34 — Phylogenetics
Before this: Ch 33 · Ch 27 · Ch 16 · Time: ~55 min
Statistics needed: S1 Probability · S3 Sampling and estimation · S6 Likelihood and Bayes
Phylogenetics is statistical inference over a discrete structure. The parameter you want is a labelled binary tree; the data are aligned sequences; the likelihood is a continuous-time Markov chain running along the branches. Everything hard about the field comes from one of two places — the parameter space is astronomically large and non-Euclidean, or the model is wrong in a way that gets more confident as you add data.
What you'll be able to do
- Read a tree correctly: distinguish topology from , rooted from unrooted, and monophyly from paraphyly and polyphyly; derive the number of unrooted topologies on n taxa and explain why exhaustive search is not an option
- Diagnose the pre-tree errors no support value can detect — mistaken for , and progressive-alignment error — and explain why each yields high confidence in a wrong tree
- Derive the distance correction from its rate matrix, and say what K80, HKY, GTR, +Γ and +I each add
- Explain the failure mode of every major method: UPGMA's clock, 's long-branch attraction, ML/Bayesian model misspecification
- State precisely what a bootstrap value and a each measure, and why they are not comparable
- Explain why time-reversibility leaves the root unidentified from sequence alone, and why a divergence date's credible interval converges on a nonzero width set by calibration rather than by sequence length
- Compute the expected coalescence times for a sample, explain why the deepest coalescences dominate a genealogy, and enumerate the causes of gene-tree/species-tree discordance, saying which are symmetric and which are not
The core idea
Sequences that share an ancestor share its errors. Two species that diverged recently have had little time to accumulate independent , so their sequences are similar; two that diverged long ago have had more. If mutations accumulated at a steady rate and never hit the same site twice, reconstructing history would be a clustering problem you could solve in an afternoon.
Both assumptions fail. Rates differ enormously between lineages and between sites, and with only four bases, a site that has changed twice can look unchanged. Observed difference saturates: it is a compressed, nonlinear function of elapsed evolutionary time. Every method in this chapter is, at bottom, an attempt to undo that compression — and every characteristic failure in this chapter is what happens when the attempt is made with the wrong model.
1. The vocabulary, and why exhaustive search is hopeless
A phylogeny is a graph. Tips (leaves, terminal nodes) are the observed sequences. Internal nodes are inferred common ancestors. Branches (edges) connect them, and a branch length is normally measured in expected substitutions per site — not in years, and not in generations.
Two things are being estimated and they are logically independent:
| What it is | How many parameters | |
|---|---|---|
| Topology | Which tips group with which — the discrete combinatorial structure | 1 discrete parameter, from an enormous set |
| Expected substitutions per site on each edge | 2n − 3 continuous parameters (unrooted, n tips) |
A rooted tree has a designated oldest node and every branch acquires a direction, so statements about ancestry and about the order of events become meaningful. An unrooted tree asserts relationship without direction. Almost all substitution models are time-reversible, which means the likelihood is identical wherever you put the root: sequence data alone cannot root a tree. Rooting is always an extra assumption (§9).
A clade (monophyletic group) is an ancestor plus all of its descendants. The two failures:
- Monophyletic: {A, B} — an ancestor and all its descendants. Also {A, B, C}.
- Paraphyletic: {A, B, D} — omits C, which shares the same ancestor. "Reptiles" excluding birds, and "prokaryotes" excluding eukaryotes, are the canonical examples.
- Polyphyletic: {A, D} grouped on some shared feature whose common ancestor is not in the group — convergence, not descent.
Only monophyly is a claim about history. The other two are claims about a naming convention.
Counting topologies
Build the tree by adding taxa. An unrooted binary tree on n tips has n − 2 internal nodes and 2n − 3 branches. Adding tip n + 1 means splitting one existing branch, so there are exactly 2n − 3 places to put it, each giving a distinct topology. With U(3) = 1:
Rooting adds one more choice — the root can go on any of the 2n − 3 branches — so R(n) = (2n − 3)!!, exactly the unrooted count for n + 1 taxa.
| n taxa | Unrooted topologies |
|---|---|
| 4 | 3 |
| 5 | 15 |
| 10 | 2,027,025 |
| 20 | 2.2 × 10²⁰ |
| 30 | 8.7 × 10³⁶ |
| 50 | 2.8 × 10⁷⁴ |
| 53 | 2.8 × 10⁸⁰ |
At 53 taxa there are more unrooted topologies than the usual order-of-magnitude estimate for atoms in the observable universe (~10⁸⁰). Real studies use hundreds to tens of thousands of taxa. Exhaustive search is over at n ≈ 11; branch-and-bound at n ≈ 20. Everything beyond that is heuristic hill-climbing over tree space using local rearrangement moves (NNI, SPR, TBR), and finding the optimal tree under either or likelihood is NP-hard. You are never guaranteed the best tree — only a good local optimum, which is why serious analyses run many independent searches from different starting points.
