RAxML-NG

What is RAxML-NG?

RAxML-NG infers phylogenetic trees from an existing multiple sequence alignment using maximum likelihood. It is the current-generation rewrite of RAxML and ExaML, built for faster likelihood calculations, more stable optimization, and large alignments where branch support still matters.

The practical point of RAxML-NG is not just tree building. It is tree building under explicit evolutionary models, with branch lengths, likelihood scores, and bootstrap support that can stand up in downstream comparative analyses. That makes it a better fit than approximate methods when the topology will be interpreted biologically or used in a manuscript.

ProteinIQ runs the verified RAxML-NG 2.0.2 release. This release includes fixes for automatic model selection during bootstrapping, bootstrap support metrics, and alignment-format detection that are not present in the older 2.0.0 binary.

This tool expects an alignment, not raw sequences. If the sequences are not already aligned, start with MAFFT, MUSCLE5, or Clustal Omega before running tree inference.

How does RAxML-NG work?

RAxML-NG searches tree space for the topology and branch lengths that maximize the likelihood of the observed alignment under a substitution model. In practice, that means comparing many candidate trees, optimizing branch lengths and model parameters, and keeping the trees that improve the score.

The rewrite improved several pieces of the original RAxML search procedure. The published implementation fixes missed topological moves from earlier versions, improves optimization for models such as LG4X, adds transfer bootstrap expectation (TBE), and reports terraces in tree space when the data structure implies many equally scoring topologies. It also incorporates site-repeat optimizations and parallelization improvements that matter on taxon-rich datasets.

One consequence of maximum likelihood inference is that alignment quality dominates the final result. RAxML-NG can optimize a tree very efficiently, but it cannot rescue a poor alignment, mixed paralogs, or a nucleotide alignment forced into a protein model.

How to use RAxML-NG online

RAxML-NG runs on ProteinIQ from a pre-aligned protein, DNA, or binary character dataset in FASTA or PHYLIP format, or use the native parser for CATG, VCF, compressed variant files, and RBA input. Submit the alignment, choose whether to run ML search, bootstrapping, or both, and the job returns every native result file written by RAxML-NG, including Newick trees, model information, binary alignment/checkpoint data, reduced alignments when needed, and the execution log.

Inputs

Input	Description
Alignment	One aligned protein, DNA, or binary (0/1) dataset in FASTA, PHYLIP, or plain text containing one of those formats. At least 4 sequences with unique, Newick-compatible taxon names are required.
Model file	Required in partition model mode. Upload native model and site-range definitions as .txt.
Constraint tree	Optional Newick topology constraint in .newick or .txt, for search or combined mode.
Site weights	Optional .txt file of positive integer alignment-column weights.
Job name	Optional label for the submitted job in the ProteinIQ interface.

Each uploaded file can contain up to 50 MiB. Standard FASTA uses the taxon name before the first space in its header; the rest is a description. Long-label FASTA mode replaces spaces with underscores. Submitted file bytes are retained.

For Native alignment parser, choose CATG, VCF, or RBA in Alignment parser. Upload CATG as .txt, VCF as .vcf, compressed variant bytes as .gz, and binary alignments as .rba. Choose the sequence type or a custom/partition model explicitly; use Model embedded in RBA for an RBA file. RAxML-NG checks these inputs before inference and requires at least four taxa. Native-check files are also available to download.

Settings

Setting	Description
Analysis mode	ML tree search, Bootstrapping only, or ML search + bootstrapping. ML search is the native default.
Sequence type	Auto-detect from alignment, Protein, DNA, Binary (0/1), or explicit multistate/10-state diploid data with a custom model. A custom GT16 model can infer its data type with automatic sequence type. Auto-detection recognizes binary matrices directly when their observed states are 0 and 1. Short nucleotide alignments made only of A, C, G, and N may need manual selection.
Model mode	Automatic (AA, DNA, or BIN) asks RAxML-NG to select a model within the chosen family. Custom model string passes an explicit model such as GTR+G, LG+G8+F, or BIN+FO. Partition mode reads the supplied model file; embedded mode reads the RBA model. The result summary reports the effective optimized model from the native bestModel file.
Custom model	Required only when Model mode is set to custom. Useful when a manuscript or lab workflow needs a specific model string instead of the family default.
Starting tree	auto, adaptive, parsimony/random selectors, or a comma-separated mixture such as pars{10},rand{10}. Native consensus starts such as cons:MR are also supported; starting-tree file paths are excluded. pars{10} starts from 10 parsimony trees, rand{10} starts from 10 random trees. This only affects ML search workflows.
Random seed	Optional positive integer used for reproducibility. Leave blank to preserve the native current-time seed.
Bootstrap trees	The native default is autoMRE{1000}, which checks convergence and can stop before 1,000 trees. Enter a fixed integer from 1 to 1,000 or native per-metric counts such as fbp{100},sh{100}. Explicit per-metric counts must be unique and total at most 1,000.
Bootstrap method	Bootstrap-only mode accepts FBP, RBS, PS, or PBS. TBE requires a reference tree and is therefore available only in the combined ML search + bootstrapping mode.
Bootstrap support metric	Combined mode accepts comma-separated fbp, tbe, rbs, ps, pbs, ebg, sh, ic1, and ica. FBP and RBS cannot be combined. Each requested native support tree is retained.
Output prefix	Optional prefix for native .raxml.* filenames. Leave blank to derive it from the submitted alignment name.

