Hi, I usually use partition finder and it works ok, but I have worked only with 56 mitogenomes. I´m not sure you need a server, It worked ok for my 56 samples and I tested 4-5 partition possibilities in my personal computer. I usually use model test or protest to confirm the models partitionfinder shows in the results.

For mammalian complete mitochondrial genomes using several partitions will influence relative branch lengths (the ratio of the length of the tips to the length of branches close to the root) more than the branching order. The branching order will be more influenced by the "long branches attract" problem.

My tree made with the complete mitochondrial genomes of vertebrates, illustrates some interesting features of the data set. After removing columns where more than 40% of the sequences were represented by gaps (uncertain regions in the alignment), I was left with 15,510 columns. The distances between fish and snakes, mammals, etc are far beyond saturation of silent sites, as illustrated by the transition:transversion plot. But within mammals, or within birds, the distances are far less.

Hi, I usually use partition finder and it works ok, but I have worked only with 56 mitogenomes. I´m not sure you need a server, It worked ok for my 56 samples and I tested 4-5 partition possibilities in my personal computer. I usually use model test or protest to confirm the models partitionfinder shows in the results.

Using just the mammals and marsupials, there is still some branching order that may not agree with the fossil record, due to the great distances involved.

The distances from Echindnas and marsupials to mammals are still in the range of being saturated with silent site mutations.

Overall, my opinion is that the large distances between mammals, for mitochondrial DNA which evolves some 10-fold faster than nuclear DNA, presents problems that partitioning will not solve. Using chromosomal genes ,with their much slower rate of evolution, may be needed to resolve some of the deeper nodes in the mammalian phylogenetic tree.

@Brian, I am looking at a mammalian clade that is fairly young (its common ancestor is estimated to have existed ca. 5 MYA), so saturation should not be a problem. In fact, I have tested for it, and found no evidence of saturation. My interest in implementing partitions for analyzing a matrix that only contains mitogenomes is because of the highly heterogeneous rates of mutations across the ca. 17000bp-long molecule. As you know, it does not make sense to use a single model of nucleotide substitution to analyze a molecule with such a high variability in rates. The paper I cited above explains this issue in a very nice way, and has a review of studies that either arbitrarily used a single model (no partition) or selected a partition scheme not analytically and without offering a justification.

@David and Brian: thank you very much to both of you for the help. Best wishes, Eliécer

@Eliécer You could also try a phylogenetic inference program that implements mixed models. I've used BayesPhylo by Mark Pagel (http://www.evolution.reading.ac.uk/BayesPhy.html) and it works great. You simply run it as many times as numbers of rates you suspect and then compare LogLikelihood scores to find which one fits the data better.

I agree with you that a very efficient way to select candidate model across several partitioned data is PartitionFinder. The main limitation of partitionfinder is t that it do not try sharing some parameters across partition: partition are totally fused or totally separated. The choice is done using AIC ( or AICc or BIC).

I am part of the BioVeL project (www.biovel.eu) and we exposed PartitionFinder as a webservice within a general workflow (http://www.myexperiment.org/packs/371.html) that perform model choice with partitionfinder, build the nexus format with MrBayes command for the model, submit a run of MrBayes then check convergence and perform a posterior predictive test to assess post hoc the fit of the model. The workflow could be used at portal.biovel.eu.

The main caveat for your specific problem is that PartitionFinder accept no more than 14 partitions as maximal set and this could be limiting for a mitogenome (generally, 13 protein genes + the trnas and 1 noncoding region often with several distinct region within) given that protein coding region probably should differentiate across the 3 coding positions. Further a coding region is generally well described by a model that use the same GTR/HKY model across the 3 positions but use 3 different base frequencies, configuration not taken in consideration with PartitionFinder. An approximate solution could be optimizing the model in several pieces with partition finder, collate manually the nexus model description and then submit the full model (you could use the variant of my workflow that accept directly nexus or the GUI that help you to design the nexus file for partitioned model). The presence of the posterior predictive test in the pipeline of analysis would allow to test post hoc the adequacy or goodness of fit of the full model.

