Neighbor-joining, Maximum Likelihood and Bayesian methods all produce very good reconstructions of known (from fossil records, from epidemiology, from other data sets, etc.) phylogenies if the data is good and adequate. In my experience, when a phylogenetic tree is not good, the cause is more than 95% of the time found in the data set, and not the method(s) used to analyze the data. There are many things which can cause problems with phylogenetic reconstructions, such as recombination, small data set size, distances too great, missing data, and misalignment of even quite small regions of the data.

Neighbor-joining, Maximum Likelihood and Bayesian methods all produce very good reconstructions of known (from fossil records, from epidemiology, from other data sets, etc.) phylogenies if the data is good and adequate. In my experience, when a phylogenetic tree is not good, the cause is more than 95% of the time found in the data set, and not the method(s) used to analyze the data. There are many things which can cause problems with phylogenetic reconstructions, such as recombination, small data set size, distances too great, missing data, and misalignment of even quite small regions of the data.

Thank you, Brian Thomas Foley for contributing to this thread. I agree on the importance of quality and adequacy of data used in the analysis. It looks like these methods use different assumptions, as a result, they may yield different outcomes. N-J method uses a distance-based approach (that uses overall similarities), which clusters similar OTUs together, therefore perhaps not useful for making inference on evolutionary relationships (phylogeny). This might need further discussion, but I believe that shared derived characters (not just the NJ distance) should be used for phylogeny reconstruction. In my opinion, maximum parsimony or model-based maximum likelihood methods are perhaps the better methods for phylogenetic inferences.

I agree with Brian

Nj normally is used for typing

because is a distance based method

ML is better because the use of evolutionary models correct the genetic distance

In this case is possible to do an evolutionary hypothesis

Madhav Nepal

If you are talking about OTUs, I dont think any of the methods would be perfect for the phylogenetic inferences. Since OTUs in itself are not specific and drawing strong conclusion based of the OTU sequences is not correct. In this regard Brian Thomas Foley answer is very relavant and very important too. If the data is bad, the tree would be bad not matter how good methods you are using. Additionally, if the accurate phylogenetic inferences are to be made, you can not just go for the gene trees, you have to go for the species trees. Moreover, if there is a practical and identifiable phylogenetic differences, most of the method can resolve it. And if not, does not matter which method you use, you can not resolve it.

All phylogentic methods whether neighbor-joining or maximum likelihood or Bayesian, or parsimony, are "distance-based". The distances in neithbor-joining can be adjusted by any model of evolution you like, just as in the other methods. You can partition the data (first, second and third codons positions for one example) and use different models on each partition if you like. All these adjustments make rather trivial changes in things such as the ratio of lengths of the tip branches to branches closer to the root of the trees while very rarely affecting the branch order or clades. A fish will never fall in the bird or mammal clade, a primate will never cluster with the rodents.

Problems with the data, on the other hand, such as a long branches attract problem, are found to create erroneous tree branching order with ALL methods. In fact, a long branches attract error will usually be more exaggerated with maximum likelihood than with neighbor-joining, and most often both trees will be wrong in the same way, but just a bit more exaggerated in the ML tree.

Measuring physical characteristics, such as limb length or body size can be very misleading. For just one example modern dog breeds from largest and fuzziest to smallest and hairless are all more similar to each other than wolves which all look the same. Likewise for modern humans compared to chimpanzees. This is because relatively minor changes in expression levels of developmental genes have huge impacts on body size and plans. E. coli look very similar to other gram negative bacteria which are more distantly related than humans are to insects. This is because the bacteri do not have much ability for morphological change.

https://www.semanticscholar.org/...neighbor-joining-method%3A...method.../13e1c3709...

I agree with the all points made by Brian Thomas Foley. In addition and based on personal experience, I have also observed that sequences that are very short and less than 200 base pairs will irrespective of the method used (e.g. neighbor-joining or maximum likelihood), NOT produce a good phylogram.

