henhaobio
Sign in

bioRxivBioinformaticsPreprint

Data mining of global sequence datasets markedly expands known diversity within the Tobamovirus genus

Neža Pajek ArambašičiD, Katarina BačnikiD, Luka KranjciD, Lana VogrineciD, Tomaž CurkiD, Denis KutnjakiD

PDFDOICode
In-depth full-text digest available · Jump to digest ↓

Abstract

High-throughput sequencing has transformed RNA virus discovery, producing vast datasets that often contain unexplored sequences of known and novel viruses across diverse sample types. Large-scale reconstruction of viral genomes within specific taxonomic groups remains a lengthy, labor-intensive process constrained by fragmented assemblies, heterogeneous workflows, and the limitations of percent identity cut-offs for taxonomic placement. Here, we focused on Tobamovirus (family Virgaviridae), a genus that contains important emerging plant pathogens. Tobamoviruses are detectable in diverse environmental matrices and are thus expected to be present across a range of datasets. We mined global short-read datasets to uncover hidden tobamovirus diversity, combining expert curation with supervised machine learning to streamline the discovery of novel tobamoviruses. Using the Serratus PalmID tool, we first identified 253 candidate datasets that are most likely to contain novel tobamoviruses. We developed a reproducible Snakemake pipeline for quality control, de novo assembly, similarity searches, and preliminary taxonomic assignment, yielding 510 candidate tobamovirus contigs across 131 datasets. Following manual curation, we confirmed 228 tobamoviral contigs from 23 datasets based on BLAST similarity searches, coding sequence content, genome organization, and assessment of potential misassemblies. To expedite the curation process, we built a two-stage machine learning classifier of tobamoviral contigs. Based on alignment-derived features and sequence properties, each ORF is first scored to determine whether it is tobamoviral. ORF-level predictions are then merged into contig-level classification using stacked generalization. Through clustering and phylogenetic analyses, we identified at least 32 putative novel tobamoviruses in datasets derived from freshwater, wastewater, plants, animal gut, and soil. Most were placed within established Tobamovirus subclades, whereas a few highly divergent sequences formed distinct basal branches related to recently identified soil-associated sequences classified as Virgaviridae sp. These findings highlight the value of aquatic and wastewater systems for expanding our understanding of tobamoviral diversity and demonstrate how data mining and targeted curation can substantially expand recognized diversity within the genus. The workflow is readily adaptable to other viral taxa and supports continuous discovery as public datasets grow.

In-depth summary

Claude Opus 5.5 · xhigh effort · 8 Nov 2026 · from the full text of version 1 (23 Oct 2026) · CC BY 4.0

Serratus-guided Snakemake mining, expert curation, clustering and phylogenetic placement identify at least 32 putative novel tobamoviruses in public short-read data, mostly Amazon River metatranscriptomes; a random-forest classifier partly automates curation.

Large-scale recovery of viral genomes for one taxon is slow, because assemblies are fragmented, workflows differ, and percent-identity cut-offs place taxa poorly. The paper combines targeted dataset selection, curation and a supervised classifier using alignment-derived features to speed up discovery within the Tobamovirus genus.

Key results

  1. ShownFunnel from public datasets to curated tobamoviral contigs: only a small share of Serratus-flagged datasets yielded confirmed tobamovirus sequences.

    6,870 candidate runs were cut to 253 SRRs (50-95% identity). The pipeline gave 510 contigs from 131 SRRs. Manual curation confirmed 228 tobamoviral contigs (45%) from 23 SRRs. · Results; The developed Snakemake pipeline narrows the set…; Manual curation confirms 45% of identified…; Targeted data mining enables the scalable…

  2. SuggestedClustering and phylogenetic placement imply a lower bound of 32 distinct novel tobamoviruses, spread over several clades of the genus. The count rests on mutual-overlap sets of partial contigs.

    At least 32 novel viruses from 15 datasets across 7 studies, against 37 ICTV-recognized species. Placement used EPA-ng on a fixed IQ-TREE reference tree and agreed broadly with the four ORF trees. · Clustering and phylogenetic analyses for contigs…; Phylogenetic analyses reveal clustering of novel…; Discovery of a diverse set of previously unknown…

  3. ShownA two-stage classifier (random forest per ORF, then logistic-regression stacking of binned ORF probabilities) separates tobamoviral from other Virgaviridae contigs well in cross-validation. The binned aggregation beat the extreme-probability rule.

    Binned: accuracy 0.9735, F1 0.9806, precision 0.9925, recall 0.9690, AUC 0.9971. Extreme: accuracy 0.9231, F1 0.9475, precision 0.9003. ORF-level accuracy was 0.767 for random forest vs 0.752 for SVM. Leave-one-virus-out, 30 x 5-fold CV. · Model selection for ORF classifier; Contig classifier, aggregation strategies of…; Evaluation of the contig classifier; Machine learning approach aids in data curation…

  4. ShownOn the study's own 510 curated contigs the classifier did worse than in cross-validation. It was sensitive but imprecise, and weakest on divergent and misassembled contigs. It mostly rejected viruses from outside Virgaviridae.

