Can I use vclust to cluster bacterial sequences?

[inspirewind]

Thank you for developing such a convenient and powerful tool!

I'm wondering why vclust emphasizes clustering viral sequences—can I use it for dereplicating bacterial sequences? Is it because viral sequences are generally shorter than bacterial ones, and vclust has specific optimizations for this?

[aziele]

Thank you for the kind words!

Vclust can be used on genome sequences in general, but it's optimized for shorter genomes, like viruses. It's very sensitive, which is great for viral genomes but most likely too sensitive for longer, bacterial genomes. Also, Vclust currently treats each sequence as a separate genome, which won't work for bacterial assemblies with multiple contigs.

We're planning a major update that will add support for multi-contig genomes and introduce sensitivity modes to better balance speed and resolution. Ultimately, we aim to support bacterial genomes as well, though we can't provide a timeline yet.

[inspirewind]

Thank you for the explanation about the algorithm principles. I think that perhaps having excessively high sensitivity on long sequences (such as bacterial ones) might actually be beneficial. In some cases, we may need very aggressive filtering strategies to obtain a smaller subset.

For example, when training a nucleic acid language model, limited by computational resources, we might need to fix the dataset to an appropriate size, while the initially compiled training data could far exceed this size.

I tried clustering the mgnify genomes using vclust, and it seems to be working well. The clustering reduced the dataset from 5,238,165 sequences to 3,365,149 subsets. Do you have any suggestions for adjusting vclust parameters to further reduce the size of the clustered dataset?