PhylaFlow: Hybrid Flow Matching in Billera-Holmes-Vogtmann Tree Space for Phylogenetic Inference

Leo Cui; Marinka Zitnik; Pardis Sabeti; Shrey Jain; Yasha Ektefaie

arxiv: 2605.21859 · v1 · pith:YSUPSD2Ynew · submitted 2026-05-21 · 🧬 q-bio.PE · cs.LG· q-bio.QM

PhylaFlow: Hybrid Flow Matching in Billera-Holmes-Vogtmann Tree Space for Phylogenetic Inference

Yasha Ektefaie , Leo Cui , Shrey Jain , Marinka Zitnik , Pardis Sabeti This is my paper

Pith reviewed 2026-05-22 03:00 UTC · model grok-4.3

classification 🧬 q-bio.PE cs.LGq-bio.QM

keywords phylogenetic inferenceBHV tree spaceflow matchinghybrid modelsBayesian phylogeneticstree topologiesposterior samplingtopology recovery

0 comments

The pith

A hybrid flow-matching model in BHV tree space learns to transport random phylogenetic trees into regions that support efficient recovery of posterior topologies during finite-budget Bayesian refinement.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces PhylaFlow to learn posterior-basin transport in BHV tree space by training on geodesic paths that mix continuous branch length changes inside fixed topologies with discrete jumps when topologies change. A sympathetic reader would care because exploring the space of possible trees and branch lengths for genetic data is computationally heavy, and better starting points or proposals could shorten the runs needed to approximate the posterior distribution. If the learned flow reaches useful regions, then initializing or guiding short Bayesian refinement from its outputs should recover trees with high posterior probability more quickly than standard random or heuristic starts. Experiments on eight benchmark datasets show that PhylaFlow lowers initial Tree-KL divergence compared to classical initializers and improves early and intermediate topology recovery after limited MrBayes refinement on most cases, with variants beating short-warmup baselines on seven of the eight.

Core claim

PhylaFlow is a hybrid flow-matching model that learns posterior-basin transport in BHV tree space. It is trained on BHV geodesic paths from random starting trees to short-run posterior samples, coupling continuous branch-length motion within orthants with learned boundary events and discrete topology transitions. When evaluated through finite-budget MrBayes refinement initialized from or guided by its terminal trees, the model recovers posterior-supported topologies more efficiently than classical initializers, as shown by reduced initial Tree-KL and improved topology-recovery trajectories on the DS1-DS8 benchmarks, with the best variant outperforming short-warmup on seven of eight datasets.

What carries the argument

The hybrid flow-matching model that couples continuous branch-length motion within orthants with learned boundary events and discrete topology transitions in Billera-Holmes-Vogtmann tree space.

If this is right

PhylaFlow substantially reduces initial Tree-KL relative to classical initializers across DS1-DS8.
Direct PhylaFlow improves early and intermediate topology-recovery trajectories after finite-budget MrBayes refinement on most datasets.
Split-guided PhylaFlow-MCMC obtains the strongest results on hard cases.
The best PhylaFlow variant outperforms short-warmup on seven of eight datasets and PhyloGFN on five of eight under the same budget.
Sequence embeddings steer posterior split recovery in joint sequence-conditioned experiments.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same flow-matching strategy on geodesic paths could supply proposals for sampling in other discrete-continuous spaces that arise in evolutionary models.
Scaling the approach to phylogenies with hundreds of taxa might reduce the mixing time problems that plague standard MCMC on large trees.
Deeper integration of raw sequence data into the flow could move toward models that infer both topology and lengths without separate alignment steps.
Learned flows on hybrid tree spaces offer a route to geometry-aware initializers that complement rather than replace existing Bayesian samplers.

Load-bearing premise

That training on BHV geodesic paths from random starting trees to short-run posterior samples produces a flow whose terminal states, when used to initialize or guide finite-budget refinement, reliably recover posterior-supported topologies more efficiently than standard methods.

What would settle it

Finite-budget MrBayes runs on the DS1-DS8 benchmarks initialized from PhylaFlow terminal trees show no reduction in initial Tree-KL and no gain in topology recovery rates relative to runs started from random trees or classical initializers.

Figures

Figures reproduced from arXiv: 2605.21859 by Leo Cui, Marinka Zitnik, Pardis Sabeti, Shrey Jain, Yasha Ektefaie.

