging 48%, 87%, and 71%, respectively. Similarly, VIBRANT identified more total viral sequence and proteins when applied to real metagenomes. When compared to PHASTER, Prophage Hunter, and VirSorter for the ability to extract integrated provirus regions from host scaffolds, VIBRANT performed comparably and even identified proviruses that the other programs did not. To demonstrate applications of VIBRANT, we studied viromes associated with Crohn's disease to show that specific viral groups, namely Enterobacteriales-like viruses, as well as putative dysbiosis associated viral proteins are more abundant compared to healthy individuals, providing a possible viral link to maintenance of diseased states.

Diverse viruses carrying genes for microbial extremotolerance in the Atacama Desert hyperarid soil

“…Six An actinobacterial MAG belonging to the Acidimicrobiia class recovered from below a boulder at site Y had acquired resistances against a unique set of diverse viruses detected across three different locations. Interestingly, 15 out of 66 matched viruses were predicted to be lysogenic by VirSorter (32) and/or VIBRANT (32,33). The likelihood for lysogenic infections in this ecosystem appears to be specific to the host taxa.…”

“…Putative viral genes in the matched viral sequences were extracted using Prodigal (84) and functionally annotated by using DIAMOND (85) against UniRef100 (41). HMM-based protein annotations including putative AMGs were conducted using VIBRANT (33) in agreement with the concept of class 1 AMGs according to (86). Viral scaffolds were dereplicated using CD-HIT v 4.6 (87) at 99% identity to identify the number of unique sequences per sample.…”

Rapid discovery of novel prophages using biological feature engineering and machine learning

“…We benchmarked PhageBoost against VirSorter(7), VIBRANT (8), and PHASTER (6). We used the default settings for PhageBoost v.0.1.2, VIBRANT v.1.2.1, and VirSorter v.1.0.5 with the database db2, while we manually submitted 54 genomes to the PHASTER web server using their URLAPI (https://phaster.ca/instructions) and chose not to use precomputed results.…”

“…We used the default settings for VIBRANT v.1.2.1 (8). To link the predictions to the regions found by viral read recruitment, and take into account the different sizes of the viral regions, we calculated the proportion of region retained.…”

“…When it comes to extracting viral sequences from (meta)genomic data, again there is a reliance on well-understood phages and the available tools are dependent on linking the reference viral genome database with sequence similarities approaches as is the case with VirSorter (7)and VIBRANT (8).…”