One Tree: The Statistics of Common Descent

Darwin's one diagram made a strict prediction: independent evidence should converge on one branching order. Cytochrome c, a Bayesian test, retroviral scars and a broken vitamin C gene all did. The prokaryote root is the one honest exception.

Every line of evidence in this article - cytochrome c, a Bayesian model, a primate retrovirus, a broken vitamin C gene - starts as tissue turned into a flow cell like this one, still mid-load and not yet run. — Image prompt and art direction by Brecht Corbeel; generation pending.

Abstract

On the Origin of Species contains one illustration, and it encodes a strict, falsifiable claim: independent character sets should recover the same nested hierarchy of life. This article follows that claim from Fitch and Margoliash's 1966 cytochrome c tree, through Penny, Foulds and Hendy's 1982 statistical test of tree agreement, to Douglas Theobald's 2010 formal Bayesian model-selection test of universal common ancestry and the Koonin-Wolf critique that followed it. It then turns to two genomic witnesses that have nothing to do with sequence-similarity statistics - shared endogenous retrovirus insertions and a shared, identically broken vitamin C gene across primates - before treating honestly the two real sources of tree conflict: incomplete lineage sorting, which the coalescent predicts in calculable amounts, and horizontal gene transfer, which makes the deep prokaryote root a network rather than a trunk. The consilience argument closes as a statistical claim, not a rhetorical one.

One diagram, one falsifiable claim

On the Origin of Species contains exactly one illustration. Buried in the fourth chapter, “Natural Selection,” it is a simple branching figure: hypothetical species labelled with letters, diverging over a marked span of generations, some lines running to extinction and others radiating into descendants. Darwin introduced it with a sentence that is easy to read past: “The affinities of all the beings of the same class have sometimes been represented by a great tree. I believe this simile largely speaks the truth” [1]. That one paragraph, illustrated by that one figure, is the whole of the book’s structural claim, and it is a far more specific claim than “species change over time.”

A tree is not the only pattern naturalists had proposed for how living things relate to one another, and it is worth being concrete about the alternatives, because a claim is only falsifiable if there are real, articulable ways for it to be wrong. Penny, Hendy and Poole later catalogued the field the tree had to beat: the medieval Great Chain of Being, a static ladder running from minerals to angels; Buffon’s degeneration theory, in which a handful of supposedly perfect archetypes decayed into modern variants; a two-dimensional map of affinities, with related forms occupying neighbouring regions rather than branch tips; the quinary system, five interlocking circles of relationship repeating at every taxonomic level; and a bare spanning tree, linking existing species to one another without implying that any one of them is ancestral to another [4]. None of these was a straw man. Naturalists holding respectable posts defended versions of most of them well into the nineteenth century, and the quinary system in particular had serious adherents among Darwin’s own contemporaries.

What common descent asserts, against that field of rivals, is unusually strict. Every organism that has ever lived is connected to every other by an unbroken chain of parent-to-offspring reproduction, which means the true pattern of relationship among species is a single, fully connected, nested hierarchy - not several disconnected hierarchies, not a two-dimensional grid, not a repeating five-fold wheel, not a chain with no branching at all. Because that pattern is a strict mathematical object, it produces a strong, checkable prediction that has nothing to do with resemblance in the loose, everyday sense: take any two independent sets of heritable characters sampled from the same set of species - a suite of skeletal measurements, say, and a century later the amino acid sequence of a single protein - and build the best tree each one supports on its own. Those two trees should agree far more than chance allows, because both are noisy readings of the same underlying genealogical event. They do not merely need to look similar. They need to make the same specific claims about which species share a more recent common ancestor with which, again and again, across characters that have no causal connection to one another beyond having been inherited down the same lineages.

That is the test the rest of this article follows: a hand-built protein tree matched against a skeleton in 1966; five independently sequenced proteins matched against each other in 1982; twenty-three universal proteins matched against a formal statistical alternative in 2010; a retrovirus’s chromosomal address and a primate’s broken vitamin C gene matched against genealogy itself, with no sequence-similarity statistics involved at all; and, honestly, the specific places where the single-tree picture stops holding and why those places are themselves predictable rather than embarrassing.

