This figure illustrates the progress of a hypothetical run of the iterative
-means algorithm. The algorithm commences with an alignment that is treated as a single subset, and for which the AICc score has been calculated (step 3 in the description in the main text; represented by the red sequence alignment at the top). During this step, each of 56 GTR + I + G submodels is fit to the alignment and the model that returns the best AICc score is chosen. Next, the algorithm calculates TIGER site rates for each site (step 4a in the description in the main text), and uses these rates to classify the sites of the alignment into fast (red) and slow (blue) sites using the k-means algorithm (step 4b in the description in the main text). The AICc score of a model in which these two subsets are treated independently is then calculated (steps 4c-d in the description in the main text). If the score improves, the split is accepted. The fast (red) and slow (blue) sites are then used to create two new alignments, and the process is repeated with each new subset. This continues until no more subset splits are accepted. The final step combines all splits that improved the AICc score to create a single partitioning scheme for the dataset.