# Learnable Topological Features for Phylogenetic Inference via Graph Neural Networks

**Authors:** Cheng Zhang  
**arXiv:** [2302.08840](https://arxiv.org/abs/2302.08840)  
**Date:** 2023-02-17  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2302.08840v1/

## Key points

- Phylogenetics is a crucial field in computational biology that aims to uncover evolutionary relationships among biological entities through sequence data analysis.
- Constructing phylogenetic trees is essential, serving as graphical models to calculate likelihood of observed sequences efficiently.
- Various statistical inference methods like maximum likelihood and Bayesian approaches are used for inferring shared evolutionary history.
- Challenges in phylogenetic inference arise from the complex parameter space involving continuous (branch lengths) and discrete components (tree topology), leading to a combinatorial explosion in possible tree topologies with increasing sequence numbers.
- Leveraging structural information of phylogenetic trees is key for developing efficient inference algorithms.
- Techniques like conditional clade distributions (CCDs) and subsplit Bayesian networks (SBNs) have been proposed to improve tree probability estimation and enhance MCMC algorithms for Bayesian phylogenetics.
- Machine learning approaches have been employed to accelerate tree-search algorithms by incorporating informative topology moves, requiring significant design effort and domain expertise.
- A novel structural representation method based on learnable topological features has been introduced, combining raw node features with graph representation learning techniques that minimize Dirichlet energy for efficient structural information adaptable to various downstream tasks without expert knowledge.

## Summary

Phylogenetics is a crucial field in computational biology that aims to uncover the evolutionary relationships among biological entities through the analysis of sequence data. This involves constructing phylogenetic trees, which serve as graphical models to calculate the likelihood of observed sequences efficiently. Various statistical inference methods, such as maximum likelihood and Bayesian approaches, are utilized in phylogenetic inference to infer shared evolutionary history. However, phylogenetic inference faces challenges due to the complex parameter space involving both continuous (branch lengths) and discrete components (tree topology), leading to a combinatorial explosion in possible tree topologies with increasing sequence numbers. To address these challenges, leveraging the structural information of phylogenetic trees becomes essential for developing efficient inference algorithms. For instance, techniques like conditional clade distributions (CCDs) and subsplit Bayesian networks (SBNs) have been proposed to improve tree probability estimation and enhance MCMC algorithms for Bayesian phylogenetics. Recently, machine learning approaches have been employed to accelerate tree-search algorithms by incorporating informative topology moves. These methods rely on heuristic features like clades and subsplits of phylogenetic trees, which often require significant design effort and domain expertise. To overcome this limitation, a novel structural representation method based on learnable topological features has been introduced in this paper. By combining raw node features with graph representation learning techniques that minimize Dirichlet energy, these learnable topological features offer efficient structural information adaptable to various downstream tasks without expert knowledge. The effectiveness and efficiency of this method were demonstrated through simulations on tree probability estimation tasks and challenging real data variational Bayesian phylogenetic inference problems. Overall, this innovative approach holds promise for advancing phylogenetic inference by automating the adaptation of structural information in a more effective and efficient manner.