Greetings,
Can anyone help me with a bootstrap resampling problem in TNT?
I am currently running an analysis with TNT in a data matrix composed by morphological characters (105 discrete and 25 continuous characters) and 87 taxa. From previous tests on TNT, I reached the conclusion that my dataset is rather complex, thus searches with default parameters for both traditional and newtech search cannot find the best result. I have been using a script with the following commands for my searches:
keep 0 ;
ttags - ;
hold 99999;
rseed 1 ;
sect:slack 1000 ;
xmult= hit 3 level 7 rat 10 drift 10; bb;
HOWEVER, I have been finding problems to run bootstrap support for this dataset: I cannot finish running the analysis. The resampling always stops doing new replicates before reaching the 500 replicates I have set. The bootstrap running analysis window will stop updating altogether for hours, until I manually stop the analysis.
These are the commands I have been using for the bootstrap search:
keep 0 ;
ttags - ;
hold 99999;
rseed 1 ;
sect:slack 1000 ;
ttags = ;
resample boot replications 500 [xmult= hit 2 level 5 rat 10 drift 10; bb] frequency gc ;
I have tried changing the parameters for the bootstrap analysis and using the default parameters from TNT, BUT they either still cannot finish 500 replicates, or when they finish the bootstrap values are exceedingly low (most likely because the resampling is not managing to find topologies close to the optimal I have found with the search).
Would anyone have any tips on how to change the parameters and commands to finish the 500 (or even 1000) replicates I am trying to achieve, while managing to do a good enough resampling of my data?
If desired, I can e-mail the matrix I am using and both scripts.
I have not used TNT per se but understand the phylogenetic bootstrap very well. What do you need to know?
Patrice Showers Corneli
I am mostly having a problem with calculating the bootstrap values for my data matrix, as the parameters I am using for the bootstrap calculations seem to be too "complex", and the program appears to be unable to calculate 500 replicates until the end. And if I use "simpler" parameters, the bootstrap values obtained are extremely low (average around 10~12), which I believe are due to the "simpler" parameters being unable to get close to the optimal topology for my dataset.
I would appreciate any tips on how to change the parameters for bootstrap calculations, so TNT can finish calculating them while still having a good search. I can also look at any alternatives to TNT for calculating bootstrap values for my dataset under parsimony, or other suggestions on different ways to calculate support values for my dataset.
You might try PAUP 4 as it is designed for parsimony analysis too and does the bootstrap as well as a few other parsimony measures. Perhaps if you ran it on a public server you could let it run longer. I like the Cipres Portal (https://www.phylo.org/restusers/login.action).
On the other hand, I suspect that you will still get very small bootstrap values and perhaps even smaller if you use the more complex model. This is because the bootstrap samples your data with replacement (so it is the same size as the original sample). Naturally some informative data will not be sampled at all for a given replicate and others may be sampled more than once. The idea is that a data set with strong phylogenetic signal will be included in many replicates while a small amount of signal which may have influenced the original tree will not be sampled in as many replicates. The bootstrap proportion is the number of times a particular pattern in your data is sampled divided by the total number number of replicates. So seldom sampled rare phylogenetic patterns will appear in only a small proportion of the replicates. With 500 replicates a value of 10 means that only 50 trees (10%) will have that branching pattern (clade). If the data is rich in phylogenetically informative data that has consistent patterns rather than conflicting patterns then you will get high bootstrap values.
Unfortunately the bootstrap does not recognized whether you model is good or not except in the sense that a complex model requires more parameters and so will require many more unique patterns to resolve the tree. As an aside, this is exactly why genomic-sized phylogenies almost always have very high bootstrap values. But as they are very large data sets, computers cannot yet compute trees with any but the most simple models with a few parameters. These models often do not adequately model the complexity of sequence evolutionary history and so the simple models are biased and provide very high support for the wrong tree. The same holds for any kind of statistical analysis. A poor model will give biased results. So your simple model may not be at all a good idea. On the other hand your data may be too limited to fully resolve the tree.
However, all is not lost, because we are often not interested in having high support for every single branch or clade. Perhaps you only want to know a-priori whether a certain clade is monophyletic. And suppose your data has enough data to resolve that particular clade. Then the support on the other clades is much less important than finding good support for your clade of interest.
Also you can in some programs (TNT I think, and PAUP I know) perform a Templeton (a paired test on the data for several trees - Kishino-Hsagawa for nucleotide data) test on your data by comparing the tree with monophyletic clade with trees that are not monophyletic. If the tree score is the same for both tree topologies then you can at least say that you monophyly cannot be rejected. Or if one tree is significantly better than another then you can say with confidence that you reject the other tree.
Should you need more advice please contact me again. This is a very common problem for phylogenetics in general.
- Badges
- Science topic
- Similar topics
- Biological Science
- Evolutionary Biology
- Phylogenetics