Resolving gTranslate and NCBI conflicts

The following is general advice on how to resolve conflicting translation table predictions between gTranslate and NCBI, using the 377 genomes in GTDB r237 with conflicting assignments as an example:

• CheckM heuristic. Results are also given for the CheckM coding density heuristic. This heuristic considers coding density under table 11 and table 4. If the coding density under table 4 is >70% and >5% higher than under table 11, this heuristic assumes a reassignment has occurred. This heuristic cannot distinguish between tables 4 and 25 so any predicted reassignment is indicated as table 4.
• Incomplete NCBI classification. The most common reason for disagreement is genomes with incomplete taxonomic classifications at NCBI. In GTDB r237, there were 107 genomes classified as just d__Bacteria at NCBI. NCBI assumes these use translation table 11. Unless there is external evidence to the contrary, the gTranslate predictions should be used for these genomes. Notably, the standard CheckM heuristic for predicting translation tables also indicate a translation table reassignment for 104 of the 107 genomes. The 3 discrepant genomes all have extremely poor CheckM 2 completeness estimates under both translation tables and would all fail standard GTDB QC.
• Incomplete NCBI classification. There were 5 p__Pseudomonadota genomes in GTDB r237 without further classification at NCBI. These were all predicted to use table 25 by gTranslate and the CheckM heuristic agreed there was a reassignment event. Notably, the coding density for these genomes were all >85% when using table 25 and <63% when using table 11. These genomes likely belong to one of the p__Pseudomonadota species known to use table 4 (e.g. Hodgkinia cicadicola). So it appears gTranslate correctly predicts a reassignment, but incorrectly predicts these are table 25 instead of table 4. These should be added to the corrections file provided to the prodigal step of the GTDB migration workflow.
• Incomplete NCBI classification. There was a single c__Bacilli genomes in GTDB r237 without further classification at NCBI. gTranslate predicts this to use table 4 which is well supported by this genome having a coding density of 91.9% under table 4 and only 72.2% under table 11. Presumably this genome belongs to the Bacillota order Mycoplasmatales which is known to use table 4.
• Unrecognized reassignment at NCBI. NCBI determines the translation table based on the taxonomic assignment of a genome. It can take time for lineages with newly identified reassignment events to be recognized by NCBI. In GTDB r237, there were 33 Eggerthellaceae genomes (27 classified as Equivita; 5 clasified as Gorillivita, 1 with no genome assignment) all indicated as using table 11. These are the genomes we published and demonstrate have genome wide reassignments to table 4. The CheckM heuristic also supports this reassignment. The gTranslate predictions should be used for these genomes. Similarly, there are 12 p__Minisyncoccota that we demonstrated used table 4, but NCBI still indicates use table 11.
• Poor quality genome assemblies. CheckM 2 completeness and contamination statistics are calculated under both the gTranslate and NCBI translation tables. Standard GTDB QC criteria are then applied to these estimates. A number of conflicts can be attributed to poor quality genome assembles for which gTranslate is not expected to perform as well. For GTDB r237, there were 71 genomes that failed QC regardless of the translation table used. These can be safely ignored.
• NCBI misclassification. Consideration should be given to the expected accuracy of NCBI assignments. In GTDB r237, there were 93 genomes classified as p__Candidatus Altimarinota. Confusingly, NCBI classifies 4 of these as using table 11 and the remaining 82 as using table 25. gTranslate and the CheckM heuristic give the opposite results: table 25 instead of 11 for 4 genomes, table 11 instead of 25 for the remaining 82 genomes. Given the agreement between gTranslate and the CheckM heuristic and the possibility that these genomes as misclassified at NCBI, it is reasonably to assume the gTranslate predictions are correct.
• NCBI misclassification. Similar results were observed for 25 other "CPR" genomes (p__candidate division CPR1; p__Candidatus Absconditibacteriota; p__Candidatus Falkowiibacteriota; p__Candidatus Parcubacteria; p__Candidatus Peregrinibacteriota). In all cases, the gTranslate and CheckM heuristic agreed, and as such are taken as more likely to be the correct translation table for these genomes.
• NCBI misclassification. There were 19 genomes classified to p__Mycoplasmatota by NCBI. Of these, 8 were classified as o__Mycoplasmatales and thus assumed to use table 4. The gTranslate and CheckM heuristic both predict table 11 for these genomes. Again, these were assumed to be NCBI misclassifications instead of incorrect predictions by gTranslate and the CheckM heuristic. The remaining 11 genomes fall into a number of cases, but again agreement between the gTranslate and CheckM heuristic was taken as evidence of misclassifications or errors at NCBI. As an example, 7 of these genomes lacked a class assignment at NCBI, but were assigned to o__Mycoplasmoidales, an order known to use table 4. Nonetheless, the NCBI data indicated table 11 for this genomes while gTranslate and the CheckM heuristic correctly indicated table 4.
• Erroneous NCBI translation table data. There was a single genome assigned to s__Candidatus Pinguicoccus supinus by NCBI. This species is known to use table 4 and this is recognized at NCBI. However, NCBI still assigned this genome to table 11 (perhaps because it has no class to family level assignments). gTranslate correctly predicts table 4.
• Erroneous NCBI translation table data. Three s__Providencia rettgeri genomes were assigned to table 4 by NCBI, but predicted to use table 11 by gTranslate and the CheckM heuristic. The coding density under table 4 and 11 are both ~85%. This species is NOT known to have a reassignment and it is unclear why the NCBI metadata indicates table 4. The gTranslate prediction of table 11 is assumed correct.
• The remaining 7 genomes are less clear:
  • A single c__Clostridia genome with 90% coding density under table 4 and 81.3% under table 11. NCBI assigns this to table 11. gTranslate and the CheckM heuristic assign this to table 4. This could be a misclassified o__Mycoplasmatales genome or simply a gTranslate misclassification. Given the strong performance of both gTranslate and the CheckM heuristic these are taken as correct in such borderline cases.
  • Three g__Candidatus Shikimatogenerans genomes assigned to table 11 by NCBI, but table 4 by gTranslate and the CheckM heuristic. The mean coding density under table 4 is 87.9% versus just 72.6% under table 11. A reassignment is not known in these genus. Perhaps these genomes belong to a new species with a genome wide reassignment event. Further exploration is required to confirm this, but again gTranslate is assumed correct for cases that cannot be easily resolved.
  • A single p__Cyanobacteriot from the genus g__Cyanobium was assigned to table 11 by NCBI, but predicted to use table 4 by gTranslate and the CheckM heuristic. The coding density under table 4 is 86.9% compared to 72.7% under table 4. Likely, this genome is just terribly misclassified. This genome fails the full GTDB QC checks so does not appear in GTDB. gTranslate is assumed correct and/or the genome of sufficiently low quality as to not warrant further attention or inclusion in GTDB.
  • A single s__Candidatus Providencia siddallii genus was assigned to table 4 by NCBI, but predicted to use table 11 by gTranslate and the CheckM heuristic. My understanding is that organisms in this species are believed to have undergone a partial reassignment. There are only 2 genomes available from this species. Ultimately, it is unclear to me how this genome should be handle as neither table 4 or 11 is correct. The coding density is very similar between these two tables. For simplicity, the gTranslate results will be used.
  • The single f__Fastidiosibacteraceae genome GCA_059737165.1 is assigned to table 4 by NCBI, but predicted to use table 11 by gTranslate and the CheckM heuristic. This is the Fastidiosibacteraceae bacterium XS4 genome published by Nakayama et al., 2025 who provide evidence that this genomes uses table 4. This genome was added to the correction table and gTranslate presumed to be incorrect.