The NUCOCOG database (nucocog.xml) was constructed by repeatedly running a set of different PERL scripts to read, create and manipulate files containing ASCII data. All files containing the information required to create the NUCOCOG database were obtained via FTP from the NCBI. Three files related to the COG-database (whog, myva, myva=gb) were taken from 7 and 159 GBK files containing the genome information of the 66 organisms currently present in the COG database were taken from 9. To obtain a maximum quality of nucleotide sequences we used current GBK files rather than the outdated ones corresponding to the original protein sequences. In principle, this may lead to errors due to the methods described below. However, those errors are very improbable and, if occurring at all, are negligible compared to all the ambiguous amino acids and nucleotides present even in current GBK versions (see table 1). The NUCOCOG database was built up by the following five main procedures: (i) The complete COG database as assembled in whog, was converted to an XML file which served as the scaffold that was subsequently stepwise extended to eventually represent the final NUCOCOG database. (ii) The amino acid sequences were extracted from myva by searching for their respective protein names as the unambiguous search keys. (iii) In the same manner, GI-numbers for all complete protein sequences were obtained from myva=gb. However, no protein fragments (domains) at all could be detected because their names are extended by an underscore followed by a consecutive number in whog which is not the case in myva=gb. We derived GI-numbers for those entries during a second run after the extensions of their names were truncated. (iv) Nucleotide sequences were then included separately for each of the 66 organisms by exclusively searching for organism specific sequences only in GBK files associated with the respective query organism. The amino acid sequences as annotated in myva were used as search keys to locate their corresponding nucleotide sequences in the GBK files. To detect all sequences, it was necessary to apply four different search approaches. The vast majority (98.4 %) of all nucleotide sequences were extracted by searching for matching amino acid sequences as annotated in the GBK files. Because some of the GBK files used in this work have been updated since the latest release of the COG database not every sequence could be detected by this method. We discovered additional sequences (1.4 %) by searching for matches with conceptual translated nucleotide sequences. Those sequences were located and their reading frames were determined according to the CDS information in the respective GBK file. In addition, the nucleotide sequences were extended by 300 adjacent nucleotides derived from the genome sequence in both directions prior to translation. Further sequences (0.1 %) could be found by searching for matching amino acid sequences in translations of whole genomes in all six reading frames. The residual 121 missing nucleotide sequences were included by manually editing the first qualifiers of the CDS feature key in the respective GBG file followed by a further search in conceptually translated nucleotide sequences. To inspect and edit GBK files, the genome annotation tool ARTEMIS 1011 was used in combination with the pairwise sequence alignment software JAligner 12. Mainly, we deleted annotated frameshifts, included new frameshifts and in some cases replaced mismatching amino acids by using coding information of the nearest matching amino acid sequence within the same reading frame. For each nucleotide sequence the respective mapping method was included to the XML-files i. e."original CDS", "extension of original CDS", "conceptual translation of whole genome", or "created or edited CDS manually". (v) After this first version of the complete NUCOGOG database had been finished, some PERL scripts were run for verification and validation purposes with the main focus on translating the nucleotide sequences and comparing the translations to their respective annotations. For four entries (MK0324, MK0315, MK0689, MK0809) even the manual search for more information e. g. by BLAST 13 assigned no GI-number and we adopted the entry "gi?" from myva=gb.

The arNUCOCOG database (arnucocog.xml) was essentially constructed as described above using the information from ar40.fa and arCOG.csv 8 to build the initial XML-file. The current version of arCOGs includes the genome of Thermoproteus tenax which has not been published at the time of its release and by request of the sequencing consortium those proteins were removed from the ar40.fa file and are also not contained in the arNUCOCOG database. In addition, 34 sequences listed in arCOG.csv could not be located in ar40.fa. These proteins for various reasons were not translated and the authors detected them by tBLASTn, using an orthologous sequence from a close relative as a query (Kira Makarova, personal communication). We included those sequences manually by reproducing her work. Searching for matching amino acid sequences in the GBK files resulted in including 99.8 % of all nucleotide sequences. The remaining sequences were detected by the alternative methods described above and only three sequences needed to be searched manually. Many of the arCOGs are new and consequently not assigned to a classical COG-number. In all those cases, we included "NO_COG" between the respective tags. Because arCOG.csv contains protein gi-numbers as the domain-ids, no unique domain-ids are assigned to all split sequences. We improved this situation by adding consecutively numbered suffixes to those gi-numbers separated by an underscore e.g. <DOMAINNAME>118430839_1</DOMAINNAME>.

We also combined NUCOCOG and arNUCOCOG resulting in nucocog_2.xml. For that purpose, we removed all sequences from the 13 archaeal genomes from NUCOCOG and included all data from arNUCOCOG instead. In addition, we added those sequences from ar40.fa that according to the information from arCOG.csv are assigned to classical COGs but are not part of any arCOG. Finally, for those amino acid sequences their corresponding nucleotide sequences were included as described above and "NO_COG" was written between the arCOG-tags.

The content of the three database files is summarized in table 1. As can be seen, some nucleotide sequences contain stop codons within coding regions and there are both ambiguous amino acids (a. a. a.) and ambiguous nucleotides (a. n.). Consequently, in those cases a distinct translation of a codon to an amino acid is impossible. For that reason, although resulting in larger database files containing redundancies, we did not delete the amino acid sequences from our files after the databases had been constructed.