The Tree Reconstruction Game: Phylogenetic Reconstruction Using Reinforcement Learning.

Area of Science:

• Computational Biology
• Bioinformatics
• Machine Learning

Background:

• Phylogenetic tree reconstruction is computationally challenging, often leading to suboptimal solutions.
• Current algorithms may get stuck in local optima, failing to find the globally best-fit tree.

Purpose of the Study:

• To develop a new paradigm for predicting maximum-likelihood phylogenetic trees using reinforcement learning.
• To overcome the limitations of local optima in traditional tree search algorithms.

Main Methods:

• Employing reinforcement learning to approximate long-term likelihood gains for optimal search strategy.
• Training an agent to navigate the phylogenetic search space towards the global optimum.

Main Results:

• Achieved log-likelihood improvements of 0.969 or higher compared to state-of-the-art methods on empirical data.
• Demonstrated a threefold speed increase on datasets with 15 sequences of 18,000 bp.
• Reduced the need for costly post-training likelihood optimizations.

Conclusions:

• Reinforcement learning offers a powerful approach to enhance phylogenetic tree reconstruction.
• The proposed method significantly improves both accuracy and efficiency in finding maximum-likelihood trees.