A molecule agrees with the skeleton

The first serious molecular test of that prediction used cytochrome c, a small protein central to cellular respiration and, for practical reasons, unusually easy to sequence in the 1960s. Walter Fitch had built a computer program that used the genetic code to calculate the minimum number of nucleotide substitutions that could account for the amino acid differences between any two cytochrome c sequences, and he had settled on that protein because ten published sequences already existed for it - more than for any other protein at the time. In the spring of 1966, Emanuel Margoliash, who had independently sequenced cytochrome c in his own laboratory, offered Fitch ten further unpublished sequences over a single lunchtime conversation in Madison, doubling the dataset overnight and extending it, in Fitch’s own later description, across “the largest part of the eukaryotic kingdom” - yeast, insects, fish, reptiles, birds and mammals, humans among them [2].

Fed through Fitch’s distance calculations and assembled into a tree by a method that allowed for unequal rates of change along different branches rather than assuming a uniform molecular clock, the twenty sequences produced a branching order that Fitch and Margoliash judged, in their own words, “quite respectable despite a few imperfections” when set against the tree comparative anatomy had already proposed from bones, teeth and soft tissue [2]. A protein nobody had ever examined with phylogeny specifically in mind had reconstructed, essentially from scratch, the branching order zoologists had spent a century assembling from skeletons. That result, published in Science in January 1967, is usually credited as the founding paper of molecular phylogenetics, and it earned that status by doing exactly what the tree metaphor demanded: an entirely independent character set converged on the same hierarchy morphology had already proposed.

One protein agreeing with a skeleton-based tree is suggestive. It is not yet a statistical test, because nobody had specified how much agreement chance alone would produce, or compared several independently sequenced proteins against one another rather than against a single hand-drawn anatomical reference. David Penny, Leon Foulds and Michael Hendy closed that gap in 1982. Working with five different protein datasets sampled from the same eleven mammal species, they built the most-parsimonious, minimal-mutation tree implied by each dataset independently, using a branch-and-bound algorithm that guaranteed the true minimal tree rather than a local approximation, and then measured how close the five resulting trees were to one another using an objective tree-comparison metric [3]. Crucially, they also worked out what that metric’s distribution would look like under a null model in which the five datasets arose from entirely separate ancestries, so that any five trees drawn at random from the vast space of possible rooted topologies on eleven taxa were as likely to resemble one another as not. The five real trees were far closer to one another than that random-tree null model allowed.

As Penny restated the conclusion two decades later, defending the test’s logic against critics who had argued that common ancestry could not be examined in a normal scientific manner, the exercise showed that “the theory of descent leads to quantitative predictions that are testable and is thus, in principle, falsifiable” [4]. That defence was not academic throat-clearing. It answered a specific, public dispute: an exchange in Nature’s own pages around 1980-81 over whether evolutionary theory qualified as normal, falsifiable science at all, with Karl Popper’s charge that it functioned more like a “metaphysical research programme” than a testable theory hanging over the field [4]. Penny, Foulds and Hendy’s answer to that charge was not to argue about whether such a test was possible in principle. It was to build one.

Figure 1. Four independently drawn samples, four independent tests - the modern version of what Fitch and Margoliash did by hand with one protein in 1966, and what Penny, Foulds and Hendy asked of five proteins at once in 1982. — Image prompt and art direction by Brecht Corbeel; generation pending.

Putting a number on how sure

Both of those tests compared trees built from a handful of datasets against what chance alone would produce. In 2010, Douglas Theobald asked a more ambitious question: given the full array of universally conserved proteins now sequenced across every domain of life, how many times more probable is a single common ancestry for all of it than the best-fitting alternative in which major lineages arose independently and only later came to resemble one another? Theobald assembled twenty-three proteins found, in recognisably homologous form, in Archaea, Bacteria and Eukarya alike, and set up a formal model-selection problem. Rather than assuming, as most earlier phylogenetic work had done implicitly, that sequence similarity itself implies shared ancestry, he built explicit alternative models in which similarity could arise without any genealogical connection at all - through convergent functional constraint, for instance - and let likelihood-ratio tests, Akaike weights and Bayes factors adjudicate between the rival models [5].

