Self-Hosted Phylogenetic Tree Inference: IQ-TREE 2 vs RAxML-NG vs BEAST 2 vs MrBayes

Phylogenetic trees — the branching diagrams depicting evolutionary relationships among species, genes, or populations — are the lingua franca of evolutionary biology. From tracking viral outbreaks to resolving deep branches of the tree of life, phylogenetics underpins our understanding of how life diversified over 4 billion years.

Modern phylogenetic inference splits into two philosophical camps: maximum likelihood (find the single best tree under an evolutionary model) and Bayesian inference (sample from the posterior distribution of trees). Four open-source tools — IQ-TREE 2, RAxML-NG, BEAST 2, and MrBayes — represent the state of the art, each with distinct strengths optimized for different research questions.

The Two Paradigms of Phylogenetic Inference

Maximum Likelihood (ML)

ML methods search tree space for the topology and branch lengths that maximize the probability of observing the sequence alignment under a specified substitution model (e.g., GTR+G+I). ML is computationally efficient, scales to thousands of taxa, and produces a single best tree with bootstrap support values.

IQ-TREE 2 and RAxML-NG are the two dominant ML implementations, each with unique optimizations.

Bayesian Inference (BI)

BI methods sample from the posterior probability distribution of trees given the data and prior beliefs. Rather than producing one “best” tree, they generate a set of credible trees with posterior probability support values. BI is more computationally intensive but provides richer uncertainty quantification.

BEAST 2 and MrBayes are the leading Bayesian packages, differentiated by their focus on time-calibrated phylogenies and model flexibility respectively.

Tool-by-Tool Analysis

IQ-TREE 2: The Swiss Army Knife of ML Phylogenetics

IQ-TREE 2, developed at the University of Vienna, has rapidly become the most popular ML phylogenetics tool with over 330 GitHub stars. Its killer feature is ModelFinder — automatic model selection that tests 286 substitution models in seconds, choosing the best-fit model by BIC before tree inference begins.

Key innovations:

• Ultrafast bootstrap approximation (UFBoot) — 10-40x faster than standard bootstrap with comparable accuracy
• Single branch tests (aLRT, SH-aLRT) for rapid branch support assessment
• Partitioned analyses allowing different genes to evolve under different models
• Built-in tree topology tests (AU test, KH test, SH test)
• Native support for phylogenomic datasets with hundreds of loci

Installation:

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11

# Conda (easiest)
conda create -n iqtree-env -c bioconda iqtree
conda activate iqtree-env

# Download binary (Linux)
wget https://github.com/iqtree/iqtree2/releases/download/v2.3.6/iqtree-2.3.6-Linux-intel.tar.gz
tar -xzf iqtree-2.3.6-Linux-intel.tar.gz
export PATH=$PATH:$(pwd)/iqtree-2.3.6-Linux-intel/bin

# Docker
docker pull quay.io/biocontainers/iqtree:2.3.6--h5b5514e_0

Example command:

1
2
3
4
5

# Auto model selection + ML tree + ultrafast bootstrap
iqtree2 -s alignment.phy -m MFP -B 1000 -T 16

# Partitioned analysis
iqtree2 -s concatenated.phy -p partition.nex -m MFP+MERGE -B 1000 -T 32

RAxML-NG: Battle-Tested Performance at Scale

RAxML-NG (Next Generation), maintained by Alexey Kozlov at the Heidelberg Institute for Theoretical Studies, represents a complete rewrite of the classic RAxML codebase. With over 460 GitHub stars, it excels at large-scale phylogenomic analyses with hundreds to thousands of taxa.

Differentiating features:

• Superior parallelization — near-linear scaling to 64+ cores
• Efficient checkpointing — resume interrupted analyses from exact stopping point
• Flexible partition models with automatic merging of similar partitions
• Robust handling of alignment gaps and missing data
• Tree inference from a starting parsimony tree, avoiding the need for separate starting tree generation

Installation:

1
2
3
4
5
6
7
8

# Conda
conda create -n raxml-env -c bioconda raxml-ng
conda activate raxml-env

# Source build
git clone --recursive https://github.com/amkozlov/raxml-ng.git
cd raxml-ng && mkdir build && cd build
cmake .. && make -j16

Example command:

1
2
3
4
5

# Standard ML search with bootstrapping
raxml-ng --msa alignment.phy --model GTR+G --threads 32

# All-in-one (model test + ML search + bootstrapping)
raxml-ng --all --msa alignment.phy --model GTR+G --threads 32 --bs-trees 1000

BEAST 2: Time-Calibrated Bayesian Phylogenetics

BEAST 2 (Bayesian Evolutionary Analysis Sampling Trees), from the University of Auckland, uniquely integrates temporal information — fossil calibrations, tip dates, and sampling times — directly into phylogenetic inference. With over 240 GitHub stars, it’s the premier tool for estimating divergence times from molecular data.

