The Many Origins

Chapter 2 — The Problem

STATUS: v0 (all seven parts drafted) Documentary parallel: chapters/02-the-problem.md (Beats 2.1–2.10, including 2.3b) Last updated: 2026-05-22

What this chapter does

Chapter 1 put a bush on the page. Multicellularity has been invented many times (at least twenty, possibly a hundred, depending how you count) on lineages that share no recent common ancestor. The picture refused the ladder. The obvious next question stayed open: if so many separate lineages all converged on the same thing, what was the thing? What does the trick consist of? Is there one problem of multicellularity, or several?

This chapter answers that. It names four functional problems any multicellular lineage has to solve, walks through how each has been solved by entirely different molecular machinery in different parts of the tree, and lands on a hard claim: the cooperation drama you have probably read (cells learning to trust each other, sacrificing autonomy for the collective) describes one kind of multicellularity and gets the rest of them wrong. By the end, you should understand:

  1. The four problems every multicellular lineage has had to solve, and why the qualifier “in some cases” in front of the fourth is structurally important.
  2. That adhesion (cells sticking to each other) has been solved at least five non-homologous ways across the tree.
  3. That cell-to-cell communication is similarly polyphyletic, with a specific factual correction to a story you may have read about a molecule called cyclic AMP.
  4. That there is a fourth communication channel (voltage across cell membranes) whose physical substrate is universal and whose molecular machinery is lineage-specific, in the same pattern as adhesion and chemical signalling.
  5. The clonal-versus-aggregative distinction, and why, under a rule written down in 1964 by W. D. Hamilton, clonal organisms get cooperation almost for free while aggregative ones do not.
  6. Why the cheater problem is a real puzzle in Dictyostelium and the myxobacteria, and a category error when imported into the animal or plant case.
  7. Why mechanism beats vocabulary, including for the cells that, at the level of voltage and ion channels, do seem to be computing something.

The chapter is in seven parts.

PartTitleApprox. length
1The four problems~1100 words
2Adhesion: at least five solutions to one problem~1500 words
3Communication: also independent, and a small correction~1750 words
4Division of labour: when one cell can’t do everything~1100 words
5Clonal and aggregative: why the cheater drama isn’t universal~2000 words
6Why “cells deciding” is the wrong story (and where the right story is open)~1500 words
7The picture so far, and the question for Chapter 3~900 words

Part 1 — The four problems

Take the bush from Chapter 1, with its bacteria, archaea, eukaryotes, and glowing dots scattered across all three branches, and ask what every dot has in common. You do not get a list of shared parts. The parts differ. You get a list of shared problems. Every lineage that crossed from single-celled to multicellular had to solve roughly the same set of them, and any account of multicellularity that doesn’t name them is missing the spine of the subject.

There are four. They are abstract, not molecules and not solutions, and the brevity of the list does some real work: the same brevity holds across every lineage we will look at, from animals to brown algae to filamentous bacteria.

Adhesion. Cells have to stay together. That sounds trivial until you remember what a cell normally does after it divides: the two daughters separate and go their own way. A multicellular organism is exactly the thing where this stops happening, where the daughters of a division remain attached, or where independent cells find each other and bind. Without some mechanism that makes “stay together” the default, no part of the rest of multicellularity can happen.

Communication. Cells that are stuck together have to coordinate. The cell at one end of a filament has to behave differently from the cell in the middle. The cell that becomes a sponge’s filter chamber lining has to know it is one, and the cell next door has to know it is something else. Coordination, at the cellular level, means information passing between cells: some physical signal that one sends and another receives.

Division of labour. Once cells can coordinate, they can specialise. One cell does this job; the cell next door does that one; together they accomplish something neither could do alone. Specialisation is what makes a multicellular body different from a heap of identical cells, and what makes multicellularity worth doing. A colony with two or three or twelve different cell types, each contributing a partial behaviour, can be a much more capable organism than any of its components.

And, in some cases, the suppression of cheating. Specialisation creates a vulnerability. If some cells in the body do the costly work, making the dispersal stalk, building the skeleton, dying so the others can survive, and other cells reap the benefits, natural selection should reward the second group. A cell that finds a way to be the one carried away as a spore rather than the one becoming the stalk that dies should leave more descendants than a cell that plays by the rules. Taken seriously, this logic threatens to collapse the whole system. If cheating is favoured, why should cooperation persist?

The “in some cases” is doing more work than it looks like, and it is the qualifier the rest of the chapter will spend most of its time on. In Part 5 we will see that the cheating problem is a real and central puzzle in some multicellular systems and barely a problem at all in others, and that the popular story conflates the two. Hold the qualifier for now; it is a structural commitment, not a hedge.

Four problems, many solutions

The strangeness of this list is that you do not usually see it. In a textbook on a single multicellular lineage, the four problems are buried inside one lineage’s particular toolkit; you see specific machinery doing each job and come away thinking you have learned about multicellularity. What you have learned about is animal multicellularity, or one other specific implementation of it.

The picture changes when you take the same four problems and ask how every lineage in the bush solved them. The labels stay the same. The solutions do not. The protein that holds animal cells together is unrelated to anything that holds plant cells together. The system that lets Dictyostelium amoebae call each other in is independent of the systems that let bacterial cells coordinate. The terminal differentiation that produces a heterocyst in a cyanobacterial filament was not inherited from a common ancestor with the Volvox somatic cells that look superficially similar; it was reinvented, with different molecules, on different parts of the tree.

Same problem, different solution, every time. The closest thing the literature has to a unified framework is Niklas and colleagues’ three-phase scheme [Niklas et al. 2013b]: adhesion with alignment of fitness across cells; then communication, cooperation, and specialisation with an export of fitness up to the organism; then, in some lineages, the further transition into complex multicellularity we met in Chapter 1 Part 5 [Grosberg & Strathmann 2007; Knoll 2011]. The four-problem framing this chapter uses is in the same spirit; suppression of cheating is broken out as its own item because it is where the popular story most consistently goes wrong.

What the four problems are not

Three clarifications. They are not a sequence. In every lineage, the four were solved simultaneously or close to it; the machinery for cells to stick together is often the same machinery that lets them signal to each other. They are not a checklist for what counts as multicellular: a young or simple multicellular organism may solve only some of them, partly, and that may be enough. And they are not the only problems: a fifth recurring solution, bioelectric signalling, fits inside communication but is universal in a different way than the chemical signals we will meet first.