Trees are serialised in Newick format: ((A:0.30,B:0.05):0.05,(C:0.30,D:0.05)); — nested parentheses, colon-prefixed branch lengths, terminating semicolon.
2. Homology: the ortholog/paralog trap
Two sequences are homologous if they descend from a common ancestral sequence. is binary — there is no "70% homologous"; you mean 70% identical. Homologs come in two kinds, distinguished by the event at their common ancestor:
| Divergence event | What the records | |
|---|---|---|
| Speciation | The history of the species | |
| Gene duplication | The history of the gene family, which is older |
This distinction is not pedantry; it is the single most common way to get a badly wrong answer. Suppose a gene duplicated in the common ancestor of all vertebrates into copies α and β. Human α, mouse α and fish α are orthologs. Human α and human β are paralogs. If your "human sequence" is α and your "mouse sequence" happens to be β, the deepest split in your tree is the duplication, not the human–mouse speciation — and it will be placed far too deep, with total confidence. The tree is not wrong as a gene tree. It is simply not the tree you thought you were building.
Practical defences: use or a curated orthology database, include enough species that paralogous clades become visible as duplications rather than as odd placements, and treat any single-gene tree whose topology contradicts well-established species relations as an orthology problem until proven otherwise. Ch 35 develops duplication and orthology properly.
3. Alignment: the error-prone first step
Every method below takes an alignment as input — a matrix in which each column is asserted to be a set of homologous positions descended from one ancestral site. That assertion is an inference, produced by a separate algorithm, and it is then treated downstream as data.
A ACGT-TTGACC
B ACGTATTGACC
C ACG--TTAACC
D ACGCATT-ACC
↑ ↑
gaps assert insertion/deletion events
Pairwise alignment is a solved dynamic-programming problem, but optimal multiple alignment under a sum-of-pairs objective is NP-hard, so practical tools are heuristic. The heuristic is progressive alignment (Clustal, MAFFT, MUSCLE): build a quick guide tree from pairwise similarities, then align the most similar pairs first and progressively merge alignments, treating each merged block as fixed. Two consequences follow directly from that description. Early mistakes are never revisited — the greedy schedule has no backtracking. And the guide tree, a crude estimate of the very thing you are about to infer, biases the alignment toward itself.
Alignment error is the most under-reported source of error in phylogenetics. It is systematic, not random, so it does not average out with more sites, and — this is the part that matters — it is invariant across bootstrap replicates, so it produces high support for the wrong answer rather than low support.
Poorly aligned regions inflate apparent divergence, and inflated divergence lands preferentially on already-divergent taxa, feeding the long-branch problem in §6. Standard mitigations: align protein sequence and back-translate for coding genes (the amino acid alphabet has 20 states, so is far less ambiguous), mask ambiguous columns with Gblocks/trimAl, or use statistical alignment methods that co-estimate alignment and tree and report the uncertainty instead of hiding it.
4. From observed difference to distance: substitution models
Model each site as a continuous-time Markov chain on the state space {A, C, G, T} with rate matrix Q; the transition probability matrix over t is P(t) = eQt.
Deriving JC69
(1969) is the simplest case: equal base frequencies, and every substitution occurs at the same rate α. Let p(t) be the probability that a site is in its original state after time t. In an instant dt the site leaves its current state at total rate 3α and arrives from the other three states at rate α each:
With p(0) = 1 this linear ODE solves to
which behaves correctly at both limits: p(0) = 1, and p(∞) = 1/4, the equilibrium frequency.
That chain follows a single lineage; what we observe is two tips. Because JC69 is time-reversible, these are the same problem: a pair of tips separated by an ancestor is statistically identical to one lineage traversed for the summed branch length. So read t as the whole path, and the probability that two sequences differ at a site is q = 1 − p(t).
