I am very new to bioinformatics. Here,i am trying to make a phylogeny reconstruction of protein sequences following methods like Maximum likelihood (MEGA 6.0 and GARLI) and Bayesian inference (Mr.bayes). At first, the corresponding protein sequences are collected and aligned using ClustalX. Later, i used gblocks to get a block for the further studies. For ML analysis tree, bootstrap of 1000 replicates was done and for Mr.Bayes, the tree was obtained after 1000000 generations with 2 parallel runs 4 chains (1cold and 3 heat). Indeed, both the methods did use the same substitution model (JTT) and rates (gamma; 6). But still, both the trees looks different and does not have the same topology to each other for the same set of proteins. I really cannot understand it. Please some could advise me to go forward in this work.
Bayesian trees are very often somewhat different fro ML because the search methods and the philosphy of what the final tree should represent is very different.
Bayesian analysis uses MCMC fro the search, which by the definition of Markov Chains mean that the next step in the search through tree space depends the present state. Hence there is necessarily a correlation running though the search and the likelihood scores of each tree that tends to bias upwards the posterior probabilities.
Furthermore the aim of a Maximum Likelihood search is to find the one maximum (highest) likelihood tree. The tree is integrated over tree space alone (conditioned on previously optimized model parameters). The aim of a Bayesian Analysis is to obtain a large sample of trees that maximizes both the model parameters and the tree. The resulting tree is a consensus tree of all of the trees sampled after the burnin. The topology sometimes is similar to the maximum likelihood tree but the Bayes tree is not a maximum likelihood tree. The credibility set usually includes the ML tree but need not include it.
A maximum likelihood tree optimizes the model parameters in a separate step and Bayes optimizes the parameters along with the tree so it is not surprising that both trees and parameter estimates differ.
For a statistician like me trained in maximum likelihood methods, I generally publish the maximum likelihood tree with ML branch lengths because Bayesian branch lengths tend to be quite inflated for various reasons. But I label the branches with the bootstrap and aLRTs (approximate Likelihood Ratio Test statistics) as well as the Bayesian posterior probabilities because the MCMC set of trees provides a genuine confidence statement on branches.
Bottom line do both types of analysis. Recognize that both trees are mere estimates (but generally very good) and that both analyses have their drawbacks and strengths.
Fro Bayesian analyses see recent Z.Yang and others from the last two years or so on choice of priors - very, very important for both topology and branch length estimation.
Yes, you are correct in that different programs may output different trees. You might want to check out the paper (link below) where they compare the same dataset on multiple platforms.
http://moodle2.lsu.edu/pluginfile.php/1390088/mod_resource/content/0/Bioinformatics2006Stamatakis.pdf
With good data, the various methods (neighbor joining, maximum likelihood, Bayesian) and models (Kimura, Felsenstein 1984, General Time-reversible etc for DNA; others for amino acid data) all give quite similar results with essentially the same topology and the differences largely in branch lengths rather than branching order. However, the methods and models can be used with any data and not all data is good for answering phylogenetic questions. There are many protein families for example which have dozens to hundreds of functional copies per organism and too little similarity remaining between them to reconstruct their evolutionary history by analysis of their sequences alone. Even when the data is sequences that diverged from each other too long ago for very accurate tree reconstruction, phylogenetic analyses are not useless or meaningless, they can still provide some types of good information. For example it may be possible to tell if a gene duplication event such as one that created myoglobin and hemoglobin gene families ocurred in vertebrates or much earlier before vertebrates diverged from invertebrate animals.
http://pharmaceuticalintelligence.com/2015/01/19/life-on-earth-is-traced-to-oxygen-binding/
Dear Brian:
I have been trying to do a phylogenetic analysis of a bacterial protein from 100 organisms. The length of the proteins varies among organisms and they range from 300-700 amino acids. In that case, i was advised to use the complete protein for MSA and so that i won´t lose any evolutionary significant information. While others advised me to use GBLOCKS to get a block of sequence with major homology. I did both and still i could not get the similar tree. Is it necessary to trim the ends which are majorly unaligned. Please give me a good strategy to refine the data for phylogenetic analysis.
Dear Artur Burzynski .
As i am very new, i am not quite sure, whether the obtained tree is affected by random noise. I have attached trees obtained from 16srDNA sequence of bacterial organisms with this one. I will be very glad if you could help me to resolve it.
The tree you uploaded there is a "consensus tree" which has meaningless branch lengths. It is far better to use the best tree and plot the bootstrap values (or just the values about 75% or 80%) onto the best tree. The tree does show that similar groups of bacteria fall on a branch together as we would expect. For example you have a clade with many Bacillus species, another clade with Halobacteria, etc. In general, 16S rRNA phylogenies work pretty well for most bacteria.
The two files that Gurusamy K.Muniasamy uploaded seem to both be trees made from 16S Ribosomal RNA (or genes encoding 16S). One is a "consensus tree" so the branch lengths have no meaning and the other is a tree with branch lengths proportional to distance. One was a PDF, and the other was a Newick format treefile. I rendered the Newick tree with FigTree to compare it to the PDF, see attached JPG screenshot.
The two trees have many clades that are shared, and as expected the ones that are shared have higher bootstrap values (the PDF seems to have bootstrap values as decimal, but I am guessing).
Anyway, they look sort of reasonable for 16S rRNA trees. The distance between the archaebacteria (Halococcus etc) and the Eubacteria (all the others) is huge as expected and likely represents misalignment as well as true distance because it is quite difficult to correctly align 16S between those groups.