**Figure 1.** Figure 1: Overview of PhylaFlow initialization in tree space. Traditional MCMC/MrBayes chains initialized from shared random starts follow long exploratory trajectories before entering the posterior basin. PhylaFlow maps the same starts into BHV tree space and transports them to diverse candidate basin-entry proposals. In contrast, maximum likelihood and maximum parsimony provide single-point initializations. Disti… view at source ↗

**Figure 2.** Figure 2: Motion in BHV tree space. Each orthant represents a fixed tree topology, with coordinates given by internal branch lengths. Moving within an orthant changes branch lengths without changing topology. When an internal branch contracts to zero, the path reaches a boundary corresponding to an unresolved tree; resolving that boundary in a different way moves the sample into a new topology. PHYLAFLOW uses this g… view at source ↗

**Figure 3.** Figure 3: Partial representative topology trajectories learned by P [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

**Figure 4.** Figure 4 [PITH_FULL_IMAGE:figures/full_fig_p013_4.png] view at source ↗

read the original abstract

Phylogenetic trees are hybrid objects: branch lengths vary continuously, while topologies change discretely through edge contractions and expansions. Billera-Holmes-Vogtmann (BHV) tree space provides a canonical geometry for this structure, representing each resolved topology as a Euclidean orthant and topological changes as motion across shared lower-dimensional boundaries. We introduce PhylaFlow, a hybrid flow-matching model that learns posterior-basin transport in BHV tree space. PhylaFlow is trained on BHV geodesic paths from random starting trees to short-run posterior samples, coupling continuous branch-length motion within orthants with learned boundary events and discrete topology transitions. We evaluate the learned geometry operationally: if the flow reaches posterior-relevant regions, finite-budget Bayesian refinement initialized from, or guided by, its terminal trees should recover posterior-supported topologies more efficiently. Across DS1-DS8 phylogenetic posterior benchmarks, PhylaFlow substantially reduces initial Tree-KL relative to classical initializers. After finite-budget MrBayes refinement, direct PhylaFlow improves early and intermediate topology-recovery trajectories on most datasets, while split-guided PhylaFlow-MCMC obtains the strongest hard-case results. The best PhylaFlow variant outperforms short-warmup on seven of eight datasets and PhyloGFN on five of eight under the same refinement budget. In a joint sequence-conditioned experiment, sequence embeddings steer posterior split recovery, although exact posterior topology recovery remains preliminary. These results show that hybrid flow matching can learn actionable transport in BHV tree space and provide a geometry-aware proposal mechanism for Bayesian phylogenetic inference.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

PhylaFlow tries hybrid flow matching in BHV tree space to generate better starting trees for phylogenetic MCMC, with benchmark gains that still need checks on whether the short-run targets actually reach the full posterior.

read the letter

The paper's main move is to train a flow-matching model that mixes continuous branch-length motion inside BHV orthants with learned jumps across topology boundaries. It generates paths from random trees to short MCMC samples and then tests whether those terminal trees speed up finite-budget MrBayes runs on eight standard datasets. The reported results show lower initial Tree-KL and faster early topology recovery than classical initializers, with the best variant beating short warmups on seven datasets and PhyloGFN on five. The sequence-conditioned experiment is a reasonable extension that lets embeddings influence split recovery. Those operational tests are a practical way to measure whether the learned transport is useful rather than just fitting an internal objective. The geometry setup itself looks cleanly motivated by the hybrid structure of trees. One real soft spot is the choice of training targets. Short MCMC runs in BHV space often mix slowly on topology, so the flow could be learning to reach the same limited regions the baselines already explore instead of broader posterior mass. If that holds, some of the refinement gains might be artifacts of matching incomplete samples rather than genuine posterior transport. The abstract does not spell out diagnostics for the quality or coverage of those short runs, which leaves the central claim harder to judge. This work is aimed at people building geometry-aware proposals or faster Bayesian phylogenetic samplers. It is worth sending to referees because the core modeling idea is coherent and the benchmark design is operational, even if the mixing concern needs direct evidence in the full methods and code.

