BUG: abinitio.py, TypeError: 'Logger' object is not callable

[Alteroldis]

Several times forgot to specify the logger object method. I looked at the lines relative to version 25.7.1 (commit 5c3b7ea). In current master branch abinitio.py: lines 455 (info), 718 (info), 918 (info), 1222 (info?), 1448 (info), 1460 (debug), 1487 (info), 1508 (info).

[nextgenusfs]

Can you paste the actual error please? I ran into a bunch of these and thought they were fixed, but obviously didn't get them all.

[Alteroldis]

For example during snap training:

Traceback (most recent call last):
  File "/home/Tools/funannotate2/bin/funannotate2", line 8, in <module>
    sys.exit(main())
             ^^^^^^
  File "/home/Tools/funannotate2/lib/python3.11/site-packages/funannotate2/__main__.py", line 28, in main
    train(args)
  File "/home/Tools/funannotate2/lib/python3.11/site-packages/funannotate2/train.py", line 240, in train
    snap_train = train_snap(
                 ^^^^^^^^^^^
  File "/home/Tools/funannotate2/lib/python3.11/site-packages/funannotate2/abinitio.py", line 918, in train_snap
    log(
TypeError: 'Logger' object is not callable

After fixing mentioned lines, all training completed successfully.

[nextgenusfs]

Thanks, can you try latest in main and see if fixed?

[Alteroldis]

Thanks, now (commit f2d52f8) the training is finished without errors.

I also decided to make a species-specific training set, because training on the metazoan set is very limited by the number of genes (~600) and I noticed an interesting thing. If I remove the terminal asterisk (stop codon) from the protein before performing the alignment, the function funannotate2.align.split_evidence_and_genes gives almost twice as few valid gene models. The same function is called after the alignment in the predict command. Doesn't this mean that using homologous/transcriptome predicted proteins without a stop codon will lead to a decrease in the number of predicted genes?
With stop codon: Generated 7358 alignments: 4718 were valid gene models
Without stop codon: Generated 7356 alignments: 2701 were valid gene models

Species-specific training set creation

Download homologous protein from ncbi with datasets. For example 17 genomes of echinoderms (5 closely related holothurians)
Align transcriptome predicted protein to homologous
Filter out predicted proteins with percent positives < 60 and subject coverage < 90
Remain only predicted proteins with hits to some number of related species. For example 5 closely related and one more distant.
Align protein set to genome: miniprot -t 24 -I --gff-only --outc=0.5 genome.fasta filtered.peps.faa >mp.pep2genome.gff. -G intron-len(for ex. 10000) --gff-only --outn=4 produces fewer gene models and depends more on species specific intron length
Activate funannotate environment: . /path/to/funannotate2-env/bin/activate
Generate funannotate compatibility gene models training set:

from funannotate2.align import split_evidence_and_genes
n_aligns, n_genes = split_evidence_and_genes(
    'mp.pep2genome.gff`', 'genome.fasta',
    'protein-alignments.gff3', 'predictions.miniprot-gene.gff3', gff_format="miniprot"
)
print(f"Generated {n_aligns} alignments: {n_genes} were valid gene models")

[nextgenusfs]

So I'm just parsing default miniprot results here and determining if there is a valid start and stop codon --> if so they got split out as "valid gene models" which means they could be considered in the consensus step as an actual gene model. So I guess it's not surprising that if you strip the stop that miniprot is basically resulting in an alignment that is likely 3 nucleotides "short" and thus missing the stop. If we wanted to try something like what I did for transcript alignments, ie try harder on each alignment to find a valid start and stop in the vicinity of the alignment it would take a little bit longer. Honestly, I'm not sure if this matters at all for the final consensus gene models -- it would really only matter if miniprot was able to find a valid gene model that the other ab initio predictors missed. However, how I have the scoring and weighting setup in the consensus module, is that at least two de novo predictions need to be in the same region in order for it to call a gene model -- this is an attempt to limit spurious gene model calls, ie those that are found only in one algorithm (ie snap which doesn't work all that well on its own).

[nextgenusfs]

Another way to train, is if there is a good public annotated genome in your species, you can actually use that annotation to generate a training set. You can pull that reference and annotation from NCBI (for example) and then run f2 train passing the annotation as the training-set, instead of running BUSCO it will use those gene models to generate the training set. You can then use that training set on your genome of interest. Note this could work also for a closely related species, I've just not tested this explicitly and not with any metazoans.