Molecular Phylogenetics
Why do humans share 98.8% of their genome sequence identity with chimpanzees but far less with mice? The answer is not a matter of opinion. It is a measurable, testable historical signal preserved in DNA itself. Every heritable mutation that survives in a lineage becomes a permanent record of descent. Because DNA and protein sequences accumulate changes at roughly predictable rates, they function as molecular clocks. They also serve as archives of relatedness simultaneously. Molecular phylogenetics is the discipline that reads this archive. It aligns sequences, models how substitutions accumulate, and reconstructs the branching history of genes and species. This reconstruction relies on patterns of similarity and difference left behind. Before genome sequencing became routine, phylogenies were built almost entirely from morphology — bones, teeth, wings, flowers — interpreted through the logic of Cladistics & Maximum Parsimony. That approach remains powerful but is limited by the sparseness of comparable traits. It is also limited by convergent evolution of form. Morphological data is also totally absent for microbes, extinct DNA-less lineages, and single-celled organisms. Molecular data changed the scale of the problem. Instead of dozens of morphological characters, biologists now compare thousands of nucleotide or amino acid positions. Each position behaves as an independent (if noisy) character with its own explicit probabilistic model of change. The big picture is this: mutation is essentially clock-like on average. Drift and selection shape which mutations survive. Shared derived changes reveal common ancestry just as shared derived traits do in cladistics. But this happens at a resolution and quantitative rigor morphology cannot match. Molecular phylogenetics therefore does not replace the logic of common descent. Instead, it supplies a far richer, more objective dataset. It also provides a statistical machinery for testing competing trees against it. This article covers how DNA and protein sequences are turned into alignments. It explains how substitution models correct for hidden multiple changes at a site. It shows how those corrected distances or character patterns are converted into trees. It also covers the classic evidence — from immunological distance experiments to genome comparisons — that validated the entire enterprise. It sets up the more advanced statistical frameworks of maximum likelihood and Bayesian inference. It also introduces the grand synthesis of the tree of life covered elsewhere.
- Molecular phylogenetics reconstructs evolutionary trees from DNA/protein sequence data rather than morphology alone.
- Homologous sequences must first be aligned, position by position, before any comparison is meaningful.
- Raw sequence differences underestimate true evolutionary change because of unseen multiple/back substitutions — substitution models (e.g., Jukes-Cantor, Kimura two-parameter) correct for this.
- Trees can be built by distance methods (e.g., neighbor-joining) or character-based methods (parsimony, and later likelihood/Bayesian approaches).
- The molecular clock hypothesis, tested and refined since the 1960s, lets sequence divergence be converted into approximate divergence times.
- Classic evidence — Sarich & Wilson's immunological clock, and decades of concordant gene trees — confirmed molecular data recover the same deep history as fossils and morphology, with far greater resolution.
DNA and protein sequences are historical documents written in a nearly-clock-like code, and molecular phylogenetics is the set of alignment and substitution-model tools that lets us read that code to reconstruct the true branching history of life.
The core idea
A molecular phylogeny is a hypothesis about the branching history of a set of homologous sequences — DNA, RNA, or protein sequences descended from a common ancestral sequence. Just as cladistics infers relationships from shared derived morphological states (synapomorphies), molecular phylogenetics infers relationships from shared derived sequence states: shared nucleotide substitutions, insertions, deletions, or amino acid replacements that arose once in an ancestor and were inherited by its descendants. The unit of comparison shifts from an anatomical character (number of vertebrae, presence of a notochord) to a single alignable position in a sequence — a site — of which a modern dataset may contain many thousands, each evolving semi-independently. Two central complications distinguish molecular data from morphological data and drive the entire methodology. First, sequences must be aligned: because insertions and deletions (indels) occur over evolutionary time, homologous nucleotides do not automatically sit in the same column across taxa, and a correct alignment must be inferred before homologous sites can be compared at all. Second, sequence evolution is not clock-perfect in a naive sense — the same site can mutate more than once, mutate and then mutate back, or undergo parallel changes in unrelated lineages (homoplasy at the molecular level, exactly analogous to convergent evolution of a morphological trait). Because DNA has only four possible states (A, C, G, T) at each site, this "multiple hits" problem is severe: two distantly related sequences that have each accumulated many substitutions can appear only modestly different simply because changes overwrite one another. Substitution models are explicit statistical models of how sites change state over time, and they are used to correct observed sequence differences into estimates of the true number of substitutions that occurred — the key innovation that separates molecular phylogenetics from simply counting mismatches. Once an alignment and a substitution model are in hand, sequences can be converted into either (a) pairwise genetic distances, used by distance methods such as neighbor-joining, or (b) treated directly as discrete characters, used by parsimony, likelihood, and Bayesian methods (the latter two are developed fully in Maximum Likelihood & Bayesian Phylogenetics). A further extension, the molecular clock hypothesis, proposes that for many genes substitutions accumulate at an approximately constant rate over time within a lineage, which — if calibrated against fossil-dated splits — converts a tree's branch lengths into estimated divergence times in millions of years. This does not hold universally (rates vary by gene, lineage, and generation time), so clock assumptions are always tested, not presumed.
