IQ-TREE
v3.1.1Build maximum likelihood phylogenetic trees with automatic model selection and ultrafast bootstrap.
Build maximum likelihood phylogenetic trees with automatic model selection and ultrafast bootstrap.
IQ-TREE online on ProteinIQ is an IQ-TREE web server that runs IQ-TREE 3.1.1, a release of IQ-TREE 3 (Wong et al., 2025). Upload an aligned FASTA, PHYLIP, NEXUS, CLUSTAL or MSF file with at least 3 sequences, and at the default settings ModelFinder chooses a substitution model by BIC, IQ-TREE searches for the maximum likelihood tree, and the tree opens in a viewer above a run summary with the model, log-likelihood, AIC, BIC and tree length.
Ultrafast bootstrap and SH-aLRT add branch support. Advanced native options reach the rest of IQ-TREE: partitioned analyses, topology tests, concordance factors, ancestral reconstruction, likelihood mapping, tree comparisons and AliSim simulation. The Files tab holds every file the analysis writes, including the .iqtree report, the .treefile, logs and compressed checkpoints.
IQ-TREE needs aligned sequences. If yours are unaligned, build a multiple sequence alignment first with MAFFT or Clustal Omega.
An IQ-TREE job costs at least 20 credits. The quote counts the alignment characters in the input, gaps included and FASTA headers and whitespace excluded, and charges one credit per 2,000 characters, rounded up, once that passes 20. The exact quote appears before submission.
| Sequences | 300 columns | 1,000 columns | 5,000 columns | 20,000 columns |
|---|---|---|---|---|
| 10 | 20 | 20 | 25 | 100 |
| 50 | 20 | 25 | 125 | 500 |
| 100 | 20 | 50 | 250 | 1,000 |
| 500 | 75 | 250 | 1,250 | 5,000 |
| 1,000 | 150 | 500 | 2,500 | 10,000 |
Each cell is a quote from the current calculator for one FASTA alignment with the given number of sequences and alignment columns. Any alignment up to 40,000 characters, such as 40 sequences of 1,000 columns, costs 20 credits. In PHYLIP, NEXUS and other formats without > headers, sequence names count as characters too, and text in other inputs, such as trees or partition files, adds to the count.
Bootstrap and SH-aLRT replicates, the substitution model and advanced native options do not change the quote, though they change how long the run takes. A batch holds up to 10 alignments, and each is quoted and run as its own job.
| Input | Accepted | Limits |
|---|---|---|
Alignment | FASTA, PHYLIP, NEXUS, CLUSTAL or MSF, or a PoMo .counts file; pasted or uploaded | 50 MiB per file; gzip-compressed .gz uploads accepted |
| Free plan | Same formats | Up to 10,000 sequence characters per job, gaps and headers excluded |
| Paid plans | Same formats | No plan cap on alignment size |
Every alignment needs at least 3 sequences, and ultrafast bootstrap needs at least 4. Some inputs are refused before submission with a message that explains the fix: FASTA records of unequal length (the sequences are not aligned), repeated sequence names, header-only FASTA records, BLAST hit tables, copied NCBI page text, NEXUS files without a DATA or CHARACTERS block, and alignment rows without a format header. CLUSTAL files written by aligners that cut names to 10 characters often trip the repeated-name check; FASTA keeps full names.
| Input | Role name | Typical use |
|---|---|---|
Additional alignments | additional-alignments | Up to 10 more loci for a partitioned analysis |
Partition definition | partition | A NEXUS or RAxML-style partition file (.nex, .nexus, .txt) |
Trees | trees | Up to 10 tree files to evaluate, constrain or compare |
Model definition | model-definition | Custom substitution models in NEXUS |
Site frequency profiles | site-frequencies | .sitefreq files for site-heterogeneous mixture models |
Tip or node dates | dates | Dates for time-tree dating (.txt, .csv, .tsv) |
Likelihood-mapping clusters | quartet-clusters | Groups for likelihood mapping |
Gene trees | gene-trees | Gene trees for concordance factors |
Other files | assets | Up to 10 extra files referenced by other options, including binary model assets |
Each file has a 50 MiB limit. Reference an uploaded file in Advanced native options with @role, or @role/filename when the role holds several files. File references also work for pasted inputs when the role contains one item. Filenames must be single names, and two different files with the same name must be renamed before upload.
