Gene number decreases when rerunning Funannotate on the same genome

[BKGiwa]

Hello,

I am using Funannotate to annotate fungal genomes belonging to the same species and I noticed an issue with reproducibility when rerunning the annotation on the same genome assembly. When I run Funannotate on the same genome multiple times using the same parameters, the total number of predicted genes decreases between runs. I expected the annotation results to be consistent when using the same input genome and settings.

I have used Funannotate ver 1 and 2. I used version 2 without providing transcript evidence and used the Busco genes from buscolite (default).

Run 1: 12951 genes
RUN 2: 12863 genes
RUN 3: 11850 genes
RUN 4: 11669 genes

The discrepancy appears to be introduced when transcripts were derived from Trinity and this produced different number of transcripts. Then the number between the runs kept increasing which was reflected in the total number of genes predicted.

Run 1 and 2 = Funannotate ver1 was used with RNA seq reads as evidence
Run 3 and 4 = Funannotate ver2, no evidence just BUSCO evidence

Could you clarify whether Funannotate is expected to produce different gene counts between repeated runs due to internal training/optimization steps, or if there are recommended settings to ensure reproducible annotations?

Thank you.

[hyphaltip]

funannotate2 is quite different set of algorithms so I would not expect the same results from comparing these two tools. You may want to more concretely compare the results in gene content produced not just numbers as a decrease is potentially a fusion of two gene models not necessarily less coding capacity predicted. Looking at overlap of the exons produced from both of these or using AED distance or other metrics to compare the similarity of the genes prediction.

but generally these are not versions of a composite gene calling pipeline that will give identical results, there is a lot wrapped up in how consensus gene calls are scored and f2 doesn't even use all the same predictors as f1 or same consensus method. You can empirically decide how much each version is producing high quality results by looking exons supported by RNAseq and proteins and some other indirect methods too to make assessments of which tool you will follow.