Distinctive capabilities:

• Simultaneous estimation of tree topology, branch lengths, substitution parameters, and divergence times
• Flexible clock models (strict, uncorrelated lognormal, random local)
• Coalescent-based phylodynamic models for viral outbreak reconstruction
• Structured coalescent for phylogeography (tracing spatial spread)
• Extensive plugin system (BEAST 2 packages) covering discrete trait evolution, species tree inference, and more
• Beautiful interactive visualization via BEAUti and Tracer

Installation:

1
2
3
4
5
6
7
8
9

# Download from official site
wget https://github.com/CompEvol/beast2/releases/download/v2.7.7/BEAST.v2.7.7.Linux.tgz
tar -xzf BEAST.v2.7.7.Linux.tgz
cd beast

# Install common packages
./bin/packagemanager -add BEASTLabs
./bin/packagemanager -add SNAPP
./bin/packagemanager -add bModelTest

Example XML (generated via BEAUti GUI): BEAST 2 uses an XML input format generated by its companion GUI, BEAUti, which specifies the sequence alignment, substitution model, clock model, tree prior, and MCMC settings. For server deployments without a GUI, XML files can be generated locally and transferred.

MrBayes: Flexible Model Specification

MrBayes, maintained by the Swedish National Bioinformatics Infrastructure, is the original Bayesian phylogenetics powerhouse with nearly 270 GitHub stars. Its strength lies in its unparalleled model flexibility — you can specify different substitution models, rate partitions, and constraints with a simple command-block syntax.

Key strengths:

• Mixed models — combine different substitution models across data partitions
• Flexible prior specification for parameters and topologies
• Posterior predictive simulation for model adequacy testing
• Stepping-stone sampling for rigorous model comparison via marginal likelihood
• Robust handling of ambiguous characters and alignment uncertainty

Installation:

1
2
3
4
5
6

# Conda
conda create -n mrbayes-env -c bioconda mrbayes
conda activate mrbayes-env

# Docker
docker pull quay.io/biocontainers/mrbayes:3.2.7--h4ac6f70_0

Example MrBayes block (in NEXUS file):

1
2
3
4
5
6

begin mrbayes;
  set autoclose=yes nowarn=yes;
  lset nst=6 rates=invgamma;
  mcmc ngen=10000000 samplefreq=1000 nchains=4;
  sumt burnin=2500;
end;

Comprehensive Comparison Table

Self-Hosted Deployment for Phylogenetics Labs

A dedicated phylogenetics server typically pairs ML and Bayesian tools on the same machine, choosing between them based on the specific research question:

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35

# docker-compose.yml for phylogenetics compute server
version: '3.8'
services:
  iqtree:
    image: quay.io/biocontainers/iqtree:2.3.6--h5b5514e_0
    volumes:
      - ./data:/data
      - ./results:/results
    working_dir: /data
    entrypoint: ["iqtree2"]

  raxml-ng:
    image: quay.io/biocontainers/raxml-ng:1.2.2--h4ac6f70_0
    volumes:
      - ./data:/data
      - ./results:/results
    working_dir: /data
    entrypoint: ["raxml-ng"]

  mrbayes:
    image: quay.io/biocontainers/mrbayes:3.2.7--h4ac6f70_0
    volumes:
      - ./data:/data
      - ./results:/results
    working_dir: /data
    stdin_open: true
    tty: true

  # Supplementary: visualization server
  figtree-server:
    image: nginx:alpine
    ports:
      - "8080:80"
    volumes:
      - ./results:/usr/share/nginx/html/results

For compute-intensive Bayesian analyses (BEAST 2, MrBayes) that may run for days or weeks, integration with HPC workload managers is essential. A Slurm submission script can launch multiple independent MCMC chains in parallel:

1
2
3
4
5
6
7
8
9

#!/bin/bash
#SBATCH --job-name=beast_tree
#SBATCH --time=168:00:00
#SBATCH --cpus-per-task=8
#SBATCH --mem=64G
#SBATCH --array=1-4

cd /data/beast_run_${SLURM_ARRAY_TASK_ID}
beast -seed ${RANDOM} -threads 8 analysis.xml

Why Self-Host Phylogenetic Inference Pipelines?

Computational demands require dedicated hardware. Phylogenetic inference at scale is computationally intensive — a Bayesian analysis of 200 taxa under a complex partitioned model with 10 million MCMC generations can consume 1,000+ CPU-hours. Cloud-based solutions charge premium rates for sustained compute, making a dedicated on-premises server with 32+ cores dramatically cheaper for any lab running more than one analysis per month. The amortized cost favors self-hosting within 6-12 months for active phylogenetics groups.