The logic of that adjudication reduces to one ratio. For sequence data compared under a universal-common-ancestry model against an independent-origins model,

K = \frac{P(D \mid M_{\text{UCA}})}{P(D \mid M_{\text{IO}})}

and K, the Bayes factor, states exactly how many times more probable the observed data D are under one model than the other - not a correlation coefficient, not a subjective sense of resemblance, but a single number with a precise probabilistic meaning that can in principle come out favouring either model. For the proteins Theobald examined, it was not close. Universal common ancestry came out favoured over the closest competing multiple-ancestry hypothesis by odds reported on the order of ten to the power of two thousand eight hundred and sixty to one, and over the specific alternative that humans arose independently of the rest of life by something on the order of ten to the power of six thousand to one - numbers large enough that Theobald himself, discussing the paper around its publication, called it beside the point to state them as ordinary betting odds. The result held, notably, even when the alternative models were allowed to include horizontal gene transfer and symbiotic fusion events rather than only a strictly bifurcating rival tree [5].

A number that large invites scrutiny, and it got it. Eugene Koonin and Yuri Wolf published a pointed methodological challenge, arguing that Theobald’s test might not cleanly measure what it claimed to measure. Running their own simulation, they showed that combining alignments built from sequences that were statistically similar for reasons that had nothing to do with shared genealogy - engineered similarity rather than inherited similarity - could still produce a likelihood-based preference for a “combined ancestry” model over separate ones. Their conclusion was carefully hedged rather than dismissive: comparative genomics as a whole, they wrote, still provides overwhelming support for universal common ancestry, but a fully general statistical proof of the kind Theobald’s framing seemed to promise remained, in their assessment, elusive, because raw sequence similarity alone can under some conditions mimic the signature that genuine common descent leaves behind [6].

Theobald’s published reply took the challenge seriously rather than waving it off, and it sharpened the original claim rather than merely restating it. For real biological sequence data, he argued, it is not bare pairwise similarity driving the model-selection result, but the specific nested, hierarchically correlated pattern of similarity across many sequences and many sites at once - the exact signature a branching genealogy predicts, and one that convergent similarity, absent shared history, does not reproduce. Koonin and Wolf’s artificial counterexample, on that reading, had reproduced raw similarity without the nested correlational structure that makes the real protein data so improbable under any non-genealogical model. That exchange is worth reading as an illustration of a field doing its job, not as evidence of unsettled doubt about common descent itself: a bold quantitative claim, a serious adversarial stress test, and a sharpened restatement of exactly which feature of the data is doing the inferential work.

The methodology kept maturing after that exchange rather than stalling on it. Bret Larget and colleagues built a new statistical framework specifically designed to separate genuine genealogical signal from the confound Koonin and Wolf had identified, by modelling site-by-site functional constraint explicitly instead of folding it into the overall similarity signal, and applied it to a large molecular alignment spanning the primates - the exact group this article’s other genomic evidence concerns. Their result: “overwhelming evidence against separate ancestry and in favor of common ancestry for orders and families of primates,” and, separately, “overwhelming evidence that humans share a common ancestor with other primate species” [12]. Different method, the same primates a critic had specifically worried about, the same conclusion - which is what a real signal, rather than a statistical artefact, is supposed to do when it is tested again with a better instrument.

Figure 2. The likelihood ratio a formal test reports - common ancestry favoured over its closest rival by odds counted in thousands of orders of magnitude - is arithmetic that has to run somewhere. — Image prompt and art direction by Brecht Corbeel; generation pending.

The genome keeps a written record of old infections

A formal statistical test is one channel of evidence. A mechanistically independent channel comes from retroviruses, and it works because of a peculiarity of how they replicate. When a retrovirus infects a germline cell - an egg, a sperm, or one of their precursors - and integrates a DNA copy of its genome into a chromosome at a particular base-pair address, that insertion becomes a permanent, heritable feature of every descendant of that individual. If the insertion has no functional consequence, as most do not, it is never removed by selection; it simply rides along, at that exact genomic position, in every descendant lineage from that point forward. These are endogenous retroviruses, and primate genomes carry thousands of them [7].

