# Molecular Phylogenetics and Evolution: Methods, Models, and Pitfalls ## Introduction to Molecular Phylogenetics and Evolution ### What is Molecular Phylogenetics? Molecular phylogenetics is the discipline that reconstructs the evolutionary relationships among organisms, genes, or populations using molecular sequence data. Rather than relying solely on morphological characters, which can be convergent, plastic, or absent in fossils, molecular phylogenetics exploits the fact that DNA and protein sequences accumulate heritable changes over time. These changes serve as a record of shared ancestry: sequences that diverged recently are more similar to one another than sequences that diverged long ago. The central goal is to infer a phylogenetic tree—a branching diagram that represents the order of divergence events and, ideally, the relative timing of those events. The field rests on a fundamental premise: homologous sequences—those derived from a common ancestral sequence—retain sufficient similarity to be aligned, yet differ enough to carry phylogenetic signal. The [Neutral Theory of Molecular Evolution](/knowledge/molecular-biology/neutral-theory-of-molecular-evolution) provides the theoretical backbone for much of this work. It posits that most observed molecular variation is selectively neutral or nearly so, and that the rate of substitution is roughly constant per unit time for a given gene. This insight transforms sequence differences into a molecular yardstick, enabling both tree reconstruction and divergence time estimation. ### Molecular Data in Evolutionary Studies The types of molecular data used in phylogenetics have expanded dramatically. Early studies relied on protein sequences and restriction fragment length polymorphisms; contemporary work uses whole genomes. The most common data types include: - **Nucleotide sequences** from protein-coding genes (e.g., mitochondrial *COI*, nuclear *RAG1*), [ribosomal RNA](/knowledge/bioinformatics/ribosomal-rna-structure-taxonomic-profiling) genes (e.g., 18S, 16S), and non-coding regions such as introns or intergenic spacers. - **Amino acid sequences** from orthologous proteins, which are useful for deep evolutionary divergences where nucleotide substitutions have saturated. - **Gene presence/absence patterns** and **gene order** data, used for bacterial and organellar phylogenomics. - **Single nucleotide polymorphisms (SNPs)** from population-level sampling, used for shallow-scale phylogenetics and phylogeography. For protein-coding genes, the choice between nucleotide and amino acid data depends on the evolutionary depth. At deep timescales (e.g., across animal phyla), synonymous substitutions—changes that do not alter the encoded amino acid—have reached saturation, obscuring phylogenetic signal at the nucleotide level. Amino acid sequences, which change more slowly, retain signal over longer timescales. Conversely, for closely related species or populations, nucleotide sequences provide the necessary variation. ## The Molecular Basis of Phylogenetic Inference ### Substitution Models Phylogenetic inference requires an explicit model of how sequences change. A substitution model is a mathematical description of the relative rates at which one character state changes into another. For nucleotides, the simplest model is the Jukes-Cantor (JC69) model, which assumes equal base frequencies and equal substitution rates among all four nucleotides. The Kimura 2-parameter (K80) model distinguishes between transitions (A↔G, C↔T) and transversions (all other changes), reflecting the observation that transitions are generally more frequent. The General Time Reversible (GTR) model is the most general reversible model, with six distinct substitution rate parameters and four base frequency parameters. These models are typically expressed as a 4×4 rate matrix **Q**, where off-diagonal entries are the instantaneous rates of change between nucleotides. The probability of change over a branch of length *t* is computed as **P(t) = e^(Qt)**, a matrix exponential. Branch lengths are measured in expected number of substitutions per site, not in time units. This distinction is critical: a branch length of 0.1 means that, on average, 10% of sites have experienced a substitution along that branch. ### The [Molecular Clock Hypothesis](/knowledge/molecular-biology/molecular-clock-hypothesis) The [Molecular Clock Hypothesis](/knowledge/molecular-biology/molecular-clock-hypothesis) states that for a given gene, substitutions accumulate at an approximately constant rate over time. This hypothesis, proposed by Zuckerkandl and Pauling in the 1960s, follows directly from the neutral theory: if most substitutions are neutral and drift is the primary force fixing them, the substitution rate equals the mutation rate, which is roughly constant per year for a given lineage and gene. The [Molecular Clock Definition](/knowledge/molecular-biology/molecular-clock-definition) in practice is a