Referee Report

2 major / 2 minor

Summary. The manuscript introduces PhylaFlow, a hybrid flow-matching model in Billera-Holmes-Vogtmann (BHV) tree space that learns posterior-basin transport for phylogenetic trees. Training uses BHV geodesic paths from random starting trees to short-run MCMC posterior samples, coupling continuous orthant motion with learned boundary crossings for topology changes. Operational evaluation on DS1-DS8 benchmarks shows reduced initial Tree-KL relative to classical initializers; after finite-budget MrBayes refinement, PhylaFlow variants improve early/intermediate topology recovery on most datasets, with the best variant outperforming short-warmup on seven of eight and PhyloGFN on five of eight. A sequence-conditioned variant is also explored.

Significance. If the results hold, the work demonstrates that flow matching can capture the hybrid continuous-discrete geometry of BHV space to produce actionable proposals for Bayesian phylogenetic inference. The operational test via refinement trajectories provides a practical, falsifiable assessment of whether terminal states reach posterior-supported regions, which could improve efficiency in topology exploration where standard MCMC mixes slowly.

major comments (2)

[§3] §3 (Training Data Generation): The targets consist of short-run MCMC samples whose mixing quality and topology coverage in BHV space are not reported. Because BHV mixing is known to be slow across topology boundaries, it is unclear whether these targets represent the full posterior or local basins; this directly affects whether the learned hybrid flow improves recovery of true posterior-supported topologies or merely artifacts of the same short-run distribution used in baselines.
[Results] Results section (benchmark tables): The reported gains (e.g., outperforming short-warmup on seven of eight datasets) lack error bars, multiple independent runs, or statistical tests. Without these, it is impossible to determine whether the topology-recovery improvements under finite-budget refinement are robust or sensitive to dataset-specific choices and data splits.

minor comments (2)

[Methods] Notation for the hybrid flow (continuous orthant motion plus boundary events) could be clarified with an explicit equation or diagram showing how the learned boundary crossings are parameterized.
[Abstract] The abstract states that 'exact posterior topology recovery remains preliminary' in the sequence-conditioned experiment; a brief quantitative comparison to the unconditional case would help readers assess the added value of sequence embeddings.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments. We respond to each major point below and indicate planned revisions.

read point-by-point responses

Referee: [§3] §3 (Training Data Generation): The targets consist of short-run MCMC samples whose mixing quality and topology coverage in BHV space are not reported. Because BHV mixing is known to be slow across topology boundaries, it is unclear whether these targets represent the full posterior or local basins; this directly affects whether the learned hybrid flow improves recovery of true posterior-supported topologies or merely artifacts of the same short-run distribution used in baselines.

Authors: We agree that characterizing the mixing and topology coverage of the short-run MCMC targets is valuable. In the revision we will add supplementary diagnostics (Tree-KL trajectories and unique-topology counts) for the chains used to generate training targets. At the same time, the paper's primary evaluation is operational rather than distributional: we test whether terminal states from the learned flow, when used to seed finite-budget MrBayes refinement, produce higher posterior-probability topologies than classical initializers on the same benchmarks. Consistent improvements under this metric would be unlikely if the flow merely reproduced uninformative local artifacts. revision: partial
Referee: [Results] Results section (benchmark tables): The reported gains (e.g., outperforming short-warmup on seven of eight datasets) lack error bars, multiple independent runs, or statistical tests. Without these, it is impossible to determine whether the topology-recovery improvements under finite-budget refinement are robust or sensitive to dataset-specific choices and data splits.

Authors: We acknowledge that the original results lack error bars and formal statistical tests. Because each MrBayes refinement trajectory is computationally expensive, the initial submission reported single runs. In the revision we will repeat the key comparisons with three independent random seeds, report means and standard deviations on the topology-recovery curves, and add paired statistical comparisons for the datasets where the ordering is consistent. revision: yes

Circularity Check

0 steps flagged

No significant circularity detected

full rationale

The paper introduces PhylaFlow as a hybrid flow-matching model trained on BHV geodesic paths from random trees to short-run posterior samples, then evaluates terminal states via independent finite-budget MrBayes refinement on external DS1-DS8 benchmarks. Reported gains in initial Tree-KL and topology recovery are measured against classical initializers, short-warmup, and PhyloGFN under the same refinement protocol rather than reducing to any fitted parameter or self-referential quantity by construction. No equations or steps in the abstract or description exhibit self-definition, fitted-input-as-prediction, or load-bearing self-citation chains; the operational claim rests on external benchmark performance and is therefore self-contained.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The approach rests on the standard geometric properties of BHV space and the assumption that short-run posterior samples are representative targets; no new free parameters or invented entities are explicitly introduced in the abstract.