Native controls

Collapsed option groups expose the native controls below. Native default or an empty optional numeric/text field leaves that argument out of the command. Automatic model selection requires model and branch optimization to remain enabled.

Setting	Description
Alignment input	Select standard FASTA/PHYLIP checks or the native alignment parser.
Alignment parser	Select auto, FASTA, long-label FASTA, PHYLIP, CATG, VCF, or RBA. Auto leaves native format detection unchanged.

Model and branch options

Setting	Description
Optimize model	Native default: on. Fixed models can disable optimization.
Optimize branches	Native default: on.
Partition branch linkage	Native default: scaled.
Minimum branch length	Native default: 1e-6 substitutions per site.
Maximum branch length	Native default: 100 substitutions per site.
Branch optimizer	Select the native branch-length optimizer.
Likelihood epsilon	Native default: 10 log-likelihood units for search/bootstrap.
Triplet likelihood epsilon	Native default: 1000 log-likelihood units.
FreeRate optimizer	Native default: auto.
Model selection options	Native options separated by /, for example criterion=AIC/rhas=G/substitution-models=GTR. Omit to preserve native defaults.
Model selection report	Native model-selection report format when emitted.

Topology search options

Setting	Description
Topology optimizer	Native default: adaptive.
Adaptive search	Native default is unchanged. Start uses adaptive starting-tree selection only.
SPR radius	Native default: auto.
SPR likelihood cutoff	A non-negative real number or off. Native default: 1.0.
Search stopping rule	Use the native SPR stopping criterion.
NNI tolerance	Native NNI tolerance.
NNI epsilon	Native NNI likelihood epsilon.
Pythia parsimony trees	Number of native parsimony trees used for alignment difficulty prediction.
Parsimony SPR	Native parsimony starting-tree SPR refinement.
Initial branch lengths	Native initialization of branch lengths.
Alignment difficulty prediction	Native Pythia control within tree inference.

Bootstrap and support options

Setting	Description
MRE convergence cutoff	Native default: 0.03.
Download bootstrap alignments	Preserve every native bootstrap alignment as a downloadable file.
SH-aLRT replicates	Native default: 1000. Hosted maximum: 1000.
SH-aLRT epsilon	Native default: 0.1 log-likelihood units.
Bootstrap starting trees	Native bootstrap starting-tree strategy.

Alignment and output options

Setting	Description
Probabilistic alignment	Forward native probabilistic CATG/VCF handling.
Pattern compression	Native default: on.
Site repeats	Native default: on.
Tip-inner optimization	Native default: off.
Rate-category scalers	Leave native automatic scalers unchanged unless explicitly selected.
Output decimal places	Native default: 6.
Outgroup taxa	Comma-separated native taxon names. Changes the tree drawing/rooting only, without rerunning inference.
Fully undetermined taxa	Native treatment of fully undetermined taxa. Default preserves the current early error.
Duplicate sequences	Native treatment of identical sequences; this does not permit duplicate taxon names.