calibration device: if the rate is known, sequence divergence can be converted into absolute time. For example, if a gene evolves at 1% divergence per million years (a rate of 0.005 substitutions per site per million years per lineage), two sequences that differ by 10% are estimated to have shared a common ancestor approximately 10 million years ago. The [Molecular Clock Model](/knowledge/molecular-biology/molecular-clock-model) is now understood to be an oversimplification—rates vary among lineages due to generation time effects, metabolic rate, and DNA repair efficiency—but it remains the foundation for molecular dating, as discussed in the section on divergence time estimation. ## [Sequence Alignment](/blog/guides/sequence-alignment-choosing-the-right-method-for-your-biological-question): The First Critical Step ### Alignment Algorithms Before any phylogenetic analysis, homologous sites must be identified across sequences. This is the [multiple sequence alignment](/blog/guides/multiple-sequence-alignment-common-pitfalls-and-quality-checks) (MSA) problem: arranging sequences so that columns correspond to positions descended from a common ancestral position. Gaps—insertions or deletions (indels)—are introduced to account for length differences. Alignment is not a trivial preprocessing step; it is a phylogenetic inference problem in its own right, and errors here propagate to every downstream analysis. The most widely used algorithms are **progressive** and **iterative** methods. Progressive alignment, implemented in Clustal Omega and MUSCLE, builds an initial guide tree from pairwise distances, then aligns sequences in order of increasing divergence, adding each new sequence to the growing alignment. The guide tree is approximate, and errors in it cannot be corrected later. **Iterative** methods, such as MAFFT and PRANK, refine the alignment by repeatedly realigning subsets and re-estimating the guide tree until a stable solution is reached. PRANK is notable for distinguishing alignment gaps that are insertions from those that are deletions, which improves accuracy in datasets with many indels. For highly divergent sequences, **profile-based** methods that use hidden Markov models (HMMs), such as those in the Pfam database, can align sequences to a conserved structural profile rather than to each other directly. This is particularly useful for protein sequences where tertiary structure is more conserved than primary sequence. ### Assessing Alignment Quality Alignment quality should be assessed before proceeding to tree inference. Poorly aligned regions—those with ambiguous gap placement or low similarity—can be identified by eye in small datasets, but for large datasets, automated filtering is essential. Programs like Gblocks and trimAl remove poorly aligned positions and divergent regions. A common rule of thumb is to remove columns where more than 50% of sequences have a gap, and to trim terminal overhangs. A more rigorous approach is to use **alignment uncertainty** directly: instead of a single alignment, generate many alternative alignments (e.g., by varying alignment parameters) and run the phylogenetic analysis on each, then summarize the results. This approach, sometimes called "phylogenetic alignment," acknowledges that alignment and tree inference are coupled problems. The practical takeaway is that alignment is not a "one-click" step; parameter choices (gap opening penalty, gap extension penalty, substitution matrix) materially affect downstream tree topology. ## Models of Sequence Evolution ### Nucleotide Substitution Models Nucleotide substitution models form a nested hierarchy. The simplest, JC69, has one parameter. The K80 model adds a transition/transversion ratio (κ). The HKY85 model adds unequal base frequencies. The GTR model adds six rate parameters. Each model can be extended with a **gamma distribution** (Γ) to model rate variation across sites, and a proportion of **invariant sites** (I) to account for positions that cannot change (e.g., those critical for [protein structure](/knowledge/bioinformatics/protein-structure-biophysical-levels-folding)). The notation GTR+Γ+I is standard. Model selection is not a matter of choosing the most parameter-rich model; over-parameterized models reduce [statistical power](/blog/guides/statistical-power-what-it-is-and-why-it-matters-in-research), while under-parameterized models can lead to systematic errors such as long-branch attraction. The standard approach is to use information criteria—the Akaike Information Criterion (AIC) or the Bayesian Information Criterion (BIC)—to balance model fit against parameter count. Software such as jModelTest and ModelTest-NG automates this process by calculating likelihood scores for a candidate set of models and ranking them by AIC or BIC. ### Amino Acid Replacement Models For protein sequences, empirical substitution matrices are used. These matrices, such as Dayhoff, JTT, WAG, and LG, are estimated from large