Mechanism
Step 1 — Choose homologous sequences. The starting material must be genuinely homologous: either orthologs (copies related by a speciation event, the right choice for a species tree) or, if care is taken, paralogs (copies related by a gene duplication event, useful for studying gene family history but dangerous if mistaken for orthologs — a classic source of "gene tree/species tree" mismatch). Mitochondrial genes (e.g., cytochrome b, COI — the standard DNA-barcoding gene), ribosomal RNA genes (18S, 16S), and single-copy nuclear genes are common choices because their orthology is relatively easy to establish.
Step 2 — Multiple sequence alignment. Sequences from different taxa are arranged into a matrix so that each column represents (a hypothesis of) one historically homologous site. Alignment algorithms (e.g., progressive alignment as in ClustalW/MUSCLE, or probabilistic hidden Markov model aligners) insert gap characters to accommodate insertions/deletions. A gap-heavy or ambiguous region is often trimmed or excluded, because misaligned columns inject false homoplasy into every downstream step. For protein-coding genes, alignment is usually done on translated amino acid sequences (which tolerate indels better and preserve reading frame) and then mapped back onto the codons.
Step 3 — Choose a substitution model. The simplest model, Jukes-Cantor (JC69), assumes all four nucleotides are equally frequent and all substitution types (any base to any other) are equally likely. More realistic models relax these assumptions:
| Model | Key feature |
|---|---|
| Jukes-Cantor (JC69) | equal base frequencies, one substitution rate |
| Kimura 2-parameter (K80) | separates transitions (A↔G, C↔T) from transversions (purine↔pyrimidine), since transitions occur more often |
| HKY85 / General Time-Reversible (GTR) | unequal base frequencies, up to 6 distinct rate classes, still time-reversible |
| GTR + Γ (gamma) | adds rate heterogeneity across sites (some sites evolve faster than others, e.g., third codon positions) |
The reason a model matters: with only 4 nucleotide states, a site can change and change back (A→G→A), or two lineages can independently hit the same state by chance. The observed proportion of differing sites (p-distance) therefore systematically underestimates the true number of substitutional events, especially over long branches. A substitution model provides a correction formula. For Jukes-Cantor, if p is the observed proportion of differing sites, the corrected distance d (expected substitutions per site) is:
d = −(3/4) ln(1 − 4p/3) At small p this correction is negligible (d ≈ p), but as p rises toward the saturation value of 0.75 (the point at which sequences look random relative to each other), the correction diverges toward infinity — a mathematical warning that highly divergent sequences carry rapidly diminishing phylogenetic signal, a phenomenon called saturation.
Step 4 — Build the tree. Two broad strategies:
- Distance methods (e.g., neighbor-joining) take the full matrix of pairwise corrected distances and algorithmically cluster taxa to minimize total tree length, producing a single tree quickly — fast, but throws away the original character-by-character information.
- Character-based methods — maximum parsimony (as in cladistics), maximum likelihood, and Bayesian inference — evaluate candidate trees directly against the aligned character matrix, asking which tree(s) best explain (or maximize the probability of) the observed pattern of sequence states. Likelihood and Bayesian methods use the substitution model explicitly to compute the probability of the data given a tree and branch lengths, an extension developed fully in Maximum Likelihood & Bayesian Phylogenetics.
Step 5 — Assess confidence. Because a single dataset gives one best tree, biologists resample the data — typically via the bootstrap: columns of the alignment are resampled with replacement to build many pseudo-replicate datasets, a tree is built from each, and the percentage of replicates recovering a given branch (bootstrap support) indicates how robust that branch is to sampling noise in the data.
