Joran Martijn (July 2024)

Automatic gene prediction algorithms/pipelines such as GeneMark, Braker2 and Funannotate can work quite nicely to get a good first estimate of gene models. However, it seems that a non-negligable number of gene models either contain introns that are not supported by the RNAseq data, or do not contain introns that are supported. These cases become obvious to the human eye when inspecting the gene models in context of the RNAseq data in for example IGV, GenomeView or similar tools.

Ignoring true introns or including false introns will lead to errors in gene models. These errors can be fixed manually (see also the other wiki page on curating gene models) but this can be quite tedious and extremely time consuming. In addition, manual edits can also introduce human error.

I've developed a python script, fix_genes_with_false_introns.py, that is available here on the GitHub repository of the lab that attempts to automatically fix gene models that contain intron features that are not supported by the RNAseq data.

Briefly, it iterates over all genes defined in an input GFF3 file, checks for each gene whether all of its introns are actually supported by the RNAseq data (provided as a BAM file). If at least one intron is not supported, the entire gene model is removed. Then, a region is defined, typically (but not always) between the upstream gene and the downstream gene, and for this region all supported intron locations are exctracted from the BAM file and subsequently spliced out of the region. This 'intron-less' sequence is then fed to a simple ORF finder, and the largest ORFs are selected. Introns are re-inserted, and the resulting ORFs are used to generate new gene features for the output GFF3 file.

This is a very straightforward but 'dumb' approach, as it does not make use of sophisticated Markov chain models. In my experience though, most new gene models are much better approximations of the true genes, but occassionally they will still be wrong, and will need to be curated by hand. For example, sometimes the shorter version of an ORF will correspond to the true gene, but the script will select the longer ORF nonetheless.