databases of aligned protein sequences and give the relative rates of replacement among the 20 amino acids. The LG matrix, estimated from a large and diverse set of protein families, is currently a common default. More recent models, such as the empirical codon models, account for the fact that amino acid replacements are mediated by nucleotide changes and are influenced by codon usage and selection. For datasets spanning very deep divergences, **mixture models** that allow different site classes to evolve under different matrices (e.g., C60, which uses 60 distinct matrices) can substantially improve fit. These models recognize that different protein regions—catalytic domains versus structural scaffolds—evolve under different constraints. ### Model Selection Criteria The likelihood ratio test (LRT) can compare nested models: if model B is a special case of model A (e.g., JC69 is a special case of K80), then twice the difference in log-likelihood is approximately chi-square distributed with degrees of freedom equal to the difference in parameter count. For non-nested models, AIC and BIC are used. AIC = 2k − 2ln(L), where k is the number of parameters and L is the maximum likelihood. BIC = k·ln(n) − 2ln(L), where n is the number of sites. BIC penalizes parameters more heavily and tends to select simpler models. In practice, the GTR+Γ model is often selected for nucleotide data, and LG+Γ for amino acid data, but this should be confirmed rather than assumed. ## Phylogenetic Tree Inference Methods ### Distance Methods Distance methods reduce sequence data to a matrix of pairwise evolutionary distances—the number of substitutions per site between each pair of sequences, corrected for multiple hits using a substitution model. The tree is then constructed from this distance matrix. The most common algorithm is **Neighbor-Joining (NJ)**, which iteratively joins the pair of taxa with the smallest branch length sum, producing a single tree. **Minimum Evolution (ME)** and **Unweighted Pair Group Method with Arithmetic Mean (UPGMA)** are alternatives; UPGMA assumes a strict molecular clock and is rarely appropriate for divergent taxa. Distance methods are computationally fast and useful for exploratory analysis or for datasets with hundreds of thousands of sequences (e.g., [viral phylogenetics](/blog/guides/viral-phylogenetics-interpreting-clusters-without-overstating-transmission)). However, they discard information: the distance matrix is a summary statistic, and information about which specific sites support which groupings is lost. Distance methods are also more sensitive to model misspecification than character-based methods. ### Maximum Parsimony Maximum parsimony (MP) seeks the tree that requires the fewest evolutionary changes to explain the observed sequences. It is a character-based method: for each site, the minimum number of changes required on a given tree is calculated, and the tree with the lowest total is chosen. Parsimony is non-parametric—it makes no explicit model of substitution—and can be statistically inconsistent under conditions of high substitution rates and unequal branch lengths, a phenomenon known as long-branch attraction (LBA). When rates are high, multiple substitutions at the same site erase signal, and parsimony can systematically group long branches together regardless of true relationships. Despite these limitations, parsimony remains useful for morphological data and for datasets with low divergence where multiple hits are rare. It is also the basis for some tree-searching heuristics in larger analyses. ### Maximum Likelihood Maximum likelihood (ML) is the gold standard for phylogenetic inference. Given a tree topology, branch lengths, and a substitution model, the likelihood of the data is calculated. The algorithm works by computing, for each site, the probability of the observed character states given the model and the tree, summing over all possible ancestral states at internal nodes using the "pruning" algorithm of Felsenstein. The product of site likelihoods (or the sum of log-likelihoods) is the likelihood of the tree. The ML tree is the topology and branch lengths that maximize this value. ML is statistically consistent: as more data are collected, the ML estimate converges to the true tree, provided the model is correctly specified. It is computationally intensive, but programs like RAxML, IQ-TREE, and PhyML use efficient heuristics (e.g., subtree pruning and regrafting, or SPR moves) to search tree space. IQ-TREE is notable for its built-in model selection (ModelFinder) and ultrafast bootstrap approximation. ### Bayesian Inference Bayesian inference (BI) treats the tree and model parameters as random variables with a posterior distribution. The posterior probability of a tree is proportional to the likelihood times the prior probability. Because the posterior cannot be calculated analytically, Markov Chain Monte Carlo (MCMC) is used to sample trees and parameters in proportion