Simulations of Sequence Evolution: How (Un)realistic They Are and Why.

Area of Science:

• Bioinformatics
• Computational Biology
• Evolutionary Biology

Background:

• Probabilistic models of sequence evolution are vital for evaluating phylogenetic inference tools and developing machine learning (ML) approaches for phylogenetic reconstruction.
• Realistic simulation of multiple sequence alignments (MSAs) is crucial for assessing the performance of phylogenetic tools on empirical data and ensuring ML models generalize well.

Purpose of the Study:

• To simulate DNA and protein MSAs using state-of-the-art methods under various evolutionary models, with and without insertion/deletion (indel) events.
• To assess the realism of simulated MSAs by quantifying the accuracy of supervised learning methods in distinguishing them from empirical MSAs.

Main Methods:

• Simulation of DNA and protein MSAs using a state-of-the-art sequence simulator under progressively complex evolutionary models.
• Application of two distinct, independently developed supervised learning classification approaches to differentiate simulated from empirical MSAs.

Main Results:

• High accuracy was achieved in distinguishing between empirical and simulated MSAs across all tested evolutionary models using both classification methods.
• Findings indicate that current evolutionary models inadequately capture key characteristics of empirical MSAs, such as site-wise evolutionary rates and nucleotide/amino acid composition.

Conclusions:

• Existing probabilistic models of sequence evolution do not fully replicate the complexity of empirical sequence data.
• Supervised learning methods can effectively identify subtle differences between simulated and empirical MSAs, highlighting areas for improvement in evolutionary modeling.