arxiv: 2605.09506 · v1 · submitted 2026-05-10 · 📊 stat.ME · q-bio.QM· stat.CO

Recognition: no theorem link

Accelerating Bayesian Phylogenetic Inference via Delayed Acceptance Sequential Monte Carlo with Random Forest Surrogates

Shijia Wang, Wentao Yu

Pith reviewed 2026-05-12 04:44 UTC · model grok-4.3

classification 📊 stat.ME q-bio.QMstat.CO

keywords Bayesian phylogeneticsdelayed acceptancesequential Monte Carlorandom forest surrogatephylogenetic tree moveslikelihood accelerationposterior sampling

0 comments

The pith

A random forest surrogate predicts likelihood changes from tree moves to let delayed-acceptance SMC reject poor proposals early while preserving the posterior.

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

The paper builds a random forest on more than thirty topological and branch-length features that forecasts the sign and rough size of the likelihood shift produced by standard phylogenetic rearrangements such as eSPR and stNNI. This cheap prediction is placed inside a delayed-acceptance kernel that first tests the surrogate; only moves that pass receive a full likelihood evaluation. The kernel is then embedded in a sequential Monte Carlo sampler that targets the same tree posterior as conventional methods. On both simulated and real sequence data the approach delivers large reductions in the number of expensive likelihood calls and corresponding wall-clock savings without detectable distortion of the posterior.

Core claim

A random forest trained on topological and branch-length features predicts the sign and rough magnitude of the log-likelihood ratio for eSPR and stNNI moves. This predictor is inserted into a delayed-acceptance kernel inside an SMC sampler, so that only moves passing the cheap surrogate test receive a full likelihood evaluation.

What carries the argument

The random forest surrogate that approximates the likelihood change for proposed phylogenetic tree moves, enabling preliminary rejection in the delayed acceptance step.

If this is right

The delayed-acceptance SMC recovers posterior distributions statistically indistinguishable from those of standard SMC.
The number of full likelihood evaluations drops substantially, producing measurable reductions in computational time.
The method performs consistently on both simulated alignments and empirical sequence data.
The surrogate can be retrained for any chosen collection of tree-move features.

Where Pith is reading between the lines

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

The same surrogate strategy could be paired with other expensive likelihood models outside phylogenetics.
If surrogate bias stays negligible, the delayed-acceptance template supplies a reusable pattern for accelerating any sampler whose proposals admit cheap side information.
Extending the feature set to datasets with hundreds of taxa would require checking whether the random forest remains unbiased at larger scales.

Load-bearing premise

The random forest surrogate, trained on topological and branch-length features, accurately predicts the sign and rough magnitude of likelihood change for standard tree moves without introducing systematic bias into the posterior.

What would settle it

Run both DA-SMC and ordinary SMC on identical data sets and compare the resulting posterior distributions over trees; any systematic shift in clade support or branch-length quantiles would show that the surrogate has biased the sampler.

Figures

Figures reproduced from arXiv: 2605.09506 by Shijia Wang, Wentao Yu.

**Figure 2.** Figure 2: An overview of the Bayesian phylogenetic DA-SMC framework. Given a list of weighted [PITH_FULL_IMAGE:figures/full_fig_p011_2.png] view at source ↗

**Figure 3.** Figure 3: The logarithm marginal likelihood (log Z) difference between K2P and GTR, K2P and JC69. the consensus tree (or the maximum-likelihood tree among the posterior samples) and the ground truth tree to assess the quality of tree estimation. 4.3 Estimate assessment and model selection using marginal likelihood In this section, we conducted a systematic comparison of marginal likelihood estimates for our propose… view at source ↗

**Figure 4.** Figure 4: Consensus log-likelihood and RF distance as a function of [PITH_FULL_IMAGE:figures/full_fig_p017_4.png] view at source ↗

**Figure 5.** Figure 5: Consensus log-likelihood, PM and RF distance as a function of different [PITH_FULL_IMAGE:figures/full_fig_p018_5.png] view at source ↗

