Comparing Cladograms, Phylograms, Dendrograms, and Ultrametric Trees

Cladograms, phylograms, ultrametric trees, and dendrograms are all branching diagrams that show relationships among organisms or other entities, but they differ in what their branch lengths represent and, consequently, what biological questions they can answer. A cladogram shows only the pattern of branching (which groups share a more recent common ancestor) without encoding any information in the lengths of its branches. A phylogram makes those branch lengths meaningful: they reflect the amount of evolutionary change, typically measured in molecular substitutions. An ultrametric tree goes further by scaling branches so that all tips line up at the same distance from the root, usually representing time. And “dendrogram” is the most general term of the four, covering virtually any tree-shaped diagram, biological or not. The distinctions sound cosmetic but carry real analytical consequences.

What Each Tree Type Actually Shows

A cladogram is the most stripped-down version of a phylogenetic tree. Its branches can be drawn at any length because those lengths carry no data. The only information a cladogram conveys is topology: the nested pattern of which lineages split from which, and in what order. If you see a cladogram with one branch drawn twice as long as another, that is a choice by the illustrator, not a statement about biology. Cladograms are common in introductory textbooks and in analyses where the goal is purely to identify evolutionary groupings without worrying about rates of change.

A phylogram retains the same branching pattern but adds a layer of information by making branch lengths proportional to the inferred amount of evolutionary change along each lineage. In molecular studies, the units are usually expected substitutions per site. A long branch in a phylogram means that lineage accumulated a lot of genetic change relative to its neighbors. This is immediately useful for spotting lineages that evolved unusually fast or slow, a phenomenon that turns out to be biologically very common. In vascular plants, for example, researchers have found that even methods designed to account for rate differences across lineages struggled to accommodate the dramatically slower rates observed in certain fern groups compared to flowering plants.

An ultrametric tree imposes an additional constraint: every path from the root to any tip has the same total length. Visually, this means all the tips line up along a straight vertical or horizontal line. When the branch lengths represent time (as they usually do in modern ultrametric trees), that alignment reflects the fact that all living species have been evolving for exactly the same amount of time since their last common ancestor at the root. The internal nodes then correspond to divergence dates, which is why ultrametric trees are sometimes called “chronograms” or “time trees.” Building one requires either assuming a molecular clock or applying a calibration strategy that converts substitution-based branch lengths into years.

“Dendrogram” is the umbrella. In its broadest sense, any tree-like branching diagram is a dendrogram, including cladograms, phylograms, and ultrametric trees. In practice, though, the term is used most often outside of phylogenetics proper, especially in cluster analysis of things like gene expression profiles or ecological community data. When a bioinformatician runs hierarchical clustering on a microarray experiment and draws a branching figure, they almost always call it a dendrogram rather than a phylogram or cladogram, even though the diagram looks similar. The key difference is that clustering dendrograms are not built from an evolutionary model; they group items by measured similarity without implying descent.

Why Branch Lengths Change Everything

The practical difference between these tree types boils down to what you can and cannot conclude from looking at a branch. On a cladogram, a branch is just a line connecting ancestors to descendants. You can read groupings (“these three species are more closely related to each other than to that one”), but you cannot say anything about how much change happened or how long ago a split occurred.

On a phylogram, you gain the ability to compare rates. If two sister lineages have very different branch lengths, one has changed much more than the other since their split. This is exactly the kind of signal that matters for detecting natural selection, identifying rapidly evolving genes, or flagging potential problems in your analysis. A suspiciously long branch can indicate either genuinely fast evolution or an artifact like a misplaced sequence. Phylograms produced by neighbor-joining, for instance, are additive trees where the distance between any two tips equals the sum of the branch lengths along the path connecting them. The neighbor-joining method was introduced as a way to reconstruct unrooted phylograms and has been shown through simulation to perform well relative to several competing distance-based methods.¹

On an ultrametric tree, branches encode time rather than raw change. This lets you read divergence dates directly from the diagram, but it comes at a cost: you must either assume that all lineages evolved at the same rate (a strict molecular clock, which is rarely true across a whole tree) or use a model that allows rates to vary while still producing time-scaled branches. Relaxed-clock models do exactly this, letting each branch have its own substitution rate while still inferring a timeline.² The result is a tree that looks neat and aligned at the tips but required much more modeling machinery under the hood than a phylogram did.