axioms (1)

domain assumption Billera-Holmes-Vogtmann tree space provides a canonical geometry in which each resolved topology is a Euclidean orthant and topological changes occur across shared lower-dimensional boundaries.
This geometric foundation is invoked to justify the hybrid continuous-discrete flow construction.

pith-pipeline@v0.9.0 · 5841 in / 1364 out tokens · 33278 ms · 2026-05-22T03:00:07.979350+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Foundation/AlexanderDuality.lean alexander_duality_circle_linking unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

hybrid flow-matching model that learns posterior-basin transport in BHV tree space... coupling continuous branch-length motion within orthants with learned boundary events and discrete topology transitions
IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

trained on BHV geodesic paths from random starting trees to short-run posterior samples

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

23 extracted references · 23 canonical work pages

[1]

Alfaro and Mark T

Michael E. Alfaro and Mark T. Holder. The Posterior and the Prior in Bayesian Phylogenetics.Annual Review of Ecology, Evolution, and Systematics, 37(V olume 37, 2006):19–42, December

work page 2006
[2]

ISSN 1543-592X, 1545-2069. doi:

work page 2069
[3]

URLhttps://www.annualreviews.org/content/journals/ 10.1146/annurev.ecolsys.37.091305.110021

1146/annurev.ecolsys.37.091305.110021. URLhttps://www.annualreviews.org/content/journals/ 10.1146/annurev.ecolsys.37.091305.110021. Louis J. Billera, Susan P. Holmes, and Karen V ogtmann. Geometry of the space of phylogenetic trees.Advances in Applied Mathematics, 27(4):733–767, 2001a. ISSN 0196-8858. doi: https://doi.org/10.1006/aama.2001.0759. URL https...

work page doi:10.1146/annurev.ecolsys.37.091305.110021 2001
[4]

ChenRui Duan, Zelin Zang, Siyuan Li, Yongjie Xu, and Stan Z

doi: 10.1371/journal.pcbi.1003537. ChenRui Duan, Zelin Zang, Siyuan Li, Yongjie Xu, and Stan Z. Li. PhyloGen: Language Model-Enhanced Phyloge- netic Inference via Graph Structure Generation, December

work page doi:10.1371/journal.pcbi.1003537
[5]

arXiv:2412.18827 [q-bio]

URLhttp://arxiv.org/abs/2412.18827. arXiv:2412.18827 [q-bio]. Yasha Ektefaie, Andrew Shen, Lavik Jain, Maha Farhat, and Marinka Zitnik. Evolutionary reasoning does not arise in standard usage of protein language models. InThe Thirty-ninth Annual Conference on Neural Information Processing Systems,

work page arXiv
[6]

doi: 10.1038/s41579-025-01159-w

ISSN 1740-1534. doi: 10.1038/s41579-025-01159-w. URL https://www.nature.com/articles/s41579-025-01159-w. James Hadfield, Colin Megill, Sidney M Bell, John Huddleston, Barney Potter, Charlton Callender, Pavel Sagulenko, Trevor Bedford, and Richard A Neher. Nextstrain: real-time tracking of pathogen evolution.Bioinformatics, 34(23): 4121–4123, December

work page doi:10.1038/s41579-025-01159-w
[7]

doi: 10.1093/bioinformatics/bty407

ISSN 1367-4803. doi: 10.1093/bioinformatics/bty407. URLhttps://doi.org/ 10.1093/bioinformatics/bty407. John P. Huelsenbeck and Fredrik Ronquist. MRBAYES: Bayesian inference of phylogenetic trees.Bioinformatics, 17 (8):754–755, August

work page doi:10.1093/bioinformatics/bty407
[8]

doi: 10.1093/bioinformatics/17.8.754

ISSN 1367-4811, 1367-4803. doi: 10.1093/bioinformatics/17.8.754. URLhttps: //academic.oup.com/bioinformatics/article/17/8/754/235132. Paschalia Kapli, Ziheng Yang, and Maximilian J. Telford. Phylogenetic tree building in the genomic age.Nature Reviews Genetics, 21(7):428–444, July

work page doi:10.1093/bioinformatics/17.8.754
[9]

doi: 10.1038/s41576-020-0233-0

ISSN 1471-0064. doi: 10.1038/s41576-020-0233-0. URLhttps://www. nature.com/articles/s41576-020-0233-0. Jinwoo Kim, Tien Dat Nguyen, Seonwoo Min, Sungjun Cho, Moontae Lee, Honglak Lee, and Seunghoon Hong. Pure Transformers are Powerful Graph Learners, October

work page doi:10.1038/s41576-020-0233-0
[10]

