whether vclust is also suitable for plant viruses assembled from transcriptomic data.

[zhangwenda0518]

Hi, I’ve carefully read your article and noticed that vclust has been tested on metagenomic and phage virus datasets. I’d like to ask whether vclust is also suitable for plant viruses assembled from transcriptomic data.

I’m currently mining plant viruses from public plant transcriptome datasets. For the species I’m focusing on, I’ve collected more than 300 samples. After assembling the reads and performing virus identification, I carried out taxonomic classification and predicted potential plant viruses based on their taxonomic ranks and host relationships. Although I used MMseqs2 to dereplicate the assembled sequences, there are still a large number of fragments remaining, and I’m stuck at this step.

I’m considering using vclust to generate vOTUs as the next step. Would that be appropriate? Because the data are transcriptomic, the assembled contigs are relatively short — their lengths range from 500 bp to 20 kb, but the vast majority fall between 500 and 3000 bp, so they are quite fragmented. Could this short fragmentation affect the results?

Below are some intermediate results from my pipeline (assembly results, virus identification results, MMseqs2 clustering results, and the plant virus candidates).

I would really appreciate any advice. Thank you very much!

Genome                                                  Size     Contigs    Max_len    N50      N90      >500bp_Num   >500bp_Ratio   >1000bp_Num  >1000bp_Ratio
megahit.mix.fasta                                       1.2G     1666981    40290      789      355      786973       71.10%         266714       40.39%
mix-cobra.merged.fasta                                  945M     1349084    40534      894      303      560222       71.73%         217827       46.26%
mix-mmseqs.cluster_rep_seq.fasta                        308M     473343     40534      741      286      183270       67.64%         59849        39.53%
Plant.classified.fasta                                  49M      49576      13672      1006     558      49371        99.79%         13800        50.26%

Furthermore, the most abundant families and genera found in the analysis results are distributed as follows:

─ Top-15 Family ─
  Alphaflexiviridae                                      62,249
  Betaflexiviridae                                       12,708
  Rhabdoviridae                                           3,815
  Caulimoviridae                                            530
  Pospiviroidae                                             191
  Bromoviridae                                               84
  Closteroviridae                                            70
  Fimoviridae                                                42
  Potyviridae                                                36
  Tombusviridae                                              25
  Secoviridae                                                23
  Atkinsviridae                                              15
  Tymoviridae                                                11
  Benyviridae                                                10
  Virgaviridae                                                9

─ Top-15 Genus ─
  Potexvirus                                             62,213
  Foveavirus                                              7,056
  Carlavirus                                              4,684
  Betacytorhabdovirus                                     3,784
  Vitivirus                                                 855
  Soymovirus                                                451
  Pospiviroid                                               191
  Ilarvirus                                                  83
  Banmivirus                                                 70
  Caulimovirus                                               51
  Emaravirus                                                 42
  Lolavirus                                                  28
  Crinivirus                                                 25
  Ipomovirus                                                 22
  Closterovirus                                              21

[zhangwenda0518]

One more question: at which step should I apply vclust clustering?
Should it be on

1. the initial assembly results,
2. the virus identification results,
3. the sequences after taxonomic classification (at the family or genus level)?

When considering options 1 and 2, my main concern is whether the length heterogeneity among contigs could lead to misclustering — for example, causing cross-family or cross-genus clusters to form. Is this a common issue?

Thank you again!

[zhangwenda0518]

Hello author, I have carefully read the previous discussion about the vclust algorithm. In generate vOTUs, you recommend using Leiden clustering with 95% ANI and 85% AF #26 #20

However, I tried to randomly select 3-5 plant viruses from the same genus and include all different genera myself, simulated random segment sizes （20-90%，SW-ON：random segments overlap by 50%, SW-OFF：random segments do not overlap）of the viral genome, and compared different algorithms using 95% ANI and 85% AF, but the results showed the advantages of the Single Linkage method. Is it a problem I simulated?

（264 species, 110K seqs）

algorithm ARI	Homogeneity	Completeness	V-measure	Purity	cluster	time
single	0.933	0.988	1.000	0.994	0.965	256	93s
set-cover	0.793	0.998	0.904	0.949	0.997	1,191	95s
cd-hit	0.775	0.999	0.899	0.946	0.997	1,058	91s
uclust	0.686	1.000	0.876	0.934	0.999	1,058	103s
leiden	0.443	0.999	0.788	0.881	0.998	2,130	102s
complete	0.039	1.000	0.588	0.741	1.000	13,826	103s

[zhangwenda0518]

@aziele Sorry to bother you, can you help me？

[aziele]

Hi!