Thank you Ayodeji O. Olarinmoye and T. H. Al-Noor for your contribution.

Thanks Abhijeet Singh! I want to make sure we are on the same page!! The abbreviation OTUs might have caused a confusion. I am using OTUs to refer to operational taxonomic units.

Yes, I was also talking about OTUs = Operational Taxonomic Units which is an arbitrary unit and there is not really a standard to define them. It depends on user to define the threshold for OTUs. Also it largely depends on the method/Pipeline/algorithm you use to generate OTUs. Now there are software which generates ASVs (Amplicon sequence variants) / ESVs (Exact sequence variants) instead of OTUs.

However, it doesn't matter what unit you take, discussion above applies to them.

Abhijeet Singh , any published literature backing up your statement "OTUs = Operational Taxonomic Units which is an arbitrary unit" ? Thanks in advance.

Madhav Nepal

Not really. But if you find any literature which defines strictly what is an OTU, let me know. I would be keen to see any literature which tries to do it. Not just arbitrary threshold of 97% similarity. (Since it is just for the 16S amplicons)

Same terminology might be used to mean different things in different disciplines. The gene sequences used in phylogenetic analysis are not "arbitrary units". At the bottom of this paragraphs are links for articles/book chapters on the history of cladistics (and phylogenetics). According to NCBI "Tree" facts site, any branch tip or a leaf of a phylogenetic tree is an operational taxonomic unit (https://www.ncbi.nlm.nih.gov/Class/NAWBIS/Modules/Phylogenetics/phylo7.html )

• Stuessy, T. F. (2009). Plant taxonomy: the systematic evaluation of comparative data. Columbia University Press.

• Article Toward a phylogenetic system of nomenclature

• Sokal, R. R., & Sneath, P. H. A. (1963). Principles of numerical taxonomy. WH Freeman. San Francisco.

• Article Defining operational taxonomic units using DNA barcode data

• Hall, B. 20111. Phylogenetic Trees Made Easy, Fourth Edition. Sinauer Associates, Oxford University Press.
• Hennig, Willi (1966), Phylogenetic systematics (tr. D. Dwight Davis and Rainer Zangerl), Urbana, IL: Univ. of Illinois Press (reprinted 1979 and 1999), ISBN 978-0-252-06814-0
• Hennig, Willi (1975), "'Cladistic analysis or cladistic classification?': a reply to Ernst Mayr" (PDF), Systematic Zoology, 24 (2): 244–256, doi:10.2307/2412765, JSTOR 2412765, responding to Mayr 1974.
• Hennig, Willi (1999), Phylogenetic systematics (3rd edition of 1966 book), Urbana: University of Illinois Press, ISBN 978-0-252-06814-0 Translated from manuscript in German eventually published in 1982 (Phylogenetische Systematik, Verlag Paul Parey, Berlin).

"Same terminology might be used to mean different things in different disciplines"

• Here we are talking about phylogeny, so we are talking about same discipline

"The gene sequences used in phylogenetic analysis are not "arbitrary units"

• Sequences are not an unit, it is the data, and something which defines the data is called unit.

"At the bottom of this paragraphs are links for articles/book chapters on the history of cladistics (and phylogenetics). According to NCBI "Tree" facts site, any branch tip or a leaf of a phylogenetic tree is an operational taxonomic unit"

• These articles are just indirectly supporting my point that there is not a strict definition/threshold to define an OTU. Its data and user specific. And of course, tree leaves are called OTU, there is no question about that, but the point of discussion is who decides what to keep on the leaves of tree.

To end the argument - I am generating OTUs at 80 % clustering similarity and you will make OTUs at 97 % clustering similarity, then are they same/similar. Would you define both kind of OTUs are just OTU, or would you say, OTU at xx % similarity? And here it becomes random/arbitrary.

Thanks Abhijeet Singh for your valuable contribution to this thread. Clarifying criteria or threshold values is very important for defining or describing what we claim in our research.