One last setup for Chapter 3. Several times in what follows, you will see the same observation: the molecular parts each lineage used to solve its version of the four problems were already there in its unicellular ancestor, doing something else, before any cell ever stayed stuck to its sister. The polyphyly we drew in Chapter 1 is mostly a story about each lineage redeploying machinery that already existed. That is Chapter 3’s whole story. For this chapter, just notice the pattern when it appears.

→ Continue with Part 2 — Adhesion: at least five solutions to one problem.

What this part draws on:

  • The recurring-problem framing across origins: content/00-framework/defining-multicellularity.md; BIG-PICTURE.md §“The recurring principles”.
  • Three-phase framework of multicellularity (closest published parallel to the four-problem framing): [Niklas et al. 2013b]; [Grosberg & Strathmann 2007]; [Knoll 2011].
  • Clonal-versus-aggregative as a distinction that will become structural in Part 5: [Bonner 1998].

Part 2 — Adhesion: at least five solutions to one problem

Take the simplest of the four problems first. Cells have to stick together. There has to be something between them, some physical material in or on the membrane or wall, that holds the daughters of a division in place, or that lets independent cells lock onto each other when they meet.

Across the bush, this has been solved at least five different ways. The solutions are non-homologous: not different versions of the same ancestral molecule (the way human haemoglobin and horse haemoglobin are different versions of one), but different molecules entirely, performing the same job by different routes. Counting them up is the easiest way to see what polyphyly looks like at the level of parts.

Animals: a protein that bridges two cells

Animal cells stick to each other principally through a family of proteins called cadherins. The mechanism, before the name: a cadherin is a transmembrane protein with one end inside the cell, the middle through the membrane, and the outside end sticking out into the space between two neighbouring cells. The outside ends of cadherins on adjacent cells bind to each other, in a calcium-dependent way, tethering the two cells through a series of these protein bridges [Brunet & King 2017]. Inside the cell, the cadherin’s other end is anchored via two more proteins (β-catenin and α-catenin) to the cell’s internal scaffolding, its cytoskeleton, a meshwork of fibres that gives the cell its shape. The result is a tissue where cells are mechanically continuous: pull on one cell, and the force transmits through the cadherin bridges to its neighbours and through to the cytoskeleton of every cell in the sheet.

That is one solution to adhesion: protein bridges, cell to cell, anchored to internal scaffolding. (If you are looking for a metaphor, “velcro” is the usual one, but it is metaphor; cadherins do not have hooks, they have very specific calcium-dependent binding geometries.)

Chapter 3 will spend a chapter on the consequences of what is about to look like a small observation: cadherin-style adhesion proteins were already present in single-celled organisms before any animals existed. The closest single-celled relatives of animals, choanoflagellates and a handful of related lineages, have cadherin genes in their genomes [Abedin & King 2008; Nichols et al. 2012]. They are not multicellular. They have the parts.

Plants: cell walls glued by sugar, and cytoplasm shared through tunnels

Plant cells do not stick to each other through cadherins, or through any cell-to-cell protein bridge at all. Plants have a different solution.

After a plant cell divides, the two daughters are separated by a thin layer of sticky material called the middle lamella, sandwiched between the new cell walls each daughter is starting to build. The middle lamella is made principally of pectin, a sugar polymer chemically nothing like cadherin, held together by calcium ions that bridge between the pectin chains. The strength of the adhesion is tuneable: enzymes in the cell wall modify the pectin to expose more sites for calcium bridging, making the wall between cells firmer or softer as the developing plant needs [Daher & Braybrook 2015]. That mechanism is what lets a stem hold up against gravity and a leaf remain a leaf.

In parallel, plant cells share something animals never do: cytoplasm. Between adjacent plant cells, small channels called plasmodesmata punch through the cell walls, leaving a continuous pipe of cytoplasm running from one cell into the next. Through them, plant cells exchange small molecules, signalling peptides, even RNA. The plant body, looked at this way, is not a community of distinct cells touching at boundaries; it is a single connected cytoplasm subdivided by walls. Plasmodesmata are independently evolved, structurally different, and more thoroughgoing than the equivalent in animals.

Brown algae: a sugar mesh, with some of the genes borrowed from bacteria

Brown algae (the long straps of kelp on a rocky beach at low tide, the bladdery wracks, the more delicate filamentous forms) sit on a completely different branch of the eukaryotic tree from animals, plants, or fungi. Their closest single-celled relatives are diatoms, the algae with glass shells.

They hold their cells together with a sugar polymer called alginate, embedded in a network of related polysaccharides outside the cell membrane. Alginate is chemically nothing like cadherin and chemically nothing like pectin. It is its own thing, a chain of sugar units that cross-link into a gel-like mesh. In a kelp blade, the alginate matrix is what lets the tissue whip back and forth in a current without tearing apart [Cock et al. 2010].

A wrinkle worth flagging: brown algae did not invent the genes for alginate biosynthesis on their own family tree. A chunk of that machinery appears to have been acquired by horizontal gene transfer, the (very common, in microbes) process by which a gene moves sideways from one lineage to an unrelated lineage rather than being passed parent-to-offspring, from a bacterial donor early in their evolution [Michel et al. 2010]. The polyphyly map stays intact (the brown-algal adhesion system is still independent of the others), but the genes that build it did not all originate on the brown-algal branch.

Fungi: a cell wall of chitin, and threads that fuse on contact

Mushrooms, moulds, yeasts. Fungi grow as hyphae: long tube-like cells joined end-to-end into branching threads, held together by a cell wall of chitin (the same polymer insect shells are made of) and β-glucan, arranged into a strong fibrous network. Where two hyphal threads grow into contact, they often fuse. The cell walls between them break down and the cytoplasm of the two threads merges into one continuous network. This is called vegetative hyphal fusion or anastomosis, and it lets a fungal colony build a connected mycelium across a forest floor, a piece of bread, the soil under a tree, as a single integrated organism.

Fungal adhesion uses no cadherin, no integrin, no pectin, no alginate. Chitin walls and hyphal fusion are their own toolkit, evolved independently from any of the others [Brunet & King 2017].

