The Many Origins

Chapter 5 — The Door Swings Both Ways

STATUS: v0 (all six parts drafted) Documentary parallel: chapters/05-the-door.md (Beats 5.1–5.10) Last updated: 2026-05-22

What this chapter does

Most popular accounts of multicellularity end at cancer. This one does too, but it asks a more specific question. What does cancer tell us about the cooperative state it breaks out of? The basic answer to what is cancer? has been settled for decades; the more interesting question is what the breakdown reveals about the thing that was working.

That takes us into one of the field’s longest-running live arguments: the atavism hypothesis, which proposes that cancer cells preferentially switch back on regulatory programs older than multicellularity itself. The hypothesis has had careful defenders and careful critics for fifteen years. In 2025, a result came out that addressed the critics’ strongest single objection, and the empirical floor under the hypothesis shifted. The conceptual disagreement remained. This chapter walks through both the result and what it leaves unsettled, using the method Chapter 1 used for the sponges-versus-ctenophores disagreement. We are not going to pick a side.

The chapter also delivers the project’s third synthesis framing: an asymmetry between two systems where multicellularity comes undone. Snowflake yeast in the lab reverts to single cells under one selection and re-evolves multicellularity under another. Cancer cells revert to single-cell-like behaviour and have never been observed re-evolving multicellular cooperation in any natural setting. Both observations are real. The difference between them is mechanistic, not moral, and it says something specific about what the body actually is.

By the end of the chapter, you should understand:

  1. What cancer is at the level of the cooperation toolkit earlier chapters have built up, and how the toolkit fails one component at a time, not all at once.
  2. The atavism hypothesis at its strongest, and the bioelectric framing alongside it as a sibling candidate, not a rival.
  3. The natural objection to atavism, why it was decisive for over a decade, and what the 2025 single-cell test actually did differently.
  4. What that test settles, and what it does not: the three open problems after 2025, and the chapter’s honest sentence.
  5. The yeast↔cancer asymmetry, and why an active enforcement environment breaks the symmetry between the two cases.
  6. Why “multicellularity” is best understood not as a finished state but as a balance, at every scale, with the same machinery running in both directions.

The chapter is in six parts.

PartTitleApprox. length
1Cancer as a starting question, and what cancer actually is~1100 words
2The atavism hypothesis at its strongest, and the bioelectric framing alongside it~1600 words
3The natural objection, and the 2025 test that controlled for it~1500 words
4What this still doesn’t settle, and the honest sentence~1300 words
5The asymmetry: yeast reverts and re-evolves; cancer reverts and stays reverted~1600 words
6Cancer, the ratchet, and the arrow that goes both ways~1400 words

Part 1 — Cancer as a starting question, and what cancer actually is

The usual cancer ending in popular treatments goes something like this: cells stop cooperating, the body’s machinery turns on itself, the price of multicellular complexity is the chance that it goes wrong. (That last clause is a metaphor; bodies do not pay prices and machinery does not turn on anything. The mechanism is what the rest of this part will lay out.) The headline is roughly right. What it leaves out is which parts of the machinery fail, in what order, and why the failure produces something that looks the way cancer looks.

What cancer is, at the level of the toolkit

The shortest accurate sentence about cancer at the level of cell biology: cancer is what multicellular cooperation looks like when the enforcement machinery fails. The specifics are not glib.

Recall the picture from earlier chapters. A multicellular animal is a population of cells that share one genome and have been built to do different jobs. Most are somatic; they participate in the body and do not pass on their genes. Only a small population, sperm and eggs, constitutes the germline. For the somatic majority, the cooperative deal is sharp. Divide on schedule. Stop dividing when neighbours signal. Perform specialised work, and die on cue when damaged or no longer needed. None of this is automatic. It is enforced.

In 2000, Doug Hanahan and Robert Weinberg published a much-cited synthesis listing the hallmarks of cancer: eight cell-level behaviours that separate a cancer cell from a normal one [Hanahan & Weinberg 2011]. Keeping growth signals on. Getting around the brakes on growth. Resisting the cell-death program. Dividing indefinitely. Recruiting blood supply. Invading other tissues. Rewiring energy metabolism. Evading immune destruction.

Athena Aktipis and colleagues [Aktipis et al. 2015] made a clean observation: the hallmarks are the photographic negative of cooperation. They list five things every multicellular body has to enforce (controlled proliferation, controlled cell death, division of labour, sharing of resources, and maintenance of the extracellular environment), and notice that each Hanahan–Weinberg hallmark is one of those foundations failing. Sustained proliferative signalling is the failure of controlled proliferation. Resistance to cell death is the failure of controlled death. Invasion is the failure of division of labour. Metabolic rewiring is the failure of resource sharing.

Cancer is the same set of cooperative duties failing one at a time, in different combinations.

The enforcement toolkit, by name

Earlier chapters introduced these components in their building roles; here they reappear in their policing roles. Five names do most of the work in the medical literature.

  • TP53, a protein commonly called p53, is the cell’s central damage-and-stress responder. When DNA is damaged or growth signals come at the wrong time, p53 either halts the cell cycle for repair, pushes the cell into a quiescent state called senescence, or triggers the controlled cell-death program. Mutations in TP53 are present in roughly half of all human cancers [Hanahan & Weinberg 2011].
  • RB1 is the gatekeeper of the transition into DNA copying. Losing RB1 takes the brake off cell-cycle entry.
  • BRCA1 and BRCA2 repair the most dangerous kind of DNA damage, the double-strand break. Losing either substantially raises lifetime risk of breast and ovarian cancers.
  • APC keeps a particular growth-signalling pathway (Wnt) from being permanently on in the intestinal lining; loss is the starting event in most colorectal cancers.
  • PTEN opposes the cell’s main “should I grow?” signal, the PI3K/AKT pathway.