to their posterior probability. The program MrBayes and the more recent BEAST2 (for dating and coalescent analyses) are standard tools. The output of a Bayesian analysis is a posterior distribution of trees. The **maximum a posteriori (MAP)** tree or the **majority-rule consensus** tree is typically reported. Bayesian support values—posterior probabilities—are the proportion of sampled trees containing a given clade. These are not the same as bootstrap values; they are generally higher and should not be compared across methods. Bayesian inference allows flexible modeling of rate variation across lineages (relaxed clocks) and the incorporation of prior information from fossils or biogeography. ## Assessing Phylogenetic Confidence ### Bootstrap Analysis The nonparametric bootstrap is the most common measure of support in ML and parsimony analyses. The procedure is: 1. Resample the alignment columns with replacement to create a new dataset of the same length as the original. 2. Infer a tree from the resampled dataset. 3. Repeat 100–1000 times. 4. For each clade in the original tree, record the proportion of bootstrap trees that contain that clade. A bootstrap value of 95% or higher is generally considered strong support; 70–95% is moderate; below 70% is weak. Bootstrap values are conservative: they measure the robustness of the result to sampling error in the alignment columns, not the probability that the clade is correct. The "ultrafast bootstrap" implemented in IQ-TREE is a computationally efficient approximation that is widely used for large datasets. ### Bayesian Support Values Bayesian posterior probabilities (PP) are the proportion of MCMC-sampled trees that contain a given clade. PP values are generally higher than bootstrap values for the same clade, and a PP of 0.95 or above is often considered strong support. However, PP values can be inflated when the model is misspecified or when the MCMC has not converged. It is good practice to run at least two independent MCMC chains and check that they converge to the same distribution (e.g., using the average standard deviation of split frequencies in MrBayes, or the effective sample size (ESS) values in BEAST2; ESS > 200 is a common threshold). Other support measures include the **approximate likelihood ratio test (aLRT)** and the **Shimodaira-Hasegawa (SH) test**, which compare the fit of different topologies. These are less commonly used than bootstrap or PP but can be informative for specific hypothesis tests. ## Molecular Dating and Divergence Time Estimation ### Relaxed Molecular Clocks The strict molecular clock—a single substitution rate across all lineages—is rarely realistic. **Relaxed clock** models allow rates to vary among lineages. The two main classes are: - **Uncorrelated relaxed clocks**, where each branch has its own rate drawn from a distribution (e.g., lognormal or exponential), with no autocorrelation between parent and daughter branches. - **Autocorrelated relaxed clocks**, where rates evolve along the tree, with daughter branch rates correlated with the parent branch rate. These models are implemented in BEAST2 and PAML (the MCMCTree program). The choice of clock model can be tested using Bayes factors or AIC. The [Molecular Clock Studies](/knowledge/molecular-biology/molecular-clock-studies) literature consistently shows that uncorrelated lognormal clocks fit most empirical datasets well. ### Fossil Calibration To convert relative times (substitutions per site) into absolute times (millions of years), at least one calibration point is needed. Fossil calibrations are the most common source. A fossil provides a minimum age for the divergence between the lineage leading to the fossil and its sister lineage. Calibrations are specified as prior distributions on node ages—for example, a lognormal prior with a hard minimum bound (the fossil age) and a soft maximum bound (based on the absence of older fossils or on geological constraints). The choice of calibration priors is critical. Using a single calibration can lead to biased estimates if the fossil is misidentified or the prior is too restrictive. Modern practice favors **multiple calibrations** across the tree, with **cross-validation** to check for conflicts. The fossilized birth-death (FBD) process is an alternative approach that integrates fossil and extant taxa in a single analysis, treating fossils as part of the tree rather than as calibration points. The [Molecular Clock in Evolution](/knowledge/molecular-biology/molecular-clock-in-evolution) concept is also applied in phylogeography and population genetics, where coalescent-based methods estimate demographic history from sequence data. These methods, implemented in BEAST2 and *msmc*, use the same clock machinery but model the genealogy of sampled lineages rather than a species tree. ## Common Pitfalls and Best Practices in Phylogenetic Analysis ### Long-Branch Attraction Long-branch attraction (LBA) is the systematic error in