I agree with Artur, that it is very counterproductive to ask questions here which are vague or misleading. It is not clear at all what hypothesis you are attempting to test, or what you hope to gain from building trees on sequences that have pretty fuzzy names and all.
As for the idea of using 100 protein sequences ranging in length from 300 to 700 amino acids, that is sure to be a complete mess, as Artur said. What type of protein is it, that can vary in length so much, and why would you want a tree?
Dear Arthur and Brian,
I would like to ask both of you to apologize me. I just took this as an other opportunity to get expertise opinion on current analyses, i was working now. I would have focused on the discussion to the actual question. in the same time, i would have also presented right data in the same format. I am very thankful for your efforts and the valuable points. As per your directions, i will make the next reconstruction analysis with less number of bacteria groups.
We are trying to study the evolutionary relationship between three different functional proteins, which are present in the cluster. In that case, we doubt that those proteins might have evolved independently. As a beginner, i took some reference paper with the similar approach to start my work. I have attached the paper with this reply. So i took 16srDNA sequences as a reference with the interesting protein sequences. If there is something i am wrong in any points, i kindly request you to correct it.
Did you throw away the trees sampled during the burnin period of the MCMC chain?
Bayesian trees are very often somewhat different fro ML because the search methods and the philosphy of what the final tree should represent is very different.
Bayesian analysis uses MCMC fro the search, which by the definition of Markov Chains mean that the next step in the search through tree space depends the present state. Hence there is necessarily a correlation running though the search and the likelihood scores of each tree that tends to bias upwards the posterior probabilities.
Furthermore the aim of a Maximum Likelihood search is to find the one maximum (highest) likelihood tree. The tree is integrated over tree space alone (conditioned on previously optimized model parameters). The aim of a Bayesian Analysis is to obtain a large sample of trees that maximizes both the model parameters and the tree. The resulting tree is a consensus tree of all of the trees sampled after the burnin. The topology sometimes is similar to the maximum likelihood tree but the Bayes tree is not a maximum likelihood tree. The credibility set usually includes the ML tree but need not include it.
A maximum likelihood tree optimizes the model parameters in a separate step and Bayes optimizes the parameters along with the tree so it is not surprising that both trees and parameter estimates differ.
For a statistician like me trained in maximum likelihood methods, I generally publish the maximum likelihood tree with ML branch lengths because Bayesian branch lengths tend to be quite inflated for various reasons. But I label the branches with the bootstrap and aLRTs (approximate Likelihood Ratio Test statistics) as well as the Bayesian posterior probabilities because the MCMC set of trees provides a genuine confidence statement on branches.
Bottom line do both types of analysis. Recognize that both trees are mere estimates (but generally very good) and that both analyses have their drawbacks and strengths.
Fro Bayesian analyses see recent Z.Yang and others from the last two years or so on choice of priors - very, very important for both topology and branch length estimation.
Patrice Cornell gave excellent advice, and Dr. Yang's papers are available from his web site: http://abacus.gene.ucl.ac.uk/ziheng/cv.html You stated that you are working with amino acid sequences, protein alignments, but the same principles apply to both DNA and protein evolution. Choosing good models of amino acid substitution is important.
I must side with Gurusamy on this one. I see no need to be patronizing when help is what is asked for. You need not answer at all if you must be critical of the question on the first place.
I do think that he deserves an apology.
Hi Gurusamy,
Looks to me like you have done just the right sort of analyses: ML and Bayes so the advice I gave earlier is the likely reason for the differences in the trees. Is it the 16SrRNA tree that is the Bayes tree? If it is open up the tree file in Figtree and you will find find that you have a choice among support values (posterior probabilities) and branch lengths which are shown on the PDF file.
I looked at the trees and here are some observations. The tree labelled 16SrRNA is shown as a cladogram which is a tree whose branch lengths have no meaning. In a cladogram you are only interested in to topology and perhaps branch length support (like bootstrap values or posterior probabilities or aLRTs). On this tree the length of the branches have no meaning.
However, on that same tree, you clearly have branch length values. Higher values mean that branches are longer and that more evolution has occurred on that tree. It is just that in this drawing the branches are not drawn to scale.
On the other hand, the bootstrap tree that you sent has all kinds of information about branch lengths and support values and was generated in Sum Trees which looks like a part of a terrific package called DendroPy.
I opened that tree file in Figtree and produced two trees (attached). On has bootstrap values (only those above 100) and one has branch lengths but unlike the other tree you sent, theses branch lengths are draw to scale.
The bootstrap values not shown are some from the base of the tree and some of these were above 70% which is not terrible. You have a surprising amount of phylogenetic information in these 16S trees.
I do note that the outgroup is very distant from the in-group. Is it possible to get rid of those taxa and use something closer for the outgroup? You might find that the rooting changes and with a shorter branch on the outgroup, the rest of the tree will be easier to read.
As I said in an earlier post, the difference in the trees is due to the different criteria and the different philosophies underlying Bayes compared to ML methods.
Dear Dr.Patrice,
I am very glad for the answers and efforts that you have put on clarifying my doubts. Of sure, i have understood so many things from your side. I would like to get your e-mail id to share some of my doubts with you. Please can you give me email id to keep in contact with you.
- Badges
- Science method
- Similar topics
- Biological Science
- Evolutionary Biology
- Phylogenetics
- Phylogenetic Analysis