These work alongside tissue-level enforcement: contact inhibition, mediated by adhesion molecules like E-cadherin, which gives a normal cell a growth-stop signal when it bumps into its neighbours. And organism-level enforcement: the immune system patrolling for cells whose surface markers have drifted.

Cancer is multistep

A persistent piece of folk biology says cancer is caused by a mutation. This is wrong, and the way it is wrong matters. A single mutation in a healthy cell almost never produces cancer. The enforcement system is redundant, with multiple layers each capable of catching what the others missed. A cell with a broken proliferation brake still gets killed by the cell-death program if it is detected as damaged. A cell that has lost the cell-death program still gets stopped by contact inhibition. And so on.

Turning a normal cell into a malignant one usually requires five to eight separate driver events, accumulated over years to decades [Hanahan & Weinberg 2011]. Each event disables one component of the enforcement system. Cancer is what happens when enough components have failed, in sequence, that no remaining layer can stop the cell from doing what an unencumbered cell would do anyway.

The cells in your body, if you stripped away the enforcement, would not start writing poetry. They would do what cells have done for several billion years: divide when nutrients are available, refuse to die. The enforcement is what prevents this default.

The polyphyly move: cancer is everywhere multicellularity is

If cancer is the failure of an enforcement system, and the enforcement system is itself a recurring feature of multicellular lineages, then cancer-like phenomena should be recurring too. Not unique to humans, not even unique to mammals. The data agrees. Tumour-like growths are documented in mammals, birds, reptiles, amphibians, fish, insects, molluscs, the cnidarian Hydra, and in plants (galls, fasciations, crown gall). Cancer-like cheater mutants have been found in the green alga Volvox [Aktipis et al. 2015; Albuquerque et al. 2018; Domazet-Lošo et al. 2014; Hanschen et al. 2020]. Where multicellularity has been looked at carefully, cancer-like cell-line defection has been found.

One further pattern. Large, long-lived animals get less cancer than the simple arithmetic of more cells × more divisions × longer lifetime would predict. This is Peto’s paradox, and it has been resolved separately in separate lineages. Elephants carry roughly twenty copies of the TP53 gene; humans have one [Sulak et al. 2016; Abegglen et al. 2015]. Naked mole rats use an unusually large version of a sugar called hyaluronan to enforce contact inhibition. Cetaceans (whales, dolphins) have lineage-specific augmentations to DNA repair [Tollis et al. 2017]. The same problem (don’t let any cell start dividing for itself) re-solved, with different molecules, in different mammalian lineages. The polyphyly thesis, on the inside of an animal body.

The rest of the chapter takes one question — what is happening, at the level of regulation, when a cell breaks the cooperative deal? — and walks it through a fifteen-year argument that has only recently moved, and only in part.

→ Continue with Part 2 — The atavism hypothesis at its strongest, and the bioelectric framing alongside it.

What this part draws on:

Part 2 — The atavism hypothesis at its strongest, and the bioelectric framing alongside it

So far the story has been mechanism: cancer is a failure of the cooperation toolkit, layer by layer, over years. A second question sits on top of that mechanism. When a cell fails the cooperative deal, what is the genetic program it defaults to?

Two contemporary research programs come at that question from different angles. They are sibling candidates working at different layers, and both can be true at once. This part presents them in turn.

What “ancient” means when applied to a gene

A piece of vocabulary first. When biologists call a gene “ancient,” they do not mean rusty or primitive. They mean something precise: closely related sequences for that gene are found across a wide swath of the tree of life. A gene found in animals, plants, fungi, bacteria, and archaea is older than a gene found only in animals, because for it to be in all those lineages, the ancestor of all those lineages must already have had it.

The method is called phylostratigraphy. You take a gene, ask what other organisms have a closely related version, and place its evolutionary origin at the deepest common ancestor of all the lineages that share it. “Old” means widely shared, hence inferred deep in the tree. A date label, not a value judgement. Most of the cell’s housekeeping machinery (DNA replication, protein synthesis, the basics of metabolism) is ancient in this sense because every living cell has needed it.

The atavism hypothesis

In 2011, the physicists Paul Davies and Charles Lineweaver proposed a particular reading of what cancer cells do at the regulatory level [Davies & Lineweaver 2011]. They called it the atavism hypothesis, atavism being an old word from comparative anatomy meaning reversion to an ancestral form. Their proposal: when a cell escapes the multicellular enforcement system, it does not just disable individual brakes at random. It preferentially switches back on regulatory programs that are older than multicellularity itself, programs from the era when the cell’s ancestors were single-celled.

Two pieces of evidence got the hypothesis moving. The first was a 2010 study by Tomislav Domazet-Lošo and Diethard Tautz [Domazet-Lošo & Tautz 2010], who applied phylostratigraphy to the genes involved in cancer and found a striking two-peak pattern. One peak sits very deep, around the origin of cellular life: genes for genome maintenance (DNA repair, cell-cycle machinery), the so-called “caretaker” genes. The second peak sits at the metazoan stem, around 700 million years ago, and corresponds to genes governing how cells signal to each other, how growth is restrained in a tissue context, and how cell death is regulated. The “gatekeepers.” Cancer-related genes cluster either in the deep-and-shared toolkit or in the metazoan-specific cooperation toolkit. Care of the cell (old) and care of the tissue (animal-stem).