Bacteria: a different system in each multicellular bacterial lineage

The bacterial cases we met in Chapter 1 (cyanobacterial filaments, Streptomyces, myxobacteria) each have their own adhesion mechanism, not even consistent across the bacterial domain.

Cyanobacterial filaments (the Anabaena-and-Nostoc style chains we met in Chapter 1) hold their cells together intimately. After a cyanobacterial cell divides, the daughters do not fully separate the outer cell-wall material, a layer of a polymer called peptidoglycan. The two cells remain connected through a shared outer wall, and small channels called septal junctions punch through the boundary, letting small molecules pass directly from one cell to the next [Mariscal et al. 2007; Flores & Herrero 2010]. The molecules involved (SepJ and FraD, among others) have no homology to animal cadherins, plant pectin enzymes, fungal chitin synthases, or brown-algal alginate enzymes; the same function is implemented in completely independent machinery.

Streptomyces uses yet another approach: shared peptidoglycan along a continuous hyphal cell wall, with the cells held in a syncytial-style arrangement. The threads form, in effect, a tube divided incompletely by cross-walls into cellular compartments [Flärdh & Buttner 2009]. Myxobacteria use surface-protein contacts and structures called type IV pili (small protein hairs the cell can extend and retract) to grip each other during their swarming and fruiting [Kaiser 2003]. Three different adhesion systems in three different bacterial lineages, on the same branch of the tree, none inherited from the others.

The same architecture, lineage-specific parts

That is at least five independent adhesion systems across the multicellular lineages, with the bacterial count alone adding three more if you split them out [Brunet & King 2017; Claessen et al. 2014; cross-summary in content/03-molecular-toolkit/cell-adhesion.md]. Look across them and ignore the molecules, and a consistent picture emerges. Every adhesion system has three pieces. There is something outside the membrane that does the actual sticking, a polymer or a protein. There is some kind of link to the inside of the cell, either to the cytoskeleton or to the cell-wall machinery. And there is some kind of regulation, a way to make the adhesion stronger when needed, weaker when the cell needs to divide or move. The convergence is on architecture, not molecules. Outside, inside, control. The architecture is what made adhesion solvable. The molecules were whatever each lineage had to hand.

→ Continue with Part 3 — Communication: also independent, and a small correction.

What this part draws on:

Part 3 — Communication: also independent, and a small correction

Adhesion gets you a clump. To get from a clump to an organism, the cells have to talk to each other, to send signals that change what their neighbours do. Communication is the second of the four problems, and like adhesion it has been solved many separate ways. There is a wrinkle: the same molecule sometimes shows up in two very different lineages doing different jobs, and that coincidence has been the source of one of the more durable misconceptions about how multicellular signalling works. This part walks through the major systems, lands the correction, and then turns to a different kind of signal, voltage across cell membranes, that fits the polyphyly pattern in a productive new way.

Bacteria: small molecules in the water around the cells

Bacteria invented chemical signalling between cells. The mechanism, before the name: a bacterium continuously releases a small molecule into the surrounding fluid. As the local population grows, the concentration of that molecule in the water builds up. At some threshold, receptors on every nearby bacterial cell detect it and trigger a coordinated change in gene expression. The cells switch on a new set of behaviours together [Waters & Bassler 2005].

This is quorum sensing: sensing the quorum, the local population, by the accumulation of a chemical the cells themselves are producing. It controls everything from bioluminescence (the original case, in marine Vibrio bacteria) to virulence-gene activation to biofilm formation. Different bacterial lineages use different chemistries: small acyl-homoserine lactones in some, modified peptides in others, a separate family called AI-2 in others again, with receptor systems to match [Waters & Bassler 2005]. There is no single “bacterial language”. Many independent signalling systems run side by side, sometimes inside the same biofilm.

(One flag, because the language is tempting: bacteria are not “talking” in any sense that means what it means for humans. They are emitting molecules and responding to molecules. Talking is metaphor, and the metaphor is the right one only if you immediately name the mechanism: chemical released, chemical accumulated, chemical detected, threshold crossed, behaviour switched.)

Dictyostelium: a chemical wave that calls cells inward

The slime mould we have met by name, Dictyostelium discoideum, the canonical aggregative multicellular eukaryote, coordinates its life cycle with a different kind of chemical signal, in an entirely independent invention.

When food runs short, Dictyostelium amoebae enter a developmental program. A small number of “pacemaker” cells start secreting pulses of a small molecule called cyclic AMP. Other amoebae detect the cyclic AMP at receptors on their cell surface, move up the gradient toward the source, and themselves release cyclic AMP to relay the signal further. The result is a propagating wave of chemical attraction, visible under the microscope as concentric or spiral patterns of inward-streaming cells, pulling ten thousand to a hundred thousand amoebae together into a single mound, which then becomes the slug and the fruiting body [Schaap 2011; Cai & Devreotes 2011]. Signalling at a different scale from quorum sensing: a propagating wave of attractant, with cells acting as both emitters and detectors, drawing distant cells together over distances thousands of times their own size.

A small correction about cyclic AMP

You may have heard a different version of the previous two paragraphs: that Dictyostelium uses cyclic AMP “the same molecule bacteria use for signalling,” now repurposed for communication between cells. That framing presents Dictyostelium’s cyclic AMP system as bacterial signalling, inherited or borrowed. It is wrong on both ends.

What bacteria actually do with cyclic AMP is, overwhelmingly, use it as an intracellular second messenger: a molecule inside one cell that conveys information from one part of that cell to another, never released. The classic example is the cyclic AMP–CRP system in E. coli, which switches on alternative carbon-utilisation genes when the cell’s preferred sugar runs out (broader context in [Sebé-Pedrós et al. 2017]). The chemicals bacteria do release as cell-to-cell signals are the quorum-sensing autoinducers we just walked through, different molecules altogether. Bacteria don’t use cyclic AMP as an extracellular cell-to-cell signal in the first place.

