Important update included at bottom of this post
Arabidopsis thaliana has one of the best curated eukaryotic genome sequences so you might expect working with Arabidopsis data to be a joy? Well, not always. Consider gene AT1G79920. This gene has two coding variants (AT1G79920.1 and AT1G79920.2) which only seem to differ in their 5' UTR regions:
The primary source for Arabidopsis genome information is The Arabidopsis Information Resource (TAIR) but another site called Phytozome has started collating all plant-related genome data (often pulling it from sites like TAIR).
Both sites allow you to download coding sequences for each gene from their FTP sites, or view the same sequence information on their web sites. You could also download GFF files and using the coordinate information, extract the coding sequences from the whole genome sequence.
A simple sanity check when working with coding sequences is that the length of a coding sequence should be divisible by 3 otherwise there might be a frameshift. There seems to be an issue with AT1G79920 in that it sometimes has a frameshift. I say sometimes because it depends on where you look at the data.
Here is a sequence alignment for exons 5 and 6 of both coding variants of this gene. The sequence identifiers include a 'P' or 'T' to indicate whether the information came from Phytozome or TAIR and they also denote whether it was taken from their web or FTP site. I also insert 'gtag' into one sequence to illustrate where the intron between these exons would occur.
The first boxed rectangle highlights a base that looks like a SNP, but these are coding variants with UTR variation, so the underlying genome sequence in the coding regions should be identical.
The second box is perhaps more disturbing. The web site versions of exon 7 all have a 1 bp deletion which would lead to a frameshift. I guess this error started at TAIR and propagated to Phytozome, but the fact that both sites also have a correct version available on their FTP site is confusing and troubling.
My boss first discovered this by looking at the GFF files for the Arabidopsis genome and this was one of 25 genes with a 'not-divisible-by-3' length error. So it pays to always check — and double check — your data.
Time to send an email to TAIR and Phytozome to report this.
I heard back from TAIR and Phytozome and it seems that there are a small number of likely genome sequence errors in the latest (TAIR10) release of the A. thaliana genome. When these affect genes and would introduce a frameshift error, TAIR make manual edits to the resulting files of CDSs, transcripts, and proteins. They do this when there is clear homology from genes in other species that suggests the change should be made.
So if you work with downloaded FASTA files from the FTP site, you won't see these errors. If you work from GFF files (which is presumably what some of their web displays do), you'll run into these issues. There is a small file (TAIR10sequenceedits.txt) included as part of the TAIR10 release which documents these changes.
Thanks to a speedy and helpful response from both sites. Perhaps I should retitle this post The Joys of Dealing with Fast and Knowledgable Genome Curators?