ARTreeFormer: A faster attention-based autoregressive model for phylogenetic inference

Area of Science:

• Computational Biology
• Evolutionary Biology
• Machine Learning

Background:

• Phylogenetic inference faces challenges in modeling large tree topology spaces.
• Existing methods like ARTree offer flexibility but struggle with scalability due to inefficient computations.
• There is a need for faster and more scalable probabilistic models for phylogenetic tree inference.

Purpose of the Study:

• To develop a novel deep learning approach, ARTreeFormer, to accelerate phylogenetic tree topology modeling.
• To enhance the computational efficiency of autoregressive models for phylogenetic inference.
• To maintain high approximation performance while significantly improving speed for large datasets.

Main Methods:

• Introduced ARTreeFormer, combining fixed-point iteration and attention mechanisms.
• Developed a fixed-point iteration algorithm for fast, vectorized computation of topological node embeddings, optimized for CUDA.
• Implemented an attention-based global message passing scheme for efficient representation learning.

Main Results:

• ARTreeFormer significantly accelerates the computation speed of ARTree.
• The method demonstrates effectiveness and efficiency on challenging real-world phylogenetic inference problems.
• Maintained excellent approximation performance comparable to existing methods.

Conclusions:

• ARTreeFormer offers a scalable and efficient solution for probabilistic modeling of tree topologies in phylogenetic inference.
• The integration of fixed-point iteration and attention mechanisms overcomes the scalability limitations of previous autoregressive models.
• This advancement has the potential to significantly impact large-scale phylogenetic analyses.