The second piece of evidence, in 2017, was a study by Anna Trigos and colleagues [Trigos et al. 2017]. Across seven solid tumour types (breast, lung, colon, others) they asked a question. When you compare tumour cells to matched normal cells, what kinds of genes are up-regulated, what kinds are down-regulated? The pattern was consistent. Unicellular-origin genes were preferentially turned up; metazoan-origin gatekeepers were preferentially turned down. Different cancers, same direction. Cancer cells, at the level of which genes are switched on, were behaving more like single cells than like tissue cells.

Davies, Lineweaver, and colleagues have refined the hypothesis since. The Serial Atavism Model [Lineweaver et al. 2021] proposes that cancer progression is an ordered sequence rather than a single backwards step. Multicellularity lost first, then more recent eukaryotic features (mitochondrial energy chemistry, adaptive immunity) lost in roughly the reverse order in which they were originally acquired. Kimberly Bussey and colleagues, separately, treat the unicellular gene-expression state as an attractor, a configuration the cell’s regulatory network tends to settle into when the constraints holding it elsewhere are removed [Bussey et al. 2017; Bussey et al. 2021; Vinogradov & Anatskaya 2023].

The hypothesis at its strongest: when multicellular cooperation breaks down inside a cell, the regulatory state it falls into is preferentially the regulatory state of a free-living single cell. Not random, and not a continuation of normal animal physiology.

This is a contested claim. The objections come in Part 3. First, the other framework.

The bioelectric framing, as a sibling

The second research program comes at the question from a different angle. It looks at the signalling layer that, in healthy tissue, governs the cooperation the genes serve.

That layer is bioelectric. Every cell membrane carries a small electrical gradient (a voltage), set up by the unequal distribution of charged ions, mostly potassium and sodium, inside and outside the cell. In an excitable tissue like a nerve or a muscle, you have heard of this voltage in the form of action potentials. Every cell membrane has a voltage, including cells outside the nervous system, including bacteria, including plant cells.

In animal tissue, neighbouring cells couple their voltages through small protein-lined channels called gap junctions: direct physical conduits between the insides of two adjacent cells. A gap junction lets ions and small molecules pass from one cell to the next. Two consequences follow. Neighbouring cells share metabolites, and their membrane voltages tie together, so that across a sheet of cells you get a tissue-level voltage pattern. Different regions hold different stable voltages, like a low-resolution electrical map of where one tissue type ends and another begins [Pezzulo et al. 2019; Cervera et al. 2023; Whited & Levin 2019].

These tissue-level patterns are not only consequences of what the genes are doing. In some experimental systems they appear to be partly instructive. Manipulate the voltage and you can change how the tissue grows. The clearest work was done in flatworms called planaria, which regenerate from small fragments. Altering the bioelectric pattern across a fragment can change the morphology of what regenerates from it, even with the same DNA in the same cells.

For cancer, the relevant finding is twofold. Tumours routinely display abnormal membrane voltages and disrupted gap-junction coupling. The depolarised-voltage signature of cancer cells has been documented across tumour types, in a literature stretching back decades [Yang & Brackenbury 2013].

The second finding distinguishes “marker” from “mechanism.” In a 2013 study in the frog Xenopus, Brook Chernet and Michael Levin found that experimentally induced tumour-like structures showed abnormal voltage at their precursor sites before the tumours were histologically visible. The voltage abnormality came first, not as a downstream consequence. When they experimentally restored the normal voltage pattern (by misexpressing certain ion-transporting proteins to repolarise the affected cells), tumour formation was significantly reduced [Chernet & Levin 2013]. Biophysical models propose that the loss of bioelectric regionalisation (the loss of crisp tissue-level voltage boundaries) is itself part of the mechanism by which a region of tissue stops behaving like a coordinated whole [Riol et al. 2021].

Why they are siblings, not rivals

Atavism is about which genes get re-activated in a cell that has escaped cooperation. Bioelectric pattern memory is about which signalling layer governs the tissue-level coordination that the gene-expression changes break out of. They sit at different layers of the same system. A cell could have its membrane voltage decouple from its neighbours (a bioelectric failure) and fall into an attractor of ancient gene expression (an atavism signature). Both can be true simultaneously. Neither has refuted the other.

A note on something the chapter is deliberately not engaging with. There is a broader interpretive framework, developed by Levin and others, that uses words like cognition, memory, learning, and decision to talk about what tissues and cells are doing when they coordinate. That framework, sometimes called basal cognition or scale-free cognition, is contested in ways the mechanistic bioelectric claims above are not [Newman 2023]. The mechanistic claims about voltage gradients, gap junctions, and pattern memory are well-supported. Whether the right vocabulary for those mechanisms is cognitive or purely physical is a separate and live disagreement. This chapter cites only the mechanistic work.

→ Continue with Part 3 — The natural objection, and the 2025 test that controlled for it.

What this part draws on:

Part 3 — The natural objection, and the 2025 test that controlled for it

The atavism hypothesis has had a serious counter-argument since the day it was proposed, and the counter is not a strawman. It is the version a thoughtful sceptic would build. Through 2024, it was widely treated as the decisive problem with the hypothesis [Daignan-Fornier & Pradeu 2024].