which rapidly evolving lineages are grouped together regardless of their true relationships. It arises because multiple substitutions at the same site erase signal; two long branches may share the same derived state by chance, and parsimony (and, to a lesser extent, ML with a poor model) will interpret this as shared ancestry. The classic example is the placement of the microsporidia (highly reduced intracellular parasites with accelerated rates) as a basal eukaryotic lineage, an artifact that was resolved when more slowly evolving genes and better models placed them within the fungi. Mitigations for LBA include: - Adding more taxa to break up long branches. - Using models that account for rate variation across sites (Γ). - Removing the fastest-evolving sites or genes. - Using amino acid data instead of nucleotides for deep divergences. ### Alignment and Model Misspecification Alignment errors are a major source of phylogenetic error. Misaligned sites create false synapomorphies—shared derived characters that are artifacts of alignment, not ancestry. This is particularly problematic for divergent sequences with many indels. Model misspecification—using a model that is too simple—can also cause systematic errors. For example, ignoring rate variation across sites (not using Γ) can lead to LBA, because sites that evolve fast are treated as if they evolve at the average rate. The best practice is to: 1. Use a modern, iterative alignment program (MAFFT or PRANK). 2. Trim poorly aligned regions. 3. Select a model using ModelFinder or jModelTest. 4. Run both ML and Bayesian analyses and compare results. 5. If results conflict, investigate the cause before trusting either. ### Best Practices Checklist 1. **Data quality**: Check for contamination, paralogy, and sequencing errors. Use BLAST to verify the identity of each sequence. 2. **Alignment**: Use MAFFT or PRANK with default parameters, then trim with trimAl (e.g., -automated1). 3. **Model selection**: Use ModelFinder (IQ-TREE) or jModelTest for nucleotides; use ModelFinder for amino acids. 4. **Tree inference**: Run ML with IQ-TREE (with 1000 ultrafast bootstrap replicates) and Bayesian inference with MrBayes (two runs, four chains, until ESS > 200). 5. **Convergence**: Check that independent runs converge to the same topology and similar likelihood scores. 6. **Sensitivity analysis**: Repeat the analysis with a different alignment, a different model, or a subset of taxa. If the key clades are stable, the result is robust. 7. **Report**: Report the model, the alignment method, the trimming parameters, and the support values. Provide the alignment and tree files as supplementary data. ## Frequently Asked Questions ### What is molecular phylogenetics? Molecular phylogenetics is the reconstruction of evolutionary relationships among organisms, genes, or populations using molecular sequence data (DNA, RNA, or protein). It relies on the principle that homologous sequences accumulate changes over time, and the degree of difference between sequences reflects the time since their last common ancestor. The [Evidence of Evolution Molecular](/knowledge/molecular-biology/evidence-of-evolution-molecular) is abundant: shared genes across distantly related taxa, conserved functional domains, and the nested hierarchy of sequence similarities all support common ancestry. ### How do you choose the best model of sequence evolution? Model selection is performed using information criteria (AIC or BIC) or likelihood ratio tests. Software such as ModelFinder (in IQ-TREE) or jModelTest evaluates a set of candidate models and ranks them by fit to the data. The selected model balances fit against parameter count. For most nucleotide datasets, GTR+Γ is a reasonable starting point; for amino acid data, LG+Γ is common. The [Concept of Neutral Evolution](/knowledge/molecular-biology/concept-of-neutral-evolution) underpins the assumption that most substitutions are neutral, which justifies the use of these empirical models. ### What is the difference between maximum likelihood and Bayesian inference? Maximum likelihood (ML) finds the single tree (topology and branch lengths) that maximizes the probability of the data under a given model. It produces a point estimate with bootstrap support values. Bayesian inference estimates the posterior distribution of trees—the probability of each tree given the data and priors—using MCMC sampling. It produces a set of trees, from which a consensus tree and posterior probabilities are derived. Bayesian methods can incorporate prior information (e.g., fossil calibrations) and flexible models (e.g., relaxed clocks), but they are computationally more intensive and require checking MCMC convergence. ### What is long-branch attraction? Long-branch attraction (LBA) is a systematic error in phylogenetic inference where rapidly evolving lineages (long branches) are incorrectly grouped together. It occurs because multiple substitutions at the same site erase