Now the key change of variable. The evolutionary distance d is the expected number of substitutions per site along the whole path connecting the two sequences — for two sequences separated by an ancestor, the sum of both branches. Since the total substitution rate per site is 3α, d = 3αt, so 4αt = 4d/3 and
Ch 33 derives the same estimator the long way, tracking both lineages jointly (dp/dt = 6α − 8αp) with t measured per lineage, so d = 2 × 3αt = 6αt there against d = 3αt here. Identical formula; the only difference is whether t counts one branch or the whole path. Reversibility is what lets you choose.
Read the properties off the formula. For small q, d ≈ q — few sites hit twice, so raw difference is fine. As q → 3/4 the logarithm diverges: at 75% difference the sequences are saturated, indistinguishable from random, and the distance is undefined, not merely large. Above 3/4 the estimator is undefined outright, which is exactly the right behaviour and a good reason to prefer it to a raw p-distance. The delta method gives Var(d̂) = q(1−q) / [n(1 − 4q/3)²], which blows up in the same limit.
The model hierarchy
Each model relaxes one assumption. They are nested, so likelihood-ratio tests apply.
| Model | Adds | Free parameters |
|---|---|---|
| JC69 | — equal rates, equal frequencies | 0 |
| K80 (Kimura 2-parameter) | transition/transversion ratio κ | 1 |
| HKY85 | unequal base frequencies πA, πC, πG, πT | 4 |
| GTR | all six exchangeability rates free (time-reversible) | 8 |
Transitions (A↔G, C↔T — ↔purine, ↔pyrimidine) outnumber transversions typically 2–5 fold, for the chemical reasons in Ch 16. Ignoring that, as JC69 does, systematically underestimates distance.
Two additions matter more than the choice of exchangeability matrix:
+Γ — rate heterogeneity across sites. Sites do not evolve at one rate: third positions, Ch 07, change far faster than first and second. Let the relative rate at a site be a draw from a Gamma(α, α) distribution — mean fixed at 1, variance 1/α, so small α means extreme heterogeneity and α → ∞ recovers a uniform rate. In practice the continuous distribution is discretised into 4 equal-probability categories (Yang 1994) and the site likelihood is averaged over them, at 4× the compute. This is usually the single most important term in the model. Ignoring rate heterogeneity underestimates multiple hits at fast sites, compressing long distances more than short ones — the same compression that drives long-branch attraction.
+I — invariant sites. A proportion of sites cannot change at all (structurally constrained). Note that +I and +Γ are partly confounded: a Gamma with small α already places mass near zero. Fitting both is common but the two parameters trade off against each other, and the individual estimates should not be over-interpreted.
Model selection is ordinary statistics, covered in S6 §4 and §8: fit the candidate set on a fixed reasonable tree, compare with AIC, BIC or LRTs for nested pairs (ModelTest, jModelTest, ModelFinder). Under- parameterisation is a systematic bias that grows with data; over-parameterisation merely costs variance. When in doubt, richer. The bigger decision is usually partitioning — allowing separate models per gene or per codon position — not the exchangeability matrix within a partition.
5. Distance methods: UPGMA and neighbour-joining
Reduce the alignment to a matrix of pairwise corrected distances, then build a tree from it. Fast (O(n³) or better), and the only realistic option for tens of thousands of sequences. The information loss is real: n(n−1)/2 numbers replace an alignment of n × L characters.
Two properties of a distance matrix determine what you can do with it:
- Additive: distances are exactly path lengths in some tree. Equivalent to the four-point condition — for any four taxa, the three pairwise sums d(A,B)+d(C,D), d(A,C)+d(B,D), d(A,D)+d(B,C) include a smallest one and two equal larger ones. The smallest identifies the true pairing.
- Ultrametric: additive and all tips equidistant from the root — a .
UPGMA repeatedly joins the closest pair and places their node at half their distance. It reconstructs the true tree only if the matrix is ultrametric. Without a clock it fails, and it fails in a specific direction: two slowly evolving lineages look similar and get joined even when they are not sisters. The worked example below shows exactly this. UPGMA is fine for a quick dendrogram and wrong as a phylogenetic method.
Neighbour-joining (Saitou & Nei 1987) drops the clock. It joins the pair minimising
The r terms are the correction: a taxon on a long branch has a large r and is penalised for it, so NJ asks "are i and j close relative to how far each is from everything else", not "are they close". NJ is statistically consistent — on an additive matrix it recovers the true tree exactly, and since corrected distances converge to additive ones as sequence length grows, NJ converges on the truth. Consistency is inherited from the distance correction, which is why the correction is not optional. NJ greedily minimises tree length, an approximation to the minimum-evolution criterion.