The objection, in plain English

Old genes are not just any genes. Many of the things a cell has to do to grow (copy its DNA, build new proteins, generate energy, take up nutrients) are tasks the very earliest cellular life on Earth had to solve four billion years ago, when there was nothing but single cells. The molecular machinery for those jobs is old by simple necessity: every living cell has needed it for the entire history of life. Old, in the phylostratigraphy sense, correlates with core in the cell-biology sense.

Now consider what a cancer cell does. A cancer cell, by definition, grows. It divides when it shouldn’t. It builds new proteins to support that division. It copies its DNA. It generates the energy required. More intensely than a normal differentiated cell, it does exactly the things the old core machinery is built to do.

So what does the phylostratigraphy result actually show? Tumour cells over-express old genes. Of course they do; old genes encode growth machinery, and tumour cells are growing. The pattern would look the same if cancer cells were not reverting to anything at all, if they were just growing, and the growth machinery happened to be old. The signature would look like atavism while being nothing more interesting than rapid cell division on a phylostratigraphic background.

This is the proliferation confound. Any time you observe ancient-gene up-regulation in a tumour, you have to ask: would I see this pattern in any rapidly dividing cell? A rapidly dividing embryonic stem cell uses the same growth machinery. A regenerating liver hepatocyte uses the same growth machinery. If the signature is just cells dividing fast, the atavism interpretation loses its empirical leg.

This was the state of the argument from 2011 through early 2025. The most pointed recent version came in a 2024 BioEssays paper by Bertrand Daignan-Fornier and Thomas Pradeu [Daignan-Fornier & Pradeu 2024], who argued that the hypothesis was hard to falsify in its strong form. Any expression-pattern shift toward older genes can be retro-fitted to look like “ancestral program reactivation,” and the alternative explanation (cancer cells preferentially exploiting core growth machinery because it is readily available, regardless of evolutionary age) was at least as well supported by the data.

What changed in 2025

Earlier this year, a paper appeared in the International Journal of Cancer by Alexander Vinogradov and Olga Anatskaya, researchers who had been working on the atavism question for years and who had developed the unicellular-attractor framing from Part 2 [Vinogradov & Anatskaya 2025]. They set out to run exactly the test the critique called for.

One more piece of vocabulary. The standard way of measuring gene expression in tissues, since the early 2000s, has been bulk RNA sequencing. You grind up a tumour sample, extract all the RNA from it (RNA is the working copy of a gene’s instructions), and measure how much of each gene’s RNA is present on average across the whole sample. The result is one expression number per gene per tumour. A tumour is a heterogeneous community of millions of cells in different states, and bulk measurement averages them together.

Single-cell transcriptomics does the same thing per cell. The tissue is dissociated; each cell’s RNA is captured separately; you end up with a separate gene-expression profile for every single cell, thousands or tens of thousands of profiles from one tumour. Instead of “the tumour, on average, expresses these genes,” you can now see “in this tumour, this fraction of cells looks like one thing and this fraction looks like another.” A different kind of microscope.

The Vinogradov and Anatskaya team ran 38 paired comparisons drawing on published single-cell datasets, more than 18,600 cells in total. Each comparison was a tumour sample vs. a closely matched normal sample (same tissue, same general context), so what they were measuring was the difference between cancer and normal, not the baseline expression of either alone.

They applied a cell-cycle correction, which is what addresses the objection directly. Cells in different phases of the cell cycle have very different expression profiles, because dividing cells run a particular suite of genes that resting cells do not. The mathematical correction effectively asks: if these two cells had been at the same point in their cell cycle, would I still see this difference? If the answer is no, the apparent atavism signature was just proliferation, and the critique wins. If the answer is yes (even after matching, cancer cells still over-express old genes and under-express the metazoan-stem ones), the critique is answered.

The answer was yes.

What makes the result interesting is the structure of the contrast. The team compared two different kinds of pairing in the same study. Pairing A was stem cells vs. differentiated cells: developmental immaturity vs. maturity, both within normal tissue. Pairing B was cancer vs. matched normal. In Pairing A, what dominates is a developmental (ontogenetic) signature; the unicellular signature drops out. In Pairing B, what dominates is the unicellular signature; the developmental signature drops out.

That contrast is where the result gets its bite. If the developmental signature were the real explanation of cancer’s expression patterns (a popular alternative reading: cancer cells as dedifferentiated, recapitulating an embryonic state), the method would have detected it. It did detect it, in the stem-vs-differentiated comparison. It just did not detect it in the cancer-vs-normal comparison; what it detected there was something else.

The careful language about what this result is and is not

The atavism hypothesis passed the most direct test that had been attempted. That phrasing is doing real work.

It is not “atavism was proven.” Single experiments, even careful ones with large sample sizes, do not prove conceptual claims. They put down evidence. This is a strong piece of evidence, and it removes the proliferation confound as a sufficient alternative explanation for the data the field has accumulated since 2011.

It is also not “the question is settled.” Part 4 is about what the 2025 result does not settle. The 2023, 2025, and 2025b papers from the same group are one continuous research programme. Not three independent confirmations from three different labs. Other labs will replicate, or fail to. Conceptual questions about whether the empirical pattern constitutes “reversion” remain.

