Research map / Probabilistic and scientific modeling / Molecules and proteins

PhyloGen: Language Model-Enhanced Phylogenetic Inference via Graph Structure Generation

NeurIPS 2024 · Probabilistic and scientific modeling › Molecules and proteins · approach: Generative latent models for single-cell data · 2 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes PhyloGen, which uses a DNABERT-2 genomic language model plus neighbor-joining initialization and a variational tree-structure generation module with a scoring function to jointly infer topology and branch lengths for phylogenetic trees.

Problem

Phylogenetic inference requires jointly handling discrete tree topologies and continuous branch lengths; MCMC converges slowly, and existing variational methods either need pre-generated topologies or require aligned sequences and optimize topology and branch lengths separately.

Approach

Treats phylogenetic inference as conditionally constrained tree structure generation: DNABERT-2 embeddings feed an NJ-initialized tree, then a TreeEncoder/Decoder and DGCNN-based branch-length module are co-optimized end-to-end with an MLP scoring function added to the ELBO, without alignment or evolutionary-model assumptions.

Key results

• Reports the highest MLL on all eight datasets, e.g., DS1 MLL -6910.02 vs. MrBayes -7108.42 and ARTree -7108.41.
• Achieves the highest ELBO on all eight datasets, e.g., DS1 -7005.98 vs. ARTree -7110.09 and SBN -7110.24.
• Ablation on DS1 shows removing the KL loss sharply reduces ELBO and MLL; robustness tests with 4 deleted or added species show improvements and lower runtime (6h32min vs. 18h28min for PhyloGFN in setting 1).

Measured effects reported in the paper

Surprising findings

• Removing the scoring function S has a large effect on MLL but a relatively small effect on ELBO.
• Reported MLL/ELBO gains over MrBayes are large in absolute terms, while the paper notes MrBayes stepping-stone values serve as gold standard, so the comparison depends on estimator choice.

Limitations

• Neighbor-Joining construction is computationally intensive, per the authors.
• Evaluated primarily on genomic (DNA) data; applicability to protein or single-cell data is untested.
• Expressiveness of the latent distributions and the scoring function's contribution are acknowledged as areas for improvement.

Experimental setup

Datasets: DS1 DS2 DS3 DS4 DS5 DS6 DS7 DS8

Benchmarks: DS1-DS8 phylogenetic inference benchmarks

Metrics: ELBO MLL Simpson's diversity index bipartition frequency

Baselines: MrBayes SBN VBPI VBPI-GNN ARTree phi-CSMC GeoPhy PhyloGFN

Builds on: variational Bayesian phylogenetic inference DNABERT-2 neighbor-joining DGCNN GeoPhy

Closest papers by meaning

• PhyloGFN: Phylogenetic inference with generative flow networks (ICLR 2024 · similarity 0.93)
• Likelihood-free inference of phylogenetic tree posterior distributions (NeurIPS 2026 · similarity 0.92)
• PhyloVAE: Unsupervised Learning of Phylogenetic Trees via Variational Autoencoders (ICLR 2025 · similarity 0.92)
• Deep Generative Models for Phylogenetic Inference with Complex Evolutionary Processes (NeurIPS 2026 · similarity 0.91)
• PhyloLM: Inferring the Phylogeny of Large Language Models and Predicting their Performances in Benchmarks (ICLR 2025 · similarity 0.90)
• Variational Phylogenetic Inference with Products over Bipartitions (ICML 2025 · similarity 0.87)