6. Parsimony, and long-branch attraction
Maximum picks the topology requiring the fewest substitutions, scored per site by Fitch's algorithm (one post-order pass, union/intersection of state sets, count the unions). It has no explicit model, which was long sold as a virtue. It is not one — it is an implicit model, and a badly specified one.
Consider four taxa with the true topology ((A,B),(C,D)), where A and C sit on long branches and B, D and the internal branch are short. Let p be the probability of a visible change on a long branch and q on the internal branch, with q ≪ p.
- The site pattern supporting the true grouping requires a change on the short internal branch: frequency ≈ q.
- The pattern supporting the wrong grouping (A with C) arises from parallel changes on the two long branches landing on the same base: frequency ≈ p² × 1/3.
So whenever p²/3 > q, the misleading pattern is genuinely more common in the data. With p = 0.35 and q = 0.02: 0.041 > 0.02, twice as frequent. Parsimony counts patterns, so it picks the wrong tree — and since the ratio of pattern frequencies is a constant, more sites give more confidence in the wrong answer. Parsimony is statistically inconsistent in this region of parameter space (the Felsenstein zone): it converges on the wrong tree with probability → 1.
This is long-branch attraction, and the mechanism generalises. Any method that under-corrects for multiple hits compresses long branches more than short ones and pulls them together — including distance methods with a too-simple model, and ML with a misspecified one. Diagnostics: check whether the suspect grouping is the pair of longest branches; break the long branches by adding taxa that subdivide them; refit under a richer model with +Γ; compare parsimony and ML topologies and treat disagreement as informative rather than as a tie.
7. Maximum likelihood and Bayesian inference
Statistics: the likelihood function, maximum likelihood and the prior → posterior update are covered in S6 §1–§5; what a posterior does and does not license is §6.
Maximum likelihood treats the tree and as parameters and maximises P(alignment | topology, branch lengths, model), assuming sites are independent so the log-likelihood sums over columns.
The engine is Felsenstein's pruning algorithm, which is dynamic programming and will look familiar. Define Lk(s) = P(data at tips below node k | node k is in state s). For an internal node with children u, v at branch lengths tu, tv:
Tips initialise to an indicator vector; the site likelihood is Σs πs Lroot(s). Cost is O(n s²) per site with s = 4 states — linear in taxa, which is why ML is feasible at all. Summing over all 4n−1 ancestral state assignments naively is not.
ML with a correctly specified model is consistent and efficient; it is not immune to long-branch attraction when the model is wrong, which is the usual situation. RAxML and IQ-TREE are the standard implementations.
Bayesian inference targets P(tree | data) ∝ P(data | tree) P(tree), and since the normalising constant sums over all topologies and integrates over all branch lengths, it is computed by MCMC — Metropolis–Hastings with proposals that both perturb continuous parameters and jump between topologies (MrBayes, BEAST, RevBayes). The output is a sample of trees; the of a clade is the fraction of sampled trees containing it.
The standard MCMC convergence diagnostics all apply, and matter more here than in most applications, because tree space is multimodal: chains get stuck on topological islands separated by low-likelihood valleys. Run multiple chains from different starts, check the standard deviation of split frequencies between them, check ESS, use Metropolis-coupled MCMC (heated chains) to cross valleys. Priors on branch lengths are not innocuous — a badly chosen branch-length prior can inflate tree length and distort posteriors, and it will not announce itself.
8. Support: bootstrap and posterior probability are not the same quantity
The nonparametric bootstrap (Felsenstein 1985) resamples alignment columns with replacement to sequence length L, rebuilds the tree, and repeats 100–1,000 times. The bootstrap proportion for a clade is the fraction of replicates containing it.
Statistics: the bootstrap in general — why resampling your own data substitutes for a formula, and why the choice of resampling unit is the design decision — is S3 §6.
What that measures: how evenly the signal for that split is spread across sites. A high value means the split does not depend on a handful of columns. That is a real and useful thing to know, and it is not the probability that the clade is correct.
The bootstrap cannot detect systematic error, because every replicate contains it. If your model is wrong, your alignment is wrong, or your "" are , that error is present in all 1,000 replicates and you get 100% support for the wrong tree. Bootstrap measures precision. It says nothing about accuracy.