Dictyostelium’s extracellular cyclic AMP relay is an independent invention on the eukaryotic side, traceable within the slime-mould family tree itself. The dictyostelid slime moulds have four major groups [Schaap 2011]. In the older three, cyclic AMP is used inside cells during the later stages of building the fruiting body, a job that has nothing to do with bringing distant cells together. Only in the youngest, derived group 4, which includes D. discoideum itself, was cyclic AMP rerouted outside the cell to function as a long-range aggregation attractant [Alvarez-Curto et al. 2005; Schaap 2011]. Bacteria have nothing to do with it.

The general lesson is one we’ll see again in Chapter 3. Cyclic nucleotides, the chemical family cyclic AMP belongs to, are a deeply old class of small molecules used as cellular signals across bacteria, archaea, and eukaryotes. The chemistry is ancient and shared; what each lineage does with it is independent.

Animals and plants: lineage-specific kits doing the same kinds of job

Animal cells coordinate with a longer list of signalling systems. Hormones are carried in blood over long distances. Growth factors are released by one cell and detected by receptors on another. A small set of cell-to-cell signalling pathways (Notch, Wnt, Hedgehog, TGF-β) each end in changes to the receiving cell’s gene expression. There is also a large family called receptor tyrosine kinases: surface receptors that, when bound by their signal, add small chemical tags (phosphate groups) to proteins inside the cell to switch them on. None of these pathways has a real bacterial homolog; they are animal-lineage inventions. Many of the parts, though (receptor tyrosine kinases especially), were already present in single-celled animal relatives. The choanoflagellate Monosiga brevicollis, an animal cousin we will meet properly in Chapter 3, has on the order of 128 tyrosine kinases in its genome [Manning et al. 2008]. The single-celled relative had the parts.

Plants coordinate with a system architecturally unlike anything in animals or Dictyostelium. Phytohormones (auxin, cytokinin, gibberellin, and others) are made in one part of the plant and travel through the body. Signalling RNAs, mobile proteins, and peptides are trafficked between adjacent cells through the plasmodesmata we met in Part 2. And there is an enormous family of surface receptors of their own, receptor-like kinases (over six hundred members in the model plant Arabidopsis), that handle most of what receptor tyrosine kinases handle on the animal side, but evolved independently with a different architecture [Brunet & King 2017]. Same kinds of job, different parts. Same problem, lineage-specific solutions, no shared inheritance.

A fourth channel — voltage across the membrane

There is one more communication system, and it fits the polyphyly pattern in a particularly useful way. Every cell, prokaryotic or eukaryotic, multicellular or single-celled, has a membrane. Across that membrane, the cell maintains a small electrical voltage: a difference of charge between inside and outside, set up by ions (charged atoms, usually potassium, sodium, calcium, chloride) being pumped one way and allowed to flow back the other way through specific channels. That voltage is itself a signal. Cells can change it, sense changes in their neighbours, and use those changes to coordinate behaviour.

In bacteria, biofilms of Bacillus subtilis propagate waves of voltage change across the colony, mediated by potassium ions through specialised channels [Prindle et al. 2015]. The wave synchronises the cells’ metabolic state in a way that lets the outside layer feed the inside layer; cells of other bacterial species can be attracted into the biofilm by the same signal (reviewed in [Nunes et al. 2025]). The phenomenon has been extended to other species: [Akabuogu et al. 2024] in E. coli.

In plants, an injury at one leaf (a chewed edge, an insect wound) sets off an electrical signal that propagates through the body to other leaves in seconds, and distant leaves begin producing defensive chemicals before the insect arrives. The molecular machinery is plant-specific [Klejchová et al. 2021], but the idea is the same as in bacteria: a voltage change crossing from cell to cell, carrying information.

In animals, the most famous use of voltage signalling is the nervous system. Action potentials fire down neurons and underlie everything from sensing to muscle contraction to thought. Animal cells outside the nervous system also use voltage as a signal. A regenerating flatworm, a planarian, coordinates its regeneration in part by the pattern of voltages across its tissue. Experimentally manipulate that voltage pattern, and the worm regenerates with the wrong number of heads [Pezzulo et al. 2019; Whited & Levin 2019]. The molecular details are mammalian-style ion channels and gap junctions; the pattern level is something more general.

This is not polyphyly with an exception attached. The physics is universal: every cell has a membrane, every membrane has a voltage, so the substrate for bioelectric signalling exists by default. The molecular machinery is lineage-specific: bacterial voltage-using channels are different from plants’ which are different from animals’ which are different from fungi’s [Nunes et al. 2025; cross-summary in content/03-molecular-toolkit/bioelectric-signaling.md]. Universal substrate, lineage-specific parts. A productive complication: multicellular lineages did not have to invent communication channels out of thin air, but each still had to invent the molecular machinery that turned the universal physical layer into a usable signal. That observation is going to matter again in Chapter 3.

→ Continue with Part 4 — Division of labour: when one cell can’t do everything.

What this part draws on:

Part 4 — Division of labour: when one cell can’t do everything

Adhesion gives you cells stuck together. Communication gives them a way to coordinate. Division of labour, the third of the four problems, is what makes multicellular bodies do anything single cells cannot.

A division of labour means cells with different jobs. The two cells may have identical genomes (usually they do), but they are expressing different genes, making different proteins, taking different shapes, and behaving differently. Three independent inventions of multicellularity, three independent inventions of division of labour, lay this out.

Cyanobacterial filaments: terminal commitment, in a bacterium, two billion years ago

We met the cyanobacterial filaments (Anabaena, Nostoc, and their relatives) in Chapter 1 as the oldest known case of multicellularity. A cyanobacterium does two things for a living: photosynthesis (producing sugars and oxygen) and, in many species, nitrogen fixation (pulling N₂ gas out of the air and turning it into ammonia). The catch: the nitrogen-fixing enzyme is destroyed by oxygen, the same oxygen the cell produces while photosynthesising. The Nostocales solve this by separating the two jobs into different cells along a single filament. About one cell in every ten to fifteen transforms into a heterocyst: a thick-walled, larger, paler cell that has switched off photosynthesis, switched on nitrogen fixation, and shut its oxygen-producing machinery down [Flores & Herrero 2010; Rossetti et al. 2010]. The rest of the filament keeps photosynthesising and feeds the heterocyst sugars; the heterocyst feeds them fixed nitrogen back through the septal junctions we met in Part 2.