For native directory inputs, use @role/*. Upload several files in Additional alignments and use ["-p", "@additional-alignments/*"] to let IQ-TREE build partitions from those files. File contents and names are preserved.
Tree-only analyses can omit the alignment. Upload a tree file and use ["-t", "@trees", "--tree-dist-all"]. For AliSim, choose a suitable custom model and use ["--alisim", "OUTPUT", "-t", "@trees", "--length", "100"], where OUTPUT selects the managed output prefix. To use a supplied root sequence, use ["--root-seq", "@alignment,sequence-name"].
| Parameter | Type | Default | Description |
|---|---|---|---|
Sequence type | enum | Auto-detect | Protein, Nucleotide (DNA/RNA), Codon, Binary, Morphology, Translate DNA to protein or Native sequence type (native -st). |
Native sequence type | string | conditional | Required with Native sequence type: a genetic code or morphology specification such as CODON2 or MORPH{10}. |
Substitution model | enum | default (auto ModelFinder) | A model-selection mode, a named model, or Custom native model (native -m); the choices are listed below. |
Custom model | string | conditional | Required with Custom native model: an unchanged IQ-TREE model string such as GTR+F+I+G4 or LG+C20+F+G. |
Ultrafast bootstrap replicates | integer | 0 | 0 disables UFBoot; nonzero values must be 1,000 to 10,000 (native -B). |
SH-aLRT replicates | integer | 0 | 0 disables SH-aLRT; nonzero values must be 1,000 to 10,000 (native --alrt). |
Random seed | integer | optional | 1 to 2,147,483,647. Blank uses IQ-TREE's clock-derived seed; set a value to reproduce a run. |
Advanced native options | string | [] | IQ-TREE arguments as a JSON array of tokens or as command-line text, such as --ancestral --rate -o seq1. |
default (auto ModelFinder) leaves the model to IQ-TREE's default behavior, which selects a best-fit model for the alignment and then infers the tree.Standard ModelFinder runs the standard TEST search, similar to jModelTest or ProtTest, and then infers the tree. Standard model selection only (TESTONLY) stops after the selection.ModelFinder (FreeRate search) runs MF, which includes FreeRate heterogeneity candidates and stops after model selection, so no tree is produced. The ranking is in the .iqtree report.ModelFinder Plus runs MFP: the same selection, then the tree with the best-fit model. This is what the default does.Partition model selection and merging runs MF+MERGE on a partitioned analysis: it selects models, merges partitions that can share one, and stops without inferring a tree.JTT, WAG and LG are empirical amino acid substitution matrices for protein alignments.GTR is the general time reversible nucleotide model, with six substitution rates and four base frequencies. HKY distinguishes transitions from transversions with unequal base frequencies, and K2P assumes equal base frequencies.Selection-only modes (MF, MF+MERGE, TESTONLY and any model string containing ONLY) cannot be combined with bootstrap, and the form says so before submission.
Each scientific parameter is set either in its field or through advanced options, not both. Output locations, threads, memory, checkpoint scheduling and reruns are managed by ProteinIQ; IQ-TREE picks its thread count automatically up to 4. Other scientific values are passed unchanged, and IQ-TREE reports unsupported combinations. The official command reference explains each option.
["--ancestral", "--rate"]["-p", "@partition"]["-z", "@trees", "-zb", "1000", "-au"]The first reconstructs ancestral sequences and writes site rates, the second runs a partitioned analysis from the uploaded partition file, and the third evaluates the uploaded trees with 1,000 RELL replicates and the approximately unbiased test.
Start with the default model unless you have a strong reason to fix one; ModelFinder tests the rate-heterogeneity part of the model as well as the matrix. For branch support, set Ultrafast bootstrap replicates and SH-aLRT replicates to 1,000 each. Counts from 1 to 999 are refused by IQ-TREE 3.1.1, so use 0 to explore an alignment without resampling. A classic count such as 100 belongs to the standard bootstrap: leave Ultrafast bootstrap replicates at 0 and add -b 100 to Advanced native options, which repeats the tree search for every replicate and takes much longer.
