The Global Virome, in One Network (VIRION) is an open atlas of the vertebrate virome.

The VIRION database was built by, and is curated by, an interdisciplinary team of virologists, ecologists, and data scientists as part of the Verena Consortium, an effort to predict which viruses could infect humans, which animals host them, and where they could someday emerge.

VIRION is the most comprehensive database of its kind, drawing data from scientific literature and online databases, and is updated automatically with new data. Today, it includes over 2.8 million interactions that capture roughly 25% of known mammal diversity, 10% of bird diversity, and 6% of total vertebrate diversity. Unlike many other databases, VIRION has undergone a taxonomic reconciliaton process using a backbone provided by NCBI, with a total (at last count) of 3,787 valid vertebrate host species and 9,003 valid viruses, of which 1,658 are ICTV-ratified valid names.

VIRION is part of an ongoing effort to build an open data ecosystem in viral ecology, and has been released prior to a preprint, publication, or peer-review process. As such (and because it includes a mix of known and predicted interactions), we strongly discourage the reproduction of the data in other databases at this time.

We encourage researchers to review this entire guide before using these data.

How we built VIRION

VIRION aggregates five major sources of information, three of which are dynamic (*):

• CLOVER, a Verena-curated database, which reconciles four static datasets on host-pathogen interactions.
• The public data released by the USAID Emerging Pandemic Threats PREDICT program.
• GLOBI*, the Global Biotic Interactions database.
• NCBI GenBank*, specifically the entirety of NCBI Virus accessions stored in the Nucleotide database.
• NCBI Sequence Read Archive*, which includes metagenomic samples that have undergone taxonomic analysis. In a sense, these interactions are predictions, not observations, as viral identity has not been confirmed through genome assembly, and a number of contaminants could lead to false positives (see below). As such, we recommend the exclusion of these data except in cases where users are able to navigate these complexities.

   

These methods will be further described in a forthcoming preprint / publication by Carlson et al.

How to use VIRION

VIRION can be used for everything from deep learning to simple biological questions. For example, if you wanted to ask which bats a betacoronavirus (like SARS-CoV or MERS-CoV) has ever been isolated from, you could run this R code:

> library(tidyverse); library(vroom)
> 
> virion <- vroom("Virion/Virion.csv.gz")
> 
> virion %>% 
+   filter(VirusGenus == "betacoronavirus",
+          HostOrder == "chiroptera",
+          DetectionMethod == "Isolation/Observation") %>% 
+   pull(Host) %>% 
+   unique()
[1] "chaerephon plicatus" "pipistrellus abramus" "rhinolophus affinis"      
[4] "rhinolophus ferrumequinum" "rhinolophus macrotis" "rhinolophus pearsonii"    
[7] "rhinolophus sinicus" "rousettus leschenaultii" "tylonycteris pachypus"

It's that simple! Here's a few small tips and tricks you should know:

• All resolved taxonomy is lowercase (see the above example); original metadata retains case, and non-taxonomic metadata is not all lowercase
• Some valid records have NA's in their taxonomy; for example, if an unclassified Betacoronavirus is found in a mouse, it might be recorded as NA in the "Virus" field. This is an intentional feature, as it enables researchers to talk about higher-level taxonomic patterns, and some studies may not need fully-resolved data.
• Sometimes, you'll see taxonomy that's outdated or strange. If you think there's an error, please leave an issue on the Github. Before you do, it may be worth checking whether a given name is correctly resolved to the NCBI taxonomy; for example, in R, you can use taxize::classification("Whateverthe latinnameis", db = "ncbi"). If the issue is related to that taxonomic backbone, please label your issue ncbi-needed
• Different databases may have overlapping records. For example, some PREDICT records are deposited in GenBank, and some GenBank records are inherited by EID2. As different data has passed between these sources, they've often lost some metadata. Presence in different datasets therefore does not indicate stronger / weaker evidence, and conversely, conflicting evidence between databases may not be indicative of any biological evidence.
• If you don't read the rest of this README.md, cut all SRA records from the dataset before beginning any analysis.

File organization and assembly

For now, VIRION lives on Github in a 100% open and reproducible format. Downloading the data directly from this website, or cloning the repository, is the easiest way to access the data. To avoid relying on the Large File Storage system, the VIRION database itself is stored in two file formats:

1. The entire database is available in Virion/Virion.csv.gz which can be easily read as-is using the vroom package.
2. The NCBI-matched components of the database are also available in a disaggregated format with a backbone (Edgelist.csv), two taxonomic metadata files (HostTaxonomy.csv, VirusTaxonomy.csv), and three sampling metadata files (Provenance.csv.gz, Detection.csv.gz, Temporal.csv.gz). The taxonomy files can be joined to the backbone with the HostTaxID and VirusTaxID fields, while the metadata files can be joined by the AssocID field (which must first be separated into unique rows). For simple tasks, not every join will be needed.