Outputs

Output file	When it appears	Meaning
prefix.raxml.bestTree	ML search or combined mode	Best maximum-likelihood tree in Newick format.
prefix.raxml.bestTreeCollapsed	ML search or combined mode	Best tree with near-zero branches collapsed.
prefix.raxml.startTree	ML search or combined mode	Starting trees generated or used by the search. This can contain more than one tree.
prefix.raxml.mlTrees	ML search or combined mode	Candidate ML trees from the starting-tree searches.
prefix.raxml.bootstraps	Bootstrap or combined mode	Complete bootstrap replicate tree collection. The result reports the number actually generated when automatic stopping is used.
prefix.raxml.support or metric-specific support tree	Combined mode	Final support-annotated tree, including suffixes such as supportTBE, supportRBS, supportPS, or supportPBS.
prefix.raxml.bestModel	When emitted by RAxML-NG	Optimized model parameters and the effective model reported in the result summary.
prefix.raxml.rba	When emitted by RAxML-NG	Native binary alignment used for efficient reuse or diagnosis.
prefix.raxml.reduced.phy and prefix.raxml.reduced.partition	When RAxML-NG removes fully undetermined columns	Reduced alignment and matching partition definition.
prefix.raxml.ckp	When retained by RAxML-NG	Native checkpoint for resuming or diagnosing an analysis.
prefix.raxml.log	When emitted, including failed runs	Full native log, including final log-likelihood, AIC, and BIC when available.

Failed or timed-out runs retain the native log, any partial/checkpoint files, and captured command output as diagnostic downloads. They do not produce a successful scientific result.

The tree viewer opens only a single final tree. Bootstrap-only runs return a multi-tree collection and are shown as downloadable files instead of being misread as one tree.

Understanding the results

RAxML-NG results are easiest to interpret in three layers: topology, branch lengths, and branch support.

Topology

The branching pattern is the inferred evolutionary hypothesis. Internal nodes define clades. A change in topology is biologically meaningful, so any region of the tree with weak support should be treated as unresolved rather than overinterpreted.

Branch lengths

Branch lengths are reported in substitutions per site. Longer branches indicate more inferred evolutionary change. Extremely long terminal branches often signal problematic sequences, alignment issues, contamination, or highly divergent taxa.

Support values

Branch support is available in combined mode. FBP, RBS, TBE, PS, and PBS use resampled data or parsimony trees; EBG predicts bootstrap support and SH reports an SH-like aLRT. IC1 and ICA quantify internode certainty. These metrics have different meanings and should not be treated as interchangeable. The percentage conventions below concern bootstrap support, rather than internode certainty.

Support value	Interpretation
>= 95	Strong support for that clade in most datasets
80-94	Reasonable support, often usable but worth checking against alignment quality
70-79	Weak to moderate support, commonly treated with caution
< 70	Unstable split, usually not strong evidence for that relationship

Those cutoffs are conventions, not laws. Deep trees with uneven taxon sampling can show lower classical bootstrap support even when the overall signal is meaningful. That is one reason TBE exists.

FBP vs TBE vs RBS

Metric	What it emphasizes	When it is useful
FBP	Exact recovery of the same bipartition across replicates	Standard reporting and direct comparison with older phylogenetics literature
TBE	Similarity between splits rather than exact matching	Large trees where a few unstable taxa would otherwise deflate support for deep branches
RBS	Rapid bootstrap workflow	Faster support estimation when runtime matters more than strict comparability with classical bootstrap

TBE is not a valid metric for bootstrap-only generation in RAxML-NG 2.0.2. Use the combined mode when TBE support is required.

When to use RAxML-NG vs alternatives

RAxML-NG occupies the middle ground between very fast approximate tree builders and feature-rich ML packages with their own model-selection ecosystems.

Tool	Best use case	Tradeoff
RAxML-NG	Maximum-likelihood inference when a solid ML search and standard bootstrap workflow are the priority	Requires a pre-aligned dataset and is slower than approximate methods
IQ-TREE	Analyses where built-in model selection and ultrafast bootstrap are central	Different search heuristics and support framework, often more feature-focused on model testing
FastTree	Very large alignments where speed matters more than exact ML optimization	Much faster, but approximate

RAxML-NG is a strong default when the alignment is already prepared and the goal is a conventional ML tree with explicit bootstrap support. IQ-TREE is often the better choice when systematic model testing is part of the analysis plan. FastTree is better for exploratory work on very large datasets where approximate topology is good enough.

Related tools

• MAFFT: Builds multiple sequence alignments before phylogenetic inference.
• MUSCLE5: Alternative MSA method for protein or nucleotide datasets.
• Clustal Omega: Scalable alignment tool that fits naturally before tree building.
• IQ-TREE: Another maximum-likelihood phylogeny tool with strong model-selection features.
• FastTree: Approximate maximum-likelihood trees for very large alignments.

Based on the RAxML-NG paper by Kozlov et al. (2019) in Bioinformatics and the official RAxML-NG source repository.