Research map / Probabilistic and scientific modeling / Variational and generative sampling

PhyloVAE: Unsupervised Learning of Phylogenetic Trees via Variational Autoencoders

ICLR 2025 · Probabilistic and scientific modeling › Variational and generative sampling · approach: Structured latent VAEs

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 PhyloVAE, a non-autoregressive variational autoencoder over phylogenetic tree topologies using a linear-time integer encoding, a GNN-based inference model, and a 2D/10D latent space for representation and fast generation.

Problem

Learning informative, high-resolution representations and generative models of phylogenetic tree topologies, which are discrete and non-Euclidean, while existing distance-based MDS embeddings are metric-sensitive and low-resolution and autoregressive density estimators (ARTree) are slow.

Approach

Encode each tree topology into an integer edge-decision vector via a linear-time decomposition/reconstruction procedure, model that vector with factorized (non-autoregressive) categorical decoders conditioned on a Gaussian latent, and infer the latent with a GNN over learnable topological node features; the latent mean serves as a representation.

Key results

• PhyloVAE (d=10) achieves the best KL divergence on DS2, DS3, DS4-level comparisons vs SBN-EM and comparable results to ARTree; e.g., DS3 KL 0.0477 vs ARTree 0.0548, and DS5 0.5744 vs ARTree 0.6266.
• ARTree is better on DS1 (0.0045 vs 0.0189), DS4 (0.0299 vs 0.0469), DS7 and DS8 for PhyloVAE (d=10).
• Training and generation time of PhyloVAE (d=10, K=32) are reported as significantly reduced compared to ARTree due to non-autoregressive factorization.

Measured effects reported in the paper

Surprising findings

• PhyloVAE uses more memory than ARTree on most benchmark data sets (e.g., 635.67 MB vs 400.37 MB on DS1), despite being faster.
• A two-dimensional latent space performs worse than d=10 on several sets but is reported as reliable for visualization.

Limitations

• Only unrooted tree topologies with a fixed leaf count and order are handled; extension to varying leaf sets is left to future work.
• Accuracy is not uniformly better than ARTree; ARTree wins on several benchmark data sets (e.g., DS1, DS4, DS7, DS8).
• Higher memory usage than ARTree on benchmark data sets.

Experimental setup

Datasets: DS1-DS8 phylogenetic benchmark sequence sets simulated eight-leaf topologies (JC model, 500bp DNA) mammal gene trees (five genes, 44 leaves) rabies genomes (290 sequences, BEAST runs) fossil crinoid data (42 sequences)

Benchmarks: DS1-DS8 tree topology density estimation (KL divergence to ground truth) runtime comparison vs ARTree

Metrics: KL divergence training time generation time memory usage Pearson correlation with RF and PD distances

Baselines: ARTree SBN-EM SBN-EM-α multidimensional scaling

Builds on: ARTree IWAE Phylo2Vec variational autoencoder learnable topological features

Closest papers by meaning

• PhyloGen: Language Model-Enhanced Phylogenetic Inference via Graph Structure Generation (NeurIPS 2024 · similarity 0.92)
• Deep Generative Models for Phylogenetic Inference with Complex Evolutionary Processes (NeurIPS 2026 · similarity 0.89)
• PhyloGFN: Phylogenetic inference with generative flow networks (ICLR 2024 · similarity 0.88)
• Likelihood-free inference of phylogenetic tree posterior distributions (NeurIPS 2026 · similarity 0.86)
• PhyloLM: Inferring the Phylogeny of Large Language Models and Predicting their Performances in Benchmarks (ICLR 2025 · similarity 0.85)
• Variational Phylogenetic Inference with Products over Bipartitions (ICML 2025 · similarity 0.85)