Pure transformers are powerful graph learners

URLhttp://arxiv.org/abs/2207.02505. arXiv:2207.02505 [cs]. Clemens Lakner, Paul van der Mark, John P. Huelsenbeck, Bret Larget, and Fredrik Ronquist. Efficiency of markov chain monte carlo tree proposals in bayesian phylogenetics.Systematic Biology, 57(1):86–103,

work page arXiv
[11]

URLhttps://www.science.org/ doi/full/10.1126/sciadv.adk7623

doi: 10.1126/sciadv.adk7623. URLhttps://www.science.org/ doi/full/10.1126/sciadv.adk7623. Takahiro Mimori and Michiaki Hamada. Geophy: Differentiable phylogenetic inference via geometric gradients of tree topologies,

work page doi:10.1126/sciadv.adk7623
[12]

Luca Nesterenko, Luc Blassel, Philippe Veber, Bastien Boussau, and Laurent Jacob

URLhttps://arxiv.org/abs/2307.03675. Luca Nesterenko, Luc Blassel, Philippe Veber, Bastien Boussau, and Laurent Jacob. Phyloformer: Fast, accurate, and versatile phylogenetic reconstruction with deep neural networks.Molecular Biology and Evolution, 42(4):msaf051, 03

work page arXiv
[13]

doi: 10.1093/molbev/msaf051

ISSN 1537-1719. doi: 10.1093/molbev/msaf051. URLhttps://doi.org/10.1093/molbev/ msaf051. Lam-Tung Nguyen, Heiko A. Schmidt, Arndt von Haeseler, and Bui Quang Minh. Iq-tree: A fast and effective stochastic algorithm for estimating maximum-likelihood phylogenies.Molecular Biology and Evolution, 32(1):268–274, 11

work page doi:10.1093/molbev/msaf051
[14]

doi: 10.1093/molbev/msu300

ISSN 0737-4038. doi: 10.1093/molbev/msu300. URLhttps://doi.org/10.1093/molbev/msu300. Megan Owen and J. Scott Provan. A fast algorithm for computing geodesic distances in tree space.IEEE/ACM Trans. Comput. Biol. Bioinformatics, 8(1):2–13, January

work page doi:10.1093/molbev/msu300
[15]

doi: 10.1109/TCBB.2010.3

ISSN 1545-5963. doi: 10.1109/TCBB.2010.3. URLhttps: //doi.org/10.1109/TCBB.2010.3. Vincent Ranwez, Frédéric Delsuc, Sylvie Ranwez, Khalid Belkhir, Marie-Ka Tilak, and Emmanuel Jp Douzery. Or- thoMaM: a database of orthologous genomic markers for placental mammal phylogenetics.BMC Evol. Biol., 7:241, November

work page doi:10.1109/tcbb.2010.3 2010
[16]

Yatish Turakhia, Bryan Thornlow, Angie S

doi: 10.1093/sysbio/sys029. Yatish Turakhia, Bryan Thornlow, Angie S. Hinrichs, Nicola De Maio, Landen Gozashti, Robert Lanfear, David Haus- sler, and Russell Corbett-Detig. Ultrafast Sample placement on Existing tRees (UShER) enables real-time phylo- genetics for the SARS-CoV-2 pandemic.Nature Genetics, 53(6):809–816, June

work page doi:10.1093/sysbio/sys029
[17]

doi: 10.1038/s41588-021-00862-7

ISSN 1546-1718. doi: 10.1038/s41588-021-00862-7. URLhttps://www.nature.com/articles/s41588-021-00862-7. William M. Woodman and Tom M. W. Nye. Brownian motion, bridges and bayesian inference in phylogenetic tree space,

work page doi:10.1038/s41588-021-00862-7
[18]