Sorry for the delayed reply - I hope this is still helpful.

For your use case, I would recommend applying Vclust after virus identification, once you have a dataset of putative viral sequences. Plant viruses use somewhat different criteria for species/genus classification, but for vOTU generation I'd suggest the commonly used threshold of ANI ≥ 95% over ≥ 85% of the length of the shorter sequence (in Vclust: --ani and --qcov). These vOTUs are typically closer to species-level units, though this can vary between virus groups.

One important caveat is that the current version of Vclust does not support multi-contig genomes. Each contig is treated as a separate genome, so highly fragmented transcriptome assemblies can influence clustering. This is something to keep in mind when interpreting results.

Regarding your benchmarking, the behavior you see is actually expected. With fragmented genomes, single-linkage often performs well because it can connect sequences through chains of overlaps (AB and BC → A, B, C in one cluster), even if A and C are not directly similar enough. This tends to help recover fragmented species-level groups and can increase completeness. In contrast, Leiden or complete-linkage are more conservative. With strong fragmentation, they may split sequences that belong to the same species into multiple clusters, which likely explains the lower completeness and ARI in your simulation. If you have at least one reasonably complete genome per species, these methods usually tend to produce cleaner clusters, as fragmented genomes can then map back to the complete reference.

Could you share a bit more about your final goal? For example, are you aiming to:

• dereplicate similar assemblies,
• define species-level vOTUs,
• or also approximate genus-level grouping?

The best clustering strategy really depends on the downstream use case.

[zhangwenda0518]

Thank you, still very helpful.

As you said, the problem I am currently encountering is based on conventional transcriptome assembly. Most virus assemblies are incomplete. When using Leiden clustering, limited by the incompleteness of the longest fragment, many small clusters will be produced. These clusters are usually less than 1000bp. In fact, some of them can be removed by blastn.

My downstream wants to define species-level vOTUs and reduce vOTUs as much as possible. Do you have any suggestions?

Thank you, sorry for causing trouble！

contigs         <1000 bp    >1000 bp
Potexvirus.fasta      44494    17719
all.cluster.ref.fasta (Leiden)     4673    1011
all.cluster.ref.fasta (single-linkage)    80   14
genome.RD.fasta (blastn removes less than 1000 bp)    1777   17719
all.cluster.ref.fasta (Leiden)    646   1678
all.cluster.ref.fasta (single-linkage)    206   27

Regarding this result, what I don't understand is that when only looking at the value greater than ani95, shouldn't the value of the Y-axis be bigger and better? How did you choose Leiden in the end?

[aziele]

No trouble at all.

Given your goal, I might try a two-step approach:

1. Extend your dataset with complete plant virus genomes from the ICTV Virus Metadata Resource and NCBI Virus. Plant virus genomes can be downloaded relatively easily from these resources. Then run CD-HIT clustering with ANI ≥ 95% and QCOV ≥ 85% on the combined dataset. Since CD-HIT uses a representative sequence (the longest one) and requires all cluster members to satisfy the threshold relative to that representative, it may be suited for grouping fragmented contigs with known species. The goal of this step would be to associate as many transcriptome-derived fragments as possible with already characterized viruses. After clustering, I would focus on the clusters that contain both your sequences and ICTV/NCBI reference genomes.
2. For the remaining sequences (i.e., those that do not cluster with any ICTV/NCBI references), you could perform a second round using Leiden with ANI ≥ 95% and QCOV ≥ 85%. This would provide a more classical species-level vOTU definition for potentially novel diversity. If the results are still unsatisfactory because of the degree of fragmentation, I would consider moving to amino-acid-level comparisons?

As for the figure, we did this analysis specifically for clustering at the 95% ANI threshold (and 85% query coverage). As described in the figure caption, we compared the ANI distribution between genome pairs assigned to the same cluster and genome pairs assigned to different clusters. The goal was not simply to maximize the number of pairs above 95% ANI within clusters. Rather, we wanted to balance two effects: (1) minimizing the inclusion of pairs with ANI below 95% within the same cluster, and (2) minimizing the number of pairs with ANI ≥95% that were assigned to different clusters. Among the algorithms we tested, Leiden provided a good compromise between these two objectives.