What the test does is shift the empirical floor. The strongest objection through 2024 — it’s just proliferation — does not survive the proliferation-corrected single-cell comparison. The hypothesis now has to be engaged with on its merits, not dismissed on its confound.

The method from Chapter 1, Part 6 applies. We are not going to pick a side. The reader is in possession of the same evidence the field is in possession of, and the field has not converged. Some careful researchers find the post-2025 evidence sufficient to take the atavism framing seriously as their working interpretation. Others, including Daignan-Fornier and Pradeu, continue to hold that the conceptual claim is under-determined by the data. Both positions are defensible.

→ Continue with Part 4 — What this still doesn’t settle, and the honest sentence.

What this part draws on:

Part 4 — What this still doesn’t settle, and the honest sentence

The 2025 result is significant, and the previous part deliberately did not undersell it. Overstating it would be worse. Three open problems survive the test, and they are not minor footnotes. They are the live edge of the field after 2025.

Open problem one: cancer metabolism does not fit cleanly

Cancer cells handle energy strangely. The most famous strangeness is the Warburg effect, named for Otto Warburg, who noticed in the 1920s that tumours often produce energy by breaking sugar down to lactic acid even when plenty of oxygen is available, instead of using oxygen-based respiration the way most healthy cells do. Wasteful per glucose molecule, but fast.

The Warburg effect fits the atavism framing reasonably well. Anaerobic glycolysis (the lactic-acid pathway) is older than oxygen-based metabolism. Earth’s atmosphere did not contain significant oxygen until around two and a half billion years ago. Cancer cells reverting to a metabolism that predates oxygen looks like exactly what the Serial Atavism Model predicts.

The problem is that cancer metabolism is not just the Warburg effect. A 2024 review by Eric Fanchon and colleagues found that the metabolic states of real tumours span a wider spectrum than reversion-to-anoxic-glycolysis can explain [Fanchon et al. 2024]. Some tumours use oxidative respiration normally. Some switch between modes depending on context. Some develop dependencies on specific amino acids that have no obvious mapping onto any ancestral metabolic state. The Warburg effect is consistent with atavism; the full metabolic picture is not.

Fanchon’s conclusion narrows the domain in which atavism applies cleanly rather than rejecting it. The framework works well for the cooperation-defection signatures (the up-regulated unicellular genes, the down-regulated gatekeepers from the Trigos and Vinogradov studies). It does not work as cleanly for the metabolic phenotypes.

Open problem two: polyploid giant cancer cells reverse the direction

The second open problem is the strangest, and the most recent. Under severe stress (typically a course of chemotherapy strong enough to kill the bulk of a tumour), a small fraction of cancer cells survive by doing something peculiar. They keep duplicating their DNA without dividing, so a single cell ends up containing multiple copies of the genome. Sometimes four, sometimes eight, sometimes more. The result is one large, multi-nucleated, often-quiescent cell: a polyploid giant cancer cell (PGCC).

These cells were dismissed for decades as oddities. They turn out to matter quite a lot, appearing to be the resistant fraction that re-seeds the tumour after treatment, fragmenting back into individual cells over days to weeks.

In 2025, the same Vinogradov and Anatskaya group looked at what these polyploid giant cells do at the level of gene expression and found something that complicates the serial-atavism picture in a useful way [Vinogradov & Anatskaya 2025b]. When cancer cells switch into the polyploid-giant state under stress, they suppress the unicellular-attractor genes (the ancient ones atavism is built around) and up-regulate multicellular-coordination genes, particularly the ones for intercellular communication. They transiently start behaving more like cells in a coordinated multicellular collective than like cells in a defection state. When the stress lifts, they fragment back into single cells and resume the standard cancer pattern.

Cancer cells taken as a whole run serial atavism. Polyploid giant cells transiently reverse it. The same cell line can occupy multiple regulatory attractors depending on context. The unicellular attractor is one of several, not the only one cells default to.

Open problem three: the conceptual interpretation remains contested

The third problem is the deepest, and the 2025 result does not touch it.

Suppose we accept everything the data shows. Cancer cells over-express genes whose closest relatives across the tree of life are in single-celled lineages. They down-regulate genes specific to the animal stem. The pattern survives proliferation correction. The pattern is shared across many tumour types. What does this mean?

There are at least two readings of the same data.

Reading one is the strong-atavism reading. The cell, freed from multicellular enforcement, falls back to a regulatory state that actually was its single-celled ancestors’ state, hundreds of millions of years before animals existed. The match is not a coincidence; it is a memory of an earlier mode of life, written into the genome’s architecture, and reactivated when the constraints holding it suppressed are removed.

Reading two is more cautious. The cell falls into the regulatory state that the architecture of the genome makes most accessible. The genes most deeply embedded in the cell’s core machinery (the most connected, most central) happen to be the old ones, because old genes have had longer to accumulate connections. Take the brakes off, and the network gravitates toward its centre of gravity. That centre happens to be ancient. The cell is not remembering its single-celled past; the network’s centre of gravity simply sits where the old genes are.

These two readings make the same predictions about gene expression. The 2025 single-cell test cannot distinguish them. A 2017 paper by Frédéric Thomas and colleagues argued that the choice between them is not even binary; cancer phenomena likely span a continuum between pure atavism and pure de novo selection on cell-population fitness, with different tumours sitting at different points [Thomas et al. 2017]. The Daignan-Fornier and Pradeu critique partly survives 2025 in this form. The conceptual question of whether the empirical pattern constitutes reversion is not resolved by improving the experiment that establishes the empirical pattern. It is a question about the meaning of “reversion,” not about whether the pattern is real.