Set Random seed when a run has to be reproduced exactly. Without it, two runs on the same alignment can end on different trees of near-equal likelihood.
| Output | Contents |
|---|---|
Tree viewer | The maximum likelihood tree in a rectangular or radial layout, with support and branch-length labels, Newick copy and download, and a run summary |
Files | Every file the analysis writes, such as the .iqtree report, .treefile, .log, .contree, .splits.nex, checkpoints, and output.command.log |
Analyses that produce several trees or no single tree, such as ModelFinder-only runs, tree distances or AliSim, show their results in the Files tab instead of the viewer. The Newick tree can be sent to other tools from the Files tab.
The Tree viewer draws the maximum likelihood tree (the .treefile, in Newick format) in a rectangular or radial layout. Support labels sit on internal nodes, and the Support values and Branch lengths switches in the display options add the support label and the exact length to each branch.
When ultrafast bootstrap is enabled, IQ-TREE also writes the consensus tree (.contree) and split support file (.splits.nex) alongside the maximum likelihood tree and the .iqtree report.
A label such as 92.4/98 on a node means SH-aLRT 92.4 and ultrafast bootstrap 98, in that order. With only one test enabled the label is a single number, and with neither the branches carry no label. The two tests ask different questions, so each has its own rule of thumb.
| Test | Setting | What it measures | Rule of thumb |
|---|---|---|---|
| UFBoot | Ultrafast bootstrap replicates | Percentage of approximate bootstrap trees that contain the branch | 95 or higher (Hoang et al. 2018) |
| SH-aLRT | SH-aLRT replicates | Likelihood-ratio test of the branch against the alternative arrangements around it | 80 or higher (IQ-TREE FAQ; test from Guindon et al. 2010) |
| Standard bootstrap | -b in Advanced native options | Percentage of trees rebuilt from resampled alignments that contain the branch | Not on the UFBoot scale |
A branch is treated as well supported when it passes both, SH-aLRT at 80 or higher and UFBoot at 95 or higher. A branch that passes only one is borderline and is better reported with both numbers.
These cut-offs were set for single-gene alignments. In concatenated phylogenomic data support tends to reach 100 on most branches, and the IQ-TREE FAQ recommends concordance factors (--gcf and --scf, Minh et al. 2020) to show where genes and sites disagree.
UFBoot and standard bootstrap are not interchangeable. In the UFBoot2 simulations with a correct model, standard bootstrap was conservative and UFBoot2 was almost unbiased, so a UFBoot value of 95 does not match a standard bootstrap value of 95. UFBoot2 can overestimate support when the model is severely violated, and the -bnni advanced option corrects for this at about twice the runtime (Hoang et al. 2018).
IQ-TREE always writes an unrooted tree. The viewer draws it from the left, but that point is only where the drawing starts: by default the first sequence in the alignment is placed at the root, and the .iqtree report notes that the tree is unrooted.
To redraw it with another taxon at the root, add ["-o", "Taxon_name"] to Advanced native options. Which taxa cluster together can be read from an unrooted tree, but which lineage branched first cannot until a justified outgroup is chosen.
Branch lengths are expected substitutions per site (the scale bar unit), not time. A branch of 0.10 carries on average one change per ten sites. For codon alignments the unit is per codon site, so lengths are about three times those of the same data analysed as nucleotides (IQ-TREE FAQ).
The Tree length is the sum of all branch lengths. The report also gives Sum of internal branch lengths as a share of it, and a small share means most change sits on terminal branches, which carry no information about how the taxa are related.
The results include the selected substitution model and its parameters. GTR+F+I+G4, for example, combines the GTR matrix (GTR) with empirical base or amino acid frequencies (+F), a proportion of invariant sites (+I) and gamma-distributed rates with 4 categories (+G4). A +R4 in place of +G4 is FreeRate heterogeneity with 4 categories, which relaxes the gamma assumption.
The .iqtree report opens the model section with Best-fit model according to BIC, then List of models sorted by BIC scores with the columns Model, LogL, AIC, w-AIC, AICc, w-AICc, BIC and w-BIC. The best model has the lowest score, and each w- column is that model's weight in the criterion.