**Figure 6.** Figure 6: Left: Runtime ratio of DA-SMC to ASMC as a function of the number of taxa. A lower [PITH_FULL_IMAGE:figures/full_fig_p019_6.png] view at source ↗

**Figure 7.** Figure 7: The RMSE (Root Mean Square Error) and R2 obtained with decreasing number of features for Random Forest training. The table at the bottom displays the feature composition within each set of features along with their importance, as determined by the importance-decreasing order. 5.2 Dataset M336 We used K = 500 and β = 5.3 for the DA-SMC algorithm, which results in 8 · 106 iterations. For the model selectio… view at source ↗

**Figure 8.** Figure 8: The majority-rule consensus trees for the M336 dataset estimated by (a) DA-SMC and [PITH_FULL_IMAGE:figures/full_fig_p022_8.png] view at source ↗

read the original abstract

In Bayesian phylogenetics, our goal is to estimate the posterior distribution over phylogenetic trees. Markov chain Monte Carlo methods are widely used to approximate the phylogenetic posterior distributions. For large-scale sequence data, repeated evaluation of the likelihood function incurs a high computational cost. In this article, we propose a machine-learning algorithm with over 35 topological and branch-length features to predict the changes in the likelihood function caused by tree moves (\eg,~eSPR, stNNI) used in standard MCMC approaches. This algorithm is then used to design a delayed acceptance MCMC kernel, which utilized the predicted surrogate function for preliminary rejection, to accelerate tree space searches. Furthermore, we integrate our proposed MCMC kernel into the sequential Monte Carlo sampler framework. We validate the proposed delayed-acceptance sequential Monte Carlo approach (DA-SMC) on simulation and real data sets. Our delayed acceptance kernel can maintain robust estimation while reduces the number of likelihood evaluations significantly, yielding substantial computational time savings. We develop a Python package that is available at https://github.com/wentYu/DAphyloSMC.

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

The paper packages a random-forest surrogate for quick rejection of tree moves inside delayed-acceptance SMC, which can cut likelihood calls in phylogenetic sampling if the predictions stay unbiased.

read the letter

The main thing to know is that this work trains a random forest on over 35 topological and branch-length features to predict the sign and size of likelihood change under eSPR and stNNI moves, then uses those predictions for early rejection in a delayed-acceptance kernel that sits inside an SMC sampler. The final acceptance still uses the true likelihood, so the target posterior is preserved in principle while many expensive evaluations are skipped. They ship a Python package and report time savings on both simulated and real data sets. That combination and the public code are the concrete advances. The approach is straightforward engineering that builds on existing delayed-acceptance ideas but applies them specifically to phylogenetic tree space with a feature set tuned to the moves that matter. The package lowers the cost of trying it out. The soft spot is the lack of visible surrogate diagnostics. The central assumption is that the forest does not systematically misclassify proposals in ways that distort the regions of tree space that reach the true-likelihood stage. If errors are correlated with topology size or branch lengths, the effective proposal distribution before correction could shift, and the SMC weights might not fully recover the correct posterior. The abstract asserts robust estimation but does not show calibration plots, confusion matrices, or direct posterior-distance comparisons to a standard run. If the full paper contains those checks and they look acceptable, the practical claim holds; otherwise the speed-up comes with an unquantified risk. This is for people who already run large-scale Bayesian phylogenetics and want to reduce wall time without rewriting their likelihood engine. A methods reader who works on surrogate or delayed-acceptance techniques will see the implementation details quickly and can test the bias question themselves. I would send it to peer review. The algorithm is clearly scoped, the code is available, and the potential payoff for routine analyses is large enough that referees should check the bias controls and run their own verification.

Referee Report

2 major / 3 minor

Summary. The paper proposes a delayed-acceptance sequential Monte Carlo (DA-SMC) sampler for Bayesian phylogenetic inference. A random forest surrogate trained on >35 topological and branch-length features predicts the sign and rough magnitude of likelihood changes under standard tree moves (eSPR, stNNI). The surrogate is used only for a first-stage rejection step inside a delayed-acceptance MCMC kernel; the true likelihood is still evaluated in the second stage. This kernel is embedded in an SMC framework to reduce the number of expensive likelihood evaluations while targeting the same posterior. Validation on simulated and real data is reported to show substantial computational savings with maintained estimation quality; an open-source Python package is provided.

