DNA Viruses

We built ColabFoldDB by expanding BFD/MGnify with metagenomic sequences from various environments. To update the database we searched the proteins from the SMAG (eukaryotes)^{14 (link)}, MetaEuk (eukaryotes)^{13 (link)}, TOPAZ (eukaryotes)¹⁵ , MGV (DNA viruses)^{16 (link)}, GPD (bacteriophages)^{17 (link)} and an updated version of MetaClust^{18 (link)} against the BFD/MGnify centroids using MMseqs2 and assigned each sequence to the respective cluster if they have a 30% sequence identity at a 90% sequence overlap (-c 0.9 --cov-mode 1 --min-seq-id 0.3). All remaining sequences were clustered using MMseqs2 cluster -c 0.9 --cov-mode 1 --min-seq-id 0.3 and appended to the database. We remove redundancy per cluster by keeping the most 10 diverse sequences using mmseqs filterresult --diff 10. The final database consists of 209,335,865 million representative sequences and 738,695,580 members (see the Data Availability section for the input files). We provide the MMseqs2 search workflow used in the server (Methods 2.2.1) as a standalone script (colabfold_search).

To generate training and testing datasets, sequences representing bacteria, archaea, plasmids, and viruses were downloaded from the National Center for Biotechnology Information (NCBI) RefSeq and Genbank databases (accessed July 2019) (Additional File 1: Table S1). For bacteria/archaea, 181 genomes were chosen by selecting from diverse phylogenetic groups. Likewise, a total of 1452 bacterial plasmids were chosen. For viruses, NCBI taxids associated with viruses that infect bacteria or archaea were used to download reference virus genomes, which were then limited to only sequences above 3 kb. This included viruses with both DNA and RNA genomes, though RNA genomes must first be converted to complementary DNA. Sequences not associated with genomes, such as partial genomic regions, were identified according to sequence headers and removed. This resulted in 15,238 total viral partial and complete genomes. To be consistent between all sequences acquired from NCBI, proteins and genes were predicted using Prodigal (-p meta, v2.6.3) [50 (link)]. All sequences were split into non-overlapping, non-redundant fragments between 3 and 15 kb to simulate metagenome-assembled scaffolds. These simulated scaffolds are hereafter called fragments and were used throughout training and testing VIBRANT. For RNA virus detection, 33 viral (bacteriophage) genomes from NCBI RefSeq and 37 from Krishnamurthy et al. were used [51 (link)], and for archaeal virus detection, all genomes were acquired from NCBI RefSeq. The RNA and archaeal viral genomes were represented in both the training and testing datasets as genomic fragments, and recall evaluation was performed on whole genomes. These were the only datasets in which training and evaluation datasets were semi-redundant. See Supplemental Methods (Additional File 16) for additional datasets and sequences used.
Integrated viruses are common in both bacteria and archaea. To address this for generating a dataset devoid of viruses, PHASTER (accessed July 2019) was used to predict putative integrated viruses in the 181 bacteria/archaea genomes. Using BLASTn [52 (link)], any fragments that had significant similarity (at least 95% identity, at least 3 kb coverage, and e-value < 1e−10) to the PHASTER predictions were removed as contaminant virus sequence. The new bacteria/archaea dataset was considered depleted of proviruses but not entirely devoid of contamination. Next, the datasets for bacteria/archaea and plasmids were annotated with KEGG, Pfam, and VOG HMMs (hmmsearch (v3.1), e-value < 1e−5) [53 (link)] to further remove contaminant virus sequence (see next section for details of HMMs). Plasmids were included because it was noted that the dataset appeared to contain virus sequences, possibly due to misclassification of episomal proviruses as plasmids. Using manual inspection of the KEGG, Pfam, and VOG annotations, any sequence that clearly belonged to a virus was removed. Manual inspection was guided first by the number of KEGG, Pfam, and VOG annotations and then by the annotations themselves. For example, sequences with more VOG than KEGG or Pfam annotations were inspected and removed if multiple viral hallmark genes were found or if the majority of annotations represented viral-like genes. The final datasets consisted of 400,291 fragments for bacteria/archaea, 14,739 for plasmids, and 111,963 for viruses. Total number of fragments for all datasets used can be found in Additional File 2: Table S2.

DNA extraction from pure cultures of the Norwegian KpSC pig isolates was performed by using a MagNA Pure 96 instrument (Roche) with the DNA/Viral NA SV 2.0 kit and the pathogen universal 200 4.0 protocol. Sequencing library preparation was performed by using the Illumina Nextera DNA Flex library prep kit (Illumina), followed by sequencing on an Illumina MiSeq instrument, producing paired-end reads with a length of 300 bp.
The genomes of the KpSC originating from pig isolates from Thailand were available from a previous study [16 (link)]. A total of 87 genomes were included.

Virus isolation was performed using C6/36 Aedes albopictus cells. Mosquito homogenates were centrifuged at 4°C for 20 min at 12,000 rpm in a high-speed refrigerated centrifuge, and 40 μL of supernatant was aspirated and inoculated into a single layer of C6/36 cells in a 24-well plate and grown for 1 day (2 wells/sample). Two wells per plate of cell controls were set up and incubated at 32°C. Cytopathic effects (CPE) were observed and recorded continuously for 4–7 days. Isolates with obvious or suspected CPE were freeze-thawed once and transferred to 25-cm² cell bottles for blind transmission to 3–6 generations. Viral isolates with regular CPE were freeze-thawed once and then aspirated into a freezing tube at −40°C for identification. During virus isolation, the group of C6/36 Aedes albopictus cell without mosquito supernatant was cultured as negative control.
Virus-specific RT-PCR was performed to identify the viral isolates. The viral nucleic acid was extracted using the ZYBIO Magnetic Beads Virus DNA/RNA Extraction Kit (ZYBIO, Chongqing, China) and the QIAamp® Viral RNA Mini Kit (Qiagen, Hilden, Germany) and reverse transcribed using M-MLV Reverse Transcriptase (Promega, Madison, WI, United States). Based on the literature, we synthesized PCR, detection, and Sanger sequencing primers (Supplementary Table S2) for a variety of arboviruses, such as flaviviruses, alphaviruses, and viruses in the family Reoviridae, and then designed primers from newly discovered viral sequences from Yunnan and Guangxi provinces published in recent years (GoTaq® Green Master Mix, Promega). Next, next-generation sequencing (NGS) was performed to detect viruses in the isolates with CEP, but negative for RT-PCR using the designed primers. A sequencing library was constructed using the NEBNext Ultra II Directional RNA Library Prep Kit (Illumina, San Diego, CA, United States). The library was subsequently sequenced on an Illumina NovaSeq 6000 (PE150) sequencing platform. Trimmomatic v0.36 (Bolger et al., 2014 (link)) was used to remove low-quality and short reads, and SPAdes v3.13.0 (Nurk et al., 2013 (link)) was used for metagenomic assembly. Blastn and Blastx were used to search for viral contigs. PCR detection and Sanger sequencing were performed using redesigned primers based on the viruses detected in samples with unknown viruses.