The honest sentence

The sentence below, in the corpus this chapter draws on, expresses the post-2025 state of the field.

Cancer preferentially switches back on evolutionarily ancient regulatory programs. After two decades of debate, the most carefully controlled test now supports this. Whether that constitutes “reversion” depends on what you mean by “reversion.”

The empirical claim is supported. The conceptual interpretation is not settled. That is the honest description of the field in 2026.

The voice rule from Chapter 1 Part 6 applies again. When serious researchers disagree, we show each side at its strongest, and we tell you where the live edge is. The live edge here is not the test result; that has moved. The live edge is what to call the thing the test result establishes. Reversion is one available word; attractor dynamics is another; exploitation of network centrality is a third. They make different demands on how we describe what cancer is.

Both atavism and its critics survive 2025, in different ways. The hypothesis survives because the strongest objection has been answered. The critics survive because the conceptual interpretation remains under-determined. Neither has been refuted. This is what a live science looks like.

→ Continue with Part 5 — The asymmetry: yeast reverts and re-evolves; cancer reverts and stays reverted.

What this part draws on:

  • Cancer metabolism not fitting cleanly into the reversion model: [Fanchon et al. 2024].
  • Polyploid giant cancer cells transiently re-acquiring multicellular coordination: [Vinogradov & Anatskaya 2025b].
  • The conceptual interpretation as a continuum, not a binary: [Thomas et al. 2017]; the post-2025 survival of the under-falsifiability critique: [Daignan-Fornier & Pradeu 2024].
  • The honest sentence and the steelmanning method invoked a second time: content/05-breakdown-and-fragility/cancer-as-cooperation-breakdown.md; chapters/05-the-door.md Beat 5.6; explainer/01-the-bush.md Part 6.

Part 5 — The asymmetry: yeast reverts and re-evolves; cancer reverts and stays reverted

The chapter has so far walked through one case of multicellularity coming undone: cancer, in real time, inside an animal body. The corpus contains another case. Snowflake yeast, in a flask, evolving cluster-formation under one selection regime and losing it under another. We met that system in Chapter 4. The two cases look superficially similar; both are multicellular states giving way under stress. Looked at carefully, they are not the same kind of thing.

What follows is project synthesis, and it should be flagged as such. The yeast literature and the cancer literature each stand on their own. The comparison across them is what is novel. The corpus assembles this comparison; the primary literature has not yet stated it quite this way. Both halves of the comparison are well-attested. The asymmetry is the synthesis.

What yeast does

Recall the snowflake yeast system. William Ratcliff and colleagues, starting in 2012, took ordinary baker’s yeast (Saccharomyces cerevisiae, a single-celled fungus) and grew it in liquid culture under selection for fast settling [Ratcliff et al. 2012]. Each day they let the cells settle, kept the bottom fraction, and discarded the rest. After about sixty days, roughly three hundred generations, the cells were no longer growing as singletons. They had evolved a mutation in a gene called ACE2 that prevents mother and daughter cells from separating after division, producing branched clusters that descend from one founder. Cluster-cells settle faster than singletons, so under settling selection they win. The lineage evolved a simple, clonal form of multicellularity in about two months.

What is relevant here is what happens when you reverse the selection. Gabe Bozdag and others showed that if you switch the cells into a regime that favours singletons (for example, by selecting for cells that stay suspended in the upper part of the flask), the clusters can lose their clustering mutation again, within a comparable number of generations [Bozdag et al. 2020]. The reversion does not retrace the forward path molecularly; different mutations dissolve the cluster than were responsible for building it. At the level of phenotype, the cells revert to a single-celled lifestyle. The door swings closed; the door also swings open. Multicellularity in this system is not, in the strict sense, irreversible.

A 2019 follow-up by María Rebolleda-Gómez and Michael Travisano [Rebolleda-Gómez & Travisano 2019] worked out what kinds of selection are sufficient to drive reversion in this system, and how much history the lineage carries (whether the route in determines the route out). The same phenotypic endpoint of single cells can be reached from cluster-bearing populations along somewhat different molecular paths.

The newest result is from 2025, by Khey and colleagues [Khey et al. 2025]. They asked a question one step further. Once a snowflake-yeast lineage has reverted to single cells, can it re-evolve multicellularity if the settling selection is switched back on? The answer was yes in some lineages and not in others. Some reverted lineages re-evolved cluster formation under settling selection, sometimes via the same kind of mutation, sometimes via a different mechanism (including via selfing, the lineage’s ability to make new combinations via genetic recombination from variants in the population). Other reverted lineages did not re-evolve clusters even under prolonged selection. They had lost too much, in too specific a way, to easily get back. Historical contingency (the dependence of the future on the path taken) is real but bounded. The door swings both ways for some lineages; for others, the hinges have rusted.

Yeast reverts under different selection. Some yeast lineages, after reverting, can re-evolve multicellularity if the original selection is reapplied. The transition is bidirectional in this experimental system. The ratchet, in the lab, is not absolute.