Tianyu Xie and Cheng Zhang

URLhttps://arxiv.org/abs/2506.22135. Tianyu Xie and Cheng Zhang. Artree: A deep autoregressive model for phylogenetic inference. InAdvances in Neural Information Processing Systems, volume 36, 2023a. Spotlight. Tianyu Xie and Cheng Zhang. ARTree: A deep autoregressive model for phylogenetic inference. InThirty-seventh Conference on Neural Information Proc...

work page doi:10.1371/journal.pcbi.1013768
[19]

arXiv:2012.00459 [q-bio]

URL http://arxiv.org/abs/2012.00459. arXiv:2012.00459 [q-bio]. Cheng Zhang and Frederick A. Matsen IV . Variational bayesian phylogenetic inference. InInternational Conference on Learning Representations,

work page arXiv 2012
[20]

Mingyang Zhou, Zichao Yan, Elliot Layne, Nikolay Malkin, Dinghuai Zhang, Moksh Jain, Mathieu Blanchette, and Yoshua Bengio

URLhttps://proceedings.neurips.cc/ paper_files/paper/2018/file/b137fdd1f79d56c7edf3365fea7520f2-Paper.pdf. Mingyang Zhou, Zichao Yan, Elliot Layne, Nikolay Malkin, Dinghuai Zhang, Moksh Jain, Mathieu Blanchette, and Yoshua Bengio. PhyloGFN: Phylogenetic inference with generative flow networks. InInternational Conference on Learning Representations,

work page 2018
[21]

A Appendix A.1 Existing approaches to phylogenetics Bayesian phylogenetic inference aims to characterize the posterior distribution over tree topologies and branch lengths, p(T, B|S, θ) = p(S|T, B, θ)p(T, B|θ) p(S|θ) , whereTdenotes the tree topology,Bdenotes the branch lengths,Sdenotes the observed sequence alignment, andθ denotes additional evolutionary...

work page 2001
[22]

The final generated tree is the output of this deterministic rollout

The rollout terminates when the maximum number of phases is reached, the rollout-step budget is exhausted, the autoregressive event budget is exhausted, no selected first-hit edge has negative velocity, or no valid topology merge can be found. The final generated tree is the output of this deterministic rollout. After this we then run the branch-length re...

work page 2008
[23]

We found sub-sampling PhyloGFN trees did not substantially change results

For each dataset, every initialization method is evaluated using the same number of starting trees and the same downstream MRBAYES configuration, except for PhyloGFN where we used 1000 trees for each dataset due to the recommendation of the paper. We found sub-sampling PhyloGFN trees did not substantially change results. D.2 Random start tree generation A...

work page arXiv 2014

[1] [1]

Alfaro and Mark T

Michael E. Alfaro and Mark T. Holder. The Posterior and the Prior in Bayesian Phylogenetics.Annual Review of Ecology, Evolution, and Systematics, 37(V olume 37, 2006):19–42, December

work page 2006

[2] [2]

ISSN 1543-592X, 1545-2069. doi:

work page 2069

[3] [3]

URLhttps://www.annualreviews.org/content/journals/ 10.1146/annurev.ecolsys.37.091305.110021

1146/annurev.ecolsys.37.091305.110021. URLhttps://www.annualreviews.org/content/journals/ 10.1146/annurev.ecolsys.37.091305.110021. Louis J. Billera, Susan P. Holmes, and Karen V ogtmann. Geometry of the space of phylogenetic trees.Advances in Applied Mathematics, 27(4):733–767, 2001a. ISSN 0196-8858. doi: https://doi.org/10.1006/aama.2001.0759. URL https...

work page doi:10.1146/annurev.ecolsys.37.091305.110021 2001

[4] [4]

ChenRui Duan, Zelin Zang, Siyuan Li, Yongjie Xu, and Stan Z

doi: 10.1371/journal.pcbi.1003537. ChenRui Duan, Zelin Zang, Siyuan Li, Yongjie Xu, and Stan Z. Li. PhyloGen: Language Model-Enhanced Phyloge- netic Inference via Graph Structure Generation, December

work page doi:10.1371/journal.pcbi.1003537

[5] [5]

arXiv:2412.18827 [q-bio]

