Hi,
I see a lot of people constructing maximum likelihood phylogenetic trees in their studies instead of neighbor joining trees. I checked the web and found no clear definition on when to use what method. Could anyone elaborate a little bit?
I know NJ is much faster than ML, is it right to say that as long as the resulted phylogeny is consistent with previous determined phylogenetic relationships between different species, then it doesn't matter which method to use?
By the way, is MEGA acceptable for journal publications? Thanks!
Hi Yong,
I think you have a good part of the answer in your question. NJ is faster in term of computing but quite weaker compare to ML... ML uses more complex evolution model, so "closer to the real life" and the mathematics behind ML/bayesian methods are know to be stronger than NJ for reconstructing sequence histories.
Personally, I used in first step the NJ method for defining a good dataset or testing new stuff and I go to ML to have a robust phylogeny for publishing.
I think now, It will be quite hard to published within a pure phylogeny paper a simple NJ tree... Or only as a pedagogic figure....
Yes, MEGA can be used for publishing phylogeny.
Best
Guillaume
Neighbour Joining is not a phylogenetic method, but a phenetic one. It establish relationships between sequences according to their genetic distance (a phenetic criteria) alone, without taking into account an evolutionary model. Thus, ancestry is never considered.
You will need a true phylogenetic tree (Maximum Likelihood, Bayesian, maybe even dynamic parsimony on POY) to get your work published. A Neighbour-Net in Splitstree gives you the same information as a NJ tree but more visual, a very graphic depiction of the distribution of genetic divergence in your samples (for example, for considering species boundaries).
@Daniel Thanks! Seems that I got a big problem, the thing is that I performed both NJ and ML, and the NJ is more consistent with the life tree in public website, and the ML is a bit mess and very hard to interpret, is it because I didn't find the right ML model?
I did downstream natural selection modeling with the NJ tree and it works well, I detected positive selection sites and very consistent with enzyme kinetic suggestion. is it okay to use NJ tree to perform natural selection modeling with codeml in PAML? Thanks a lot!
NJ assumes that species with lower difference in the given matrix must be closer than those species with any other on the matrix. You can use NJ as your phylogenetic method, but usually only matrices with low distance between sequences have good results with NJ. MEGA can provide you with p distances for each species and if the numbers are equal to/greater than 0,3 you should probably use ML.
In ML, modeltest can be used to adress which model is best for you data. Also, having results similar to proposed phylogenies is not really a good point, since those phylogenies could be wrong.
Thanks Barbara! I understand that the phylogeny inferred from a single gene family might not reflect the real relationship between species! But I guess the NJ tree might be better to resolve the relationship between paralogous genes which result from duplication before the species radiation? At least this seems true with the gene family I am working on! I will check the p distance value for my NJ, thanks a lot! Am going to submit my manuscript to a journal, really hope the reviewers trust it!
Thanks @David! You made a point! I will re-examine the ML tree I got! I think the "disorder" my ML tree displayed should be due to the asymetric evolutionary rates different species have. Some plants genomes are deemed to retain more ancient characteristics. The case with my target gene family is that two paralogous groups exist, they split before species radiation, one of them was lost in some genomes, now might be due to the unequal evolutionary rates, some higher eudicot species were clustered with lower eudicots in my ML tree, and the two paralogous groups did not separate well. However, the NJ tree showed two distinct lineages. I want to detect positive selection pressure in one of the two lineages, I don't know how to deal with this. Anyone has any comment? Thanks!
NJ is not a phenetic method, and does not simply cluster species pairs associated with shorter genetic distances. The algorithm includes a correction of the initial distance matrix, that takes into account the distance of each species to all others, thereby taking into account the fact that some branches are longer than others on the phylogenetic tree. It is a true phylogenetic method, and has been shown to be more robust than maximum parsimony to the problem generated by the juxtaposition of long and short branches on the same phylogenetic tree (sometimes referred as the “long-branch attraction” problem).
On the other hand, the NJ algorithm produces a single tree, without giving you any means to compare it to other potential trees. ML, like MP or a Bayesian method, investigates the space of all possible phylogenetic trees, trying to identify those that are best, i.e., in the case of ML, those that are associated with the highest likelihood. The trees produced by a ML search can be compared to other trees, e.g., those generated by another ML search or by a NJ run, using the likelihood criterion. The superiority of the ML method (compared to NJ) therefore lies in its consideration of all possible trees to identify the best ones (i.e., exploring the space of all possible trees) and in the use of a clear criterion to compare them.
@Patrick Thanks Patrick! From my understanding NJ is more about a distance method, however, it can also include a few different models, such JTT etc, what does these models do in NJ method then? I noticed these models occur in the ML method often, Do they apply some reality to the NJ tree also?
Hi Yong,
In NJ, nucleotide substitution models are used to estimate genetic distances among pairs of sequences, to generate the distance matrix. Genetic distances are then considered as estimates of the number of mutations that have accumulated along the branches leading to each sequence since they split from their common ancestor. In ML, the same substitution models are used to calculate the likelihood of a specific phylogenetic tree, given the sequence data.
Cheers,
Patrick
That said, NJ is more versatile because of the very fact that it doesn't have built into it any assumptions about the way sequences evolve. All you need is a distance matrix. Linguists have recently started using NJ to track the evolution of languages, for example, and it's sparked a revolution within the field revealing new language families etc. You can't just do that out of the box with ML, MP and BI.
What I like about NJ is that you can adopt it to many different needs. For instance, if you want to make a whole genome phylogenetic tree, you can make a distance matrix based on how many genes differ between your species, and use that to construct an NJ tree. Often, such a tree will be much more accurate than an ML tree based on an alignment of your gene of choice.
"doesn't have assumptions ... all you need is a distance matrix" - but that distance matrix is inferred by a specific algorithm and the metric needs to be chosen, e.g. Euclidean or city-block. In fact in many cases the distances are inferred under a sequence evolution model, eg. GTR. It is NOT true that distance based methods have fewer assumptions than MP or ML.
For latest information regarding the topic of discussion;
https://academic.oup.com/mbe/article/35/6/1547/4990887