[Question] Can this be used broadly for all microeukaryotes (e.g., fungi, protists, etc)?

[jolespin]

I'm thinking about trying this out for a backend for my VEBA eukaryotic binning module (https://github.com/jolespin/veba) as an alternative to MetaEuk.

Looking at the examples here:

Helixer.py --lineage land_plant --fasta-path Arabidopsis_lyrata.v.1.0.dna.chromosome.8.fa  \
  --species Arabidopsis_lyrata --gff-output-path Arabidopsis_lyrata_chromosome8_helixer.gff3

I won't have the lineage or species at this point in the pipeline.

• How important is the --lineage parameter in the gene calling process?
• Can --species be anything? For example, can I give it a bin ID such as S1__METABAT2__bin.1?
• Will this be available on bioconda any time soon?
• Are there any plans to implement this with BUSCO lineages to do autodetection of lineage and then use models based on BUSCO lineages?

[BjoernUsadel]

Hi the lineage selects the trained models currently one of

• land plants
• invertebrates
• vertebrates
• fungi hence you need this or alternatively model filepath set with "--model-filepath"

Do you want to align it to the BUSCO lineages? This could be done by a simple lookup. However as gene paramters change for the gene detection depending on these broadf lineages these are needed. Cheers b

[jolespin]

I meant if Helixer had a model for each eukaryotic BUSCO lineage, then it does auto detection in the backend to know which model to use.

I guess I'm asking if there are plans to build out the models available.

[alisandra]

Hi Jolespin,

Thanks for your interest!

This is a cool idea on auto-detection. One would have to test it, but I'd hazard a guess that auto detection might work off of the confidence of Helixer's raw predictions.

That said, I don't currently have much capacity for larger schemes, so in summary

• yes, --lineage is critical for current models, else a poorly fitting model would be used; and prediction quality would suffer massively.
• --species should be any reasonable length non-white space string. It's going to end up in the gff3 file's gene names. A prefix such as 'At' or 'A_thal' for Arabidopsis thaliana would be reasonable. The bin ID should work without issue.
• not in the immediate future
• not in the immediate future

That said on the no's, if any one wants to see this enough to give it a shot, I'm happy to do what I can at a high-level; drop a line.

[jolespin]

Understood the bandwidth issue. I'm in the same boat right now w/ my VEBA package.

I would be using this for microeukaryotic organisms (e.g., protists). Do you have any instructions on training a custom model? If so, what is required for this? Genomic and CDS sequences or could this be done directly from protein sequences?

[alisandra]

Yes, I'm currently working on the latest instructions here: https://github.com/weberlab-hhu/Helixer/blob/cleanup/docs/training.md (I will merge them to main soon).

You will need fasta and gff3 files for the training species (it's supervised training). I am not on top of the availability of good references in protists, I can imagine it might be a challenge.

Drop questions whenever!

[jolespin]

For the training data, are genes with alternative start codons used or discarded?

[alisandra]

In the current default implementation, genes with non-ATG start codons are used, but the upstream region is masked. So the network will learn there's a gene there, but not receive feedback on where exactly it started.

This behavior is of course not designed for alternative start codons, but is a side effect of the partial-gene-model detection and masking, which assumes the standard genetic code.

In general, supporting genetic code variants is on the "think about it" list; not doing so yet has been a rare--but still noticeable--issue in fungi already.

[jolespin]

Any updates on whether or not this should be able to handle protists soon?

I've compiled this huge protein set for protists and fungi: https://zenodo.org/records/10139451 Though, the genomes aren't available in this dataset unfortunately.