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

Deep Generative Models for Phylogenetic Inference with Complex Evolutionary Processes

NeurIPS 2026 · Probabilistic and scientific modeling › Variational and generative sampling · approach: Score-based generative modeling analysis

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.

Trains iterative generative models over phylogenetic trees (bottom-up, middle-out, top-down) with a discrete diffusion ELBO on simulated data, giving simulation-based posteriors that beat Phyloformer 2 and match or beat IQ-Tree under complex evolutionary models.

Problem

Bayesian phylogenetic posterior inference is intractable or too expensive for complex evolutionary models (indels, site-heterogeneous CAT-style models), and existing simulation-based methods use a single forward pass that limits posterior expressivity.

Approach

Model the tree posterior as an iterative generative process (bottom-up merging, middle-out taxon insertion, or top-down percolation) with a graph-neural network conditioned on sequences and the current partial tree, trained via a discrete diffusion ELBO with closed-form Rao-Blackwellized losses, plus covariance pooling to avoid representation collapse on long sequences.

Key results

• Bottom-up, middle-out and top-down models achieve substantially better NELBO than Phyloformer 2 across JC, CAT and InDel data (true entropy about 10 nats on JC).
• On real TreeBASE datasets, models recover MrBayes posteriors much more closely (lower MMD) than Phyloformer 2, though synthetic-data performance is better than real.
• On CAT and InDel data at N=45, L=500, the models achieve lower normalized RF than IQ-Tree, while matching IQ-Tree on JC data; bottom-up performs best among paradigms.

Surprising findings

• Phyloformer 2's reported advantage over IQ-Tree is attributed to test leakage: shuffling input sequence order raised its normalized RF from 0.046 to 0.238, and after fixing the leak Phyloformer 2 no longer beats IQ-Tree.
• Models are under-confident in their posterior predictions on realistic out-of-distribution data.

Limitations

• Only tree topology is inferred; branch lengths and evolutionary parameters are not handled.
• Scope limited to N up to 65 and L up to 2600; scaling to larger datasets untested.
• Under-confident posteriors and reduced accuracy on real out-of-distribution data relative to simulated data.

Experimental setup

Datasets: simulated JC, CAT, and InDel tree/sequence data (N=25-65, L=300-2600) 8 TreeBASE real datasets (DS1-8)

Benchmarks: real TreeBASE datasets DS1-8 vs MrBayes posteriors simulated JC, CAT, InDel data vs IQ-Tree and QuartFormer

Metrics: ELBO / NELBO MMD^2 split Spearman correlation coverage normalized Robinson-Foulds distance

Baselines: Phyloformer 2 IQ-Tree IQ-Tree Flex QuartFormer MrBayes

Builds on: Phyloformer 2 ARTree discrete diffusion (SCUD / Campbell et al.) EvoPF axial attention architecture muP

Closest papers by meaning

• Likelihood-free inference of phylogenetic tree posterior distributions (NeurIPS 2026 · similarity 0.92)
• PhyloGen: Language Model-Enhanced Phylogenetic Inference via Graph Structure Generation (NeurIPS 2024 · similarity 0.91)
• PhyloGFN: Phylogenetic inference with generative flow networks (ICLR 2024 · similarity 0.90)
• PhyloVAE: Unsupervised Learning of Phylogenetic Trees via Variational Autoencoders (ICLR 2025 · similarity 0.89)
• PhyloLM: Inferring the Phylogeny of Large Language Models and Predicting their Performances in Benchmarks (ICLR 2025 · similarity 0.86)
• Variational Phylogenetic Inference with Products over Bipartitions (ICML 2025 · similarity 0.86)