ATGC: Phyml 3.0 benchmarks

PhyML 3.0 Benchmarks

Efficiency of filtering strategies

Comparison of distance-based PhyML SPR program (W. Hordijk and O. Gascuel, 2005) and PhyML 3.0 SPR , using different parsimony filtering with various intensities.

See DNA benchmark - See protein benchmark - Download DNA medium-size data sets - Download protein medium-size data sets

Distribution of relative computing times: for each of the 2 sets of alignments (50 DNA and 50 protein medium-size alignments) we measured the base-2 logarithm of the ratio of the computing time of the given method, and that of the fastest approach with the corresponding alignment. Thus, a log-ratio equals to X corresponds to a method being 2^X times slower than the fastest approach; e.g. distance-based method is much faster than others, and the unfiltered 3.0 version is by far the slowest one. The PT=0 and PT=5 ones seem to be close and a good compromise.

DNA	Av. LogLk rank	Delta>5	P-value<0.05	Av. RF distance
PhyML SPR	3.49	15	1	0.24
PhyML 3.0 SPR (PT=0)	2.43	3	0	0.09
PhyML 3.0 SPR (PT=5)	2.17	2	0	0.06
PhyML 3.0 SPR (PT=infinity)	1.91	2	0	0.05

PROTEIN	Av. LogLk rank	Delta>5	P-value<0.05	Av. RF distance
PhyML SPR	2.85	7	2	0.18
PhyML 3.0 SPR (PT=0)	2.69	2	0	0.12
PhyML 3.0 SPR (PT=5)	2.25	2	0	0.06
PhyML 3.0 SPR (PT=infinity)	2.12	1	0	0.04

Performance of the parsimony filter. PhyML-SPR filter (Hordijk and Gascuel, 2005) uses distance-based minimum evolution principle, while PhyML 3.0 SPR filter uses parsimony. When PT=infinity all SPRs are evaluated with likelihood, without any preliminary filtering. On the opposite, PT=0 corresponds to strong filtering (see text). The column ‘Av. LogLk rank’ gives the average log-likelihood ranks for the different methods. These ranks are corrected by taking into account information on tree topologies (see text). ‘Delta>5’ gives the number of cases (among 50) for which the difference of log-likelihood between the method of interest and the highest log-likelihood for the corresponding data set is greater than 5. The column ‘p-value<0.05’ displays the number of cases for which the difference of log-likelihood when comparing the method of interest to the corresponding highest log-likelihood is statistically significant (SH test). The ‘Av. RF distance’ values are the average Robinson and Foulds topological distances between the trees estimated by the method of interest and the corresponding most likely trees (0 corresponds to identical trees, while 1 means that the two trees do not have any clade in common).

Data sets

The benchmark contains 50 protein alignments and 50 DNA alignments.

DNA alignments
We selected the 50 most recent alignments from Treebase with at least 50 sequences, less than 200 sequences and less than 2000 sites.
Protein alignments
We selected the 50 most recent alignments from Treebase with at least 5 sequences, less than 200 sequences and less than 2000 sites.

Hardware

All programs have been run on a cluster Intel(R) Xeon(R) CPU 5140 @ 2.33GHz, 24 computing nodes, with 8GB of RAM for one bi-dualcore unit. Times can be compared because we've only considered effective computing time for the CPU.

Programs

4 programs and options have been compared. All programs were configured with the GTR model for DNA sequences, with WAG for proteins, and with 4 discrete gamma rate categories (alpha estimated from the data).

PhyML SPR
Previous version of PhyML SPR (W. Hordijk and O. Gascuel, 2005), PhyML, optimizing the topology with SPR (and NNI) operations, and using a BioNJ starting tree.
PhyML 3.0 SPR PT=0
PhyML, optimizing the topology with SPR (and NNI) operations, and using a BioNJ starting tree. Only SPR with a parcimony score at least equal to the current solution are considered.
PhyML 3.0 SPR PT=5 (default option)
PhyML, optimizing the topology with SPR (and NNI 3.0) operations, and using a BioNJ starting tree. Only SPR with a parsimony score at most 5 points worst than the current solution are considered.
PhyML 3.0 SPR PT=infinity
PhyML, optimizing the topology with SPR (and NNI) operations, and using a BioNJ starting tree. SPR are not filtered by parcimony.

Results

Resulting trees are compared regarding topology, log-likelihood and computing time.

Computing time ranks
The four methods are ranked for each of the alignments, based on the computing time. First rank contains methods with computing time ranging from the best (B) computing time to 1.25 X B (i.e. nearly best computing time). Remaining methods are ranked in the same way, until all methods are ranked. Ties are accounted for; e.g. if the first and second group contains 2 methods each, the ranks will be 1.5 ( (1+2)/2 ) and 3.5 ( (3+4)/2 ). To summarize these results, we provide the median and average ranks for all DNA and protein alignments.
Topology ranks
The four methods are ranked for each of the alignments using a similar principle, based on the tree likelihood. First rank contains all methods which find the same best topology. And so on. Moreover, we provide the median and average ranks for all DNA and protein alignments.
Robinson and Foulds distances
RF is the Robinson and Foulds (bipartition) distance between the best topology and the given topology.
Delta>5
Another variable of interest is the number of times a method fails to find a phylogeny which log-likelihood is close to the highest log-likelihood found by any of the methods being compared. We thus counted the number of data sets for which the log-likelihoods returned by a given method was smaller than the highest log-likelihood found on the corresponding alignments minus 5.0. While this boundary of 5.0 points of log-likelihood is arbitrary, we believe that it provides a simple and practical way to tell the methods apart at first sight.
SH tests
We used the Shimoidara-Hasegawa (SH) test to assess the statistical significance of the likelihood differences. Every result displays the P-value between its logLk and the logLk of the best result for the same data. As a summary, we provide the number of times each method is significatively worst than the best one.