Model control and method transparency. The choice of substitution model (GTR, HKY, WAG, LG) and rate heterogeneity parameters directly affects phylogenetic conclusions. Many cloud-hosted tools obscure model selection behind “smart defaults” that may be inappropriate for non-model organisms or rapidly evolving viral sequences. Self-hosting lets you explicitly specify models, examine likelihood scores at each step, and document the full analytical pipeline from raw alignment to final tree — essential for reproducible research practices.

Integration with genomic data pipelines. Phylogenetic trees are rarely the final output — they feed into downstream analyses including ancestral state reconstruction, gene family evolution, and comparative genomics. Self-hosted pipelines integrate naturally with scientific data management systems that track provenance from raw sequencing reads through alignment to final tree. This is particularly important for collaborative projects where multiple labs contribute data.

Custom model implementation. Both BEAST 2 and MrBayes support user-defined substitution models and tree priors via scripting. For researchers studying organisms with unusual evolutionary dynamics — codon-position-specific rates in RNA viruses, strand-specific substitution bias in mitochondrial genomes, or heterotachy in deep phylogenetics — custom models are essential and only available through self-hosted deployments.

Sensitive data handling. Phylogenetic analyses of pathogens (viral outbreaks, antimicrobial resistance surveillance) often involve sequences from clinical samples with privacy implications. Self-hosting ensures that pathogen genomic data remains within institutional security boundaries, compliant with public health data governance requirements. For viral outbreak analysis, BEAST 2’s phylodynamic models combined with self-hosted infrastructure have been used by major public health agencies to track transmission chains without exposing patient-associated sequence data to external services.

Choosing the Right Tool for Your Research Question

FAQ

Maximum likelihood or Bayesian — which is better?

Neither is categorically better — they answer different questions. ML produces a point estimate (best tree) with bootstrap support, suitable for hypothesis testing and exploratory analysis. Bayesian inference produces a posterior distribution, useful when quantifying uncertainty matters (divergence time estimation, ancestral state reconstruction). In practice, many researchers run both on the same dataset — ML for a quick best tree, Bayesian for rigorous uncertainty quantification and time calibration.

How many bootstrap replicates do I need?

For ML analysis with IQ-TREE 2, 1,000 ultrafast bootstrap replicates are standard and computationally cheap. For RAxML-NG with standard bootstrapping, 100-500 replicates are common given the higher computational cost. Bayesian posterior probabilities from BEAST 2 or MrBayes do not require replicates — they are derived from the MCMC sample directly.

Can I run BEAST 2 on a headless server?

Yes, but the primary interface (BEAUti) requires a GUI for setting up analyses. Workaround: generate the XML file on a local machine with BEAUti, then transfer to the server for headless execution. Alternatively, BEAST 2 XML can be written programmatically — several R packages (babette, beautier) provide programmatic interfaces for generating BEAST 2 input files.

How do I know if my MCMC chain has converged?

Check ESS (Effective Sample Size) values > 200 for all parameters, and examine trace plots for mixing. Tracer (from the BEAST 2 ecosystem) is the standard tool for MCMC diagnostics. For MrBayes, the .pstat and .tstat output files contain convergence diagnostics. Run at least two independent chains and compare — their posterior probability estimates should agree within 5%.

How long will my analysis take?

IQ-TREE 2 with UFBoot on 200 taxa: ~2-4 hours on 16 cores. RAxML-NG on 1,000 taxa: ~8-12 hours. BEAST 2 on 100 taxa with molecular clock: ~24-72 hours for 100 million generations. MrBayes on 200 taxa: ~48-168 hours for 10 million generations. All times scale with alignment length, model complexity, and number of taxa.

Can I combine results from ML and Bayesian analyses?

Yes — and this is recommended. Present the ML tree (from IQ-TREE 2 or RAxML-NG) as your primary result with bootstrap support values, and the Bayesian consensus tree (from BEAST 2 or MrBayes) with posterior probabilities as a supplementary figure. Agreement between ML and Bayesian topologies strengthens confidence in evolutionary relationships; disagreement flags branches that deserve further investigation with additional data or alternative models.

💰 想测试你的市场判断力？我用 Polymarket 做预测市场交易——这是全球最大的预测市场平台，从大选结果到技术监管时间线，什么都可以押注。和赌博不同，这是真正的信息市场：你懂的信息越多，胜率越高。我靠预测技术相关事件的走向已经赚了不少。用我的邀请链接注册：Polymarket.com