What cancer does, by comparison

Cancer cells, inside an animal body, also exit the multicellular state. They escape the regulation of their tissue context, shed the cooperative behaviour the body requires, become in many functional senses single cells again. The metastatic ones physically leave the tissue, travel through blood vessels, and lodge elsewhere, where they continue dividing as effectively unicellular parasites of the body’s resources. The transmissible cancers (CTVT in dogs, DFTD in Tasmanian devils, the bivalve transmissible neoplasias) take this further still: cell lineages that have left their original host entirely and now propagate from individual to individual, as asexually reproducing unicellular parasites with a mammalian or molluscan genome [Murchison et al. 2014; Metzger et al. 2016; Ujvari et al. 2016]. These are, in a precise sense, reversions to unicellularity in lineages whose host organisms are otherwise obligately multicellular.

Cancer cells, in any of these forms, have never been observed re-evolving multicellularity in any natural setting. The cells leave the cooperative state and do not return to it. Even the polyploid giant cancer cells from Part 4, which transiently up-regulate multicellular-coordination genes under treatment stress, are not building a new cooperative body; they are surviving as resistant individuals and then fragmenting back into single cancer cells. Cancer reverts and stays reverted.

Yeast in a flask: out, and back, and sometimes out and back again. Cancer in an animal body: out, and out only. That is the asymmetry.

Why the asymmetry — and why it is mechanistic, not moral

Why? Not because cancer cells are worse at it, or because their machinery is broken in a different way. The reason is structural, and it has to do with what the surrounding environment is doing.

A snowflake-yeast lineage reverts in an empty environment. The flask contains nutrient broth and other yeast cells and not much else. Whatever a cell does is judged solely by the selection regime the experimenter has set up. The environment is not actively penalising any particular cellular state beyond what the selection regime imposes.

A cancer cell reverts in a hostile environment. Every part of the body around it is running the enforcement toolkit from Part 1. Contact inhibition tells normal cells to stop dividing when they touch their neighbours. The immune system patrols for cells with abnormal surface markers. Tissue-organisation signals try to slot every cell back into the place its tissue type demands. A cancer cell can only persist by avoiding these signals, by not reintegrating with the cooperative tissue. Reintegration would mean re-acquiring the differentiation markers that the cancer escape required disabling, and at the level of population-genetic selection within the body, any cancer-cell variant that successfully started reintegrating would be the variant the body could most easily eliminate.

That sentence describes a mechanism, not a moral. The body is not “fighting back” in any sense that requires intent. The enforcement system from Part 1 makes one direction of motion (toward defection) compatible with cell-line survival in this environment, and the other direction (toward re-cooperation) lethal to it. The asymmetry comes from the environment being an active selector, not an empty one.

The yeast flask is the empty environment. The mammalian body is the active environment. The same kind of cell-line process, in the two settings, looks bidirectional in one and one-directional in the other, because the surrounding selective regime is different.

The comparison rhymes with what is happening at lineage scale, on a much longer timescale, in the polyphyly map you have been carrying since Chapter 1. Reversion to unicellularity has been documented at the lineage level multiple times. In the budding yeasts (the Saccharomycotina, derived from filamentous fungal ancestors) [Naranjo-Ortiz & Gabaldón 2020]. In Helicosporidium, a derived unicellular green alga now an insect parasite, descended from photosynthetic multicellular relatives [Pombert et al. 2014]. In Gloeocapsopsis, a unicellular cyanobacterium that retains genomic signatures of multicellular filamentous ancestry [Urrejola et al. 2020]. And in cyanobacterial lineages more broadly [Schirrmeister et al. 2011]. Some of these reversions are tens or hundreds of millions of years old. The polyphyly bush from Chapter 1 contains reversion dots on the same lineages. Cancer is the same phenomenon on a much shorter timescale, inside one body, in one lifetime, in a particular kind of selective environment set up to penalise it.

A note before moving on

Active selection against re-cooperation is qualitatively different from no selection for re-cooperation. A yeast lineage in a flask without settling selection has no force pushing it back into clusters. A cancer cell line in a body has a force pushing it away from anything that looks like re-cooperation. Different mechanical situations, different observed outcomes.

This framing also explains a small puzzle that might otherwise nag. Why have biologists not seen cancer cells re-evolving multicellularity even in the long-running culture systems where cancer cell lines are grown in dishes for years? The dish, while less hostile than a body, is also not selecting for re-cooperation. The conditions that would actively favour re-cooperation in a population of cancer cells (the conditions a multicellular body would have provided to the ancestral metazoan stem cell, six hundred million years ago) are absent.

The asymmetry is real. Both halves are well-documented. The comparison across them is novel here: two independent literatures placed next to each other.

→ Continue with Part 6 — Cancer, the ratchet, and the arrow that goes both ways.

What this part draws on:

Part 6 — Cancer, the ratchet, and the arrow that goes both ways

Same toolkit, different direction

Cancer is what multicellularity looks like when the cooperation toolkit fails one component at a time. The cadherins and other adhesion molecules that hold cells in tissue; the signalling pathways that tell cells when and where and whether to divide; the contact-inhibition machinery; the programmed-cell-death machinery; the immune surveillance that watches for cells whose surface markers have drifted — these are cancer when they break, in the right combinations. The same molecular components that built the cooperative state.

There is no body-cooperation pathway and, in parallel, a cancer pathway. One set of molecular components performs the cooperative job when it works and produces cancer-like phenotypes when it fails. Cancer is the same machinery running in the other direction.