The evidence
Long before genome sequencing was cheap, the first rigorous molecular phylogenetic evidence came from immunological distance studies. In the late 1960s, Vincent Sarich and Allan Wilson at Berkeley measured the immunological cross-reactivity of blood serum albumin between primate species: an antibody raised against human albumin was tested against albumins from other species, and the strength of the immune cross-reaction was used as a proxy for how similar (and therefore how recently diverged) the albumin protein sequences were. Assuming this molecular difference accumulated at a roughly constant rate — the molecular clock hypothesis, which Sarich and Wilson's data were central in establishing — they calibrated the albumin clock against fossil-dated splits within the primate tree and calculated that the human/chimpanzee/gorilla lineage had split from the orangutan lineage around 10–13 million years ago, and that humans and chimpanzees themselves diverged startlingly recently, only around 4–6 million years ago. This was a direct, quantitative challenge to prevailing paleoanthropological opinion of the time, which — based on isolated fossil interpretations (notably of Ramapithecus, since reassigned) — held that the human lineage had split from other apes 15–30 million years ago or more. The molecular result was initially met with considerable skepticism from paleontologists. But subsequent fossil discoveries (better-dated hominin and ape fossils) and, decisively, whole-genome comparisons beginning in the 1990s–2000s vindicated the molecular estimate: humans and chimpanzees are now understood to share a common ancestor roughly 6–7 million years ago, squarely within the range the molecular clock had predicted decades before direct genome sequencing was possible. This episode is one of the clearest cases in evolutionary biology where a molecular method made a falsifiable, numerically precise prediction that was later confirmed by independent fossil and genomic evidence, and it is often cited as the moment molecular data earned full standing alongside morphology as primary phylogenetic evidence. A second, complementary line of evidence is the congruence test: independent genes, evolving under different functional constraints, in different genomic compartments (mitochondrial vs. nuclear), analyzed with different substitution models, overwhelmingly recover the same branching topology for a given set of species — and that topology, in turn, overwhelmingly matches trees built independently from morphology and from the fossil record's temporal order of first appearances. This congruence across independent data sources is precisely the pattern expected if all these sequences and traits are tracking one true underlying history of common descent, and it is exceedingly difficult to explain under any alternative hypothesis. Modern genome-wide studies extend this: the human genome project and subsequent primate genome sequencing confirmed at the level of millions of aligned base pairs what Sarich and Wilson had inferred from a handful of protein comparisons — human and chimpanzee genomes differ at roughly 1.2% of aligned single-nucleotide sites, consistent with the multi-million-year divergence time the immunological clock had proposed.
Worked examples
Example 1 — Correcting an observed distance for multiple hits (Jukes-Cantor). Two lineages have been evolving independently since a common ancestor. Across a 1,000 bp aligned mitochondrial gene region, 250 sites differ between them (p = 0.25). A naive biologist might report "25% divergence" and stop there. Applying the Jukes-Cantor correction: d = −(3/4) ln(1 − 4(0.25)/3) = −(3/4) ln(1 − 0.3333) = −(3/4) ln(0.6667) = −(3/4)(−0.4055) ≈ 0.304 So the true estimated number of substitutions per site is about 0.304, roughly 22% higher than the raw 0.25 observed — because some sites have mutated more than once (or mutated and reverted), silently erasing evidence of change. Now suppose a third, more distantly related lineage shows p = 0.60 relative to the first. Plugging in: d = −(3/4) ln(1 − 0.8) = −(3/4) ln(0.2) ≈ −(3/4)(−1.609) ≈ 1.207 — more than double the raw p-value, illustrating how the correction grows sharply nonlinear as sequences approach saturation (p → 0.75, where the natural log term goes to −∞ and the corrected distance is undefined). This is exactly why highly divergent, ancient splits (e.g., resolving relationships among bacterial phyla) require slower-evolving genes (like ribosomal RNA) rather than fast-saturating ones (like mitochondrial control regions), which retain phylogenetic signal only over shorter timescales.
Example 2 — Reading a small molecular tree and reconciling it with morphology. A researcher sequences the same 800 bp nuclear gene in four bird species — Falcon, Hawk, Owl, and Duck — plus an outgroup, Ostrich. After alignment and GTR-model distance correction, neighbor-joining produces the topology: (Ostrich, ((Owl,(Falcon,Hawk)),Duck)) This says Falcon and Hawk are sister taxa (most recently diverged), that clade's sister is Owl, and Duck is the next outgroup within the ingroup, with Ostrich as the true outgroup rooting the tree. Historically, "raptors" (falcons, hawks, owls) were grouped together largely on the shared derived morphology of hooked beaks and talons — a classification going back to Linnaeus. But bootstrap support for a Falcon+Hawk clade nested with Owl (rather than falcons instead grouping with songbirds and parrots) turned out, in real large-scale genomic studies of birds, to be weak or contradicted: falcons are now understood (via genome-scale molecular phylogenetics) to be more closely related to parrots and songbirds than to hawks, with the raptorial morphology having evolved convergently. This is a genuine historical case where molecular data overturned a morphology-based classification, because talons and hooked beaks are adaptations to a raptorial lifestyle that can evolve independently (homoplasy) in unrelated lineages, while sequence data — properly modeled — is far less prone to this particular kind of convergent bias. The lesson: a well-supported molecular tree can and should override a morphology-based tree when the morphological characters are plausibly under strong, convergent selective pressure.
