TreeFlow: Probabilistic Modeling and Automatic Differentiation for Phylogenetics
TL;DR
TreeFlow is a new software library that embeds phylogenetic trees into the TensorFlow Probability framework to enable probabilistic modeling and automatic differentiation for fixed tree topologies.
Problem / question
General probabilistic modeling frameworks cannot easily handle phylogenetic problems because of the unique computational properties of phylogenetic tree objects.
Methods
The authors developed TreeFlow, a software library that integrates phylogenetic trees into the TensorFlow Probability framework and implements gradient-based inference algorithms for phylogenetic models with fixed tree topologies.
Key findings
TreeFlow successfully enables the rapid implementation and assessment of new phylogenetic models. Furthermore, it achieves reasonable performance for gradient-based inference algorithms when compared to existing specialized computational libraries for phylogenetics.
Why it matters
It allows researchers to leverage modern probabilistic programming and automatic differentiation tools for phylogenetic analysis, speeding up the development and testing of new evolutionary models.
Limitations
The inference algorithms in TreeFlow are currently limited to models with a fixed tree topology.
Takeaway
TreeFlow bridges the gap between modern machine learning frameworks and phylogenetics by embedding trees into TensorFlow Probability for gradient-based inference on fixed topologies.