Learning to explore tree neighbourhoods for phylogenetic inference.

Area of Science:

• Computational Biology
• Evolutionary Biology
• Bioinformatics

Background:

• Phylogenetic inference is crucial for understanding evolutionary relationships and comparative genomics.
• The balanced minimum evolution problem (BMEP) is a key formulation but is computationally intractable for large datasets.
• Existing methods struggle with scalability and adaptability for complex phylogenetic reconstruction.

Purpose of the Study:

• To develop a reinforcement learning (RL) framework to address the computational challenges of the BMEP.
• To improve the efficiency and accuracy of phylogenetic tree reconstruction for large instances.
• To investigate the generalization capabilities of the RL approach across diverse datasets and evolutionary models.

Main Methods:

• Proposed a novel RL formulation specifically designed for phylogenetic inference and the BMEP.
• Trained an RL agent to perform local search within the space of phylogenetic trees.
• Integrated the trained RL agent into a search-based framework for adaptive evaluation.

Main Results:

• The RL agent successfully solved BMEP instances with up to 100 taxa.
• The proposed method outperformed traditional greedy heuristics.
• The RL framework demonstrated competitive performance against state-of-the-art algorithms, especially under distributional shifts.

Conclusions:

• Reinforcement learning offers a promising avenue for tackling complex phylogenetic inference problems like the BMEP.
• The developed RL framework enhances the scalability and adaptability of phylogenetic tree reconstruction.
• This work highlights the potential of RL in advancing computational biology and evolutionary analysis.