Conventional readings — >70% "reasonable", >95% "strong" — are rules of thumb calibrated on simulations under correct models, and are conservative in some regimes and anticonservative in others. The ultrafast bootstrap in IQ-TREE is differently calibrated again and its thresholds (≥95%) are not interchangeable with the classical ones.
Posterior probabilities are what people want bootstrap values to be: P(clade | data, model, prior), a direct probability statement. The catch is in the conditioning. That probability is well calibrated only if the model is right, and is empirically much more confident than bootstrap on the same data — 1.00 posteriors accompanying 60% bootstrap proportions are routine. Under model misspecification, posteriors are known to be systematically overconfident, for the same structural reason as the bootstrap: the model error is common to every MCMC sample.
| Bootstrap proportion | ||
|---|---|---|
| Question answered | Would resampling these columns give the same split? | What is P(split | data, model, prior)? |
| Interpretable as P(correct)? | No | Only if the model is correct |
| Sensitive to model error | Not at all — invisible to it | Not at all — invisible to it |
| Typical magnitude | Lower | Higher |
| Cost | k full tree searches | One MCMC run |
Report which one you used. Never compare a 0.95 posterior with a 95% bootstrap as if they were the same statement.
9. Rooting, and molecular clocks
Outgroup rooting is standard: include a taxon known on independent grounds to fall outside the group of interest, infer the unrooted tree, then place the root on the branch leading to it. The outgroup must be close enough to align reliably and far enough to be genuinely outside — a tension with no general solution. A too-distant outgroup sits on a very long branch and is attracted to the fastest-evolving ingroup lineage (§6), putting the root in the wrong place with high support. Midpoint rooting places the root at the middle of the longest tip-to-tip path; it assumes a clock and fails exactly when the clock does.
Under a strict , substitutions accumulate at a constant rate r, so d = 2rt between two tips separated by divergence time t, and the tree becomes ultrametric. The clock is a hypothesis, testable by likelihood-ratio test against the unconstrained model with n − 2 degrees of freedom, and across any broad set of taxa it is rejected — generation time, metabolic rate, DNA-repair efficiency and all vary.
Converting into time requires an external rate. Two routes:
Fossil calibration. A fossil assigned to a lineage shows that lineage existed by that date, so a fossil gives a minimum age for a divergence, never the divergence date itself. The true split is older by an unknown amount depending on preservation and sampling. Modern practice encodes this as a prior with a hard minimum and a soft, long-tailed maximum, rather than a point .
Tip dating. When sequences are sampled at known, spread-out times — pathogens, ancient DNA — the sampling dates calibrate the rate directly. Root-to-tip regression of divergence against sampling date estimates the rate, and its R² is a useful check for whether there is enough "measurable evolution" to date at all.
Relaxed clocks let rates vary between branches: uncorrelated (each branch draws a rate iid from a lognormal) or autocorrelated (a branch's rate is drawn near its parent's, modelling heritable rate determinants). These fit far better, at the cost of many more parameters.
Now the statistical point to carry away. Divergence-time estimates are the product of rate × time, and only their product is identified by sequence data. Yang and Rannala showed that as sequence length → ∞, the posterior credible interval on a divergence time does not shrink to zero: it converges to a finite interval determined entirely by calibration uncertainty. Sequencing more genomes will not narrow it. This is why published dates for the same event differ by twofold between studies with overlapping data, and why any divergence date without an explicit statement of its calibrations should be read as an order of magnitude.
10. Gene trees, species trees, and the coalescent
The tree you compute is a tree of sequences, not a tree of species. Every locus has its own genealogy, and those genealogies genuinely differ from one another and from the species tree — not because of estimation error, but because that is how populations work.
Four sources of discordance, with different signatures:
| Source | Mechanism | Signature |
|---|---|---|
| Incomplete lineage sorting (ILS) | Ancestral polymorphism survives a speciation and sorts randomly afterwards | Symmetric — the two discordant topologies occur equally often |
| / | Gene flow after divergence | Asymmetric — one discordant topology is in excess |
| Horizontal gene transfer | Transfer between distant lineages ( in bacteria) | Single genes deeply misplaced; often with atypical composition |
| Duplication and loss | mistaken for (§2) | Splits that are too deep; resolved by adding taxa |
The coalescent
Statistics: expectations of random variables, and the linearity that lets these waiting times simply be added up, are covered in S1 §6.