Significance. If the surrogate filter does not introduce systematic bias into the effective proposal distribution, the method offers a practical route to accelerate tree-space exploration for large alignments without altering the target posterior. The explicit separation of surrogate and true-likelihood stages, together with the released code, supports reproducibility and further benchmarking.

major comments (2)

[Results] Results section: the claim that DA-SMC 'maintains robust estimation' is supported only by qualitative statements of reduced evaluations and time savings; no quantitative diagnostics (bias in posterior tree summaries, coverage probabilities, ESS ratios, or posterior-distance metrics relative to standard SMC) are supplied. Without these, it is impossible to verify that surrogate errors do not distort the stationary distribution before the second-stage correction.
[Methods] Surrogate model description (Methods): the random forest is trained on independent features and used only for preliminary rejection, yet no calibration plots, confusion matrices, or topology/branch-length stratified error rates are reported. If the surrogate systematically under-predicts positive likelihood deltas on large trees or certain topologies, the effective proposal distribution becomes biased even though the final acceptance uses the true likelihood.

minor comments (3)

[Methods] Notation for the delayed-acceptance kernel (e.g., the two-stage acceptance probability) should be written explicitly with the surrogate and true likelihood distinguished, rather than left implicit.
[Abstract and Results] The abstract and results would benefit from a short table summarizing wall-clock time, number of likelihood calls, and at least one posterior quality metric for DA-SMC versus baseline SMC on each dataset.
[Introduction] A few sentences placing the work against existing surrogate-assisted MCMC or SMC methods in phylogenetics would clarify the incremental contribution.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their thoughtful and constructive review. The major comments correctly identify areas where additional quantitative evidence and surrogate diagnostics would strengthen the manuscript. We address each point below and will incorporate the suggested revisions.

read point-by-point responses

Referee: [Results] Results section: the claim that DA-SMC 'maintains robust estimation' is supported only by qualitative statements of reduced evaluations and time savings; no quantitative diagnostics (bias in posterior tree summaries, coverage probabilities, ESS ratios, or posterior-distance metrics relative to standard SMC) are supplied. Without these, it is impossible to verify that surrogate errors do not distort the stationary distribution before the second-stage correction.

Authors: We agree that quantitative diagnostics are necessary to rigorously demonstrate that the delayed-acceptance correction preserves the target posterior. Our current validation shows comparable posterior summaries and substantial reductions in likelihood evaluations on both simulated and real data, but we did not report formal bias, coverage, or distance metrics. In the revised manuscript we will add, for the simulation experiments, (i) bias and coverage for key posterior quantities such as clade probabilities, (ii) ESS ratios between DA-SMC and standard SMC, and (iii) posterior-distance metrics (e.g., average Robinson-Foulds distance between independent runs). These additions will appear in the Results section and will be supported by the released code. revision: yes
Referee: [Methods] Surrogate model description (Methods): the random forest is trained on independent features and used only for preliminary rejection, yet no calibration plots, confusion matrices, or topology/branch-length stratified error rates are reported. If the surrogate systematically under-predicts positive likelihood deltas on large trees or certain topologies, the effective proposal distribution becomes biased even though the final acceptance uses the true likelihood.