Common misconceptions & traps
| Misconception | Why it's wrong |
|---|---|
| "More sequence difference always means more evolutionary time." | True only if rates are constant (clock-like) and sites aren't saturated; different genes and lineages evolve at very different rates (e.g., mitochondrial DNA evolves faster than most nuclear genes). |
| "Percent sequence identity is the actual number of mutations that occurred." | Raw p-distance systematically underestimates true substitutions due to unseen multiple hits and reversions — this is exactly what substitution models correct for. |
| "A gene tree is automatically the species tree." | Gene duplication, incomplete lineage sorting, and horizontal transfer can make an individual gene's history differ from the true species-splitting history; multiple independent genes must be compared. |
| "Alignment is a trivial bookkeeping step." | Alignment is itself a phylogenetic hypothesis about indel history; a bad alignment guarantees a bad tree regardless of how sophisticated the downstream model is. |
| "Molecular data always trumps morphology." | Molecular data is powerful but not infallible — long-branch attraction, saturation, and model misspecification can produce confidently wrong molecular trees; congruence across multiple independent genes and with fossil calibration is what builds real confidence. |
The scientist behind it
Allan Wilson (1934–1991), a New Zealand-born biochemist working at UC Berkeley, was among the first to argue — against considerable initial resistance from paleontologists — that molecules themselves could serve as a primary, quantitative record of evolutionary history rather than a mere supplement to fossils. Working with graduate student Vincent Sarich, Wilson used immunological cross-reactivity of blood proteins to build the primate molecular clock, producing the startling and initially controversial estimate that humans and chimpanzees diverged only a few million years ago. Wilson went on to help pioneer mitochondrial DNA phylogenetics, co-authoring the 1987 "mitochondrial Eve" study that used human mtDNA diversity to argue for a recent African origin of modern humans — a direct methodological descendant of the same molecular-clock logic he had championed for primates two decades earlier. His career exemplifies the shift, over the second half of the twentieth century, from morphology-only systematics toward a synthesis in which molecular sequence data became a primary, independently testable line of phylogenetic evidence.
Origins · 1 minAllan C. Wilson and how this was discovered
Where this matters today
Molecular phylogenetics is now the backbone of practical work across biology and medicine. In public health, real-time phylogenetic trees built from viral genome sequences (e.g., SARS-CoV-2, influenza, HIV) let epidemiologists trace transmission chains, identify superspreading events, and monitor the emergence of new variants almost as fast as the virus spreads — a direct application of the same alignment-plus-substitution-model machinery described above, applied to viral rather than species evolution. In conservation biology, DNA barcoding (short, standardized gene regions such as COI) and phylogenetic placement let regulators identify illegally traded wildlife products (ivory, shark fin, bushmeat) down to species, and let taxonomists discover previously unrecognized "cryptic species" hidden within what looked morphologically like a single species. In agriculture, phylogenetic analysis of crop wild relatives guides breeders toward useful genetic diversity for disease resistance and drought tolerance, and tracks the spread of crop pathogens and pesticide-resistance alleles. In medicine and biotechnology, phylogenetic methods reconstruct the history of gene families (revealing which genes are true orthologs suitable for functional inference across species, as opposed to paralogs that have diverged in function), inform the design of vaccines against rapidly evolving pathogens by identifying conserved epitopes, and underlie forensic and paternity applications that rely on the same distance and substitution logic at a population level. All of these applications sit downstream of the more powerful statistical frameworks — Maximum Likelihood & Bayesian Phylogenetics — that build directly on the alignment and substitution-model foundations introduced here, and all ultimately feed into the grand synthetic project of the Tree of Life.
Ask the tutor: How do maximum likelihood and Bayesian methods use substitution models to actually calculate which tree is most probable, rather than just correcting distances?
Why can't we just count the number of differing nucleotides and call that "the distance" between two sequences?
Because DNA has only four possible states per site, a single site can be hit by more than one substitution over long evolutionary time — it can mutate and then mutate again, or mutate and then revert to its original state. Both events leave no visible trace in a simple comparison of two present-day sequences: a site that changed twice, or changed and reverted, looks identical to a site that never changed. Raw percent difference (p-distance) therefore systematically underestimates the true number of substitutional events, especially for anciently diverged or fast-evolving sequences. Substitution models such as Jukes-Cantor or GTR provide a mathematical correction, converting the observed proportion of differing sites into a statistically corrected estimate of actual evolutionary distance, which grows increasingly important — and increasingly uncertain — as sequences become more divergent.
