Wednesday, October 31, 2007

October 31, 2007

Our next meeting will be on November 14th at 11am and we will be headed back to population genetics with an overview of methods by Violaine. She will talk with us about what she has done in the past few years, developing microsatellites and subsequent analysis. We will be reading the following (I still have to finish scanning the references, but the text of the paper is now up):

**BE PATIENT...IT IS A BIG FILE**
KT Scribner & JM Pearce. (2000). Microsatellites: Evolutionary and methodological background and empirical applications at individual, population and phylogenetic levels. In: Molecular Methods in Ecology (ed. AJ Baker), pp235-273. Blackwell Science, Malden, USA.

Violaine also recommends the following paper if you are looking for more information:
Selkoe KA & Toonen RL (2006). Microsatellites for ecologists: a practical guide to using and evaluating microsatellite markers. Ecology Letters. doi: 10.1111/j.1461-0248.2006.00889.x

Today we discussed the article from last week, but before I summarize, I thought I would put Felsenstein's webpage up. It is crazy useful, having a large part (all?) of current phylogenetics programs at hand.

Today we went back over the article "Phylogeny Estimation: Traditional and Bayesian Approaches" by Holder and Lewis. We began discussing the general dichotomies of phylogenetic analyses:
  • Cladistics
    • such as Parsimony methods and does not assume a model of evolution
  • Phenetics (usually good to start with the use of MODELTEST to determine a model of evolution)
    • Distance methods
      • Neighbor Joining, Minimum Evolution
  • Probabilistic methods
    • Maximum Likelihood, Bayesian

We then generally discussed some issues in determining which methods might be used in different cases and what should be taken into consideration with each. We jumped around a bit, so the following reflects that.

MODELS OF EVOLUTION
* a model of evolution informs your tree building with how changes occur. The simplest models of nucleotide evolution treat all changes (e.g. A-->T, A-->C, A-->G) as equally weighted, with more complex models weighting changes differently.

The first increase in complexity of the model of evolution usually deals with different weights being attributed to transitions (purine to purine OR pyrimidine to pyrimidine changes) and transversions (changes between purines and pyrimidines).
purine=A, G
pyrimidine=C,U,T (and is easily remembered by CUT the Pie (Py) :)

Subsequent increases to the complexity of the model can attribute different weights to each type of base change, as well as how changes in each position differ (i.e. do some sites change frequently while others tend to be conserved?).

NEIGHBOR JOINING (MEGA):
* With short internal branches relative to terminal branches, this method is not advisable
* When your tree is ladderized, this method is also not advisable
* It is limited to dealing with one gene at a time, unless you have previously determined that different genes are homogeneous with regard to their model of evolution (this can be checked by doing a hompart (Partition Homogeneity (congruence)) test in PAUP). If two or more genes are homogeneous, then they may be concatenated.
* This method is often the best place to start to generate a beginning tree- one that can be used in subsequent methods.

MAXIMUM PARSIMONY (PAUP):
* No model of evolution assumed
* considers gaps as information
* also limited to one gene at a time (see above)

MAXIMUM LIKELIHOOD (PHYML):
*strictly estimates parameters, searching until highest likelihood (lnL) is found. Can therefore give undue weight to nuisance parameters.
* Requires multiple runs and bootstrapping to arrive at a consensus tree with supported nodes.
* also limited to one gene at a time (see above)

BAYESIAN (MR BAYES):
* can deal with multiple genes, with each having a different model of evolution. A very powerful tool in this respect
* uses a window (prior probabilities) of parameter values in order to estimate the highest likelihood, versus ML which optimizes each parameter value.
* The method (multiple runs, multiple generations, multiple chains) results in a long string of high likelihood values that are then used together to provide support for the nodes of your tree using Bayesian posterior probability (BPP).

SUPERTREES
* I am not entirely familiar with this method (or whether the above link is correct), but it requires you to do single gene analyses in order to determine the best topologies for each gene and then to develop a 1/0 matrix for each tree. These can be converted from a NEWICK format by using the programs RAINBOW, RAD-CON, and SUPERTREE or can be done manually. (Baum and Reagan) and Purvis (no dependent columns) have different methods of doing this and although Purvis' method is more time consuming, significant differences between them have not yet been found.

GAPS (or INDELS)
* Only in Parsimony methods can this be automatically done (as a 5th or 21st option for nucleotide and amino acids respectively). This "missing" data can sometimes be very phylogenetically informative. It is important to consider how gaps are dealt with, particularly in non-coding regions, where the incidence should be higher. In methods where gaps are not considered, (i.e. all methods other than parsimony) indel (gap) information can sometimes be manually added.
* Nadir mentioned that there might be a program/website that can generate indel information lines for you from your data, FAST GAP or GAP RECODER.
* No model of gap evolution exists- which would seem to me to make their inclusion chancy in other methods.

Again, please contact me or add comments in the side panel if you have comments, questions, additions or corrections!