binsplitting not happening with different separator?

[xvazquezc]

Hi there,

I've been using Vamb for a while (esp. TaxVamb since 4.1.4.dev150+g8fa3280) and just decided to update recently. I installed Vamb from master (5.0.5.dev20+g8a13cf5f8) and after dealing with the more strict checks for the taxonomy files (now every contig needs to be listed) I just run into something a bit odd, i.e. the vaevae_clusters_split.tsv file is basically empty (only column headers) but there is no error or warning in the log regarding the binsplitting not happening.

(base) z3382651@katana1:.../Functional/binning $ ll -rht taxvamb5_out/
total 881M
-rw-rw----+ 1 z3382651 ferrari 288M Nov  5 00:52 composition.npz
-rw-rw----+ 1 z3382651 ferrari  30M Nov  5 00:53 abundance.npz
-rw-rw----+ 1 z3382651 ferrari  32M Nov  5 02:46 predictor_model.pt
-rw-rw----+ 1 z3382651 ferrari 169M Nov  5 03:25 results_taxometer.tsv
-rw-rw----+ 1 z3382651 ferrari 202M Nov  5 10:26 vaevae_model.pt
-rw-rw----+ 1 z3382651 ferrari  87M Nov  5 10:28 vaevae_latent.npz
-rw-rw----+ 1 z3382651 ferrari   23 Nov  5 10:35 vaevae_clusters_split.tsv
-rw-rw----+ 1 z3382651 ferrari  53M Nov  5 10:35 vaevae_clusters_unsplit.tsv
-rw-rw----+ 1 z3382651 ferrari  22M Nov  5 10:35 vaevae_clusters_metadata.tsv
drwxrws---+ 2 z3382651 ferrari  44K Nov  5 10:36 bins
-rw-rw----+ 1 z3382651 ferrari 159K Nov  5 10:36 log.txt

I'm reusing already "labelled" contigs in the form of {SAMPLE}-{CONTIG}, i.e., using - as separator (and indicated as such in the vamb bin taxvamb command), and seemingly being detected (it shows at the end of the log):

2025-11-05 10:28:06.293 | INFO    | Clustering
2025-11-05 10:28:06.294 | INFO    |     Windowsize: 300
2025-11-05 10:28:06.294 | INFO    |     Min successful thresholds detected: 15
2025-11-05 10:28:06.294 | INFO    |     Max clusters: None
2025-11-05 10:28:06.294 | INFO    |     Use CUDA for clustering: True
2025-11-05 10:28:06.294 | INFO    |     Binsplitter: "-"
2025-11-05 10:28:06.456 | INFO    |       10 % of contigs clustered
2025-11-05 10:28:33.912 | INFO    |       20 % of contigs clustered
2025-11-05 10:28:40.737 | INFO    |       30 % of contigs clustered
2025-11-05 10:28:51.559 | INFO    |       40 % of contigs clustered
2025-11-05 10:29:09.755 | INFO    |       50 % of contigs clustered
2025-11-05 10:29:43.168 | INFO    |       60 % of contigs clustered
2025-11-05 10:30:36.292 | INFO    |       70 % of contigs clustered
2025-11-05 10:31:44.854 | INFO    |       80 % of contigs clustered
2025-11-05 10:32:54.945 | INFO    |       90 % of contigs clustered
2025-11-05 10:34:03.634 | INFO    |      100 % of contigs clustered
2025-11-05 10:35:11.088 | INFO    |     Clustered 1120675 contigs in 497327 split bins (480100 clusters)
2025-11-05 10:35:11.090 | INFO    |     Wrote cluster file(s) in 424.8 seconds.
2025-11-05 10:36:08.708 | INFO    |     Wrote clusters above 200000 bp to FASTA files in 57.62 seconds.

I only found out because if I run vamb recluster off the Vamb 5 output, it doesn't generate any bins, whether you indicate the binsplitter or not.

As a side note, the same data was successfully run with the old Vamb 4.1.4.dev150+g8fa3280, with both the binsplitting and recluster working without issues despite the alternate binsplitter character:

(base) z3382651@katana1:.../Functional/binning $ ll -rht taxvamb_out/
total 865M
-rw-rw----+ 1 z3382651 ferrari 286M Nov  1 09:47 composition.npz
-rw-rw----+ 1 z3382651 ferrari  30M Nov  1 09:47 abundance.npz
-rw-rw----+ 1 z3382651 ferrari  36M Nov  2 10:14 predictor_model.pt
-rw-rw----+ 1 z3382651 ferrari 127M Nov  2 11:09 results_taxometer.tsv
-rw-rw----+ 1 z3382651 ferrari 216M Nov  5 10:43 vaevae_model.pt
-rw-rw----+ 1 z3382651 ferrari  87M Nov  5 10:47 vaevae_latent.npz
-rw-rw----+ 1 z3382651 ferrari  30M Nov  5 12:46 vaevae_clusters_metadata.tsv
-rw-rw----+ 1 z3382651 ferrari  25M Nov  5 12:46 vaevae_clusters_unsplit.tsv
-rw-rw----+ 1 z3382651 ferrari  30M Nov  5 12:46 vaevae_clusters_split.tsv
drwxrws---+ 2 z3382651 ferrari  40K Nov  5 12:47 bins
-rw-rw----+ 1 z3382651 ferrari 151K Nov  5 12:47 log.txt

The commands used for both runs were the same, with the exception of --cuda removed from the Vamb 4 run as the university HPC has a 12h limit on the GPU queue and was going overtime:

vamb bin taxvamb -p 8 --cuda -o - \
--outdir taxvamb5_out --fasta merged_contigs.fasta \
--abundance_tsv abundance.tsv \
--taxonomy mmseqs-easy/taxonomy_lca.taxconv.tsv \
--minfasta 200000

vamb recluster -p 8 --cuda -o - \
--outdir recluster5_out \
--fasta merged_contigs.fasta \
--abundance taxvamb5_out/abundance.npz \
--latent_path taxvamb5_out/vaevae_latent.npz \
--taxonomy taxvamb5_out/results_taxometer.tsv \
--clusters_path taxvamb5_out/vaevae_clusters_split.tsv \
--hmm_path /srv/scratch/ferrari/utils/vamb/vamb/marker.hmm \
--minfasta 200000

[xvazquezc]

As I haven't got any response, I reran the dataset with the latest TaxVamb using the default "C" binsplitter. While I got very different number of bins (1540 vs 1759 in the prior run), which might relate to the random seed number(?), the vaevae_clusters_split.tsv file is still empty.

According to the Changelog, the split cluster are not output if the binsplitting is not happening. To be honest, I have no idea if it's not happening, if it's happening but it's not splitting anything or what's the reason as I don't see anything in the log nor I see any option for increased verbosity that may help figuring out

I include the log.txt in hopes anyone may have answers

[jakobnissen]

I'll take a look now and see if I can identify the issue. It's almost certainly due to a recent refactor of reclustering (commits 6480f6c..7826426) which may have introduced a bug.

Until I find the issue, you can binsplit manually. Simply look at your unsplit bins, and for each bin, partition it by bin_name.partition('C')[0].

[jakobnissen]

@xvazquezc I think I found the error.

Can you check if the file vae_unsplit.txt accidentally contains both the split and unsplit clusters? You can count the lines with wc -l and see if it fits with the expected ~1120676 lines, or if it has about the double amount.
I'll push a fix shortly.

[xvazquezc]

Yeah, It contains about twice as many lines