That formulation also dissolves a tempting framing the chapter has been refusing. Cancer is sometimes described as a failure of the body’s plan, as if the body had a design specification and cancer was the deviation from it. That framing imports a teleology the biology does not contain. The body does not have a plan. What the body has is a population of cells that are, generation by generation, kept on the cooperative side of a balance by a stack of enforcement components, each of which can fail. Cancer is what the same components produce when the balance tips the other way.

Reversion at every scale

Multicellular cooperation giving way is not a curiosity that only happens inside animal bodies. It happens at every scale on which multicellularity exists.

At the lineage scale, in geological time, multicellular lineages have lost their multicellularity multiple times. The budding yeasts, derived from filamentous fungal ancestors that were complex enough to form mushroom-like fruiting bodies, collapsed back to a single-celled lifestyle and stayed there. Helicosporidium, an obligate parasite of insects, descends from a photosynthetic multicellular green-algal ancestor and now lives as a single cell inside the gut of beetles. Gloeocapsopsis, a unicellular cyanobacterium, still carries in its genome signatures of the filamentous multicellular ancestry it has otherwise abandoned [Naranjo-Ortiz & Gabaldón 2020; Pombert et al. 2014; Urrejola et al. 2020]. The polyphyly bush you have been carrying since Chapter 1 is not just a map of inventions; it has reversion dots flickering on the same lineages.

At the cell-line scale, inside a single animal body, in a single lifetime, cancer is the same phenomenon on a much shorter timescale. The transmissible cancers (CTVT in dogs, the two DFTD lineages in Tasmanian devils, the bivalve transmissible neoplasias) carry it one step further: somatic-cell lineages that escaped their original body and now propagate from host to host, as effectively unicellular parasites with a mammalian or molluscan genome [Murchison et al. 2014; Metzger et al. 2016; Ujvari et al. 2016]. The same kind of event as a yeast lineage losing its mushrooms, compressed into the lifetime of a single dog or a single Tasmanian devil.

These are not the same event. They are the same kind of event, cooperative-state-to-defection-state transitions, at radically different scales, with the regime-difference from Part 5 determining whether reversion is bidirectional or one-way.

Cancer, the ratchet, and the central correction

Chapter 4 introduced the ratchet: the set of mutations and developmental commitments that, accumulated over time, raise the cost of reverting to a single-celled life and lock a lineage into a multicellular one. The ratchet has Type 1 components (mutations that improve fitness within the multicellular context but reduce fitness for an isolated revertant cell) and Type 2 components (mutations that lower the probability that a future mutation will produce a viable revertant). Together they make reversion progressively harder. They do not abolish it [Libby & Ratcliff 2014].

Cancer is what a ratchet looks like when it slips a tooth. Not all the teeth; a single mutation in TP53 does not produce cancer (Part 1). The enforcement system is redundant, and that redundancy is one of the most important Type 1 ratchets the animal body has. But over decades, over chains of mutations, the ratchet slips. A cell that has shed the constraints multicellularity placed on it does what a free single-celled cell would do, and its regulatory state moves toward the ancient programs that did that job for billions of years before there was any multicellular cooperative state at all (Part 3).

Multicellularity is not a finished state. It is a balance. The cooperative arrangement that an animal body looks like from the outside is at every moment being maintained: within a generation against somatic mutation and within-host selection, across geological time against the lineage-scale reversion processes that have happened in fungi and algae and cyanobacteria. Ratchets bias the balance toward staying cooperative. They make reversion progressively harder in lineages where they have accumulated. They do not make it impossible.

The popular framing of multicellularity as a one-way achievement, an arrow pointing from simple-and-old to complex-and-new, imports the very ladder Chapter 1 worked to dismantle. The data shows multicellularity arising at many places on the tree (Chapter 1’s bush), the cooperation problem solved each time using lineage-specific molecular tools (Chapter 3), ratchets accumulating in some lineages and not in others (Chapter 4), and reversion documented at every scale at which multicellularity exists (this chapter). The arrow goes both ways. Cells, lineages, kingdoms have all been observed reversing direction under the right conditions, and those conditions are mechanical: enforcement systems, selective regimes, and the ratchets that bias which way a balance tips.

The bridge

Chapter 6 takes the bush back out. Same picture you have been looking at since Chapter 1, but with everything Chapters 2 through 5 have added: the cooperation problem every multicellular lineage had to solve (Chapter 2), the molecular toolkit each lineage solved it with — sometimes by inventing molecules from scratch and sometimes by re-purposing parts already there (Chapter 3), the ratchet that locks a lineage in when it accumulates (Chapter 4), and the failure modes (the reversion dots) that flicker on the same lineages where the cooperative state was first invented (this chapter).

The arrow goes both ways, always, because there is one set of molecular machinery doing both jobs: building the cooperative state and, when it slips, producing the defection state. The same toolkit, in the same lineage, can produce both. Which one shows up at a given moment depends on the balance and the enforcement and the selective regime.

That is Chapter 6.

What this part draws on:

End of Chapter 5 (draft state).

When complete, this chapter should be readable in one sitting (~9,000 words across six parts) by someone with no prior biology. It is the explainer’s most editorially careful chapter — the live debate about the atavism hypothesis required steelmanning both sides, the yeast-versus-cancer asymmetry was flagged in the prose as project synthesis rather than published finding, and the bioelectric framing was held to its mechanistic claims rather than its broader cognitive-vocabulary extensions. Chapter 6 takes the bush from Chapter 1 back out, with the cooperation problem, the molecular toolkit, the ratchet, and the reversion dots from this chapter all overlaid on the same picture.