Johnson and Coffin’s 1999 survey exploited a further quirk of the mechanism that makes endogenous retroviruses unusually rigorous phylogenetic markers rather than merely suggestive ones. Each integrated provirus is flanked by two long terminal repeats, identical to each other at the moment of integration and then accumulating mutations independently of one another afterward - so a single locus yields two separate, internally checkable estimates of the same phylogenetic relationship, since both flanking repeats should show a mutually consistent pattern of divergence if the shared-ancestry model is correct [7]. Because most of these loci are selectively neutral, their distribution across species is not being shaped or homogenised by natural selection toward some pattern independent of history, which is exactly what makes their pattern, where one exists, close to a direct readout of descent rather than an artefact of function.

The distribution itself is the evidence, and it is specific rather than merely qualitative. Johnson and Coffin identified three endogenous retrovirus loci present in Old World monkeys and hominoids but absent from New World monkeys, consistent with integration into a shared ancestor at least thirty-one million years ago, after the platyrrhine and catarrhine lineages had already diverged; and three further loci found only in humans, gorillas, chimpanzees and bonobos - present in none of the more distantly related apes or monkeys - consistent with integration roughly five million years ago, within the African great-ape lineage [7]. The logic that makes this compelling is straightforward and does not depend on any tree-building algorithm at all: the probability that the same retrovirus would integrate, independently, at the identical base-pair position in the genomes of two or more separately originating lineages is negligible against genomes billions of base pairs long. A shared insertion at a shared address is, for practical purposes, proof of shared inheritance from one ancestral integration event, not of two coincidental ones.

The method is not naive about its own pitfalls, either, which is worth noting because it speaks to the field’s discipline rather than to any weakness in the evidence. The two flanking repeats at a locus can occasionally undergo gene conversion, homogenising them and destroying their independence as two separate phylogenetic estimates. Johnson and Coffin flag exactly this for one locus in gorilla, where eleven substitution sites shared identically between the two otherwise-independent repeats make truly independent post-integration evolution implausible, and point instead to a conversion event that has to be screened out before the locus is treated as two independent data points rather than one [7]. That kind of internal quality control is itself a mark of a field checking its own inferences rather than accepting convenient ones.

Figure 3. A panel like this queries the same chromosomal address across several primate genomes at once; a retrovirus that inserted there only once, in one shared ancestor, is why the same scar turns up in every descendant lineage. — Image prompt and art direction by Brecht Corbeel; generation pending.

The same gene broken the same way

Most mammals synthesise their own vitamin C from glucose, through a short pathway that ends with an enzyme called L-gulonolactone oxidase, GULO, oxidising gulonolactone into ascorbate. Humans cannot. Neither can any other haplorhine primate - the group that includes monkeys, apes and humans, as distinct from the lemurs and lorises that retain the full pathway - which is why vitamin C is a dietary requirement for us and scurvy is a specifically primate condition rather than a general mammalian vulnerability. The reason is that the GULO gene itself is still present in the human genome, recognisably homologous to the working copy in other mammals, but disabled: a pseudogene rather than a gene.

Lachapelle and Drouin’s 2011 analysis of that pseudogene is a small case study in what actually makes a shared loss of function evidence for common ancestry, as opposed to merely evidence that several species happen to share a trait. Comparing the GULO region across human, chimpanzee and macaque, they found that all three genomes are missing exons one and two at the identical genomic position, and that all three additionally carry two identical seven-base-pair insertions in the region immediately upstream of exon one - not merely three broken genes, but the same break, at the same coordinates, in three different genomes [8]. Using a method that estimates elapsed time from the pattern of neutral substitution accumulated since a sequence stopped being functionally constrained, they dated that inactivation to roughly sixty-one million years ago, in the common ancestor of the haplorhine primates, well before the group’s later diversification into monkeys and apes [8].