A heterocyst is permanently committed. It cannot revert. It can no longer divide. In some species, the cell physically rearranges its DNA during this commitment, deleting genes it no longer needs: an irreversible genome edit locking the cell into its new job for the rest of its life [Rossetti et al. 2010]. Terminal differentiation, a cell committing irreversibly to a specialised non-reproductive role, is not something eukaryotes invented. The bacterial lineage has been doing it for about 2.5 billion years.

Volvox: a sphere of cells, two job descriptions

Now jump nine-tenths of the way up the timeline. Volvox is a microscopic spherical green alga, just barely visible to the naked eye, made of a few thousand cells embedded in a transparent matrix, the whole sphere rolling through pond water by the coordinated beating of small whip-like extensions on each surface cell.

Volvox has two cell types. The majority, typically around 2000 cells on the outer surface, are somatic cells: they cannot reproduce; their job is to beat their flagella, propel the colony toward light, and keep it alive. The minority, typically around 16 in the sphere’s interior, are reproductive cells, which divide to make new colonies but cannot swim. The somatic cells die when their job is done; the reproductive cells are released to become new little spheres of their own [Kirk 2005].

This is germ–soma division: a permanent split between the cells that reproduce (germ) and the cells that handle everything else and die without leaving descendants (soma). In Volvox, the split is enforced by a single master regulator, regA, expressed only in somatic cells, that switches off the genes those cells would otherwise need to reproduce [Michod 2007]. regA did not appear from nowhere, and this is the thing Chapter 3 will return to: it descends from a gene that, in single-celled green-algal ancestors, was part of the day-night circadian rhythm machinery, doing a completely unrelated job. The toolkit was already there before Volvox repurposed it.

Animal sponges: a filter chamber, and many cells specialised for parts of one job

Now jump to animals. Take the simplest animal, a sponge. A sponge sits on the seabed and feeds by pulling seawater through itself and filtering out food particles. The pulling-through is done by a cell type called a choanocyte: a cell with a single waving extension (a flagellum) surrounded by a collar of finger-like projections. Choanocytes line the sponge’s internal water channels in tight chambers; their coordinated flagellar beating draws water through, and the collars trap and ingest bacteria from it. That is all a choanocyte does. It does not move around the body; it does not lay down the skeleton; it does not reproduce. Other cells do those things, and they do nothing else.

A sponge body has a handful of distinct cell types, each specialised for a partial job. A typical animal has on the order of a few hundred. The red blood cell that carries your oxygen has no nucleus. The neuron that fires when you read these words is permanently post-mitotic; it will never divide again. Each cell is doing one piece of one job; each, on its own, would die.

A multicellular body of this kind is, from one angle, a system of mutual hostages. Every cell in your skin would die alone within hours if not surrounded by the rest of you; every cell in the rest of you would die without skin. (“Hostages” is metaphor; there is no captor. The shape of the dependence is what the metaphor is gesturing at.)

Reversible versus terminal

Not all division of labour is permanent. A bacterium in a biofilm may switch between making the biofilm matrix and not making it depending on conditions; an amoeba in a Dictyostelium aggregation may, until late in the process, still revert to a solitary state if food returns. Reversible specialisation, where a cell takes a job, does it, and changes back when conditions change, is the more common kind across the bush.

What is striking is how many lineages also have terminal differentiation: cells that take a job and cannot ever return to anything else. Heterocysts. Volvox somatic cells. Animal neurons and red blood cells. Dictyostelium stalk cells (we will meet these in Part 5). Terminal differentiation is everywhere in the more committed cases, and it is not a eukaryotic monopoly. Cyanobacteria invented it billions of years before any eukaryote did.

→ Continue with Part 5 — Clonal and aggregative: why the cheater drama isn’t universal.

What this part draws on:

  • Cyanobacterial heterocysts as terminal differentiation, ~2.5 Gya: content/02-model-systems/cyanobacteria.md; [Flores & Herrero 2010]; [Rossetti et al. 2010].
  • Volvox germ–soma division and the regA somatic regulator: [Kirk 2005]; [Michod 2007].
  • Cooperation and cell-type comparison across origins: content/00-framework/kin-selection-and-cooperation.md §“Comparative summary across origins”.

Part 5 — Clonal and aggregative: why the cheater drama isn’t universal

Most of the cooperation drama you may have read about multicellular life comes apart in this part.

We are going to set up the case for cheating in a specific organism honestly, using exactly the kind of intuition the popular framing uses, because the intuition fits the case. Then we are going to step over to a different kind of multicellular organism and apply the same logic. It does not work. The cooperation problem people usually have in mind, when they talk about how multicellularity is hard, only applies to one kind of multicellular life. The animals and plants and Volvox and cyanobacteria, the lineages most readers have in mind when they imagine multicellularity, are a different case, and the popular story is the wrong story for them.

A real puzzle in Dictyostelium

Return to Dictyostelium, the slime mould we met in Part 3 as the eukaryote that uses cyclic AMP waves to call its cells together. Now look at the rest of its life cycle.

When food runs out, anywhere from ten thousand to a hundred thousand amoebae from the local soil patch stream into a single aggregation centre. They form a mound; the mound elongates into a multicellular slug; the slug crawls some distance through the soil; finally it rears up into a small stalked structure called a fruiting body. About twenty percent of the cells become the stalk, which terminally differentiates, lays down a cellulose tube, and dies: a vertical platform of dead cells. The other roughly eighty percent become spores, sitting on top of the stalk, ready to be carried away on the foot of an insect to a fresh patch of soil where they will hatch and start a new colony [Strassmann, Zhu & Queller 2000; Kessin 2001].

The catch is in the word aggregation. The amoebae converging on a single mound do not all descend from one founder cell. In the wild they are often distinct lineages, cells from genetically different parent amoebae sharing the same patch of soil. They aggregate together, and then a fraction of them dies to lift the rest.

Imagine, for a moment, you are one of these amoebae. Imagine your child is among the cells streaming inward toward the aggregation centre. Now imagine your child being asked to be a stalk cell, to die so that someone else’s children are carried away as spores.

Selection should favour not doing that. It should favour any cell that arrives, joins the aggregation, and avoids the stalk fate, letting somebody else’s lineage do the dying. Cells that pulled this off would leave more descendants than cells that played by the rules; the “do not become stalk” gene would spread, the stalk gene would shrink, the whole cooperative system should collapse.