How These Trees Get Built

The type of tree you end up with depends partly on your data and partly on the method you choose to analyze it. Some methods naturally produce one tree type, while others can be adapted to produce several.

Distance-based approaches start with a matrix of pairwise distances between taxa (often calculated from sequence alignments). The classic UPGMA algorithm, which groups taxa by their average pairwise similarity, always produces an ultrametric tree. It does this by assuming a constant rate of evolution across all lineages, which makes it straightforward but unrealistic for most real datasets. Neighbor-joining, by contrast, does not assume equal rates and produces an unrooted additive tree (a phylogram). When researchers compared the two approaches alongside several other distance methods in simulations, neighbor-joining consistently recovered the correct tree more often.¹

Character-based methods work directly from aligned sequences rather than pre-computed distances. Maximum parsimony evaluates trees by counting the minimum number of character-state changes needed to explain the data. It does not use an explicit evolutionary model, which makes it conceptually simple but potentially misleading when evolution is heterogeneous. Maximum likelihood and Bayesian approaches, by contrast, evaluate trees by calculating the probability of the data given a specific model of sequence evolution. These probabilistic methods have increasingly replaced parsimony in the literature because they can accommodate complex substitution patterns and rate variation.³ Both maximum likelihood and Bayesian methods typically produce phylograms, but their output can be converted into ultrametric trees through downstream time-calibration.

Bayesian methods have an additional advantage when estimating divergence times because they can jointly infer the tree topology, branch lengths, substitution rates, and divergence dates in a single framework. Extensions of Bayesian approaches now allow relaxed clocks and the multispecies coalescent model to operate together, making it possible to estimate species divergence times even when the strict clock is violated and when gene trees disagree with the species tree.⁴

Turning a Phylogram into a Time Tree

A phylogram’s branch lengths are in units of substitutions per site, which is not the same as years. To get a time tree (an ultrametric tree with a time scale), you need calibration points: nodes in the tree whose approximate age is known from an independent source, usually the fossil record or a well-dated biogeographic event. The dating problem then becomes one of assigning ages to every node on the tree in a way that is consistent with the calibration points and the observed branch lengths.⁵

This conversion is far from mechanical. Age estimates can shift depending on which genes are analyzed, which parts of the genetic code are used, and whether the analysis assumes a constant rate across all lineages or allows rates to vary. In a study of land plants, researchers found that even sophisticated methods designed to handle rate variation across lineages could not fully account for groups with dramatically different evolutionary tempos.⁶ When one lineage has been evolving in slow motion for tens of millions of years while its relatives raced ahead, forcing all of them onto the same time scale introduces uncertainty that no amount of clever modeling fully eliminates. This is one reason researchers sometimes present phylograms and time trees side by side for the same dataset: the phylogram shows what the data look like in substitution space, and the time tree shows the best estimate of when events happened.

Historically, UPGMA-based trees (called phenograms in the numerical taxonomy tradition) were among the first molecular phylogenies. As one historical review notes, a tree constructed by UPGMA was originally used to represent phenotypic similarity, but it can serve as a molecular phylogeny when the rate of substitution is roughly constant.⁷ That “when” is doing a lot of work. In practice, the assumption of rate constancy rarely holds across a broad tree, which is why UPGMA has been largely supplanted by neighbor-joining for distance-based phylograms and by relaxed-clock Bayesian methods for time trees.

How Confident Should You Be in Any of These Trees

No matter which tree type you are looking at, the question of how well supported each branching point is matters enormously. The standard approach has long been nonparametric bootstrapping, where the dataset is resampled many times, a tree is built from each resampled dataset, and the proportion of resampled trees that recover a given branch is reported as its “bootstrap support.” A branch with support above roughly 70% is usually considered reasonably well supported, though the precise threshold is debated.

