Fig. 4. Topology-dependent asymmetry of the ability of model-violating site-homogeneous models to reconstruct the correct tree.

(A) A total of 100 datasets were simulated using a site-heterogeneous model for each of the topologies shown (orange/gray boxes). (B) For the datasets based on the whole alignment, site-heterogeneous (top) and site-homogeneous models (bottom) were used to reconstruct a maximum likelihood tree. The proportion of times the orange or black tree was reconstructed is shown in the bar charts. Data simulated under the Nephrozoa and the Ctenophora-first trees always yield the correct topology regardless of the model. Data simulated under the Xenambulacraria and Porifera-first topologies mostly yield the correct topology under the site-heterogeneous model but an incorrect topology under the site-homogeneous model. The incorrect tree is always Nephrozoa and the Ctenophora-first, respectively. (C) The experiments were repeated for the datasets based on the sets of genes best and worst at reconstructing known clades. For the best genes under both models, the inference is improved for data simulated under Xenambulacraria and Porifera-first topologies. A decrease in the performance of both models is observed using the worst data.