This is the cheater problem, and in Dictyostelium it is real. Mix two genetically distinct strains of D. discoideum in the lab and let them aggregate, and about half of two-strain chimeras show a measurable bias: one strain ends up over-represented in the spores and under-represented in the stalk [Strassmann, Zhu & Queller 2000]. The cheating is not just predicted; it is observed. Dictyostelium has been a model system for social evolution for decades because of exactly this: it lets biologists watch the cooperation problem play out in a Petri dish.

Now make the same argument about your own body

Run the same thought experiment about the cells in your own body.

You started as a single fertilised egg. That cell divided into two; the two became four; through about forty rounds of division you became a body of tens of trillions. Every one of those cells is descended, by an unbroken chain of mitotic divisions, from that single starting cell. They share, with small adjustments from somatic mutation along the way, identical DNA.

Imagine the same thought experiment as before. A cell somewhere in your skin is being “asked” by its developmental programme to die, to flake off, to make way for new cells. Should selection favour a variant of that cell that refuses, that sneaks past the apoptosis pathway, that lives on and divides instead of dying?

The question is structured to sound similar to the Dictyostelium case, but it isn’t. The cells in your body are not strangers. They are descendants of one founding cell, your zygote. The cell being asked to die and the cell that will benefit from its dying are genetic copies of each other. From the perspective of any single gene in your genome, the cell that dies is carrying exactly the gene the surviving cell is. Cooperation here is cells of one lineage doing different things on behalf of one shared genome, not cells of different lineages working out a deal.

Hamilton’s rule, verbally

This is, almost exactly, the situation William Hamilton was thinking about in 1964 when he wrote down what is now called Hamilton’s rule.

Hamilton was trying to explain why apparently selfless behaviour, animals helping others at cost to themselves, should ever exist under natural selection. His answer was a formal one [Hamilton 1964]. He showed that a costly cooperative behaviour can be favoured by selection when the benefit it gives to the recipient, weighted by how closely related actor and recipient are, is large enough to outweigh the cost to the actor. The relevant kind of related is genetic: what fraction of the genome the two individuals share by recent common descent.

Hamilton’s rule is the one equation-shaped idea in this chapter, and we are going to keep it verbal: the benefit times the relatedness exceeds the cost. When the cooperator’s behaviour helps someone whose genes are the cooperator’s own genes, helping that someone is helping the cooperator’s genes survive, and selection finds the cooperative behaviour even if the individual cooperator suffers for it.

Now apply this to multicellularity. When relatedness between two cells is one, as it is between any two cells in your body, the rule is satisfied trivially. Benefit times one exceeds the cost as long as the benefit is positive. Helping another cell is, gene-for-gene, helping yourself. There is no defection problem to solve because there is no defector in any meaningful sense.

When relatedness drops, as it does in a Dictyostelium aggregation that pools amoebae from genetically distinct parents, the rule is no longer trivially satisfied. Relatedness becomes a variable that may be high (in the wild Dictyostelium aggregations often are near-clonal [Gilbert et al. 2007]) or low. When it is low, cooperation has to be worth more in benefit terms than it costs the cooperator, and a variant that takes the benefits without paying the cost has an immediate selective advantage. The system needs something else to keep cooperation stable.

Two regimes, one chapter’s worth of difference

So multicellular life splits into two cooperation regimes. The names go back at least to Bonner’s foundational work in the late 1990s [Bonner 1998].

Clonal multicellularity. A single founder cell divides; the daughters stay together; every cell in the resulting body descends from the same starting cell. Relatedness between any two cells in the organism is essentially one. Animals (you started as one fertilised egg), land plants (a seed is, mostly, a single starting cell), Volvox, cyanobacterial filaments, Streptomyces: the complex multicellular cases we met in Chapter 1 are all of this type. The cheating problem largely does not arise here, because the cells are not strangers to begin with.

Aggregative multicellularity. Cells of independent origin (possibly different parents, possibly different genotypes) come together into one multicellular body. Relatedness is variable, depending on how genetically homogeneous the local population happens to be. Dictyostelium and the other slime moulds (we met six independent eukaryotic origins of aggregative multicellularity in Chapter 1 Part 4); myxobacteria, on the bacterial side. The cheating problem is real here, and has to be solved by something other than relatedness alone.

The popular story (cells learning to cooperate, sacrificing autonomy, struggling against the temptation to defect) fits the aggregative case well. It fits the clonal case badly, because in the clonal case there is no question the cells are working out. The relatedness is already one. Cooperation is the default. The genes in the cell that dies and the genes in the cell that lives are the same genes.

How aggregative systems actually police themselves

When relatedness is low, active machinery has to evolve to control cheating. In Dictyostelium, the cleanest example is a pair of cell-surface genes called tgrB1 and tgrC1: a polymorphic ligand-and-receptor pair where cells displaying matching versions co-aggregate cleanly, and cells with mismatched versions segregate from each other in the soil [Benabentos et al. 2009; Hirose et al. 2011]. This is kin recognition at the molecular level, a chemical handshake. Cells whose handshakes match find each other and aggregate together; cells whose handshakes don’t match end up in different mounds. The system is not perfect (it reduces cheating rather than eliminating it), but it shifts the relatedness inside any given fruiting body upward, moving Hamilton’s rule back toward favourable territory. The aggregative cases have mechanical, molecular enforcement systems; the clonal cases largely do not need them.

A small thing about you and your gut

One complication will get its own treatment later. Hamilton’s rule applies within a clonal lineage. You and the bacteria in your gut are not clonal with each other, and the relatedness between you and any individual gut microbe is essentially zero, so cooperation across the host-microbe boundary has more in common with the aggregative case than with the clonal one, and is policed by similar kinds of enforcement mechanisms when it is policed at all.

Forward to a chapter about doors

One more consequence of the clonal-versus-aggregative distinction, and it is going to drive Chapter 4. Clonal lineages can, and animal and plant ones have, locked multicellularity in permanently. The cells in your body do not get to choose, generation after generation, whether to be a single cell or part of a body. They were committed from the moment they descended from the zygote. A Dictyostelium cell does not have this commitment: when food returns, a slug’s cells happily go back to being solitary amoebae. The aggregative life cycle has multicellularity as one of its modes; it can always come back out of it. Clonal lineages, by contrast, have evolved machinery that makes the multicellular state a one-way door. Chapter 4 will pick this up in detail, and will name the things that lock the door — call them ratchets.