Bootstrapping is computationally expensive for large datasets, which has spurred the development of faster approximations. One widely used approach, called ultrafast bootstrap, dramatically reduces computation time while still providing meaningful support values for groups in maximum-likelihood trees.⁸ Bayesian methods offer an alternative: posterior probabilities for each branch, derived from the posterior distribution of trees sampled during the analysis. Recent work has suggested that for datasets with low levels of homoplasy (conflicting signal), Bayesian support values tend to be higher than frequentist bootstrap values and may better match the usual statistical interpretation of a confidence level, with a selection threshold around 85% to 90% performing well.⁹

Support values can be displayed on any of the four tree types, but their meaning interacts with tree type in subtle ways. A strongly supported branch on a cladogram tells you that the grouping is robust, but says nothing about how much change occurred along it. A strongly supported but very short branch on a phylogram might indicate a real but rapid divergence event, the kind of thing that is hard to resolve and prone to conflict among different genes. And a well-supported node on an ultrametric tree still has uncertainty in its date even if the topology is solid, because date estimation involves additional modeling assumptions beyond those that determine branching order.

Reading Trees Wrong

One of the most persistent problems in how people interpret phylogenetic trees, regardless of type, is what researchers have called the “primitive lineage fallacy.” It happens when someone reads a tree from left to right (or bottom to top) and assumes that lineages appearing earlier in the diagram are more “primitive” or “ancestral” than those appearing later. This mistake leads to misleading inferences about trait evolution and misrepresents the evolutionary process.¹⁰ A lineage that branches off near the base of a tree is not frozen in time; it has been evolving just as long as every other living species on the tree. On an ultrametric tree, this is visually obvious because all tips align at the present. On a cladogram with arbitrary branch lengths, it is easy to forget.

Another common misreading involves phylograms specifically. When someone sees a long branch, it is tempting to assume that lineage is “more evolved.” But a long branch on a phylogram only means more substitutions accumulated along it. That could reflect fast evolution driven by strong selection, a large effective population with lots of genetic drift, or simply a longer stretch of unsampled time between a lineage and its nearest relative in the dataset. Without additional context, you cannot distinguish these explanations from the branch length alone.

Ultrametric trees bring their own trap. Because they look clean and orderly, with all tips neatly aligned, they can give a false sense of certainty about divergence times. Every node on a time tree has a confidence interval, and for deep divergences those intervals can span tens of millions of years. The tidy diagram hides that uncertainty unless the researcher explicitly annotates it.

When No Tree Tells the Full Story

All four tree types share a fundamental limitation: they assume evolution is strictly bifurcating, meaning every ancestral lineage splits into exactly two daughter lineages and those daughters never exchange genetic material again. Real evolution violates this assumption regularly. Hybridization, introgression (where genes leak from one species into another through backcrossing), and horizontal gene transfer (common in bacteria) all create patterns that a simple branching tree cannot represent. Reticulate processes like these cannot be fully captured by a bifurcating tree.¹¹

The practical consequence is gene tree discordance: different genes in the genome tell different stories about which species are most closely related. This is now routinely observed in phylogenomic datasets. In the nightshade family, researchers found that incomplete lineage sorting (where ancestral genetic variation sorts randomly into descendant lineages) was the primary source of gene tree conflict, with hybridization playing a secondary role.¹² In the amaranth family, a combination of incomplete lineage sorting, ancient hybridization, and uninformative genes all contributed simultaneously.¹³ And in the large tropical order Malpighiales, the combination of very short internal branches, incomplete lineage sorting, estimation error, and ancient gene flow created what researchers described as a “perfect storm” of discordance that neither coalescent-based nor concatenation methods could fully resolve.¹⁴

When reticulation is important, the appropriate representation is not a tree at all but a phylogenetic network, which allows branches to merge as well as split. Networks are harder to estimate, harder to visualize, and harder to interpret than trees, which is part of why tree-based thinking still dominates most published studies. But for groups with frequent hybridization or horizontal transfer, presenting only a bifurcating tree, no matter how carefully estimated, can be actively misleading.