BIC is the default criterion. In the ModelFinder simulations it inferred the optimal model even when the tree was wrong, while AIC and AICc showed a 2 to 3% bias toward richer rate models (Kalyaanamoorthy et al. 2017).
When the top models have similar weights, the data do not separate them clearly. The rate part of the choice (+G4 against +R4) depends more on the tree than the substitution matrix does, since models are scored on a starting tree before the final search (Kalyaanamoorthy et al. 2017).
The run summary under the tree shows Sequences, Sites, Model, Type, Log-likelihood, AIC, BIC and Tree length. The Bootstrap and aLRT entries are the replicate counts you requested, not support values.
Log-likelihood (the report's Log-likelihood of the tree) is the log probability of the alignment on the final tree, so it is always negative and grows more negative with longer alignments. Compare it only between trees or models on the same alignment. The report also lists Unconstrained log-likelihood (without tree), a ceiling that no tree can exceed.
AIC and BIC (and Corrected Akaike information criterion (AICc) score in the report) are lower for better models, after a penalty for the Number of free parameters (#branches + #model parameters). They rank models on one alignment and say nothing about whether the best model fits well in absolute terms. Tree length is the sum of branch lengths in substitutions per site, a measure of total divergence.
The alignment block of the report adds Number of constant sites, Number of parsimony informative sites and Number of distinct site patterns. Few informative sites means little signal for resolving branches, which shows up as low support even for a correct tree.
The alignment check in the .log file (also in output.command.log) lists each sequence with Gap/Ambiguity, Composition and p-value, marks it passed or failed, and ends with a count of sequences that failed the composition chi2 test. A sequence fails when its residue composition differs from the alignment average at p below 5%, so about 1 in 20 sequences would fail by chance even if composition were uniform.
The IQ-TREE FAQ calls the test explorative and advises against removing failing sequences by default, but suggests checking the alignment when the tree looks unexpected.
A warning that sequences contain more than 50% gaps or ambiguity points to fragments that add little signal. For a group of identical sequences, IQ-TREE keeps the first two for the search and adds the others back afterwards (IQ-TREE FAQ), with a note naming each duplicate. Remove duplicates yourself if you want a tree without them.
Support is not the probability that a clade is real. It measures how stable a branch is under resampling or against nearby alternatives, given your alignment and model.
A model that is wrong in the same way for every replicate, or a contaminated or misaligned sequence, can produce 100 on a branch that is false. Support values from different programs or methods should not be compared with each other, and a low value does not make a branch wrong: it can simply mean few informative sites separate the groups.
The viewer draws the .treefile, whose topology is the single best tree. The .contree is a consensus of the bootstrap trees and can have unresolved nodes where the bootstrap trees disagree, so the two files may differ within one run. State which file a figure or support value comes from.
If support is uniformly low, the alignment may be too divergent or contain paralogs rather than orthologs. Alignment quality matters more than the tree method, since poorly aligned regions add noise that no search can remove; MAFFT with L-INS-i is a common choice for difficult sequences.
Maximum likelihood phylogenetics finds the tree topology and branch lengths that maximize the probability of observing the alignment under a given evolutionary model. For each tree, IQ-TREE calculates the likelihood that the observed alignment arose from that tree:
where is the pattern at site , is the tree topology, and represents model parameters like substitution rates and branch lengths. IQ-TREE explores tree space with a stochastic perturbation algorithm that keeps rearrangements which raise the likelihood (Minh et al. 2020).
ModelFinder tests a set of substitution models and selects the best one using information criteria. For each model , it calculates:
where is the number of free parameters and is the alignment length. BIC penalizes complex models more heavily than AIC, reducing the risk of overfitting. The original ModelFinder paper combined 22 DNA or 36 protein substitution models with 13 models of rate variation across sites (gamma-distributed rates, FreeRate categories and invariant sites), and scored each combination on a parsimony starting tree.
In its benchmarks, ModelFinder was 39 to 289 times faster than jModelTest on DNA data and 16 to 52 times faster than ProtTest on protein data (Kalyaanamoorthy et al. 2017).
