For the phylogenomic dataset from birds we first removed all sites in the alignment that were removed by the original authors [43 (link)], and then defined data blocks based on each intron, and each codon position in each exon. This resulted in a total of 168 data blocks. We then performed a total of 12,002 searches for partitioning schemes on this dataset, described below.
We performed 2 searches for optimal partitioning schemes using the greedy algorithm [2 (link)]: one with the AICc, and one with the BIC.
We performed 2000 searches for optimal partitioning schemes using the strict hierarchical clustering algorithm described above. The 2000 searches comprise 1000 searches using the BIC and 1000 using the AICc, where each search used one of 1000 distinct clustering weights (the ‘--weights’ commandline option in PartitionFinder). The clustering weights are defined by a vector of four numbers that specify the relative importance of four parameter categories (the overall subset rate, the base frequencies, the GTR model parameters, and the alpha parameter of the gamma distribution; see above). Analysing 1000 sets of weights allows us to empirically compare the performance of different weighting schemes, and to determine the relative importance of the different parameter categories when searching for partitioning schemes, as well as the variation in the algorithm’s performance under different weighting schemes. The first 15 sets of weights comprise all possible combinations of setting at least one weight to 1.0, and other weights to 0.0 (setting all weights to 0.0 is nonsensical, as it would lead to all subsets appearing to be equally similar). These represent 15 of the 16 corners of a four dimensional hypercube, and allow us to compare the 15 cases where either all parameter categories are given equal weight (i.e. --weights “1, 1, 1, 1”) or where one or more parameters are given zero weight (e.g. --weights “1, 0, 0, 1”). The other 985 points were chosen using Latin Hypercube Sampling in the ‘lhs’ package, version 0.1 in R [48 ]. This procedure ensures that the sampled points are relatively evenly distributed in four-dimensional space, and is a more efficient way of sampling high-dimensional space than using a grid-based sampling scheme.
We performed 10,000 searches for optimal partitioning schemes using the relaxed clustering algorithm described above. These 10,000 searches comprised 5000 searches using the AICc, and 5000 using the BIC, each of which was performed with 1000 different clustering weights, and at 5 different values of the parameter P. The 1000 weighting schemes we used were identical to those used above, and the values of the parameter P (which defines the percentage of possible partitioning schemes that are considered at each step of the relaxed clustering algorithm) that we used were 1%, 2%, 5%, 10%, and 20%.
The results of all 12002 analyses presented here are available at figShare (http://dx.doi.org/10.6084/m9.figshare.938920 ).
We performed 2 searches for optimal partitioning schemes using the greedy algorithm [2 (link)]: one with the AICc, and one with the BIC.
We performed 2000 searches for optimal partitioning schemes using the strict hierarchical clustering algorithm described above. The 2000 searches comprise 1000 searches using the BIC and 1000 using the AICc, where each search used one of 1000 distinct clustering weights (the ‘--weights’ commandline option in PartitionFinder). The clustering weights are defined by a vector of four numbers that specify the relative importance of four parameter categories (the overall subset rate, the base frequencies, the GTR model parameters, and the alpha parameter of the gamma distribution; see above). Analysing 1000 sets of weights allows us to empirically compare the performance of different weighting schemes, and to determine the relative importance of the different parameter categories when searching for partitioning schemes, as well as the variation in the algorithm’s performance under different weighting schemes. The first 15 sets of weights comprise all possible combinations of setting at least one weight to 1.0, and other weights to 0.0 (setting all weights to 0.0 is nonsensical, as it would lead to all subsets appearing to be equally similar). These represent 15 of the 16 corners of a four dimensional hypercube, and allow us to compare the 15 cases where either all parameter categories are given equal weight (i.e. --weights “1, 1, 1, 1”) or where one or more parameters are given zero weight (e.g. --weights “1, 0, 0, 1”). The other 985 points were chosen using Latin Hypercube Sampling in the ‘lhs’ package, version 0.1 in R [48 ]. This procedure ensures that the sampled points are relatively evenly distributed in four-dimensional space, and is a more efficient way of sampling high-dimensional space than using a grid-based sampling scheme.
We performed 10,000 searches for optimal partitioning schemes using the relaxed clustering algorithm described above. These 10,000 searches comprised 5000 searches using the AICc, and 5000 using the BIC, each of which was performed with 1000 different clustering weights, and at 5 different values of the parameter P. The 1000 weighting schemes we used were identical to those used above, and the values of the parameter P (which defines the percentage of possible partitioning schemes that are considered at each step of the relaxed clustering algorithm) that we used were 1%, 2%, 5%, 10%, and 20%.
The results of all 12002 analyses presented here are available at figShare (
Free full text: Click here
