Long-branch attraction bias and inconsistency in Bayesian phylogenetics

Area of Science:

• Evolutionary biology
• Computational biology
• Bioinformatics

Background:

• Bayesian inference (BI) and maximum likelihood (ML) are probabilistic methods for phylogenetic tree reconstruction.
• BI is generally assumed to share ML's statistical properties, including unbiased topology inference and increasing reliability with more data.

Purpose of the Study:

• To investigate potential biases in Bayesian inference (BI) for phylogenetic relationships.
• To compare the performance of BI and ML in inferring evolutionary trees, particularly concerning branch length estimation.

Main Methods:

• Experimental simulation studies were conducted to assess BI's performance.
• Numerical and mathematical analyses were employed to understand the behavior of BI under varying conditions.
• Empirical sequence data analyses were performed to validate simulation findings.

Main Results:

• BI exhibits a bias favoring topologies that group long branches together (long-branch attraction).
• This bias is exacerbated by increasing amounts of data, leading to incorrect tree inference with high support.
• BI is less efficient and robust than ML, especially with heterogeneous sequence evolution or incorrect evolutionary models.

Conclusions:

• Bayesian inference (BI) is susceptible to long-branch attraction bias, unlike maximum likelihood (ML).
• The integration of uncertainty in branch lengths, a feature of BI, contributes to this bias.
• Phylogenetic trees inferred using BI warrant caution; ML may offer a more reliable framework for phylogenetic analysis.