Authors: We acknowledge that explicit surrogate performance diagnostics were not included. Although the two-stage delayed-acceptance construction guarantees that the true likelihood is used for final acceptance and therefore the target posterior remains exact, reporting surrogate accuracy is valuable for assessing efficiency and potential bias in the first-stage filter. In the revision we will add (i) calibration plots of predicted versus true likelihood change, (ii) confusion matrices for sign prediction, and (iii) error rates stratified by tree size and move type (eSPR versus stNNI). These will be placed in the Methods section or a supplementary figure. revision: yes

Circularity Check

0 steps flagged

No circularity: surrogate accelerates but final posterior uses true likelihood

full rationale

The paper trains a random forest on topological and branch-length features to predict likelihood deltas for eSPR/stNNI moves, then deploys it only for first-stage rejection inside a delayed-acceptance kernel whose second stage always evaluates the exact likelihood. Because the target stationary distribution is recovered from the true-likelihood accept/reject decisions (and the SMC weights are likewise computed with the true likelihood), the fitted surrogate never enters the final posterior by construction. Validation is performed on held-out simulations and real datasets rather than on the training data itself, and no self-citations or uniqueness theorems are invoked to justify the method. The workflow therefore remains independent of its own fitted parameters.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The approach rests on standard domain assumptions of Bayesian phylogenetics and supervised learning; no new entities are postulated and no free parameters are explicitly fitted beyond routine random-forest hyperparameters.

axioms (2)

domain assumption Standard phylogenetic tree moves (eSPR, stNNI) and the likelihood function under common substitution models are well-defined and computable.
Invoked when the surrogate is trained to predict changes caused by these moves.
domain assumption A random forest trained on hand-crafted topological and branch-length features can generalize to unseen tree proposals within the same data regime.
Central modeling assumption required for the delayed-acceptance step to be unbiased.

pith-pipeline@v0.9.0 · 5491 in / 1378 out tokens · 34176 ms · 2026-05-12T04:44:07.107670+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

49 extracted references · 49 canonical work pages

[1]

Harnessing machine learning to guide phylogenetic-tree search algorithms

Azouri, D., Abadi, S., Mansour, Y., Mayrose, I., and Pupko, T. Harnessing machine learning to guide phylogenetic-tree search algorithms. Nature Communications, 12 0 (1): 0 1983, 2021. doi:10.1038/s41467-021-22073-8. URL https://doi.org/10.1038/s41467-021-22073-8

work page doi:10.1038/s41467-021-22073-8 1983
[2]

J., Lee, A., and Drovandi, C

Bon, J. J., Lee, A., and Drovandi, C. Accelerating sequential monte carlo with surrogate likelihoods. Statistics and Computing, 31 0 (5): 0 62, 2021. doi:10.1007/s11222-021-10036-4. URL https://doi.org/10.1007/s11222-021-10036-4

work page doi:10.1007/s11222-021-10036-4 2021
[3]

Bouchard-C\^ot\'e, A., Sankararaman, S., and Jordan, M. I. P hylogenetic inference via sequential M onte C arlo. Systematic Biology, 61: 0 579--593, 2012

work page 2012
[4]

G., Barido-Sottani, J., Duch \^e ne, S., Fourment, M., Gavryushkina, A., Heled, J., Jones, G., K \"u hnert, D., De Maio, N., et al

Bouckaert, R., Vaughan, T. G., Barido-Sottani, J., Duch \^e ne, S., Fourment, M., Gavryushkina, A., Heled, J., Jones, G., K \"u hnert, D., De Maio, N., et al. BEAST 2.5: An advanced software platform for B ayesian evolutionary analysis. PLoS computational biology, 15 0 (4): 0 e1006650, 2019

work page 2019
[5]

Random forests

Breiman. Random forests. MACH LEARN, 2001,45(1) 0 (-): 0 5--32, 2001

work page 2001
[6]

Using early rejection markov chain monte carlo and gaussian processes to accelerate abc methods

Cao, X., Wang, S., and Zhou, Y. Using early rejection markov chain monte carlo and gaussian processes to accelerate abc methods. Journal of Computational and Graphical Statistics, 34 0 (2): 0 395--408, 2025. doi:10.1080/10618600.2024.2379349. URL https://doi.org/10.1080/10618600.2024.2379349

work page doi:10.1080/10618600.2024.2379349 2025
[7]

Central limit theorem for sequential M onte C arlo methods and its application to B ayesian inference

Chopin, N. Central limit theorem for sequential M onte C arlo methods and its application to B ayesian inference. The Annals of Statistics, 32 0 (6): 0 2385--2411, 2004

work page 2004
[8]