phylogenetic signal, and the shared derived states that remain are more likely to be homoplasious (convergent) than homologous. LBA is exacerbated by simple models that ignore rate variation across sites and by sparse taxon sampling. It can be mitigated by adding taxa, using gamma-distributed rates, and removing fast-evolving sites. ### How do you estimate divergence times from a phylogeny? Divergence times are estimated by converting branch lengths (in substitutions per site) into absolute time using a molecular clock. The strict clock assumes a single rate; relaxed clocks allow rates to vary among lineages. Calibration points—typically from fossils—provide absolute age constraints on specific nodes. Software like BEAST2 and MCMCTree jointly estimate the tree, the clock model, and the node ages. The [Molecular Clock Model](/knowledge/molecular-biology/molecular-clock-model) is the foundation of this approach, and the accuracy of the estimates depends on the quality of the calibrations and the clock model. ### What is a bootstrap value in phylogenetics? A bootstrap value is the proportion of bootstrap replicate datasets (resampled from the original alignment) in which a given clade appears in the inferred tree. It measures the robustness of a clade to sampling error in the alignment columns. A value of 95% or higher is considered strong support, 70–95% moderate, and below 70% weak. Bootstrap values are not the probability that the clade is correct; they are a measure of repeatability under resampling. ### Why is sequence alignment important for phylogenetics? Sequence alignment determines which positions in different sequences are homologous—derived from a common ancestral position. If alignment is incorrect, the phylogenetic analysis compares non-homologous sites, generating false signal and potentially incorrect trees. Alignment errors are particularly problematic for divergent sequences with many insertions and deletions. Accurate alignment is therefore a prerequisite for any meaningful phylogenetic inference, and alignment uncertainty should be assessed, for example by trimming or by using multiple alignments. ## Key Takeaways - Molecular phylogenetics reconstructs evolutionary relationships from sequence data, relying on the neutral theory and the molecular clock as theoretical foundations. - Substitution models (JC69, K80, HKY85, GTR, and amino acid matrices like LG) describe how sequences change; model selection via AIC/BIC is essential to avoid systematic errors. - [Multiple sequence alignment](/blog/guides/multiple-sequence-alignment-common-pitfalls-and-quality-checks) is the first critical step; errors here propagate to all downstream analyses, so iterative methods (MAFFT, PRANK) and trimming are recommended. - Maximum likelihood and Bayesian inference are the preferred tree-building methods; distance methods and parsimony are faster but less robust. - Bootstrap values and Bayesian posterior probabilities measure support but are not interchangeable; convergence checks are mandatory for Bayesian analyses. - Divergence time estimation requires relaxed clocks and fossil calibrations; multiple calibrations with cross-validation improve accuracy. - Long-branch attraction, alignment artifacts, and model misspecification are the main pitfalls; sensitivity analyses and best-practice checklists mitigate these risks. ## Further Reading - Lynch Alfaro JW et al. *Special issue: Comparative biogeography of Neotropical primates*. Molecular phylogenetics and evolution. 2015. [PubMed 25451803](https://doi.org/10.1016/j.ympev.2014.09.027) - Ting N, Sterner KN. *Primate molecular phylogenetics in a genomic era*. Molecular phylogenetics and evolution. 2013. [PubMed 22960143](https://doi.org/10.1016/j.ympev.2012.08.021) - Edwards SV et al. *Implementing and testing the multispecies coalescent model: A valuable paradigm for phylogenomics*. Molecular phylogenetics and evolution. 2016. [PubMed 26518740](https://doi.org/10.1016/j.ympev.2015.10.027) - Zimmer E, Wildman D. *Editorial for 25(th) Anniversary Issue of Molecular Phylogenetics and Evolution*. Molecular phylogenetics and evolution. 2017. [PubMed 29157553](https://doi.org/10.1016/j.ympev.2017.10.024) - Lin CP, Danforth BN, Wood TK. *Molecular phylogenetics and evolution of maternal care in Membracine treehoppers*. Systematic biology. 2004. [PubMed 15503670](https://doi.org/10.1080/10635150490445869) - Krenz JG et al. *Molecular phylogenetics and evolution of turtles*. Molecular phylogenetics and evolution. 2005. [PubMed 15964217](https://doi.org/10.1016/j.ympev.2005.04.027)
## Related Clinical & Scientific Guides * [MAPK Pathway: Mechanism, Function, and Clinical Relevance](/knowledge/molecular-biology/mapk-pathway) * [Mammalian Cell Culture Bioreactors: A Practical Guide](/knowledge/molecular-biology/mammalian-cell-culture-bioreactor) * [Nucleotide Formation: Biosynthesis and Assembly of DNA/RNA Building Blocks](/knowledge/molecular-biology/nucleotide-formation)