Reverse time. Instead of asking how a population evolves forward, take the sample you actually have and trace its lineages backwards until they merge. For a population of N individuals — 2N gene copies — the chance that two specific lineages coalesce in the previous generation is 1/(2N). With k lineages there are C(k,2) pairs, so the waiting time Tk until the first coalescence is geometric, well approximated by an exponential:
Summing over k = n down to 2 telescopes cleanly, since 1/[k(k−1)] = 1/(k−1) − 1/k:
Three things fall straight out of this.
The deepest coalescence dominates. E[T₂] = 2N — the final interval, with only two lineages left, alone takes half the expected time to the MRCA, however large the sample. The first coalescence among n = 100 samples takes on average 4N/9900 ≈ 0.0004N generations. Genealogies are shaped like a broom: a burst of coalescences near the tips, then a long lonely wait.
Sampling more individuals barely deepens the tree. E[TMRCA] → 4N as n → ∞, and n = 10 already reaches 90% of that. Deep history is a property of the population, not of your sample size.
Total grows only logarithmically. Interval k has k lineages, so contributes k × 4N/[k(k−1)] = 4N/(k−1), giving E[Ttotal] = 4N Hn−1 with H the harmonic number. Since land on branches at rate μ, E[S] = θHn−1 with θ = 4Nμ — which is exactly Watterson's estimator from Ch 33, now derived rather than asserted. The fraction of total branch length in the single deepest interval is 1/Hn−1: 35% at n = 10, still 19% at n = 100. A fifth of all the mutational opportunity in a sample of 100 sits on two branches.
Quantifying ILS
For a species tree ((A,B),C) with an internal branch of T generations, the A and B lineages fail to coalesce in the ancestral AB population with probability e−T/(2N). If they fail, three lineages enter the deeper ancestral population and all three topologies are then equally likely, so two-thirds of that probability yields a discordant :
| τ ( units) | P(gene tree ≠ species tree) |
|---|---|
| 0.25 | 0.52 |
| 0.5 | 0.40 |
| 1 | 0.25 |
| 2 | 0.090 |
| 3 | 0.033 |
Short internal branches relative to Ne — rapid radiations, large ancestral populations — mean most genes disagree with the species tree, and that is the correct answer, not an error. Worse, Degnan and Rosenberg showed there are "anomaly zones" where the most probable gene tree differs from the species tree, so concatenating loci or taking the majority gene tree is not merely inefficient but positively misleading. The fix is a method that models the coalescent explicitly (ASTRAL, *BEAST, SVDquartets), taking gene trees as observations of a species tree rather than as replicates of it.
Because ILS is symmetric, the asymmetry test is powerful: count sites with pattern ABBA and pattern BABA across a tree (((P1,P2),P3),Outgroup). Under ILS alone their counts are equal in expectation, so D = (nABBA − nBABA) / (nABBA + nBABA) has expectation zero, and a significant deviation indicates gene flow between P3 and one of P1/P2. This is Patterson's D, and it is how archaic introgression was established.
Nothing in that derivation mentions hominins: any four populations with a known topology will do, and D is the standard detector of introgression wherever two lineages still exchange genes. Ch 35A §5 generalises it — the block jackknife that supplies its significance, the ghost-lineage and ancestral-structure caveats that D ≠ 0 does not exclude, and three cases where the introgressed was adaptive — and sets it beside the other way of measuring gene flow between diverging lineages, the .
11. What it is used for
Phylogeography places genealogies on a map, treating location as a character evolving along the tree, to infer where lineages originated and how they moved. Applied to human mtDNA and , it produced the out-of-Africa framework; applied to whole genomes, it maps migration and (Ch 28).
Pathogen phylogenetics works in real time because fast-evolving viruses accumulate detectable change on a timescale of weeks. SARS-CoV-2 evolves at roughly 10⁻³ substitutions per site per year — order of a couple of changes per 30 kb genome per month — so a tree built from sequences sampled a fortnight apart is informative. Trees are used to date the origin of an outbreak, distinguish a single introduction from many, identify transmission chains within a hospital, and detect variants as unusually long branches with excess non-synonymous change. Platforms like Nextstrain rebuild global trees continuously.