That asymmetry has a darker version worth planting for later. The cells in your body have no option to go back to being unicellular. When one of them tries (divides without permission, fails to die when it should, ignores the signals from its neighbours and lives on its own behalf), we have a name for what’s happening. We will get to that in Chapter 5.

→ Continue with Part 6 — Why “cells deciding” is the wrong story (and where the right story is open).

What this part draws on:

Part 6 — Why “cells deciding” is the wrong story (and where the right story is open)

You will read, in many places, that cells “learned” to cooperate. That they “decided” to live together. That they “chose” to specialise. That cells “trust” each other, or “negotiate” with each other, or “remember” their commitments to the body.

The language is everywhere: in popular books, in documentary voiceovers, in the more accessible textbooks, sometimes in the papers themselves when researchers reach for an intuition they can share with a wider audience. It is one of the warmest, most natural ways of talking about cells, and the warmth is precisely the problem.

What the language imports, smuggled in under the words, is the human version of the things it names. Deciding, when you do it, involves the felt weighing of options, the awareness of having something at stake, the sense that you could have done otherwise. Trusting involves beliefs about another person, an emotional commitment to their reliability, an accepted vulnerability. Negotiating involves a back-and-forth of stated positions and offered concessions. Remembering involves retrievable storage of representations of past events. Apply these words to cells and you bring all of this content along with them. The vocabulary stops describing biology and starts describing us.

At the level the chapter has been working at (genes, populations, selection), the story does not need any of that. The cooperation we walked through in Part 5 happens by a mechanism that has no felt content. Variants of cells that produce certain phenotypes leave more descendants; the variants accumulate; the population shifts. The cell that “sacrifices itself” did not deliberate; there was no felt sacrifice. The lineage that produces cells which die at the right time outcompeted lineages that did not, because, in the clonal case, the genes in the cell that dies are the same genes as those in the cell that lives, and gene-frequency accounting does not require deliberation in any cell. No felt decisions; no remembered loyalties. The mechanism is mechanical, and the mechanism is sufficient.

The popular framing of cellular cooperation as a moral drama is, at the level the framing is operating at, wrong. That is the easy renunciation. There is a separate question lurking underneath, and it deserves more care, because it is open in a way the moral framing is not.

A second level — something is happening that looks computational

We saw in Part 3 that voltage across cell membranes is a real signal. Manipulate a regenerating planarian’s bioelectric pattern and the worm regenerates with the wrong number of heads [Pezzulo et al. 2019]. The pattern level (the spatial arrangement of voltages across a sheet of cells) is doing something the cells respond to. Networks of cells coupled by voltage-gated channels and gap junctions can implement what, mathematically, look like logic-gate computations (reviewed in [Cervera et al. 2023]; [Whited & Levin 2019]). Pattern states can persist for periods longer than any single stimulus that produced them.

This is not a metaphor. Voltage patterns are real physical states; the responses cells give to them are reproducible; the persistence is measurable. Something computational is happening in cells, at least in some specific systems, at least in some characterisable sense of “computational.”

Whether you should call that cognitive is the open question.

Two careful positions, neither of which the chapter picks

A research programme sometimes called basal cognition, associated most prominently with Michael Levin’s lab and allied groups, argues that the persistence of voltage patterns, the response of cells to those patterns, and the way the network states implement computations on that substrate all earn the same kind of vocabulary brains have monopolised: words like decision, memory, learning. The strongest version of the argument [Levin 2019; Whited & Levin 2019] is not the cartoon version (“cells have brains,” “tissues are conscious”). It says, more carefully, that decision describes a graded biological property: a property of any system that integrates inputs, persists in some states, transitions to others depending on those inputs. Nervous systems are an elaboration of this graded property, not its only instance. On that reading, refusing to extend the vocabulary to single cells or small networks of non-neural cells is itself a metaphysical commitment, not a neutral mechanism-only stance.

A different position [Newman 2023] is that morphogenesis depends on inherent physical and material properties of tissues (on the rheology of cell collectives, on the mechanical and chemical generic behaviours of soft active matter), and that fitting these properties into a computational frame is a category error. Cognition, on that view, is a term that carries specific phenomenological content (integrated information, intentional content, the felt character of mental states), and it does not generalise just because some sub-processes of biological systems can be modelled computationally. The strongest version is not that biology is “just chemistry.” It is that the words borrowed from cognitive science come pre-loaded with content biology has not yet shown is present in cellular substrates, and that the borrowing makes empirical claims it does not have the evidence for.

Both positions are held by serious researchers, both engage with the same mechanistic findings, and they reach different judgements about what vocabulary the findings warrant. The chapter is not going to pick a side.

The chapter’s editorial commitment

The mechanism is what matters; what you choose to call it is a separate question.

What that sentence is committing to: the mechanistic claims are well supported. Cells maintain voltages. Voltage patterns are real signals. Pattern states persist past the stimuli that produce them. Cell networks coupled by ion channels and gap junctions implement logical computations in a defensible technical sense [Prindle et al. 2015; Pezzulo et al. 2019; Cervera et al. 2023; Nunes et al. 2025]. None of this is in doubt.

What it is refusing is a verdict about the right vocabulary, in both directions. The basal-cognition programme is serious; the Newman-style counter is serious; whether the right words are cognitive or purely physical is the live disagreement, and the chapter is not going to resolve it. That refusal also excludes the lazy counter-framing that says “stop being mystical, cells are just chemistry.” That move is exactly as confident-without-evidence as the over-extension on the other side. Both over-extending the cognitive vocabulary and foreclosing it pretend a settled answer where the field does not have one.

(One worry worth flagging. There is a way of reading “something computational is happening in cells” that turns it into “tissues are conscious,” “your liver is making decisions,” “an embryo is a thinking thing.” That extension is not what the careful versions of the research programme claim, and it is one of the misconceptions the underlying corpus expressly flags [Misconception 1 in content/03-molecular-toolkit/bioelectric-signaling.md]. The chapter has to be careful not to accidentally licence the cartoon version on its way to refusing the moral-drama version.)

