Partitioned models in GARLI
Partitioned models have been available for some time in a testing version of the program (GARLI-PART0.97). GARLI 2.0, the first "official" release including partitioned models. It is a merging of official release 1.0 and beta version GARLI-PART 0.97. Briefly, it includes models for nucleotides, amino acids, codons, and morphology-like characters, any of which can be mixed together and applied to different subsets of data.
If you just want to details of how to set up a partitioned analysis, see the information on this page: Using_partitioned_models.
Features of this version
- Partitioning of data into subsets, each of which may be assigned separate evolutionary models, with parameters independently estimated. Each model can be at the level of nucleotides, amino acids, codons or discrete morphology.
- Any submodel of GTR with gamma rate heterogeneity and a proportion of invariant sites may be assigned to each data subset.
- Different subsets can have different models (e.g., GTR for one, HKY for another)
- One set of branch-length parameters may be shared by all subsets, or one set shared with independent rate multipliers for each subset (adding #subsets - 1 parameters). I term this "subset-specific rates".
- A single set of model parameters may be shared by all subsets, but subset-specific rates still estimated. This is akin to the site-specific rates model in PAUP. Other forms of rate heterogeneity may still be inferred on top of this.
- Mk/Mkv morphology models, including mixing of sequence and morphology data. See specific details here.
- Most features that worked with earlier "normal" versions of the program still work with partitioned models, including checkpointing, output of site likelihoods, providing or fixing model parameter values, internal state reconstruction, etc.
What is not in this version
- Linking of particular model parameters between subsets (e.g., two subsets with independent base frequencies but sharing a single rate matrix) as implemented in MrBayes.
- Fully independent branch-length parameters for each partition subset. A single set of branch lengths in inferred, along with the rate of each subset relative to these global branch lengths.
What is the history of partitioned models in GARLI?
- May 2008: GARLI Version 0.96b8 was released. This was a beta, and the first version to allow amino acids and codon models.
- Dec 2008: The first partitioned version was posted for limited testing. Because it was derived from the 0.96b8, I called it Partitioned GARLI Version 0.96b8. It continued to be called that until March 2010. This was an oversight on my part.
- Dec 2009: GARLI Version 1.0 was released. This was a finalization of version 0.96b8, and does NOT include the partitioned models.
- March 2010: To avoid further confusion, I rename the partitioned version GARLI-PART Version 0.97.
- April 2011: Version 2.0 is released, bringing partitioned models into an official release.