Le and Gascuel introduced an improved general amino acid replacement matrix (LG) for protein phylogenetics. The LG matrix was estimated using a larger and more diverse dataset (3,912 Pfam alignments, ~50,000 sequences, ~6.5 million residues) compared to the WAG matrix. The LG matrix incorporates site rate variability and uses an adaptation of the XRATE software for estimation. The LG matrix was evaluated against WAG and JTT using independent test alignments from TreeBase and Pfam. The results showed that LG significantly improved likelihood values compared to WAG and JTT. LG outperformed WAG in 38 out of 59 alignments and was significantly worse in 2. Tree topologies inferred with LG, WAG, and JTT frequently differed, indicating that LG impacts both likelihood and tree topology. Results with Pfam test alignments were similar. LG and a PHYML implementation are available at http://atgc.lirmm.fr/LG. The LG matrix accounts for site rate variability and uses a more comprehensive dataset than previous methods. The LG matrix was estimated using a two-step approach: first, phylogenies were inferred, then the replacement matrix was estimated using an expectation-maximization algorithm. The LG matrix was compared with WAG and JTT using AIC values and tree topologies. LG showed a significant improvement in AIC values compared to WAG and JTT, with an average AIC gain per site of 0.25 and 0.21 for TreeBase and Pfam, respectively. LG was significantly better than WAG in 38 out of 59 alignments and significantly worse in 2. The LG matrix was found to be more accurate in capturing evolutionary patterns and capturing more hidden substitutions, resulting in longer trees. The LG matrix was also compared with WAG' (a matrix estimated from Pfam alignments using the same procedure as WAG). LG showed a significant improvement over WAG' in AIC values. The LG matrix was found to be more accurate in capturing amino acid exchangeability and was better at distinguishing between rare and common substitution events. The LG matrix was also compared with other models, including JTT and JTT+F, which were found to be worse than WAG and WAG+F. The LG matrix was found to be more accurate in capturing evolutionary patterns and was better at distinguishing between rare and common substitution events. The LG matrix was also found to be more accurate in capturing amino acid exchangeability and was better at distinguishing between rare and common substitution events. The LG matrix was found to be more accurate in capturing evolutionary patterns and was better at distinguishing between rare and common substitution events. The LG matrix was found to be more accurate in capturing amino acid exchangeability and was better at distinguishing between rare and common substitution events. The LG matrix was found to be more accurate in capturing evolutionary patterns and was better at distinguishing between rare and common substitution eventsLe and Gascuel introduced an improved general amino acid replacement matrix (LG) for protein phylogenetics. The LG matrix was estimated using a larger and more diverse dataset (3,912 Pfam alignments, ~50,000 sequences, ~6.5 million residues) compared to the WAG matrix. The LG matrix incorporates site rate variability and uses an adaptation of the XRATE software for estimation. The LG matrix was evaluated against WAG and JTT using independent test alignments from TreeBase and Pfam. The results showed that LG significantly improved likelihood values compared to WAG and JTT. LG outperformed WAG in 38 out of 59 alignments and was significantly worse in 2. Tree topologies inferred with LG, WAG, and JTT frequently differed, indicating that LG impacts both likelihood and tree topology. Results with Pfam test alignments were similar. LG and a PHYML implementation are available at http://atgc.lirmm.fr/LG. The LG matrix accounts for site rate variability and uses a more comprehensive dataset than previous methods. The LG matrix was estimated using a two-step approach: first, phylogenies were inferred, then the replacement matrix was estimated using an expectation-maximization algorithm. The LG matrix was compared with WAG and JTT using AIC values and tree topologies. LG showed a significant improvement in AIC values compared to WAG and JTT, with an average AIC gain per site of 0.25 and 0.21 for TreeBase and Pfam, respectively. LG was significantly better than WAG in 38 out of 59 alignments and significantly worse in 2. The LG matrix was found to be more accurate in capturing evolutionary patterns and capturing more hidden substitutions, resulting in longer trees. The LG matrix was also compared with WAG' (a matrix estimated from Pfam alignments using the same procedure as WAG). LG showed a significant improvement over WAG' in AIC values. The LG matrix was found to be more accurate in capturing amino acid exchangeability and was better at distinguishing between rare and common substitution events. The LG matrix was also compared with other models, including JTT and JTT+F, which were found to be worse than WAG and WAG+F. The LG matrix was found to be more accurate in capturing evolutionary patterns and was better at distinguishing between rare and common substitution events. The LG matrix was also found to be more accurate in capturing amino acid exchangeability and was better at distinguishing between rare and common substitution events. The LG matrix was found to be more accurate in capturing evolutionary patterns and was better at distinguishing between rare and common substitution events. The LG matrix was found to be more accurate in capturing amino acid exchangeability and was better at distinguishing between rare and common substitution events. The LG matrix was found to be more accurate in capturing evolutionary patterns and was better at distinguishing between rare and common substitution events