The force of that result depends on a fact easy to lose sight of: there is no shortage of ways to break a gene. A single point mutation converting almost any of dozens of codons into a premature stop would do it; so would a deletion of almost any size at almost any position; so would a single mutation at a splice site. If human, chimpanzee and macaque ancestors had each independently lost the capacity to synthesise vitamin C - as, entirely separately, guinea pigs, some bats, some passerine birds and several teleost fish lineages genuinely have, each losing GULO function through its own distinct mutations in its own lineage - the overwhelming expectation is that each lineage’s dead gene would be broken in its own distinct way. Instead the primate lesion is identical across three genomes down to the exact missing exons and the exact upstream insertions, which is the signature of a single ancestral event, inherited once and carried by everything descended from the ancestor in which it happened, not of the same disability being reinvented three separate times.

The comparison case makes the point sharpest. The guinea pig’s independent loss of GULO function, dated by the same divergence-based method, falls at roughly fourteen million years ago - a different lesion, at a different position, in an unrelated lineage, at a wholly different point in time [8]. Convergent loss of a function is common enough in evolution; convergent loss executed by the identical mutation at the identical coordinate is not, and it is not what these genomes show anywhere except within the primate clade that actually shares the ancestor in which the break first occurred.

Figure 4. Amplifying the same stretch of a broken gene across species is how the identical deletions became visible: human, chimpanzee and macaque carry the same two missing exons at the same address. — Image prompt and art direction by Brecht Corbeel; generation pending.

Where the branches genuinely disagree, and why the disagreement has a shape

None of the preceding sections implies that every gene tree ever built agrees with every other one, and pretending otherwise would misrepresent the evidence in the opposite direction from creationist folklore. Real disagreement between trees built from different genes, or between a gene tree and the species tree those genes supposedly track, is well documented and routine. The point worth making carefully is that this disagreement arises from at least two quite different, independently well-understood sources - and distinguishing which is operating in a given case is itself a test of whether the field understands its own data or is merely explaining away inconvenient results.

The first source is simple statistical noise. Rokas, Williams, King and Carroll examined a hundred and six orthologous genes across eight yeast species and found that trees built from individual genes disagreed with one another often, sometimes sharply [10]. But when they concatenated the genes into a single combined dataset, the disagreement dissolved: as few as twenty concatenated genes were enough to recover one fully resolved species tree with maximum statistical support, and the full set of a hundred and six did the same [10]. The natural reading is that most single-gene incongruence in that dataset reflected genes simply being too short, individually, to carry enough phylogenetic signal to overcome sampling noise, rather than a genuine conflict in evolutionary history. That reading is itself a falsifiable claim rather than an assumption: if the true history really is one tree obscured by per-gene noise, adding independent genes should let the shared signal accumulate while noise, being independent from gene to gene, cancels out. That is exactly what concatenation did.

The second source is not noise at all, and it is more interesting: incomplete lineage sorting, a real population-genetic process with a mechanism and a quantitative theory attached to it, the coalescent. When two speciation events occur close together in time relative to the size of the ancestral population, genetic variants segregating as ordinary polymorphism in that ancestral population can sort into descendant lineages in a way that does not track the order in which the species actually split - so an individual locus’s own history can, with a specific, calculable probability, disagree with the species tree even though the species tree itself is a clean, real, single bifurcating history. Hobolth and colleagues modelled exactly this across the human, chimpanzee and gorilla genomes using a coalescent hidden Markov model that allows different genomic windows to support different local topologies, and recovered real, non-random structure rather than uniform disagreement: an ancestral effective population size for the human-chimpanzee ancestor of roughly sixty-five thousand, for the human-chimpanzee-gorilla ancestor of roughly forty-five thousand, and a human-chimpanzee speciation time around 4.1 million years ago [11].

