I have been obsessing over errors and contamination in our public sequence databases. This week I was trying to use UniProt as a set of reference sequences for fungi. Our goal is fairly simple: To find the fungal DNA in a metagenomic shotgun sequence sample - which is just a mixture of all the DNA present in a scraping from mouth, throat, or any other body site.
UniProt makes it quite easy to sort all their proteins by taxonomy, and to download a subset of the data clustered at 100% (combining all exact duplicate sequences), 90%, or 50% amino acid identity. One might expect that fungal genes should not match bacteria at more than 50% identity. But surprisingly there are quite a lot of 50% and 90% clusters that contain both bacterial and fungal sequences (about 3000 of the 90% fungal clusters also contain bacterial proteins).
The UniProt support staff provided some very useful help to build a query on their system that finds only those clusters of 90% identical proteins that contain fungal genes, but NO (NO!) bacterial genes. In case you like this sort of thing, here is the exact query:
uniprot:(taxonomy:"Fungi [4751]") NOT taxonomy:"Bacteria [2]" AND identity:0.9
[Note the careful use of quote marks parenthesis and square brackets, this stuff is rather tricky]
So I downloaded this set of putative fungal proteins (UniProt very helpfully creates a single 'representative' UniRef sequence in FASTA format for each cluster). I tested the fungal proteins against all the gene coding sequences (CDS) from the E.coli genome using BLASTx. Once again, there are far too many high similarity matches.
One of the top matches is to a gene (Guanosine-3',5'-bis(diphosphate) 3'-pyrophosphohydrolase) from the fungus Beauveria bassiana that has 98% identity to E.coli. Since I am in an obsessive mood about this sort of thing, I decided that for this one example, I would collect some evidence to decide if we have strong sequence homology between bacteria and fungi for this gene, if Beauveria bassiana has a horizontal gene transfer, or if E. COLI CAN BE A CONTAMINANT IN GENOME SEQUENCES (!!!) [emphasis mine]
I put this Beauveria gene into a generic NCBI BLAST against all 'nr' proteins, and I got a very interesting result. There are exactly two matches to eukaryotes (Beauveria and a nematode), and 11,858 matches to bacteria, including lots of E.coli.
So I traced the Beauveria bassiana protein in UniProt back to its source as a whole genome shotgun sequence uploaded to GenBank on Nov 3, 2014 by the Institute of Plant Protection, Jilin Academy of Agricultural Sciences, Accession PRJNA178080, WGS ANFO00000000, Assembly GCA_000770705.1.
I downloaded the whole genome assembly and BLASTed it with the E.coli hydrolase gene from above. This very quickly pinpointed a contig 00271 (ANFO01000251.1 Beauveria bassiana D1-5 contig00271) that contains the matching sequence. The contig is 72,232 bases long. I then put this conting into NCBI BLAST against Bacteria. I get matches that correspond to lots of bacterial genes (POL I, RecG, iPGM, XanP, CpxA, GTP binding protein, GSI beta, and my ppGpp hydrolase) all with >90% identity and BLAST e-value 0.0.
Final answer: This is a contaminant. There was some E.coli DNA sequenced and assembled with the Beauveria DNA, and nobody checked before loading these sequences into GenBank.
My recommendation to GenBank and de novo genome sequencers everywhere is to check all predicted proteins from new genomes for matches to bacteria and human before loading them into a trusted database.