Dendrograms Outside Evolutionary Biology

The term “dendrogram” has a life of its own outside phylogenetics. Hierarchical clustering, one of the oldest unsupervised learning techniques, produces dendrograms as its natural output. You start with individual data points, merge the two most similar ones into a cluster, then merge the next most similar pair, and repeat until everything is joined in a single tree. The result is a dendrogram where branch heights reflect the distance or dissimilarity at which clusters merged.

In genomics, this is especially common for gene expression studies. When researchers cluster genes by their expression profiles across different conditions, they typically use algorithms like hierarchical clustering (with either Euclidean or correlation-based distances), K-means, or self-organizing maps, and the hierarchical approaches produce dendrograms.¹⁵ These dendrograms are often ultrametric by construction (because UPGMA or similar agglomerative algorithms enforce equal root-to-tip distances), but they carry no evolutionary meaning. The branch lengths represent similarity in expression pattern, not time or substitutions. Assessing whether the clusters in such dendrograms are statistically meaningful, rather than just artifacts of the algorithm, requires separate testing beyond the clustering itself.¹⁶

This overlap in visual form between evolutionary trees and clustering dendrograms is a genuine source of confusion. A student looking at a dendrogram of gene expression data and a phylogram of species relationships might struggle to articulate the difference, because the diagrams look almost identical. The distinction is entirely in the data and the model: one assumes an evolutionary process with descent and modification, while the other assumes only pairwise similarity and a merging algorithm.

File Formats and Software Considerations

If you work with any of these tree types in practice, you quickly encounter the question of how to store and share them digitally. The most widely used text format is Newick, which represents a tree as a nested series of parentheses with optional branch lengths and labels. A cladogram in Newick format simply omits the branch-length numbers; a phylogram includes them; and an ultrametric tree includes them with the constraint that they produce equal root-to-tip distances. The format is human-readable (barely), but it cannot encode rich annotations like support values, taxonomic metadata, or node colors without extensions.

NEXUS and phyloXML are more expressive formats that can store additional information alongside the tree. Several software libraries now provide unified tools for reading and writing all three formats, so that researchers can convert between them and visualize trees regardless of which format their analysis pipeline produced.¹⁷ Reference guides to tree visualization tools have documented the strengths and limitations of various software packages for displaying different tree types.¹⁸

One practical point worth knowing: most tree-viewing software lets you toggle between cladogram and phylogram display of the same underlying tree. If the tree file contains branch lengths, the software can render it as a phylogram (showing the lengths) or as a cladogram (ignoring them and spacing branches evenly). This means the distinction between cladogram and phylogram is sometimes a display choice rather than a data choice. An ultrametric tree, though, is a genuinely different object: it requires either that the tree was built under an equal-rate assumption or that it went through an explicit time-calibration step. You cannot just switch a phylogram to ultrametric display without changing the branch lengths themselves.

The Ultrametric Inequality

There is a clean mathematical way to tell whether a tree is ultrametric. For any three taxa on the tree, consider their three pairwise distances. In an ultrametric tree, the two largest of those distances are always equal. This property, sometimes called the ultrametric inequality, states that for any three entities i, j, and k, the distance between i and j is less than or equal to the maximum of the distances between i and k, and between j and k.¹⁹ If your distance matrix satisfies this condition for every possible trio, it corresponds to an ultrametric tree. If it does not, but satisfies a looser condition called the four-point condition, it corresponds to an additive tree (a phylogram). And if it satisfies neither exactly, as real biological data almost never do, then the tree is an approximation, and the method you use to build it determines how it handles the mismatch.

This mathematical distinction matters less for everyday interpretation than for understanding why different algorithms give you different tree types. UPGMA forces the ultrametric inequality to hold, which is why it always produces an ultrametric tree. Neighbor-joining permits the more relaxed additive condition, which is why it produces phylograms that can have unequal root-to-tip distances. The biology lives in the gap between these mathematical ideals and messy real-world data, and no tree type perfectly captures it.