IQ-TREE 3: phylogenomic inference software using complex evolutionary models

1IQ-TREE is a popular open-source software package for phylogenetic and phylogenomic analysis using a likelihood framework. IQ-TREE has been used by thousands of researchers and practitioners to understand evolution and biodiversity across many species and time scales. These studies range from inferring ancient relationships within Bacteria, Archaea, and eukaryotes; to more recent divergences among plants, animals, birds, gum trees, and primates; and even viral divergences such as SARS-CoV-2 virus. The two major releases of IQ-TREE have been cited more than 44,000 times (Google Scholar; 17th March 2026) and downloaded more than 1.3 million times (combined data from the GitHub repository and Bioconda package). IQ-TREE is also an integral component of many open-source biomedical applications such as CIPRES, GTDB, QIIME2, OrthoFinder, Nextstrain, Galaxy, PhyloSuite, TBtools, and Get_PhyloMarkers.

2Since the release of IQ-TREE version 2 in 2020, we have substantially expanded its capabilities, allowing users to perform data analyses under additional models and scenarios. Here, we introduce IQ-TREE version 3 and highlight its key developments, including mixture models, concordance factors, sequence simulation, protein model estimation, divergence time estimation, and pathogen tree inference.

3Phylogenetic analyses are often conducted with partitioned models, which assign different substitution models to different a priori defined groups of sites to account for heterogeneous evolutionary processes along genomes and/or within genes (e.g. among codon positions). Partitioned models are available in RAxML, RAxML-NG, and IQ-TREE 2; and tools like PartitionFinder can merge partitions to reduce over-parameterization.

4IQ-TREE 2 implemented mixture models, which assume a user-defined number (k) of substitution models (or classes), like partition models. However, unlike partition models, users do not need to specify how to partition the data in mixture models. Rather, model-based probabilities of sites are conditional upon classes, and each site has a certain probability (or weight) of belonging to each class. The site likelihood is calculated as the weighted average of the conditional probabilities, giving the marginal probability of the site pattern. Models of rate heterogeneity among sites based on a Gamma distribution or free-rates are versions of mixture models. Mixture models in IQ-TREE 2 have been used to address many deep phylogenetic questions (e.g.; ). In particular, such models help to avoid artefacts such as long-branch attraction, which can be exacerbated with simpler models.

5IQ-TREE 3 substantially improves the IQ-TREE 2 implementation by automatically determining the number of classes (k) and the substitution models for each class via the MixtureFinder algorithm for DNA data. In this way, the usage of DNA mixture models is now even simpler than partitioned models because no partition file is needed. This also removes user-subjectivity in determining the initial partitioning scheme, although it does come at the cost of additional computational resources: the amount of time and memory required by mixture models increases linearly with k. For protein data, in addition to profile mixtures and structure-based mixture models available in version 2, IQ-TREE 3 allows users to estimate amino-acid exchangeability matrices under complex profile mixture models, the so-called GTRpmix model. Notably, IQ-TREE 3 can also relax the assumption of a single tree common to many phylogenetic analyses by using the Mixtures Across Sites and Trees (MAST) model, which generalizes the General Heterogeneous evolution On a Single Topology (GHOST) mixture model of branch lengths. The MAST model can be useful to examine controversies between competing phylogenetic hypotheses and to pinpoint which sites in an alignment are best explained by which topologies; currently, it requires a user-provided set of input tree topologies.

6While the MAST model can summarize the contributions of a limited set of alternative tree topologies, IQ-TREE 3 also allows users to summarize discordance on all branches of a reference tree (e.g. an estimate of the species tree). We previously introduced methods to calculate two measures, called the gene concordance factor (gCF) and the site concordance factor (sCF), which quantify the level of concordance between each branch of the reference tree and the genes/loci (for the gCF) and sites (for the sCF). As the original sCF is based on parsimony and can be affected by homoplasy (similarities due to processes other than common ancestry) and taxon sampling, we introduced a likelihood-based version of the sCF that partially overcomes these issues. All of these methods and related summary statistics are available in IQ-TREE 3.