URLhttp://arxiv.org/abs/2412.18827. arXiv:2412.18827 [q-bio]. Yasha Ektefaie, Andrew Shen, Lavik Jain, Maha Farhat, and Marinka Zitnik. Evolutionary reasoning does not arise in standard usage of protein language models. InThe Thirty-ninth Annual Conference on Neural Information Processing Systems,

work page arXiv

[6] [6]

doi: 10.1038/s41579-025-01159-w

ISSN 1740-1534. doi: 10.1038/s41579-025-01159-w. URL https://www.nature.com/articles/s41579-025-01159-w. James Hadfield, Colin Megill, Sidney M Bell, John Huddleston, Barney Potter, Charlton Callender, Pavel Sagulenko, Trevor Bedford, and Richard A Neher. Nextstrain: real-time tracking of pathogen evolution.Bioinformatics, 34(23): 4121–4123, December

work page doi:10.1038/s41579-025-01159-w

[7] [7]

doi: 10.1093/bioinformatics/bty407

ISSN 1367-4803. doi: 10.1093/bioinformatics/bty407. URLhttps://doi.org/ 10.1093/bioinformatics/bty407. John P. Huelsenbeck and Fredrik Ronquist. MRBAYES: Bayesian inference of phylogenetic trees.Bioinformatics, 17 (8):754–755, August

work page doi:10.1093/bioinformatics/bty407

[8] [8]

doi: 10.1093/bioinformatics/17.8.754

ISSN 1367-4811, 1367-4803. doi: 10.1093/bioinformatics/17.8.754. URLhttps: //academic.oup.com/bioinformatics/article/17/8/754/235132. Paschalia Kapli, Ziheng Yang, and Maximilian J. Telford. Phylogenetic tree building in the genomic age.Nature Reviews Genetics, 21(7):428–444, July

work page doi:10.1093/bioinformatics/17.8.754

[9] [9]

doi: 10.1038/s41576-020-0233-0

ISSN 1471-0064. doi: 10.1038/s41576-020-0233-0. URLhttps://www. nature.com/articles/s41576-020-0233-0. Jinwoo Kim, Tien Dat Nguyen, Seonwoo Min, Sungjun Cho, Moontae Lee, Honglak Lee, and Seunghoon Hong. Pure Transformers are Powerful Graph Learners, October

work page doi:10.1038/s41576-020-0233-0

[10] [10]

Pure transformers are powerful graph learners

URLhttp://arxiv.org/abs/2207.02505. arXiv:2207.02505 [cs]. Clemens Lakner, Paul van der Mark, John P. Huelsenbeck, Bret Larget, and Fredrik Ronquist. Efficiency of markov chain monte carlo tree proposals in bayesian phylogenetics.Systematic Biology, 57(1):86–103,

work page arXiv

[11] [11]

URLhttps://www.science.org/ doi/full/10.1126/sciadv.adk7623

doi: 10.1126/sciadv.adk7623. URLhttps://www.science.org/ doi/full/10.1126/sciadv.adk7623. Takahiro Mimori and Michiaki Hamada. Geophy: Differentiable phylogenetic inference via geometric gradients of tree topologies,

work page doi:10.1126/sciadv.adk7623

[12] [12]

Luca Nesterenko, Luc Blassel, Philippe Veber, Bastien Boussau, and Laurent Jacob

URLhttps://arxiv.org/abs/2307.03675. Luca Nesterenko, Luc Blassel, Philippe Veber, Bastien Boussau, and Laurent Jacob. Phyloformer: Fast, accurate, and versatile phylogenetic reconstruction with deep neural networks.Molecular Biology and Evolution, 42(4):msaf051, 03

work page arXiv

[13] [13]

doi: 10.1093/molbev/msaf051

ISSN 1537-1719. doi: 10.1093/molbev/msaf051. URLhttps://doi.org/10.1093/molbev/ msaf051. Lam-Tung Nguyen, Heiko A. Schmidt, Arndt von Haeseler, and Bui Quang Minh. Iq-tree: A fast and effective stochastic algorithm for estimating maximum-likelihood phylogenies.Molecular Biology and Evolution, 32(1):268–274, 11

work page doi:10.1093/molbev/msaf051

[14] [14]

doi: 10.1093/molbev/msu300

ISSN 0737-4038. doi: 10.1093/molbev/msu300. URLhttps://doi.org/10.1093/molbev/msu300. Megan Owen and J. Scott Provan. A fast algorithm for computing geodesic distances in tree space.IEEE/ACM Trans. Comput. Biol. Bioinformatics, 8(1):2–13, January