    AUC 0.899; 212 TP, 73 FP, 209 TN, 16 FN; precision 74.4%, recall 93.0%, F1 82.7%. Accuracy: tob1 96.4%, tob2 96.3%, tob3 75.0%, mas 40.9%. 39 of 41 non-Virgaviridae viral sequences went to the outgroup. · Machine learning approach aids in data curation…; Machine learning–assisted curation: strengths…

  5. ShownFreshwater (Amazon River metatranscriptomes) supplied most novel tobamoviral clusters. Wastewater was rich in mostly known tobamoviral contigs. The most divergent sequences sit near basal positions, but that placement is not stable across ORF trees.

    Of 23 SRRs with tobamoviral sequences, 10 were Amazon River (PRJNA237345). Wastewater held 43 of 228 contigs (39 known). Divergent tob3 clusters (2, 16, 25, 31, 33) were detected in the Amazon datasets. Clusters 2, 16, 31 and 33 sat at basal positions in the placement tree. In the ORF trees, clusters 2, 31 and 33 fell within known tobamoviral clades, often on long branches. · Novel tobamoviral contigs are associated with…; Phylogenetic analyses reveal clustering of novel…; Discovery of a diverse set of previously unknown…; Highly divergent tobamoviral sequences

Methods, data and reproducibility

Serratus PalmID hits (50-95% RdRp identity) selected SRA runs. These were assembled and screened in a Snakemake pipeline, curated by an expert into six categories, and classified by a random-forest and stacked-generalization model. Novel contigs were clustered and placed phylogenetically.

  • SRA short-read runs flagged by Serratus PalmID against RdRp of 35 tobamovirus species · virus (tobamoviruses; mixed environmental and host samples) · n = 6,870 runs, 253 after identity filter, 131 with candidate contigs
  • Amazon River metatranscriptomic datasets, main source of novel clusters · mixed environmental community (freshwater) · n = 10 of the 23 SRRs with tobamoviral sequences · PRJNA237345
  • Curated candidate contigs from the pipeline, expert-labelled in six categories · virus · n = 510 contigs; 228 tobamoviral
  • Reference Virgaviridae genomes (clustered at 90% identity) for training with synthetic contigs · virus (Virgaviridae) · n = 61 tobamoviral and 27 outgroup genomes; 34,492 ORFs, 148 features
  • External test: NCBI viral representatives deduplicated by genus, passing a DIAMOND prefilter · virus · n = 1,398 sequences; 41 passed the prefilter
  • Code and supplementary archives on Zenodo · 10.5281/zenodo.23366618; 10.5281/zenodo.23366548; 10.5281/zenodo.23366154
Methods & tools
Serratus PalmID, Snakemake v7.32.4, SPAdes / MEGAHIT, DIAMOND, MEGAN6, scikit-learn, MAFFT, IQ-TREE v3.0.1, EPA-ng, NetworkX, BLAST+, orfipy
Software & environment
Tool versions given in Methods (e.g. Snakemake v7.32.4, MAFFT v7.520, SciPy v1.14.1, IQ-TREE v3.0.1, EPA-ng v0.3.8). No container or requirements file mentioned.
Reproducibility gaps
The GitHub address of tobamo-supp-data is printed with spaces, so no verified URL is listed; the Zenodo DOI is given instead. No compute resources reported.
Limitations (5)
  • Short or fragmented contigs obscure evolutionary relationships. Placements of highly divergent sequences differ between the placement tree and individual ORF trees. · stated by the authors
  • The minimum count of 32 novel viruses is a lower bound that assumes non-recombinant genomes. · stated by the authors
  • The classifier was trained mainly on Virgaviridae and excluded the movement protein ORF, which caused many errors. It over-called (precision 74.4%) and failed on misassembled contigs. Expert review remains essential. · stated by the authors
  • The approach suits environmentally stable viruses like tobamoviruses; less stable viruses are likely under-represented in public repositories. · stated by the authors
  • Evaluation on the 510 contigs uses the authors' own manual curation as ground truth, with no independent labels or wet-lab confirmation of the novel viruses. · noted in this summary

Read this if you want a reusable data-mining and contig-classification workflow for lineage-specific RNA virus discovery in public sequencing data, or the tobamovirus diversity results.

Citations
1

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISChuprikova '26Rothman '22Bejerman '21Gilbert '19Lood '26Olendraite '22
62% semantic matchBioinformatics
Liubov Chuprikova, Li Chuin Chong, Sergej Ruff+2 more·bioRxiv·Jan 28, 2026