Christen, J. A. and Fox, C. Markov chain monte carlo using an approximation. Journal of Computational and Graphical Statistics, 14 0 (4): 0 795--810, 2005. doi:10.1198/106186005X76983. URL https://doi.org/10.1198/106186005X76983

work page doi:10.1198/106186005x76983 2005
[9]

Feynman-Kac Formulae: Genealogical and Interacting Particle Systems with Applications

Del Moral, P. Feynman-Kac Formulae: Genealogical and Interacting Particle Systems with Applications. Springer, New York, 2004

work page 2004
[10]

Sequential M onte C arlo samplers

Del Moral, P., Doucet, A., and Jasra, A. Sequential M onte C arlo samplers. Journal of the Royal Statistical Society: Series B (Statistical Methodology), 68 0 (3): 0 411--436, 2006

work page 2006
[11]

E., and Matsen IV, F

Dinh, V., Darling, A. E., and Matsen IV, F. A. Online B ayesian phylogenetic inference: theoretical foundations via S equential M onte C arlo. Systematic Biology, 67 0 (3): 0 503--517, 2017

work page 2017
[12]

Sequential M onte C arlo methods in practice

Doucet, A., de Freitas, N., and Gordon, N. Sequential M onte C arlo methods in practice . Springer-Verlag, New York, 2001

work page 2001
[13]

and Suchard, M

Drummond, A. and Suchard, M. Bayesian random local clocks, or one rate to rule them all. BMC biology, 8 0 (1): 0 114, 2010

work page 2010
[14]

G., Culliford, R., Medina-Aguayo, F., and Wilson, D

Everitt, R. G., Culliford, R., Medina-Aguayo, F., and Wilson, D. J. Sequential B ayesian inference for mixture models and the coalescent using sequential M onte C arlo samplers with transformations. Statistics and Computing, 30 0 (3): 0 663--676, 2020

work page 2020
[15]

Maximum likelihood and minimum-steps methods for estimating evolutionary trees from data on discrete characters

Felsenstein, J. Maximum likelihood and minimum-steps methods for estimating evolutionary trees from data on discrete characters. Systematic Biology, 22 0 (3): 0 240--249, 1973

work page 1973
[16]

Evolutionary trees from DNA sequences: a maximum likelihood approach

Felsenstein, J. Evolutionary trees from DNA sequences: a maximum likelihood approach. Journal of molecular evolution, 17 0 (6): 0 368--376, 1981

work page 1981
[17]

Foster, P. G. Modeling compositional heterogeneity. Systematic Biology, 53 0 (3): 0 485--495, 06 2004. ISSN 1063-5157. doi:10.1080/10635150490445779. URL https://doi.org/10.1080/10635150490445779

work page doi:10.1080/10635150490445779 2004
[18]

C., Dinh, V., McCoy, C., Matsen, I., Frederick, A., and Darling, A

Fourment, M., Claywell, B. C., Dinh, V., McCoy, C., Matsen, I., Frederick, A., and Darling, A. E. Effective online B ayesian phylogenetics via sequential M onte C arlo with guided proposals. Systematic biology, 2017

work page 2017
[19]

Scalable inference on K ingman's coalescent using pair similarity

G\" o r\" u r, D., Boyles, L., and Welling, M. Scalable inference on K ingman's coalescent using pair similarity. Journal of Machine Learning Research, 22: 0 440--448, 2012

work page 2012
[20]

and Teh, Y

G\" o r\" u r, D. and Teh, Y. W. An efficient sequential M onte C arlo algorithm for coalescent clustering. In Advances in Neural Information Processing Systems (NIPS), 2009

work page 2009
[21]

Efficient continuous-time M arkov chain estimation

Hajiaghayi, M., Kirkpatrick, B., Wang, L., and Bouchard-C\^ o t\' e , A. Efficient continuous-time M arkov chain estimation. In International Conference on Machine Learning (ICML), volume 31, pages 638--646, 2014

work page 2014
[22]