A very different approach would be to use Bayesian factor approximated by the savage dickey ratio as proposed by Suchard 2001. I myself used the approach on a mitochondrial data set, but of only 3 genes (Steinfartz et al. 2006). The approach consist to perform a bayesian inference with the maximally distinct model (all possible partition with GTR+G) and then using several R script work out the ratio. The work is quite involving.

Finally all this approach use nucleotide model but probably a better representation would be to use codon model. But no program for model choice support codon model. MrBayes would allow to perform the inference, but you should use maximum likelihood program as HYPHY to obtain the likelihood score to calculate AIC. The posterior predictive test would be in theory available but nor webservice in the workflow nor other program (PUMA) would allow the use of codon model. Indeed the codon model would be useful if you are interested to have a more realistic model but probably this would not change the topology of your data. But you should be aware that several literature point out that base frequency vary along the mitogenome and that in mitogenome of metazoa there is a correlation between the base in the third position with the base in the first position of the next codon (the cause are unclear [trna, mutational effect of flanking region?] but the effect is visible). This makes the codon model itself quite an approximation of reality and so a model with different base frequencies for the 3 codon positions different for each genes best approximate the situation.

Very informative and helpful comments, Eduardo and Saverio, molte grazie!

With mitochondrial genomes from a young mammalian clade, members of which shared a common ancestor some 5 million years ago, as you say, the need for partitions and very complex models of evolution is debatable. If two sequences are 95% identical to each other, then very few sites have experienced multiple hits, and a simple % identity or "hamming distance" will be essentially the same as the phylogenetically corrected distance no matter which model of evolution is used. When analyzing much more distantly related sequences, such as 100 genome sampled from all clades of mammals, perhaps using marsupials as an outgroup, the simple percent identity begins to vary significantly from the best estimate of true distance, both from theoretical estimates and estimates based on fossil records and other data. When studying the mtiochondrial genomes of all vertebrates, or of all metazoans, the situation changes yet again.

Although there is no harm done in knocking yourself out with sophisticated phylogenetic methods, you are far better to spend your efforts ensuring that the data sets you use are appropriate, and that whatever the phylogeny tells you helps you to understand the biological question(s) you are asking. For example, inferring past population sizes of one or more of the species you are studying requires different data sets than inferring the date of common ancestry of all of the species.

Because there are very different rates across the mitogenome even of a single individual, you should not apply a single model to the whole molecule and instead you should use partitions. Thanks for your help on my question, Brian. Best, Eliécer

I agree that the model should include site-specific rate variation, such as a GTR (General Time Reversible) plus gamma distribution of rate variation, plus invariant sites. But I am not sure that partitioning the data by first, second, third codon position, or by gene etc, rather than (or in addition to) by observed rate variations, produces a significant gain when the sequences are more than 95% identical to each other. It is mostly not so much the rate of mutation that is variable from one gene to another, as the selection pressures on each codon of each protein. Each protein tends to have some variable and some conserved sites.

I highly recommend the paper I cited above by Powell et al., which nicely explains the need for partitioning mitogenomes even in closely related taxa.

I agree. The paper by Powell et al makes both of our points rather perfectly. If I don't spend hours with data paritioning, I can use Powell's data plus a few more species for "outgroup" and within 30 minutes have a nice tree essentially identical to Powell's tree. I then can see that the real issues might be whether or not all the various Genera of birds in this family might better be considered to be species of one genus. I also learn that searching GenBank for Agelaioides badius will not find the right entry, because Powell et al entered it as Moluthrus badius. Most biologists care less about whether my tree is 4.67% "worse" than Powell's tree. Pretty much all of the major points that could be made with either tree, are made by BOTH of these trees. Mine took about 30 minutes to build, and theirs would take me a week to reproduce.