Ancient DNA and archaic . Sequencing degraded remains produced high- Neanderthal and Denisovan genomes, and the ABBA-BABA asymmetry above showed that the discordance was not ILS. Current estimates, which move as methods improve:
| Archaic ancestry | |
|---|---|
| Neanderthal, non-African populations | ~1–2%; roughly 1.7% in Europeans, marginally higher in East Asians |
| Neanderthal, African populations | ~0.3–0.5%, mostly via back-migration from Eurasia |
| Denisovan, Papuan and Oceanian populations | ~3–5% (estimates are method-dependent and some are considerably lower) |
| Denisovan, East Asian populations | ~0.1% |
The bulk of Neanderthal gene flow is now dated to a single extended period roughly 50,500–43,500 years ago. Introgressed segments are not randomly distributed: they are depleted on the X and near genes, consistent with against archaic ancestry, and enriched at a handful of loci — immunity, skin and hair biology, high-altitude adaptation via EPAS1 in Tibetans — that were plausibly adaptive.
Common misconceptions
| What people think | What's actually true |
|---|---|
| A shows which species evolved from which | Tips are all contemporary. Extant species descend from inferred internal nodes, not from each other. Humans did not evolve from chimpanzees; both descend from an unsampled ancestor |
| A high bootstrap value means the clade is probably correct | It means the split is robust to resampling columns. Systematic error — wrong model, bad alignment, — is in every replicate and yields high support for wrong trees |
| Posterior probabilities and bootstrap values are interchangeable | Different quantities on different scales. Posteriors are routinely much higher for the same data, and are calibrated only under a correct model |
| is assumption-free because it has no model | It has an implicit model that assumes change is rare and evenly distributed. Where that fails it is statistically inconsistent — more data, more confidence, wrong tree |
| Species that look similar are closely related | Similarity can be convergence (polyphyly) or shared retained ancestral traits, neither of which is evidence of recent common ancestry. Only shared derived characters group taxa |
| Discordant mean somebody made an error | Discordance is expected under the . With short internal branches most loci disagree with the species tree, and that is the true answer |
| More sequence data will pin down the divergence date | Only rate × time is identified. Date intervals converge to a nonzero width set by calibration uncertainty, no matter how much sequence you add |
| means elapsed time | It means expected substitutions per site. Converting to time requires a clock model plus an external calibration, both of which are assumptions |
| is a percentage | Homology is binary — shared ancestry or not. The percentage is identity or similarity |
Worked example: four taxa, by hand
Four sequences, 1,000 aligned sites, with these counts of differing sites:
A B C D
A - 280 435 310
B 280 - 310 136
C 435 310 - 280
D 310 136 280 -
Step 1 — raw p-distances. Divide by 1,000: pAB = 0.280, pAC = 0.435, pAD = 0.310, pBC = 0.310, pBD = 0.136, pCD = 0.280.
Step 2 — JC69 correction, d̂ = −(3/4) ln(1 − (4/3)p̂).
For AB: (4/3)(0.280) = 0.37333; 1 − 0.37333 = 0.62667; ln(0.62667) = −0.46734; −0.75 × (−0.46734) = 0.3505.
For AC: (4/3)(0.435) = 0.58000; 1 − 0.58000 = 0.42000; ln(0.42000) = −0.86750; −0.75 × (−0.86750) = 0.6506.
For AD: (4/3)(0.310) = 0.41333; 1 − 0.41333 = 0.58667; ln(0.58667) = −0.53329; −0.75 × (−0.53329) = 0.4000. Same for BC.
For BD: (4/3)(0.136) = 0.18133; 1 − 0.18133 = 0.81867; ln(0.81867) = −0.20008; −0.75 × (−0.20008) = 0.1501.
For CD: same arithmetic as AB: 0.3505.
Note the size of the correction on the most divergent pair: 0.435 observed becomes 0.651 corrected, a 50% increase, while the closest pair moves from 0.136 to 0.150, only 10%. The correction is nonlinear and hits long branches hardest — precisely why omitting it causes long-branch artefacts. Rounding to 0.35, 0.65, 0.40, 0.40, 0.15, 0.35.
Step 3 — check the four-point condition.
| Pairing | Corrected sum | Raw p sum |
|---|---|---|
| (AB)(CD) | 0.35 + 0.35 = 0.70 | 0.280 + 0.280 = 0.560 |
| (AC)(BD) | 0.65 + 0.15 = 0.80 | 0.435 + 0.136 = 0.571 |
| (AD)(BC) | 0.40 + 0.40 = 0.80 | 0.310 + 0.310 = 0.620 |
The corrected distances are exactly additive — one small sum, two equal larger ones — and the small one identifies the topology ((A,B),(C,D)). The raw distances are not additive: the two larger sums differ by 0.049, and the shortfall lands on the pairing containing pAC, the pair of long branches. That is long-branch attraction visible in a distance matrix.
