Fig 1.
Overview of the VIRify pipeline.
(A) Starting from a set of contigs (fasta file) the pipeline preprocesses the input sequences (ID renaming, length filtering) and predicts contigs from a putative viral origin that are split into high confidence (HC), low confidence (LC) and putative prophage (PP) sets. Each selected contig is then annotated and assigned to a taxonomy, if possible. All results (annotated viral contigs) are quality-controlled with CheckV, and finally summarised and visualised. (B) The assigned taxonomy is based on the informative ViPhOG hits per contig and performed on genus, family, subfamily and finally order rank. We consider high-confidence and low-confidence ViPhOG hits and discard non-informative models where no clear taxonomy signal could be assigned. TSR—taxon-specific ratio; —taxon average CDS; σCDStaxon—taxon CDS standard deviation.
Fig 2.
Number of informative ViPhOGs identified for different viral taxonomic ranks.
31,150 ViPhOGs were searched against all entries in UniProtKB, and based on the output they were designated as specific for different viral taxa (see Methods). Purple refers to specific ViPhOGs assigned to prokaryotic viral taxa, whereas yellow indicates specific ViPhOGs for eukaryotic viral taxa.
Fig 3.
Exemplary viral contig selection and annotation procedure for the Kleiner co-assembly.
(A) Comparison of virus predictions for the Kleiner co-assembly performed with various tools run via WtP. Shown are the 35 contigs (rows) predicted as viral by at least one of the tested tools. The column with red squares highlights the contigs manually identified as viral, as described in the methods section. Blue squares indicate contigs predicted as viral by VirSorter (virome decontamination mode), VirFinder (VF.modEPV_k8.rda model) or PPR-Meta. The column with green squares indicates the viral contigs reported by the prediction approach implemented in VIRify, based on the results from WtP (see Methods section). Yellow squares indicate CheckV-quality results for contigs selected by VIRify that are either high-quality, medium-quality, low-quality or not-determined; going from dark (high-quality) to light yellow (not-determined). (B) ORFs predicted with Prodigal and annotated with the informative ViPhOGs for the 14 contigs identified as viral by VIRify. Of these, eleven were predicted as high confidence (HC) and three as putative prophages (PP). No low-confidence viral predictions were reported for the Kleiner co-assembly. The coloured contig labels indicate the CheckV scores: red—high-quality, orange—medium-quality, yellow—low-quality, and black—not-determined. Dark grey bars indicate predicted ORFs without any ViPhOG hit, while green bars indicate ViPhOG hits based on the model-specific bitscores. (C) Predicted viral sequences and corresponding taxonomic assignments based on informative ViPhOG hits for the Kleiner co-assembly.
Table 1.
VIRify’s viral prediction results for two mock community assemblies.
The analyses were conducted for all assembled contigs longer than 1.5 kb.
Fig 4.
Predicted ORFs and corresponding ViPhOG annotations and taxonomy assignments for the Neto co-assembly.
(A) ViPhOG-annotated ORFs for all contigs predicted as viral with high confidence (HC) and low confidence (LC) for the Neto assembly. Note that due to the high number of LC hits, only a selection of contigs is shown. The coloured contig labels indicate the CheckV scores: red—high-quality, orange—medium-quality, yellow—low-quality, and black—not-determined. VIRify assigned the genus Alphacoronavirus to NODE 82 based on seven informative ViPhOG model hits that are additionally shown as a circular visualization. NA—no taxonomy could be assigned due to missing model support. (B) Neto contigs predicted as viral with high and low confidence and their taxonomic assignments based on the ViPhOG model hits. Both visualizations can be automatically produced by VIRify and were only slightly manually adjusted via Inkscape.
Fig 5.
Viruses predicted and annotated by VIRify for 243 TARA Oceans assemblies.
The assembly identifiers include information about the size fractionation of the corresponding sample. For example, samples obtained by smaller filter size 0.1–0.22 μm (and so expected to be enriched for smaller viruses) are labelled with the suffix _0.1–0.22. (A) Shows a selection (filters 0.1–0.22 μm and 0.45–0.8 μm) of 41 samples and the number of predicted viruses based on high confidence (VirSorter categories 1 and 2) and low confidence (VirSorter category 3 and combined VirFinder and PPR-Meta results) hits. More viruses are found for smaller filter sizes, as expected. Assemblies based on smaller filter sizes are highlighted in bold. For visualization purposes, we summarize high and low confidence predictions for samples labelled with the filter sizes 0.1–0.22 and 0.45–0.8. (B) Selection of 0.22–3 μm filtered samples with a high number of predicted prasinoviruses, large double-stranded DNA viruses belonging to the order Algavirales. These viruses are predominantly found in the low confidence set; thus they would have been missed if only VirSorter were run on the data but are predicted by our combination of VirFinder and PPR-Meta.
Fig 6.
Taxonomic classification of representative sequences from the GPD.
One representative sequence from each of the 57,866 VCs present in the GPD was selected based on their reported CheckV quality and completion values. The representative sequences were analysed with VIRify using the ‘—onlyannotate’ option to obtain taxonomic classifications for them. The taxa displayed in the Sankey plot correspond to those in which at least 10 representative sequences were classified.
Fig 7.
(A) Comparison of annotated CDS from VIRify’s high-confidence viral predictions from all 243 TARA Oceans assemblies. Our comparison shows that the ViPhOG models comprehensively cover a large proportion of potentially viral CDS in agreement with other public databases. The RVDB had the fewest predictions, with only 15.3% annotated CDS. Results of the VPF database, where models are derived from the IMG/VR database that also includes novel viral sequences derived from metagenome approaches, comprise a large proportion of unique models that exclusively match 117,962 CDS (9.8%) that are not covered by any of the other databases. However, a significant number of CDS are annotated by models from the other databases but missed by VPF. Our model-specific bit score filtering (ViPhOG—threshold) reduces the number of CDS annotations derived from ViPhOGs by 7.3% (B) Visualization of predicted (top row) and annotated ORFs for one exemplar contig from TARA Oceans assembly CEUI01. Grey bars indicate hits against an HMM of the corresponding database, while informative ViPhOG hits with taxonomic information are shown in green. The top ViPhOG track shows hits filtered by bit score (or e-value if no bit score could be assigned for a model, see Methods) and the bottom ViPhOG track shows e-value-filtered hits. (C) While the model-specific ViPhOG bit score threshold can lead to the loss of potentially informative annotations for taxonomy assignment, it can also reduce the number of false positive model hits and thus increase the overall accuracy of VIRify.