7Sequence simulations are used in many scenarios, such as evaluating new methods, testing hypotheses with parametric bootstrap, performing approximate Bayesian computation, and generating training and testing data to help with the development of machine learning applications (, but see ). However, existing tools like Seq-Gen, Dawg, and INDELible lack support for mixture and other complex models. IQ-TREE 3 implements a built-in alignment simulator called AliSim to take advantage of all advanced models available in IQ-TREE. To optimize speed, AliSim provides two different algorithms, one to simulate sequences of low divergence and another for high divergence, and can adaptively determine which algorithm to use. AliSim substantially outperforms Seq-Gen, Dawg, and INDELible in terms of both runtimes and memory usage. We also provided a high-performance version, AliSim-HPC, which can parallelize the simulations over multicore CPUs and multinode clusters.

8Protein data have been used to address many phylogenetic questions, especially at deep time scales. In many cases, analyses rely on empirical protein models like JTT, WAG, and LG, which were estimated from large collections of alignments spanning many clades on the tree of life. The scale of modern phylogenomic datasets and powerful computers present an opportunity to estimate models tailored specifically to individual clades. To facilitate this, IQ-TREE 3 provides a new maximum likelihood tool called QMaker to estimate a protein substitution matrix for a user-provided set of protein alignments and introduces a general matrix Q.pfam and a series of new clade-specific matrices—Q.bird, Q.insect, Q.mammal, Q.plant, Q.yeast—derived from large empirical datasets. These matrices will be useful for phylogenetic analyses of these and related clades. IQ-TREE 3 also provides an extension of QMaker for nonreversible models and a series of corresponding nonreversible Q matrices, as well as two new matrices, EAL (Eukaryotes and Archaea) and ELM (Eukaryotes), to be used with profile mixture models. The nonreversible matrices allow IQ-TREE 3 to reconstruct rooted trees without the requirement of an outgroup (e.g.; ), as well as to infer the confidence of the placement of the root via the rootstrap.

9IQ-TREE 3 integrates CMAPLE, an ultra-fast method for inferring trees with millions of closely related sequences, such as those produced by global SARS-CoV-2 (the virus that caused the COVID-19 pandemic) sequencing efforts. This enables IQ-TREE 3 to perform functions that are useful in tracking large pathogen outbreaks, such as online sample placement. Additionally, IQ-TREE 3 introduces a dedicated “--pathogen” option for dynamically selecting between the CMAPLE and IQ-TREE search algorithms to optimize runtime based on sequence divergence: CMAPLE is selected for highly similar sequences, and IQ-TREE is selected for more distantly related sequences.

10Moreover, IQ-TREE 3 and CMAPLE implement subtree pruning and regrafting tree assessment (SPRTA), a fast and appropriate approach for assessing phylogenetic uncertainty in genomic epidemiology. SPRTA can be invoked via the “--sprta” option.

11IQ-TREE 3 provides two facilities for divergence time estimation. The first is the Least Squares Dating method, LSD2, which is suitable for dating very large phylogenies with tip date information, such as viral phylogenies. The second is the new IQ2MC workflow to facilitate Bayesian divergence time estimation with MCMCTree. In the IQ2MC workflow, IQ-TREE 3 is used to compute the gradient vector and Hessian matrix of the branch lengths on a fixed tree topology, which are then used for fast approximate Bayesian MCMC sampling with MCMCTree. IQ-TREE dramatically speeds up the precalculation steps needed by MCMCTree, particularly for large phylogenomic datasets, and allows the use of the wide range of IQ-TREE's complex substitution models (such as mixture models), which were previously unavailable for time tree inference, and which are essential for dating ancient phylogenies.

12Finally, IQ-TREE 3 also includes a reimplementation of fast distance-based methods such as Neighbor-Joining via the DecentTree library. It also includes a Python interface through piqtree, an open-source Python package that exposes IQ-TREE's phylogenetic inference engine as a library of Python functions, thus facilitating interactive analyses in environments such as Jupyter notebooks and simplifying the development of custom phylogenetic workflows.

13IQ-TREE version 3 presents a joint collaborative effort among researchers from across the globe, which substantially extends version 2. Particularly, we would like to advocate for the use of mixture models instead of only using partition models, especially as the single-model assumption within a locus is often violated. In addition, if the single-tree assumption common to concatenation approaches is violated, then users should employ the MAST model and/or interrogate their data by examining concordance factors.

14We strive to continuously support the user community. Every new feature is accompanied by a tutorial and the manual is frequently updated at https://iqtree.github.io/doc/. User questions, issue reports, and feature requests can be submitted at the GitHub repository.