Methods used to generate PAML data

Tim Sackton and Amanda Larracuente

11/7/2007



I. Alignments

ftp://ftp.flybase.net/genomes/12_species_analysis/clark_eisen/alignments



Alignments were produced based on the GLEANR consensus set of gene models by 

Dan Pollard and Venky Iyer for every melanogaster gene with a single ortholog in 

sim, sec, yak, ere, and ana. For genes with multiple transcripts in dmel, only 

the longest was used, leaving a total of 8668 alignments.  Details of the 

masking procedure in the supplemental materials of Drosophila 12 Genomes Consortium, 

Evolution of genes and genomes on the Drosophila phylogeny. Nature doi:10.1038/nature06341



Before running PAML, alignments were further processed as follows:

Any codon which was masked or missing in more than one species was stripped from 

the alignment.  Any alignment with a stop codon in any species, or where the total length 

of the stripped alignment was less than 30% of the total length of the full alignment, or 

where the total length of the stripped alignment was less than 60 nucleotides was filtered 

out, leaving a total of 8511 alignments for PAML.



II. PAML runs

Each aligment/model combination was run for all three possible yak/ere topologies (yak/ere sister, 

yak outgroup, and ere outgroup). In order to deal with convergence problems, most models were run 

at least twice with different initial conditions, and the run with the best likelihood for each 

model was kept.  In addition, for all cases where the LRT between the null and alternative model 

were negative, both models were rerun.  For on the order of 100 alignments, we were unable to get 

good convergence for a small number of models even after several reruns; the best run has been left 

in the data set, but where these issues cause a negative LRT,the p-value has been reported as -99.



The true null distribution for several of the tests is unknown.  In order to calculate p-values 

for those tests, data was simulated under the null model, and the p-values were calculated from 

the empirical cumulative distribution in R.



III. Choosing the best tree

In order to simplify the analysis, we only present data for the best tree for each alignment.  

For each alignment, the tree with the lowest likelihood was selected for each of the 24 models.  

For >90% of the cases, one tree had the lowest likelihood in > 50% of all models; that tree was 

then selected.  For the small fraction of cases where a single tree was not supported in the 

majority of models, the tree used was the tree with the most support (even if less than 50%), 

and the tree best supported by M0 was counted twice to break ties.



The impact of different trees is generally minimal: although the difference in likelihoods between trees 

can be substantial, the differences tend to be consistent across models, so the LRT test calculated for 

one tree tends to be similar to that calculated for different trees.



The raw data for each tree is available on request from Tim Sackton (tbs7@cornell.edu) or Amanda Larracuente (aml69@cornell.edu).