The sharpest confirmation is what coalescent theory predicts should happen on the X chromosome, which carries a smaller effective population size than the autosomes and should therefore show less discordance, because lineages sort to completion faster when the ancestral population carrying the competing variants was smaller to begin with. That is exactly what the data show: roughly seventy-eight percent of the X-chromosome alignment supports the true human-chimpanzee grouping, a higher concordance than the autosomal genome as a whole [11]. Incomplete lineage sorting is not a hole in the tree model. It is a quantitative prediction the tree model makes about exactly where and how much local disagreement should appear, confirmed at a level of numerical detail that no morphological tree, and no single protein comparison, could ever have supplied.

Figure 5. Building a tree one locus at a time, across hundreds of them, is how the disagreements turn out to have a shape: most look like noise from too little signal, some look like real incomplete lineage sorting. — Image prompt and art direction by Brecht Corbeel; generation pending.

The root is not a trunk

There is one place where the single-tree claim genuinely does not hold, and saying so plainly is more honest, and ultimately more persuasive, than quietly hoping nobody asks. It is at the very base of the tree - among bacteria and archaea - and the reason has nothing to do with the quality of the data and everything to do with how prokaryotic genomes actually acquire genes.

Molecular phylogeneticists built early “universal trees of life” from ribosomal RNA and a small number of near-universal proteins, and for a time these trees were treated as the skeleton for a single natural classification of everything alive. Doolittle’s 1999 review laid out why that project runs into a genuine obstacle once whole genomes, rather than single genes, are compared: most bacterial and archaeal genomes, and even the reconstructed ancestral eukaryotic nuclear genome, turn out to be mosaics, carrying genes acquired from multiple, disparate sources through horizontal gene transfer rather than solely by vertical inheritance from a single parent lineage [9]. If lateral transfer among prokaryotes is common rather than a rare exception - and comparative genomics has made an increasingly strong case that it is, particularly deep in prokaryotic history - then no single strictly bifurcating tree can be the one true history of a prokaryotic genome, because different genes in the very same organism can have travelled to it by different routes and therefore carry genuinely different, truly conflicting histories, not merely noisy estimates of one shared history [9].

It is worth being precise about what does and does not follow from that. Every individual gene still has its own real history, tracing back through actual copying events to actual ancestral sequences; nothing here revives the pre-Darwinian alternatives this article opened by ruling out, and no gene is arising independently of prior life. What breaks down is only the stronger claim that a whole genome, or a whole prokaryotic lineage, can be assigned a single position on one universal bifurcating tree, because the genome itself is stitched together from more than one inheritance pathway. The honest picture at the deepest, oldest part of life’s history is a network - vertical descent and horizontal transfer overlaid on each other, not a single trunk with clean branches all the way to the root.

This is a structurally different kind of disagreement from incomplete lineage sorting, and the difference is worth stating precisely rather than folding both into one vague admission that “trees sometimes disagree.” Incomplete lineage sorting says: there is one real species tree, and an individual gene’s history can locally depart from it for calculable population-genetic reasons. The prokaryotic network says something stronger: there may not be one single tree to depart from in the first place, because the unit being compared - a whole genome - was never inherited down one line to begin with. Granting that costs the wider argument nothing. The claim this article has been building - that independent character sets recover the same nested hierarchy - has never needed to hold at the prokaryotic root to hold everywhere it has actually been tested, which is overwhelmingly among eukaryotes, and most rigorously of all among the primates whose genomes and retroviral scars this article has followed most closely.

Figure 6. Mixed strains from different bacterial and archaeal lineages share genes sideways as often as they inherit them downward, which is why the deepest part of the tree of life is better drawn as a network. — Image prompt and art direction by Brecht Corbeel; generation pending.

Consilience is a number, not a mood

Put the pieces back together in order. A single hand-built protein tree matched a skeleton-based tree in 1966. Five independently sequenced proteins across the same eleven mammals matched each other, far beyond chance, in 1982. Twenty-three universally conserved proteins spanning every domain of life matched a single-ancestry model over its best specified rivals by odds counted in thousands of orders of magnitude in 2010, a number that survived a genuine adversarial challenge and emerged from the exchange sharper rather than weaker. A heritable retroviral scar, with no connection whatsoever to sequence-similarity statistics, turns up at the same genomic address in exactly the primate lineages the rest of the evidence says share that ancestor, and nowhere else. A single broken enzyme carries the identical, specific lesion in humans, chimpanzees and macaques, dated to one ancestral event tens of millions of years before any of those species existed as such. None of these four lines of evidence depends on any of the others being true. Skeletal anatomy has no access to what a Bayes factor is; a retrovirus’s integration site has no access to a pseudogene’s exon count. That is what makes their convergence a statistical claim rather than a rhetorical flourish borrowed from one dataset and repeated for effect across the others.