What is the difference between an ortholog and a paralog, and why does it matter for building a species tree?
Orthologs are gene copies in different species that trace back to a common ancestor through a speciation event — they are the correct comparison for reconstructing the history of species splitting apart. Paralogs are gene copies within a genome (or across genomes) that trace back to a gene duplication event, meaning their divergence reflects when the duplication occurred, not when the species split. If a paralog is mistakenly compared as though it were an ortholog, the resulting "gene tree" can differ substantially from the true species tree, because the duplication may have occurred well before or after the relevant speciation events. Careful ortholog identification, and ideally comparison across several independent genes, is essential before building a tree meant to represent species-level history.
What does the molecular clock hypothesis actually claim, and is it always true?
The molecular clock hypothesis proposes that, for a given gene or protein, substitutions accumulate at an approximately constant rate over time within and across lineages, so that sequence divergence can be converted into an estimated divergence time once calibrated against at least one fossil-dated split. It is a useful approximation for many genes over moderate timescales but is not universally true: substitution rates vary with generation time, metabolic rate, population size, and the strength of selection acting on a given gene, so different lineages and different genes often show different rates ("rate heterogeneity"). Modern methods (relaxed molecular clocks) explicitly allow rates to vary across branches of a tree rather than assuming strict constancy, and clock-like behavior is always tested statistically rather than simply assumed.
How did molecular data actually change our understanding of the human-chimpanzee divergence time?
Before the 1960s, paleoanthropological interpretation of ape and hominin fossils suggested humans split from other apes perhaps 15-30 million years ago or more. Sarich and Wilson's 1967 immunological distance study of blood albumin, calibrated as a molecular clock, instead predicted a much more recent split of roughly 4-6 million years ago — a figure many paleontologists initially rejected. Subsequent fossil reinterpretation and, decisively, direct whole-genome sequence comparisons beginning in the 2000s confirmed the molecular estimate, showing human and chimpanzee genomes differ at only about 1.2% of aligned sites, consistent with a divergence roughly 6-7 million years ago. This is a landmark case of a molecular method making a falsifiable numerical prediction later vindicated by independent evidence.
Why do biologists use the bootstrap, and what does a bootstrap value of 95% actually mean?
Any single dataset produces one "best" tree, but that tree is a statistical estimate subject to sampling noise, since only a finite number of sites were sequenced. The bootstrap resamples columns of the alignment with replacement, generating many pseudo-replicate datasets of the same size as the original, and rebuilds a tree from each replicate. The percentage of these pseudo-replicate trees that recover a given branch (clade) is that branch's bootstrap support value. A bootstrap value of 95% means that 95% of resampled datasets, drawn from the same underlying signal, still support that particular grouping — indicating the branch is well-supported by the data, though it does not by itself guarantee the branch reflects the true evolutionary history if the underlying model is badly misspecified.
Can molecular phylogenetics contradict a well-established morphology-based classification, and should we trust it when it does?
Yes, and this has happened in real, well-documented cases — for example, genome-scale bird phylogenies show falcons group with parrots and songbirds rather than with hawks and eagles, despite sharing a raptorial hooked-beak-and-talon body plan long used to classify them together. Morphological convergence under similar selective pressures (a predatory lifestyle) can produce misleading shared traits that are not due to common ancestry (homoplasy), while properly modeled sequence data, especially when many independent genes agree, is generally less susceptible to this particular bias. The strongest confidence comes not from molecular data automatically overriding morphology, but from congruence: when multiple independent genes, and ideally fossil timing, all point to the same topology, that convergent evidence is far more persuasive than either data type alone.
What's the practical difference between distance methods like neighbor-joining and character-based methods like parsimony or likelihood?
Distance methods first collapse the entire aligned sequence matrix into a single number per pair of taxa (a corrected genetic distance), then use an algorithm like neighbor-joining to build a tree that best fits that matrix of distances — this is fast but discards information about exactly which sites changed and in what pattern. Character-based methods (maximum parsimony, maximum likelihood, Bayesian inference) instead evaluate candidate trees directly against the original character-by-character alignment, asking which tree topology best explains, or maximizes the probability of, the actual observed pattern of shared and differing states at each site. Character-based methods generally retain more information and can be more accurate, but likelihood and Bayesian versions in particular are far more computationally intensive, which is why fast distance methods remain useful for large datasets or initial exploratory trees.