What you should probably know about the data

Like most datasets that record host-virus associations, this includes a mix of different lines of evidence, diagnostic methods, and metadata quality. Some associations will be found in every database, with every evidence standard; others will be recorded from a single serological data point with unclear attribution. VIRION can aggregate all this data for you, but it's your job as a researcher to be thoughtful about how you use these data. Some suggested best practices:

• As a starting point, you can remove any records that aren't taxonomically resolved to the NCBI backbone (HostNCBIResolved == FALSE, VirusNCBIResolved == FALSE). We particularly suggest this for data that come from other databases that also aggregate content but use multiple taxonomic backbones, which may include invalid names that are not updated.

• You should also be wary of records with a flag that indicates host identification by researchers was uncertain (HostFlagID == TRUE) or that indicates the virus in a particular metagenomic sample is a common contaminant or false positive (VirusFlagContaminant == TRUE). The former is usually derived from source data, while the latter is provided as part of VIRION's quality control efforts.

• Limiting evidence standards based on diagnostic standards (e.g., using Nucleotide and Isolation/Observation records, but no Antibodies or k-mer) or based on redundancy (i.e., number of datasets that record an association) can also lead to stronger results.

• We encourage particular caution with regard to the validity of virus names. Although the NCBI and ICTV taxonomies are updated against each other, valid NCBI names are not guaranteed to be ICTV-valid species level designations, and many may include sampling metadata. We recommend that researchers manually curate names where possible, but can also use simple rubrics to reduce down controversial names. For example, in the list of NCBI-accepted betacoronavirus names, eliminating all virus names that include a "/" (e.g., using stringr::str_detect()) will reduce many lineage-specific records ("bat coronavirus 2265/philippines/2010", "coronavirus n.noc/vm199/2007/nld") and leave behind cleaner names ("alpaca coronavirus") but won't necessarily catch everything ("bat coronavirus ank045f"). Another option is to limit analysis to viruses that are ICTV ratified (ICTVRatified == TRUE), but this is particularly conservative, and will leave a much larger number of valid virus names out.

• Finally, we encourage every researcher who uses this data to make a deliberate choice about the use of metagenomic data (see below).

A special note about the metagenomics

Unlike nearly every dataset familiar to disease ecologists, this dataset includes a mix of known interactions (e.g., PCR detection of virus) and predicted interactions (the k-mer analysis conducted on the Sequence Read Archive). As such, the data cannot be used safely and uncritically off the shelf, it should not be indexed in other datasets of known interactions, and it should be carefully evaluated by researchers with attention to the mix of data standards.

To generate the SRA component of the dataset, we used a k-mer based taxonomy tool which identifies the number of virus "hits" in a host sample. Starting from an analysis of every possible pairwise combination (given in SRA_as_Edgelist.zip for any interested researchers), we identified the maximum number of hits for a given pair. Subsequently, we use the CLOVER data as a ground truth (for "known associations" or not) to identify the most plausible associations, with a threshold selected that maximizes the kappa statistic. This returns the top ~1% of possible associations (see the figure to the right). Given that most matches are likely to be imprecise, we currently restrict the taxonomy of these matches to the genus level and above (though species matches can be reconstructed from the VirusTaxID field at this time).

Even though this analysis is incredibly conservative, it is in a preliminary state, and only predicts possible associations. (These could later be confirmed, for example, by assembling viral genomes from the original SRA samples.) As such, users should be very careful about whether or not they include these interactions in their analysis. Example problems they might encounter:

• The prediction could, simply, be a false positive. This is an imprecise method.

• The highest scoring match might be a known relative of an unknown virus (for example, in a sample with a novel bat betacoronavirus, the highest score might be returned for SARS-CoV). This is partially solved by removing the Virus names, but not entirely.

• The score might be a product of technological issues or cross-contamination (see the VirusFlagContaminant field).

As such, users may want to remove all of these records entirely from the dataset, which can be done in a single line of code using the DetectionMethod or Database columns, e.g.,

library(tidyverse); library(magrittr)
virion %<>% filter(!(DetectionMethod == "kmer")) # option 1
virion %<>% filter(!(Database == "SRA")) # option 2 (currently equivalent)

Other, more advanced users may be interested in using the entire edge list of possible host-virus associations in SRA, which is found in SRA_as_Edgelist.zip. Alternate scoring methods that are less conservative will include many more false positives, but also potentially more true positives. In the long term, we hope to develop score metrics that are more informative but still easily incorporated into the VIRION architecture.

Additional information

Citing VIRION

Please do not use VIRION for published research yet! This is only a beta release and probably contains a number of bugs that we still need to fix. (Similarly, so those bugs don't escape our orbit, please don't reproduce the data elsewhere yet!)

Contact

• For general questions about VIRION, please reach out to Colin Carlson or Gregory Albery.
• For specific questions about the SRA data, please contact Timothée Poisot or Ryan Connor.
• For specific questions about the CLOVER dataset, please contact Rory Gibb.