11
$\begingroup$

I have a large phylogenomic alignment of >1000 loci (each locus is ~1000bp), and >100 species. I have relatively little missing data (<10%).

I want to estimate a maximum-likelihood phylogenetic tree from this data, with measures of statistical support on each node.

There are many phylogenetics programs that claim to be able to analyse datasets like this (e.g. RAxML, ExaML, IQtree, FastTree, PhyML etc). Given that I have access to a big server (512GB RAM, 56 cores), what are the pros and cons of each program.

Which programme likely to give me the most accurate estimate of the ML tree for a dataset of this size within a reasonable timeframe?

Improve this question
$\endgroup$
3
  • $\begingroup$ You want us to review all the phylogenetic programs or just the ones you listed: RAxML, ExaML, IQtree, FastTree, PhyML (Which still I think is quite broad). How do you measure the accuracy of the ML tree? $\endgroup$
    – llrs
    Jun 10, 2017 at 10:51
  • $\begingroup$ I'm interested in any and all opinions, evidence, and links to comparisons between any software capable of estimating ML trees from large datasets like this. There would be many ways of measuring accuracy, including: (i) evidence from simulation; (ii) comparisons of Likelihood scores of estimated trees in a common framework. $\endgroup$
    – roblanf
    Jun 11, 2017 at 23:17
  • $\begingroup$ RAxML is the goto program. The problem with FastTree is that its accuracy is limited to the datasets used, i.e. yours might be an outlier. $\endgroup$
    – M__ ♦
    Dec 28, 2018 at 21:22

1 Answer 1

Reset to default
11
$\begingroup$

This paper claims that FastTree is almost as accurate as RAxML, while being much faster. You just have to be careful, however, that the support values output by FastTree are not bootstrap values, they are based on the Shimodaira-Hasegawa test. (Also, see this comment for the case you have very short branch lengths). [update: However, according to the recent comparison paper mentioned below FastTree performed quite poorly in comparison to RAxML or IQ-tree.]

From what I understand, you should use ExaML only if your data is too large to be handled by RAxML in a single node. ExaML should perform like RAxML but with some parallelization overhead. For all effects I treat them as the same. I don't know of relevant advantages of phyML over RAxML (for me, it's easier to use but I am very used to phyML).

I am not familiar with IQ-tree, but its authors claim that even given the same time as RAxML or phyML, IQ-tree already finds better likelihoods more often than not (although by default it takes a bit longer to converge). A recent comparison between all these programs favoured IQ-TREE for both single-gene and concatenation analyses (with RAxML very close). It may also estimate branch support through a SH-like test only, but I'm not sure. [update: IQ-tree offers 3 measures of support, standard bootstrap, aLRT, and ultrafast bootstrap. See OP's comment below for details.]

However, since you have few missing data, you might also want to try a single-locus tree inference followed by gene tree clustering (using treescape or treeCL) to see how spread your data is, or to see the effect of removal of outliers, or to use ideas similar to statistical binning.

Improve this answer
$\endgroup$
5
  • 1
    $\begingroup$ In the recent comparison you mention, I note that RAxML only performed better than IQtree when they did 10 independent searches per replicate. So one might just favour IQtree here too, noting that it's sensible to do a bunch of independent searches. Also FastTree performed very poorly in all comparisons of that ms. $\endgroup$
    – roblanf
    Jun 11, 2017 at 23:20
  • 1
    $\begingroup$ Measures of support in IQtree: (i) standard bootstrap; (ii) aLRT (approximate likelihood ratio test, which is something like asking whether a given branch length is significantly >0); (iii) Ultrafast bootstrap (not the same interpretation as a standard bootstrap, more like a posterior probability if I've understood it right). $\endgroup$
    – roblanf
    Jun 11, 2017 at 23:22
  • $\begingroup$ Note that likelihood comparisons between program outcomes also depend on the choice of model and data partitioning. I haven't looked at the details, but some programs may implement models that are not available in others. This can be a criterion of choice. $\endgroup$
    – bli
    Jun 12, 2017 at 8:33
  • 1
    $\begingroup$ Thank you for the comments @roblanf, I took the liberty of updating my answer with this info. $\endgroup$
    – Leo Martins
    Jun 12, 2017 at 8:50
  • 2
    $\begingroup$ note also that there is a newly released re-write of RAxML available here. It combines the best parts of RAxML and ExaML, while being faster in general. However not all current RAxML features have been implemented yet. $\endgroup$
    – Pierre Barbera
    Jun 12, 2017 at 15:28

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge that you have read and understand our privacy policy and code of conduct.

Not the answer you're looking for? Browse other questions tagged or ask your own question.