Holder, M. T. Dendropy: a python library for phylogenetic computing. Bioinformatics, 26 0 (12): 0 1569, 2010

work page 2010
[23]

Huelsenbeck, J. P. and Ronquist, F. MRBAYES: B ayesian inference of phylogenetic trees. Bioinformatics, 17 0 (8): 0 754--755, 2001

work page 2001
[24]

H., Cantor, C

Jukes, T. H., Cantor, C. R., et al. Evolution of protein molecules. Mammalian protein metabolism, 3 0 (21): 0 132, 1969

work page 1969
[25]

A simple method for estimating evolutionary rates of base substitutions through comparative studies of nucleotide sequences

Kimura, M. A simple method for estimating evolutionary rates of base substitutions through comparative studies of nucleotide sequences. Journal of molecular evolution, 16 0 (2): 0 111--120, 1980

work page 1980
[26]

P., Larget, B., and Ronquist, F

Lakner, C., van der Mark, P., Huelsenbeck, J. P., Larget, B., and Ronquist, F. Efficiency of M arkov chain M onte C arlo tree proposals in B ayesian phylogenetics. Syst. Biol., 57 0 (1): 0 86--103, 2008

work page 2008
[27]

J., and Suchard, M

Lemey, P., Rambaut, A., Drummond, A. J., and Suchard, M. A. Bayesian phylogeography finds its roots . PLoS Computational Biology, 5 0 (9): 0 e1000520, 2009. doi:10.1371/journal.pcbi.1000520. URL http://dx.doi.org/10.1371/journal.pcbi.1000520

work page doi:10.1371/journal.pcbi.1000520 2009
[28]

Adaptive tree proposals for bayesian phylogenetic inference

Meyer, X. Adaptive tree proposals for bayesian phylogenetic inference. Systematic Biology, 70 0 (5): 0 1015--1032, 01 2021. ISSN 1063-5157. doi:10.1093/sysbio/syab004. URL https://doi.org/10.1093/sysbio/syab004

work page doi:10.1093/sysbio/syab004 2021
[29]

Pedersen, A. G. sumt: a command-line program for computing consensus trees and other phylogenetic tree summaries , November 2023. URL https://github.com/agormp/sumt

work page 2023
[30]

H., Chan, L., Dominus, M

Piel, W. H., Chan, L., Dominus, M. J., Ruan, J., Vos, R. A., and Tannen, V. Treebase v. 2: A database of phylogenetic knowledge. In Proceedings of the e-BioSphere 2009, London, UK, 2009. Available at https://treebase.org

work page 2009
[31]

and Yang, Z

Rannala, B. and Yang, Z. Probability distribution of molecular evolutionary trees: a new method of phylogenetic inference. J. Mol. E, 43: 0 304--311, 1996

work page 1996
[32]

Robinson, D. F. and Foulds, L. R. Comparison of weighted labelled trees. In Horadam, A. F. and Wallis, W. D., editors, Combinatorial Mathematics VI, pages 119--126, Berlin, Heidelberg, 1979. Springer Berlin Heidelberg

work page 1979
[33]

and Foulds, L

Robinson, D. and Foulds, L. Comparison of phylogenetic trees. Mathematical Biosciences, 53: 0 131--147, 1981

work page 1981
[34]

L., Marin, A., and Medina, J

Rodriguez, F., Oliver, J. L., Marin, A., and Medina, J. R. The general stochastic model of nucleotide substitution. Journal of theoretical biology, 142 0 (4): 0 485--501, 1990

work page 1990
[35]

and Huelsenbeck, J

Ronquist, F. and Huelsenbeck, J. P. M r B ayes 3: B ayesian phylogenetic inference under mixed models. Bioinformatics, 19 0 (12): 0 1572--1574, 2003

work page 2003
[36]

L., Darling, A., Hohna, S., Larget, B., Liu, L., Suchard, M

Ronquist, F., Teslenko, M., van der Mark, P., Ayres, D. L., Darling, A., Hohna, S., Larget, B., Liu, L., Suchard, M. A., and Huelsenbeck, J. P. Mr B ayes 3.2: Efficient B ayesian phylogenetic inference and model choice across a large model space. Syst. Biol., 61: 0 539--542, 2012

work page 2012
[37]