A note on a more ambitious version. Some in the same research community have extended the basal-cognition argument into deeper territory: mathematical frameworks like the free-energy principle, questions about whether the distinction between living and non-living systems is sharp at all, proposals invoking quantum mechanics for biological agency [Fields & Levin 2025 is the most ambitious recent synthesis]. The chapter is not going to engage with this layer; the mechanistic claims about voltage and ion-channel logic stand on their own, and the deeper interpretive framings are contested even within the basal-cognition community itself.

What the chapter carries forward, then, is one principle: mechanism over name. The cooperation in Part 5 does not require cells deciding anything, trusting anything, or remembering loyalties, and the chapter will say so without hedging. Whether a separate, narrower vocabulary of decision and memory turns out to apply, in some careful technical sense, to what cells do at the level of voltage and information is open, and the chapter will keep it open through Chapter 4 and Chapter 5.

→ Continue with Part 7 — The picture so far, and the question for Chapter 3.

What this part draws on:

  • The mechanism-vs-vocabulary editorial commitment: content/00-framework/kin-selection-and-cooperation.md §“Common misconceptions” #1; content/03-molecular-toolkit/bioelectric-signaling.md Active Debates §1; BIG-PICTURE.md §“What this corrects about the source video”; README.md §editorial standards 7 and 8.
  • Basal cognition as a research programme (the steelmanned position): [Levin 2019]; [Whited & Levin 2019].
  • The Newman counter-position: [Newman 2023].
  • Bioelectric pattern memory and the substrate it rests on: [Pezzulo et al. 2019]; [Cervera et al. 2023]; [Prindle et al. 2015]; [Nunes et al. 2025].
  • The most ambitious FEP-in-biology synthesis (mentioned only, not endorsed): [Fields & Levin 2025].

Part 7 — The picture so far, and the question for Chapter 3

Four problems. Adhesion. Communication. Division of labour. The suppression of cheating where it is needed. Every dot on the bush from Chapter 1, every cyanobacterial filament, every brown alga, every fungal mushroom, every animal, every plant, every slime mould, every haloarchaeon, faced this same set of problems. And every dot found a solution to them.

The solutions are not the same. The molecules that hold animal cells together are unrelated to the molecules that hold plant cells together, which are unrelated to the molecules that hold brown algae or fungi or bacteria together. The chemical signals that coordinate a bacterial biofilm are unrelated to the chemical signals that coordinate a slug of slime-mould amoebae, which are unrelated again to the hormones and growth factors that coordinate an animal embryo. The terminal differentiation that builds a cyanobacterial heterocyst is unrelated to the terminal differentiation that builds a Volvox somatic cell or an animal neuron or a sponge filter-chamber cell. What recurs is the shape of the puzzle. Adhesion. Communication. Specialisation. (And the cheater control, when it is needed.) The solutions vary; the problems don’t.

And one structural fact, the one Parts 5 and 6 spent most of their length on. The cooperation drama you may have read about, cells learning to trust each other, struggling against the temptation to cheat, holding themselves together against the centrifugal pull of self-interest, describes one kind of multicellularity and not the others. In the aggregative cases (slime moulds, myxobacteria), the drama is real, because the cells in those bodies do come from different parents, relatedness is variable, and selection can favour defection. Cells in those systems have evolved mechanical, molecular kin-recognition machinery to control cheating, exactly because the cheating problem is one they actually face.

In the clonal cases (animals, plants, Volvox, cyanobacterial filaments, Streptomyces), the cells in the body all descend from a single founder cell, and they share, gene for gene, the same DNA. The relatedness between any two cells in the body is essentially one. Cooperation under Hamilton’s rule is the default in this case. There is no cheating problem in the form the popular drama assumes, because there are no cheaters in any meaningful sense: the cells “sacrificing themselves” carry the same genes as the cells benefiting. Clonal multicellularity gets cooperation almost for free. The popular drama, imported into the clonal case, is the wrong story.

Clonal multicellularity gets cooperation almost for free; aggregative multicellularity requires active machinery; the popular story conflates the two and loses the structural fact that makes the polyphyly thesis tractable.

A question, with no answer yet

The variety we have walked through raises a question that has surfaced several times in passing. The cadherin proteins that hold animal cells together were already present, in something close to their current form, in single-celled animal cousins before there were any animals. The receptor tyrosine kinases that coordinate animal development were already there, in choanoflagellates, doing whatever they do in a single-celled organism. The regA gene that locks Volvox’s somatic cells into their flagellum-beating role descends from a green-algal circadian-clock gene that worked the day-night cycle in unicellular ancestors. The septal junctions that connect cyanobacterial cells along a filament are built from bacterial wall proteins as old as bacterial walls themselves. Even the bioelectric signalling we met in Part 3: every cell on Earth already has a membrane, and every membrane already has a voltage; the substrate was there before anyone used it for signalling.

The same observation, in lineage after lineage. The parts each multicellular origin used to solve its version of the four problems were, very often, already there in the unicellular ancestor, doing something else, on the shelf.

So the question. If multicellularity has been invented twenty-some separate times, in lineages that share no common ancestor for more than a billion years, where did the parts come from? Were they invented at each origin? Or were they already there, doing something else, waiting?

That is Chapter 3.

What this part draws on:

  • The bush of independent origins, the recurring four-problem structure: BIG-PICTURE.md §“The recurring principles”; content/01-polyphyly/independent-origins.md.
  • The clonal-vs-aggregative-with-Hamilton’s-rule analytical move: Parts 5 and 6 of this chapter; content/00-framework/kin-selection-and-cooperation.md.
  • The transition question into Chapter 3: chapters/03-the-tools.md opening beat.

End of Chapter 2 (draft state).

When complete, this chapter should be readable in one sitting (~10,000 words across seven parts) by someone with no prior biology who has read Chapter 1, though Part 1 is self-contained enough that a cold reader can pick the chapter up. The chapter’s central analytical move (clonal versus aggregative, and the cooperation regimes they generate) is structurally required by everything in Chapters 3, 4, and 5; without it, the polyphyly thesis becomes harder to hold together. The next chapter, 03-the-tools.md, picks up the question Part 7 closes on: where the molecular parts each multicellular origin used were already sitting before they were repurposed.