Figure 2

BIC scores for partitioning schemes estimated during empirical testing (lower is better). The k-means methods presented here outperform traditional methods. “None” is no partitioning, “All” is the user partitioning scheme, “PF-G” is the PartitionFinder greedy algorithm, “TIGER” is iterative k-means using TIGER site rates, “Likelihood” is iterative k-means using likelihood site rates. Note: The “PF-G” score for the McCormack 2013 dataset was obtained using the PartitionFinder relaxed clustering followed by model selection with PhyML as described in the Methods, not the greedy algorithm.