Phyloformer: Fast, Accurate, and Versatile Phylogenetic Reconstruction with Deep Neural Networks.

Area of Science:

• Computational Biology
• Evolutionary Biology
• Machine Learning

Background:

• Phylogenetic inference reconstructs evolutionary history from sequence data.
• Current methods like maximum likelihood and Bayesian inference are computationally expensive.
• Realistic evolutionary models are often too slow for practical use.

Purpose of the Study:

• Introduce Phyloformer, a novel method for fast and accurate phylogenetic reconstruction.
• Address the computational limitations of existing phylogenetic inference techniques.
• Enable the use of more sophisticated and realistic evolutionary models.

Main Methods:

• Developed Phyloformer using likelihood-free inference and geometric deep learning.
• Trained a neural network to predict phylogenetic trees from multiple sequence alignments.
• Utilized graphics processing unit (GPU) acceleration for computational speedup.

Main Results:

• Phyloformer demonstrated superior speed and accuracy compared to FastME, FastTree, and IQTree on simulated data under a common protein evolution model.
• Achieved high accuracy in both topology and branch lengths using the Kuhner-Felsenstein metric.
• Outperformed other methods across all metrics for models with site dependencies and fewer than 80 sequences.
• Matched maximum likelihood methods' topological accuracy on empirical gene alignments.

Conclusions:

• Phyloformer provides a fast and accurate alternative for phylogenetic reconstruction.
• The method shows promise for incorporating complex, realistic evolutionary models.
• Facilitates broader adoption of advanced phylogenetic inference techniques.