work page doi:10.1093/molbev/msu300

[15] [15]

doi: 10.1109/TCBB.2010.3

ISSN 1545-5963. doi: 10.1109/TCBB.2010.3. URLhttps: //doi.org/10.1109/TCBB.2010.3. Vincent Ranwez, Frédéric Delsuc, Sylvie Ranwez, Khalid Belkhir, Marie-Ka Tilak, and Emmanuel Jp Douzery. Or- thoMaM: a database of orthologous genomic markers for placental mammal phylogenetics.BMC Evol. Biol., 7:241, November

work page doi:10.1109/tcbb.2010.3 2010

[16] [16]

Yatish Turakhia, Bryan Thornlow, Angie S

doi: 10.1093/sysbio/sys029. Yatish Turakhia, Bryan Thornlow, Angie S. Hinrichs, Nicola De Maio, Landen Gozashti, Robert Lanfear, David Haus- sler, and Russell Corbett-Detig. Ultrafast Sample placement on Existing tRees (UShER) enables real-time phylo- genetics for the SARS-CoV-2 pandemic.Nature Genetics, 53(6):809–816, June

work page doi:10.1093/sysbio/sys029

[17] [17]

doi: 10.1038/s41588-021-00862-7

ISSN 1546-1718. doi: 10.1038/s41588-021-00862-7. URLhttps://www.nature.com/articles/s41588-021-00862-7. William M. Woodman and Tom M. W. Nye. Brownian motion, bridges and bayesian inference in phylogenetic tree space,

work page doi:10.1038/s41588-021-00862-7

[18] [18]

Tianyu Xie and Cheng Zhang

URLhttps://arxiv.org/abs/2506.22135. Tianyu Xie and Cheng Zhang. Artree: A deep autoregressive model for phylogenetic inference. InAdvances in Neural Information Processing Systems, volume 36, 2023a. Spotlight. Tianyu Xie and Cheng Zhang. ARTree: A deep autoregressive model for phylogenetic inference. InThirty-seventh Conference on Neural Information Proc...

work page doi:10.1371/journal.pcbi.1013768

[19] [19]

arXiv:2012.00459 [q-bio]

URL http://arxiv.org/abs/2012.00459. arXiv:2012.00459 [q-bio]. Cheng Zhang and Frederick A. Matsen IV . Variational bayesian phylogenetic inference. InInternational Conference on Learning Representations,

work page arXiv 2012

[20] [20]

Mingyang Zhou, Zichao Yan, Elliot Layne, Nikolay Malkin, Dinghuai Zhang, Moksh Jain, Mathieu Blanchette, and Yoshua Bengio

URLhttps://proceedings.neurips.cc/ paper_files/paper/2018/file/b137fdd1f79d56c7edf3365fea7520f2-Paper.pdf. Mingyang Zhou, Zichao Yan, Elliot Layne, Nikolay Malkin, Dinghuai Zhang, Moksh Jain, Mathieu Blanchette, and Yoshua Bengio. PhyloGFN: Phylogenetic inference with generative flow networks. InInternational Conference on Learning Representations,

work page 2018

[21] [21]

A Appendix A.1 Existing approaches to phylogenetics Bayesian phylogenetic inference aims to characterize the posterior distribution over tree topologies and branch lengths, p(T, B|S, θ) = p(S|T, B, θ)p(T, B|θ) p(S|θ) , whereTdenotes the tree topology,Bdenotes the branch lengths,Sdenotes the observed sequence alignment, andθ denotes additional evolutionary...

work page 2001

[22] [22]

The final generated tree is the output of this deterministic rollout

The rollout terminates when the maximum number of phases is reached, the rollout-step budget is exhausted, the autoregressive event budget is exhausted, no selected first-hit edge has negative velocity, or no valid topology merge can be found. The final generated tree is the output of this deterministic rollout. After this we then run the branch-length re...

work page 2008

[23] [23]

We found sub-sampling PhyloGFN trees did not substantially change results

For each dataset, every initialization method is evaluated using the same number of starting trees and the same downstream MRBAYES configuration, except for PhyloGFN where we used 1000 trees for each dataset due to the recommendation of the paper. We found sub-sampling PhyloGFN trees did not substantially change results. D.2 Random start tree generation A...

work page arXiv 2014