The standard bootstrap resamples the alignment columns with replacement and rebuilds a tree from every resampled alignment (Felsenstein 1985), which the IQ-TREE FAQ notes may take weeks or months on large data sets. UFBoot approximates this distribution by reusing the trees that the search already visits, scoring each against every resampled alignment instead of searching again from scratch. UFBoot2 was 778 times faster than the standard bootstrap (median) and 8.4 times faster than RAxML rapid bootstrap (Hoang et al. 2018).
The SH-aLRT (Shimodaira-Hasegawa-like approximate likelihood ratio test, Guindon et al. 2010) tests each branch against the alternative arrangements around it using a likelihood ratio, so it rests on a different statistical framework than resampling. Using both tests gives two complementary support values per branch.
The official IQ-TREE web server is W-IQ-TREE, run by the Center for Integrative Bioinformatics Vienna (CIBIV) and described by Trifinopoulos et al. (2016). ProteinIQ is not affiliated with the IQ-TREE developers or CIBIV; it runs the published IQ-TREE 3.1.1 program on its own servers. When we checked the W-IQ-TREE address on 23 October 2026, it showed a notice that the server had been taken down because of server issues and that a replacement featuring IQ-TREE 3 was in final testing.
The IQ-TREE web server tutorial describes W-IQ-TREE with Tree Inference, Model Selection and Analysis Results tabs and PHYLIP, FASTA, NEXUS, CLUSTAL and MSF input. Tree inference runs ultrafast bootstrap and the SH-aLRT test by default. Jobs that need more than 24 hours of CPU time or 1 GB of RAM are stopped, and the checkpoint can be downloaded to resume the run locally.
Results come as a zip with the report, log, tree, consensus tree, and SVG and PDF tree drawings, and an email can announce when a job finishes. The IQ-TREE website also lists web servers at Los Alamos National Laboratory and the CIPRES Gateway.
On ProteinIQ, IQ-TREE runs with up to 4 threads, and a run stops if IQ-TREE has not finished within 55 minutes. Ultrafast bootstrap and SH-aLRT are off unless requested, which is IQ-TREE's own command-line default. Advanced native options and the optional file inputs reach partitioned analyses, topology tests, concordance factors, dating, ancestral reconstruction and AliSim. A batch can hold 10 alignments.
Every job is saved to the workspace's job history with its inputs, settings and files. The tree can be passed to other tools or used as a step in a workflow, and paid plans can submit jobs through the ProteinIQ API. ProteinIQ charges credits for each job, as listed under pricing.
IQ-TREE assumes the input is a correct alignment of homologous sequences, and it builds a tree from whatever it receives. Unaligned or misaligned sequences, paralogs mixed with orthologs, and contaminated records all produce confident-looking trees. Check the alignment and the composition warnings in the log before reading the tree.
Runs that need more than 55 minutes, such as very large alignments with many bootstrap replicates or rich mixture models, stop with a timeout error. For large alignments where speed matters more than model choice, FastTree builds approximate maximum likelihood trees much faster with simpler models, and RAxML-NG gives an independent maximum likelihood estimate to compare against. Both fit into a wider phylogenetic analysis.
For phylogenomic data with many genes, IQ-TREE's partition models assign separate models to separate loci, which matters when genes evolve at different rates. Upload the loci under Additional alignments or a partition file under Partition definition.
Align multiple protein or nucleotide sequences and export FASTA, Clustal, or Phylip outputs.
Build phylogenetic trees from aligned protein or nucleotide sequences using approximate maximum-likelihood methods.
Align protein or nucleotide sequences with selectable accuracy and speed trade-offs.
Align multiple protein or nucleotide sequences with high-accuracy PPP refinement.
Hardware-accelerated edit distances and global or local sequence scores
Universal structure alignment for proteins, RNA, and DNA molecules
Search and cluster protein or nucleotide sequences for homology discovery at large scale.
Align and compare whole genomes to detect SNPs, indels, and structural variants.
Predict consensus RNA secondary structures from sequence alignments.
Transcript-level RNA-seq quantification with Salmon selective alignment