Hi Tony,

You should use a different partition for each gene. RAxML, if I am not wrong, uses GTR + I + G for all genes, changing the I and G parameters for each case. But, each gene might be telling you a different history (different branch length and even tree topology) so you should use a different partition for each gene.

Hi,

In my experience (I used whole mt genomic data), branch lengths of the resultant trees are not drastically changed between the all gene partitioned model and the model you mentioned in which the genes should have "similar (statistically same)" substitution model are regarded as a single partition (I think you determined such partitions using Partitionfinder or similar software).

I think the resultant branch lengths between these models would be effected by the number of the partitions, molecular type, and/or genes of the data, and also usable substitution models. If almost all partitions have similar substitution model, the branch lengths may not be so differed. Also, in employing RAxML, GTR+G (I do not prefer to use +I) would be selected for many of gene partitions; and thus, the resultant branch lengths may not be so differed. By contrast, if you have many gene partitions with distinct preferred substitution models, the differences of the resultant branch lengths would be strengthen between the partition schemes.

(The above is just my imagination...I have not confirmed this, sorry :-)

Hello Toni

I likewise did not test this up till now, but

The default of RaxML is that it computes a joined branch length estimate for all partitions. Thus if you do not change this option your branch lengths in apartitioned model should not deviate too far from a joined analysis. You can change this setting however by activating the "-M" option. This is what the Manual writes at least.

Clearly Atsushi is right, if your estimated models do not vary too much this nevertheless will result in branch lengths corresponding to your overall analysis. Thern again, if models do not vary too much it is questionable whether partitioning is necessary (so usually they will deviate from a mean when you choose to take this option)

Greetings

Nikola

Partitioning you data will increase the computing time substantially. So it all depends on how important it is for you to get good estimates of branch lengths, or if you are only interested in the tree topology. Having I and G parameters for each gene does affect the estimation of branch lengths, which is crucial if you are going to estimate divergence time or anything of the sort.

Hi Toni,

I suggest that you look for PartitionFinder, is a software that tests different partition schemes and DNA substitution models at the same time. You can set it to test only DNA models that fit in RAxML. The site for the software is:

http://www.robertlanfear.com/partitionfinder/

Hi Toni,

PartitionFinder is a great software and it as made to address your problem!

Thank you all, this thread was very useful.

Hi Toni,

with partitionfinder you can define all the partitions you want to test, like each gene separately, all together, by function,... and you obtein which partition is the best. The problem could be if you obtain a model that it´s not in RAxML, that´s why, sometimes, I also use protest or modeltest.

David

@Osca

The most resent version of partitionfinder lets you work with raxml. Your analysis will then be confined to all models that are implemented in raxml.

Yes, you are right. I know but sometimes we prefer to know which one is really the best model to use and if it´s necessary we use other program to implement the correct model. Sorry, I was talking in general, for example bayesian models, we just implement the model if it is not in mrbayes.

I think the following paper is highly relevant to this discussion. I recently heard from one of the authors that the same subsets could be submitted to PartitionFinder, which would make the work easier.

http://www.sciencedirect.com/science/article/pii/S1055790312003636

"Mol Phylogenet Evol. 2013 Jan;66(1):69-79. doi: 10.1016/j.ympev.2012.09.006. Epub 2012 Sep 18.

Empirical evaluation of partitioning schemes for phylogenetic analyses of mitogenomic data: an avian case study.

Powell AF, Barker FK, Lanyon SM.

Source

Department of Ecology, Evolution and Behavior, Bell Museum of Natural History, 100 Ecology Building, 1987 Upper Buford Circle, University of Minnesota, St Paul, MN 55108, USA. [email protected]

Abstract

Whole mitochondrial genome sequences have been used in studies of animal phylogeny for two decades, and current technologies make them ever more available, but methods for their analysis are lagging and best practices have not been established. Most studies ignore variation in base composition and evolutionary rate within the mitogenome that can bias phylogenetic inference, or attempt to avoid it by excluding parts of the mitogenome from analysis. In contrast, partitioned analyses accommodate heterogeneity, without discarding data, by applying separate evolutionary models to differing portions of the mitogenome. To facilitate use of complete mitogenomic sequences in phylogenetics, we (1) suggest a set of categories for dividing mitogenomic datasets into subsets, (2) explore differences in evolutionary dynamics among those subsets, and (3) apply a method for combining data subsets with similar properties to produce effective and efficient partitioning schemes. We demonstrate these procedures with a case study, using the mitogenomes of species in the grackles and allies clade of New World blackbirds (Icteridae). We found that the most useful categories for partitioning were codon position, RNA secondary structure pairing, and the coding/noncoding distinction, and that a scheme with nine data groups outperformed all of the more complex alternatives (up to 44 data groups) that we tested. As hoped, we found that analyses using whole mitogenomic sequences yielded much better-resolved and more strongly-supported hypotheses of the phylogenetic history of that locus than did a conventional 2-kilobase sample (i.e. sequences of the cytochrome b and ND2 genes). Mitogenomes have much untapped potential for phylogenetics, especially of birds, a taxon for which they have been little exploited except in investigations of ordinal-level relationships."

See also:

Bayesian selection of nucleotide substitution models and their site assignments.

Wu CH, Suchard MA, Drummond AJ.

Mol Biol Evol. 2013 Mar;30(3):669-88. doi: 10.1093/molbev/mss258. Epub 2012 Dec

PMID: 23233462

Abstract

Probabilistic inference of a phylogenetic tree from molecular sequence data is predicated on a substitution model describing the relative rates of change between character states along the tree for each site in the multiple sequence alignment. Commonly, one assumes that the substitution model is homogeneous across sites within large partitions of the alignment, assigns these partitions a priori, and then fixes their underlying substitution model to the best-fitting model from a hierarchy of named models. Here, we introduce an automatic model selection and model averaging approach within a Bayesian framework that simultaneously estimates the number of partitions, the assignment of sites to partitions, the substitution model for each partition, and the uncertainty in these selections. This new approach is implemented as an add-on to the BEAST 2 software platform. We find that this approach dramatically improves the fit of the nucleotide substitution model compared with existing approaches, and we show, using a number of example data sets, that as many as nine partitions are required to explain the heterogeneity in nucleotide substitution process across sites in a single gene analysis. In some instances, this improved modeling of the substitution process can have a measurable effect on downstream inference, including the estimated phylogeny, relative divergence times, and effective population size histories.

Robert Lanfear has a new paper out on partitioning for phylogenetic analyses:

Selecting optimal partitioning schemes for phylogenomic datasets

Robert Lanfear, Brett Calcott, David Kainer, Christoph Mayer and Alexandros Stamatakis

BMC Evolutionary Biology 2014, 14:82 doi:10.1186/1471-2148-14-82

http://www.biomedcentral.com/1471-2148/14/82/abstract