Prepping input for mikado

[desmodus1984]

Hi,
I have struggled with installing mikado, but I want to ask a question about input data.
I have annotated my genome with Tiberius 1st, and then I used the output as input for annotation with EviAnn using protein, transcript, and RNA-seq data. I was recommended to use Mikado to polish the annotation.
I mapped all the reads, merged the bam files and then used Portcullis to polish the boundaries. Also, I will input the Tiberius and EviAnn gffs, but I wanted to have an addition gff based on the NCBI annotation of an individual of the same species. I checked the gff, and it has annotated pseudogenes. I tried liftoff annotation transfer _+ polishing using the gff with and without pseudogenes, and I got more genes/exons/CDS using the gff without pseudogenes - which kinda does make sense.

I have a question.
Should I filter the polished gff with gffread to keep only genes with START/STOP codons and without in-frame STOP Codons, or
should I feed Mikado the raw gff output by Liftoff?

Another question: should I use a softmasked genome or a regular fasta file is OK?

Thanks.

[swarbred]

Hi @desmodus1984
You can use mikado to select among different gene models but generally you would want to make use of additional metrics other than those mikado calculates, i.e. if you have a mix of gene predictions and liftoff / lifton models you would give these in as models but also calculate additional metrics e.g. comparing each model to all other models so you can score models that share the same exon-intron structure higher. We use mikado in this context in our minos pipeline https://github.com/EI-CoreBioinformatics/minos . You might want to check that out, though we haven't fully written a step by step walk through with detailed notes on the best configuration and scoring (though we will cover in the online workshop we are running in April https://www.earlham.ac.uk/events/genome-annotation-workshop-2026).

On "Should I filter the polished gff with gffread to keep only genes with START/STOP codons and without in-frame STOP Codons, or
should I feed Mikado the raw gff output by Liftoff?"

You can do either, mikado prepare will throw out models with internal stops anyway, models lacking the start / stop could be removed / heavily penalised in the scoring configuration for mikado pick . But if you want to simply remove these before then thats fine, as long as you are ok with that hard filter (i.e. you know you have other models that can potentially fill any gaps)

Below is an example I use for a small toy dataset for arabidopsis, where we have done similar to what you have done.

1. Ran lifton / liftoff on a number of related species
2. Ran https://github.com/EI-CoreBioinformatics/LiftClean to remove some problematic models / flag any issues (sometimes you can have LiftOn (especially) or Liftoff models that will break downstream tools)
3. As i had lots of alternative models (and could afford to safely exclude some) i also filtered with gffread -C -N -V -J
   4 Ran minos with the liftOn, liftoff, helixer and tiberius models and also giving these in as evidence (that effectively favours the concensus exon structure)

Below is the comparison to the reference Arabidopsis annotation for all of these models, including the final minos output (as mentioned it's a toy dataset and not necessarily representative of how specific tools perform in all contexts, but gives an idea of how the different gene sets compare)

[desmodus1984]

Hi,
I am following https://github.com/BaderLab/GenAnT , and I am trying to improve or refine the annotation with mikados.
I used as input the gff from Tiberius, Litoff, and the Eviann annotation.
I used EviAnn annotation as "ref, and it had the following BUSCO scores:
C:95.7%[S:25.0%,D:70.7%],F:1.0%,M:3.2%,n:13727
and the mikado one, following GenAnt tutorial got:
C:96.7%[S:50.2%,D:46.5%],F:1.2%,M:2.2%,n:13727

Since those were done using all proteins, and there are a lot of duplicates, so I used agat to get the longest isoform per gene and the stats changed to:
original: C:93.8%[S:92.8%,D:1.0%],F:1.7%,M:4.5%,n:13727
mikado: C:96.3%[S:95.3%,D:1.0%],F:1.4%,M:2.3%,n:13727
I read that a good BUSCO complete score is ~95%, at first the mikado_longest annotation looks great.

Nonentheless, it is based on "BUSCO" sets, so limited. Thus, I used OMark to assess proteome quality.
OMark uses 'peptide kmers' and does a similar to BUSCO using from hierarchical (clade) orthologues -representative clade proteins, and for all proteins.
It found that the mikado annotation had higher "unknown" and "inconsistent proteins" - based on Carnivora clade.

Consistency assessment is based on the proportion of query proteins placed in gene families of the correct lineage (consistent), gene families of an incorrect lineage (either randomly (Inconsistent) or to specific species (contamination)) and placed in no gene families at all (unknown).

#A:Consistent (taxonomically)[P:Partial hits,F:Fragmented], I: Inconsistent (taxonomically)[P:Partial hits,F:Fragmented], C: Likely Contamination[P:Partial hits,F:Fra>
A:49078[P:2501,F:1000],I:1174[P:347,F:280],C:0[P:0,F:0],U:8785
A:83.13%[P:4.24%,F:1.69%],I:1.99%[P:0.59%,F:0.47%],C:0.00%[P:0.00%,F:0.00%],**U:14.88%**

while, for the EviAnn annotation, it detected less unknown and inconsistent proteins.

#A:Consistent (taxonomically)[P:Partial hits,F:Fragmented], I: Inconsistent (taxonomically)[P:Partial hits,F:Fragmented], C: Likely Contamination[P:Partial hits,F:Fra>
A:110609[P:4657,F:9636],I:2236[P:393,F:478],C:0[P:0,F:0],U:1584
A:96.66%[P:4.07%,F:8.42%],I:1.95%[P:0.34%,F:0.42%],C:0.00%[P:0.00%,F:0.00%],U:1.38%

I was reading the scoring section, and it mentions about monoexonic genes. I got the stats from each annotation, and mikado had a lot of monoexonic genes: 9721, compared to original Eviann: 3967.

I am working with a marine mammal, and the best annotated and closest related species is the California sea lion, which has an Ensembl annotation. The gff stats are in the following file:
Zalcal.stats.txt
For mikado pick, I used the mammalian scoring in lenient mode.

I filtered the Mikado annotation with agat to retain only complete gene models, have START/STOP codon, and the BUSCO complete score reduced to 90.6%, and the unknown proteins were still very high, at least 8%- if using only complete gene models.

My main is to - if possible, increase the BUSCO complete score to >95%, but without affecting the quality scores from OMark, specially detecting a lot of unknown proteins.

I hope you can help determine a good scoring pattern.

Thank you.