A., Ionides, E

Smith, R. A., Ionides, E. L., and King, A. A. Infectious disease dynamics inferred from genetic data via sequential M onte C arlo. Molecular biology and evolution, 34 0 (8): 0 2065--2084, 2017

work page 2065
[38]

Suchard, M. A. and Redelings, B. D. BA li- P hy: simultaneous B ayesian inference of alignment and phylogeny. Bioinformatics, 22 0 (16): 0 2047--2048, 2006

work page 2047
[39]

A., Lemey, P., Baele, G., Ayres, D

Suchard, M. A., Lemey, P., Baele, G., Ayres, D. L., Drummond, A. J., and Rambaut, A. Bayesian phylogenetic and phylodynamic data integration using BEAST 1.10. Virus evolution, 4 0 (1): 0 vey016, 2018

work page 2018
[40]

W., Daum\'e III , H., and Roy, D

Teh, Y. W., Daum\'e III , H., and Roy, D. M. B ayesian agglomerative clustering with coalescents. In Advances in Neural Information Processing Systems (NIPS), 2008

work page 2008
[41]

Fox, Martin B

Wang, L., Bouchard-C\^ot\'e, A., and Doucet, A. Bayesian phylogenetic inference using a combinatorial sequential M onte C arlo method. Journal of the American Statistical Association, 110 0 (512): 0 1362--1374, 2015. doi:10.1080/01621459.2015.1054487. URL http://dx.doi.org/10.1080/01621459.2015.1054487

work page doi:10.1080/01621459.2015.1054487 2015
[42]

An annealed sequential M onte C arlo method for B ayesian phylogenetics

Wang, L., Wang, S., and Bouchard-C \^o t \'e , A. An annealed sequential M onte C arlo method for B ayesian phylogenetics. Systematic Biology, 69 0 (1): 0 155--183, 2020

work page 2020
[43]

and Wang, L

Wang, S. and Wang, L. Particle gibbs sampling for bayesian phylogenetic inference. Bioinformatics, 37 0 (5): 0 642--649, 10 2020. ISSN 1367-4803. doi:10.1093/bioinformatics/btaa867. URL https://doi.org/10.1093/bioinformatics/btaa867

work page doi:10.1093/bioinformatics/btaa867 2020
[44]

and Zhang, C

Xie, T. and Zhang, C. ART ree: A deep autoregressive model for phylogenetic inference. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. URL https://openreview.net/forum?id=SoLebIqHgZ

work page 2023
[45]

A., and Zhang, C

Xie, T., Richman, H., Gao, J., Matsen Iv, F. A., and Zhang, C. Phylovae: Unsupervised learning of phylogenetic trees via variational autoencoders. 2025

work page 2025
[46]

Improved variational bayesian phylogenetic inference with normalizing flows

Zhang, C. Improved variational bayesian phylogenetic inference with normalizing flows. In Proceedings of the 34th International Conference on Neural Information Processing Systems, NIPS '20, Red Hook, NY, USA, 2020. Curran Associates Inc. ISBN 9781713829546

work page 2020
[47]

and IV, F

Zhang, C. and IV, F. A. M. Variational bayesian phylogenetic inference. In International Conference on Learning Representations, 2019. URL https://openreview.net/forum?id=SJVmjjR9FX

work page 2019
[48]

and IV, F

Zhang, C. and IV, F. A. M. A variational approach to bayesian phylogenetic inference. Journal of Machine Learning Research, 25 0 (145): 0 1--56, 2024. URL http://jmlr.org/papers/v25/22-0348.html

work page 2024
[49]

M., and Aston, J

Zhou, Y., Johansen, A. M., and Aston, J. A. Toward automatic model comparison: an adaptive sequential M onte C arlo approach. Journal of Computational and Graphical Statistics, 25 0 (3): 0 701--726, 2016

work page 2016