An independent base composition of each rate class for improved likelihood-based phylogeny estimation; the 5rf model

Area of Science:

• Phylogenetics and evolutionary biology
• Computational biology and bioinformatics

Background:

• Time-reversible Markov processes with gamma-distributed rates and invariant sites (e.g., tr4gi) are standard for nucleotide evolution models.
• These models assume homogeneous and stationary nucleotide frequencies, which may not reflect biological reality.

Purpose of the Study:

• To explore the benefits and challenges of a new model (4gi5rf) that allows each rate category to have its own nucleotide frequencies.
• To assess the practical identifiability of the 4gi5rf model using a BEAST 2 implementation.

Main Methods:

• Developed a 4gi5rf model where each of the five rate classes has independent nucleotide frequencies at the tree root.
• Implemented and tested the model in BEAST 2 to evaluate parameter estimation accuracy.
• Analyzed vertebrate mitochondrial DNA (mtDNA) and nuclear sequences.

Main Results:

• The 4gi5rf model significantly improved marginal likelihoods compared to the 4gi model.
• While reversing some errors, the 4gi5rf model introduced new ones, suggesting complex evolutionary processes.
• Problems were linked to non-stationary DNA repair processes affecting mutation/substitution spectra across lineages and time.

Conclusions:

• Vertebrate mtDNA is challenging to model accurately, and higher likelihoods do not guarantee improved biological accuracy.
• Non-stationarity in DNA repair processes poses a significant challenge for reliable phylogenetic inference, especially for divergence times near the root of major groups.
• An open-source implementation of the 4gi5rf model is available in the beastbooster package for BEAST 2.