Consilience of this kind is also, precisely because it is quantitative, honest about its own limits, and the honesty is part of the case rather than a concession extracted from it. Where trees disagree - a handful of yeast genes before concatenation, a portion of the human-chimpanzee-gorilla genome sorted incompletely, the mosaic genomes at the deep prokaryotic root - the disagreement turns out to have identifiable causes and, in the case of incomplete lineage sorting, a quantitative theory that predicts its size in advance of measuring it. A theory whose exceptions are themselves predictable, rather than merely excusable after the fact, is doing something a myth, a metaphor, or an untestable just-so story cannot do.

Darwin’s single figure, drawn from reasoning alone in 1859 with no molecular data available to check it against, asked whether independent branches of evidence about the living world would keep confirming one shape when examined separately and in more and more demanding ways. A century and a half of increasingly independent, increasingly quantitative tests have kept answering yes - in orders of magnitude Darwin had no way to compute, and in genomic evidence, endogenous retroviruses, pseudogene lesions, coalescent statistics, that did not exist even as concepts until well over a century after he wrote the sentence down.

Sources

1. Charles Darwin. On the Origin of Species (1859, first edition). John Murray / Project Gutenberg (1859).
2. Walter M. Fitch and Emanuel Margoliash. Construction of Phylogenetic Trees. Science (1967). DOI: 10.1126/science.155.3760.279.
3. David Penny, L. R. Foulds, and Michael D. Hendy. Testing the theory of evolution by comparing phylogenetic trees constructed from five different protein sequences. Nature (1982).
4. David Penny, Michael D. Hendy, and Anthony M. Poole. Testing fundamental evolutionary hypotheses. Journal of Theoretical Biology (2003). DOI: 10.1016/S0022-5193(03)00099-7.
5. Douglas L. Theobald. A formal test of the theory of universal common ancestry. Nature (2010). DOI: 10.1038/nature09014.
6. Eugene V. Koonin and Yuri I. Wolf. The common ancestry of life. Biology Direct (2010). DOI: 10.1186/1745-6150-5-64.
7. Welkin E. Johnson and John M. Coffin. Constructing primate phylogenies from ancient retrovirus sequences. Proceedings of the National Academy of Sciences (1999). DOI: 10.1073/pnas.96.18.10254.
8. Marc Y. Lachapelle and Guy Drouin. Inactivation dates of the human and guinea pig vitamin C genes. Genetica (2011). DOI: 10.1007/s10709-010-9537-x.
9. W. Ford Doolittle. Phylogenetic Classification and the Universal Tree. Science (1999). DOI: 10.1126/science.284.5423.2124.
10. Antonis Rokas, Barry L. Williams, Nicole King, and Sean B. Carroll. Genome-scale approaches to resolving incongruence in molecular phylogenies. Nature (2003). DOI: 10.1038/nature02053.
11. Asger Hobolth, Ole F. Christensen, Thomas Mailund, and Mikkel H. Schierup. Genomic Relationships and Speciation Times of Human, Chimpanzee, and Gorilla Inferred from a Coalescent Hidden Markov Model. PLoS Genetics (2007). DOI: 10.1371/journal.pgen.0030007.
12. Bret Larget, Cécile Ané, Martin Bontrager, Steve Hunter, Noah Stenz, and David A. Baum. Statistical evidence for common ancestry: New tests of universal ancestry. bioRxiv (2016). DOI: 10.1101/036327.

Originally published at https://absolutedigitalpublishers.com/articles/one-tree-the-statistics-of-common-descent.