Step 4 — UPGMA gets it wrong. UPGMA joins the smallest distance first. That is dBD = 0.15, so it groups B with D. B and D are not sisters; they are the two slow lineages, on opposite sides of the tree. UPGMA's ultrametric assumption is violated and it fails in exactly the predicted direction.
Step 5 — neighbour-joining gets it right. Row sums: rA = 0.35+0.65+0.40 = 1.40, rB = 0.35+0.40+0.15 = 0.90, rC = 0.65+0.40+0.35 = 1.40, rD = 0.40+0.15+0.35 = 0.90. With n = 4, Qij = 2dij − ri − rj:
| pair | 2d | −(ri+rj) | Q |
|---|---|---|---|
| A,B | 0.70 | −2.30 | −1.60 |
| A,C | 1.30 | −2.80 | −1.50 |
| A,D | 0.80 | −2.30 | −1.50 |
| B,C | 0.80 | −2.30 | −1.50 |
| B,D | 0.30 | −1.80 | −1.50 |
| C,D | 0.70 | −2.30 | −1.60 |
The minimum Q is −1.60, attained jointly by A,B and C,D — the two true cherries, which must tie here because the tree is symmetric (A:0.30/B:0.05 mirrors C:0.30/D:0.05). NJ joins either one and reaches the same topology; we take C,D forward. Both beat B,D at −1.50, despite dBD = 0.15 being the smallest distance in the matrix. The r terms did the work: C is far from everything (r = 1.40), so being 0.35 from D is impressive, whereas B and D are close to everything and being 0.15 apart is not.
Join C and D at node u:
- dCu = ½(0.35) + ¼(1.40 − 0.90) = 0.175 + 0.125 = 0.30
- dDu = 0.35 − 0.30 = 0.05
- dAu = ½(0.65 + 0.40 − 0.35) = ½(0.70) = 0.35
- dBu = ½(0.40 + 0.15 − 0.35) = ½(0.20) = 0.10
Three nodes remain (A, B, u) with dAB = 0.35, dAu = 0.35, dBu = 0.10. Resolve the star at centre v:
- dAv = ½(0.35 + 0.35 − 0.10) = 0.30
- dBv = ½(0.35 + 0.10 − 0.35) = 0.05
- duv = ½(0.35 + 0.10 − 0.35) = 0.05 ← the internal branch
Final tree, in Newick: ((A:0.30,B:0.05):0.05,(C:0.30,D:0.05));
0.30 0.30
A ──────────┐ ┌────────── C
├──┤ 0.05
B ───┐ │ │ ┌─── D
0.05 └──┘ 0.05
A and C are the long branches, non-sisters, separated by an internal branch of only 0.05 — textbook Felsenstein-zone geometry. Neighbour-joining on corrected distances survives it because the correction restores additivity. Nothing here required a computer, and every step would have gone wrong at the same place if the JC69 correction had been skipped.
Connections
- Back to: Ch 33 — supplies the substitution process that makes clocks and distances meaningful, and Watterson's θ is re-derived here from the · Ch 27 — and Ne set the coalescent timescale · Ch 20A — the mechanisms behind the horizontal-transfer row of §10's discordance table: transformation, conjugation and transduction are how a bacterial comes to disagree with its species tree, and why the disagreement is concentrated in plasmid- and phage-borne genes rather than spread evenly · Ch 16 — the transition/transversion bias that K80 onward encode · Ch 12 — likelihood-ratio testing, reused for clock tests and model selection
- Forward to: Ch 35 — duplication, orthology and gene-family evolution, the §2 problem developed in full · Ch 35A §5 — §10's D-statistic restated for any four populations rather than for archaic humans, with the block jackknife, the ghost-lineage caveat and adaptive ; §10's "speciation" nodes are also what 35A opens, and its E[TMRCA] → 4N is why reciprocal monophyly is slow to arise even with zero gene flow — which is what makes both the phylogenetic and the ESU criterion contested · Ch 44 — comparative depends on correct identification · Ch 45 — bacterial and HGT · Ch 56 — tumour phylogenies are the same machinery applied to somatic clones · Ch 57 — outbreak reconstruction in public health