Currently megablast is running to test the proportion of the mistake. There are also gene transcripts that are correct and match their annotated genomic sequence. But even with smaller data set computation is intensive.

I am afraid we don't have enough information to pinpoint the source of your problem. I imagine such discrepancies could be caused by i) using a different genome assembly; ii) having some kind of alternative transcripts with intron inclusion, alternative promoter or not terminated where you would expect them to be; iii) human or computer errors in annotation. Anything else?

Ab initio gene predictors use gene models in their predictions. For that they use a set of genes to create a model to predict additional genes. Repeats, esp. complex repeats are usually of viral origin which often include genes that are compositionally and structurally different to the rest of the genome. Because of this, gene models including many of this "foreign" genes don't perform as expected and both underperforming and mispredicting many protein-coding genes.

Pretty sure that this is something that has come around in Maker's mail list.

For the coordinates issue, are you using the annotation file that corresponds to your version of the genome? Annotations with coordinates are specific for the genome they have been generated for.

Thank you for your answers, there is space for thinking now.

Different genome assembly - shouldn't be truth, as I downloaded new genome assembly with annotation files from one source/folder of public repository, and this non-model plant genome isn't sequenced elsewhere. At least there wasn’t any other annotation file. Furthermore, this second version of the assembly contains new gene ID naming index, that was common for genes in all files for the v.2. Some genes still perfectly correspond to genome coordinates and to their transcripts.

I was downloading this genome version after a year after the genome version publication. I was completely new to whole-genome bioinformatics studies, therefore some experienced researchers helped me to download and process all the data. I didn’t expect errors in the initial files. However, owner of this genome (another 10 months was passed) has changed all annotation files and transcripts, they now are re-named again completely. Looks like we took uncompleted annotation...

How one could test the genome annotation for the accuracy?

As far as I look at them in the annotation files genes, that wasn’t correct was marked with ‘AUGUSTUS’, but correct genes with 'GmapIndex_5000bpSoftMaskedNew’.

The gene prediction is always performed in the one direction: from transcripts to genome scaffolds / or there exists reverse approaches? I could imagine, that some TE transcripts were defined as a genes, so if the one family repeats have 80% similarity, but genes are separated on sim level 98%, all those repeats from the one family could have their false gene ID and corresponding coordinates to the genome.