Figure 2.

Highly constrained sites drive long-branch attraction artifacts in the Felsenstein zone. We simulated amino acid alignments with 10,000 sites exhibiting across-site compositional heterogeneity (Schrempf et al. 2020) along Felsenstein-type trees (insets in the top row; Felsenstein 1978) with different branch lengths ${\displaystyle q\text{=0}\text{.1}}$ and ${\displaystyle p\text{=0}\text{.3}}$, 0.8, and 1.2 from (a) to (c). We performed analyses with CAT-PMSF, the Poisson (Felsenstein 1973; Nei 1987), the LG (Le and Gascuel 2008), and the GTR (Tavaré 1986) models constrained to the correct topology as well as to an incorrect topology (inset in the bottom row; Farris 1999) with IQ-TREE 2 (Minh et al. 2020). The site-specific log-likelihood differences between the maximum likelihood trees of the two competing topologies binned according to the site-specific effective number of amino acids are shown. A positive value (blue background) indicates support for the true topology, a negative value (yellow background) indicates support for the incorrect topology exhibiting long-branch attraction. The LG and GTR models incorrectly infer Farris-type trees if ${\displaystyle p\ge 0.8}$.