I have a few questions about these three methods.

  1. We commonly use 16s rRNA to identify species and construct a phylogenetic tree. If there's a new isolate from a new and undiscovered ecosystem, should we confirm this new isolate identity with ANI taxonomy/ genome to genome distance calculation (GGDC)?
  • If that isolate has a high similarity with a certain species in the 16s rRNA analysis result, should we proceed to analyze further with ANI/GGDC? Or just let it be because we have a robust 16s rRNA analysis? How can we be sure with our 16s rRNA BLAST result?
  1. When should we use 16s rRNA rather than ANI/GGDC and vice versa?
  2. Is it always necessary to confirm 16s rRNA analysis with ANI/GGDC? (Assuming there's a lot of closely-related genome data available in databases)
  3. Should we confirm the ANI result with GGDC and vice versa?

1 Answer 1


I do understand that you are working with is an isolate and thus what occurs in metagenomics you feel is over doing it, particularly if 16S is e.g. >99% similarity via Blast.

Before reading further you might check my answer history to understand the angle I am approaching this question from. I would see select given methods as the 'gold-standard' other methods are compared against, in addition to single versus multi-gene trees. There is good theoretical criteria for believing this. For example, bacteria are often subject to heavy %GC bias and this makes other methods error prone, because it introduces compositional bias (below).

Summary Blast alone - thats a 'no' (theory section). The problem with any one approach is that errors can arise from many different sources stemming from both biology and methodology. Thus a combination of approaches is required. Therefore your maximum approach is a minimum standard.

I personally would go one step further than your question: full maximum likelihood and Bayesian phylogenetics with bootstraps in addition to Blast (minor), ANI and GGDC.

History From tradition 16S is the major standard, but even then multi-gene phylogenies were required. Thus genome to genome distance calculation (GGDC) with ANI is essential at a minimum. 16S is back in fashion because its easy to perform a metagenomics study of species population dynamics.

Species delineation has always been problematic in bacteria and the rule of thumb was 5% divergence defined a species boundary.

Advice I would use a multi-locus appproach under a full phylogenetics analysis, where a species boundary is defined by >5% divergence and bootstraps >80% with Bayes P values > 0.95.

Recombination is essentially ignored in taxonomy providing there isn't over reliance on a single gene tree, but using BLAST, ANI, GGDC as a method data-mining would justify a select sample size. The classification is then nailed with a phylogenetics analysis.

One of your methods (probably ANI) might be based machine learning/deep learning (its not GGDC) and the accuracy of misassigning a species under this method is high. Historically it was really high (80% accuracy), this might have improved to 90%. Its not like image recognition where the error rate is beyond human perception.

Theory A fundamental issue is that recombination in bacteria exists on a spectrum from clonality to panmixia. This makes single gene taxonomy risky, thus at theoretical level would not believe 16S was a singularly criterion, but it never was. Your bacteria is unknown (to us) so we'd have to assume some level of recombination.

Compositional bias is well known for taxonomic misassignment. It is readily demonstrated in simulation. Thus if two unrelated bacteria demonstrate e.g. 75% GC, they are more likely to artifactually grouped together on e.g. a distance criterion.

  • 1
    $\begingroup$ Agree. 16S is not reliable for classification, especially at the species level. Bacteria can be quite promiscuous with horizontal gene transfer, so a single gene on its own is not a reliable indicator of taxonomy. $\endgroup$
    – gringer
    Feb 20, 2022 at 21:11
  • $\begingroup$ Cool, I did check ANI and its a pairwise aligner and identity algorithm not machine learning. It looks locked behind a webserver so you'd need e.g. Blast to formulate the subset of bacteria to check. $\endgroup$
    – M__
    Feb 21, 2022 at 3:17

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.