I care about branch lengths and time since divergence events, not only topology and nodal support; hence, I aim to use the best-fit model I can for my data, and this requires me to use partitions. My question was focused on objective methods to assess the best partitioning schemes, which is one methodological task I am facing right now with a pilot project I am working on.

How, exactly, do you propose to find the times/dates, from the distances/branch lengths? The relative branch lengths in my tree will be less than 1% different from those in Powell's tree. Neither tree can tell us exactly which eon or era the Grackles began to diverge from a common ancestor unless we have some very solid fossil record or other events on which we can calibrate our clock. If the estimates of the dates, based on the fossils, are +/- 10% or 20%, then the 1% difference in the phylogeny branch lengths is not where our effort should be focused.

Again, I am not arguing that data partitioning is not good. I am only suggesting that people should consider many aspects of phylogenetic analyses, in addition to the models and methods. I have read several papers in recent years which used flawless phylogenetic methods, but failed to understand that the tree which was produced was de-facto evidence that the data must have been flawed to begin with.

I also want to encourage people to reproduce trees seen in papers, without worrying about exactly reproducing the methods. It is rather common in many publications (such as the one by Powell) to present cladograms with meaningless branch lengths, when the reader (as you are) is very interested in those branch lengths.

I am also not suggesting that Dr, Gutierrez does not know exactly what he is doing. I am just assuming that other people, working in other systems, might benefit from discussion of some of the other factors to consider, besides the phylogenetic method. The best data and best methods may be able to pinpoint the date of the human/chimpanzee/gorilla common ancestor within a 10% margin of error, but that does not necessarily mean that the same method, applied to another data set will give equally precise results. Some of the error comes from phylogenetic methods, but more of it comes from the data chosen, and the fossil record.

I am in search for advice on methods to objectively assess what partitioning scheme to use on mitogenomes. I welcome answers to my question.

Best.

Hey Eliécer, There´s a new method apparently to find partitions and their substitution models in a Bayesian framework:

http://mbe.oxfordjournals.org/content/30/3/669

Thank you, Eduardo! I will check it out.

A discussion of the BEAST partitioning add-on, published in the paper Eduardo cited just above, has started here in the BEAST users group: https://groups.google.com/forum/#!topic/beast-users/zeI3wyRddFw

Thank you, Brian!

Huang and Green discuss partitioning sites by rate of evolution, rather than by gene or codon position. They show that the context of each site (such as C followed by G which makes the C the target of DNA methylation enzymes) is important.

Dick G. Hwang and Phil Green.

Bayesian Markov chain Monte Carlo sequence analysis reveals varying neutral substitution patterns in mammalian evolution.

PNAS, 101(39):13994-4001, 2004.

The Huang and Green data, and some other very nice data sets, are available here:

http://www.nisc.nih.gov/data/

Morgan et al have a recent paper with a very good analysis of mammalian evolution:

Morgan CC, Foster PG, Webb AE, Pisani D, McInerney JO, O'Connell MJ (2013) Heterogeneous models place the root of the placental mammal phylogeny. Molecular Biology and Evolution 30(9): 2145-2156. doi:10.1093/molbev/mst117

But their purpose was to look at the very root of the mammalian phylogeny, with marsupials as an outgroup. My understanding is that you are more interested in the evolution of a few species within one genus of mammals. For highly related species, within a genus, it is best to look at highly variable DNA, such as mitochondrial DNA. For the deep history of all mammals, highly conserved DNA is better.

The nice part of this publication, is that they also provide a NEXUS file of the data in the DRYAD repository:

http://datadryad.org/resource/doi:10.5061/dryad.vh49j

Hi, you can avoid the problem of gene-sequence partition using a mixture model.

Pagel, M. and Meade, A. 2004. A Phylogenetic Mixture Model for Detecting Pattern-Heterogeneity in Gene Sequence or Character-State Data. Systematic Biology, 53, 571-581.

Saludos

Enrique