Life Imitating Life: CA and the Biological Sciences
Before Conway published his rules, before anyone had coined the phrase “artificial life,” Alan Turing sat in his office in Manchester and wrote an equation describing two chemicals reacting in a petri dish. He was not thinking about computation. He was thinking about how a leopard gets its spots.
The year was 1952. The paper was “The Chemical Basis of Morphogenesis,” published in Philosophical Transactions of the Royal Society B (volume 237, pages 37–72). Its central claim was audacious: that the remarkable spatial patterns of living organisms — the stripes of a zebra, the whorls of a seashell, the arrangement of fingers on a hand — could arise without any blueprint, without any central director, through the purely local interaction of chemicals diffusing through tissue. Pattern from rules, not from plan.
The structure of Turing’s mechanism — local interactions, no central plan, global pattern as output — is also the structure of Conway’s Game of Life. That convergence is worth examining carefully.
The Mechanism, Not the Metaphor
The standard account of biology’s relationship to cellular automata is that it is analogical: cells are like grid squares, chemical signals are like neighbor counts, development is like an update rule. The analogy is instructive, the account concludes, but it is only an analogy.
That framing is too modest — but the alternative requires precision. There are three distinct types of connection between biological systems and CA, and collapsing them causes more confusion than clarity.
The first is formal equivalence: systems where the mathematical description is literally a CA. Stuart Kauffman’s NK model of gene regulatory networks, where each gene is a Boolean variable that switches based on K neighboring genes, is formally a CA on a network topology. The McCulloch-Pitts neuron — a binary threshold unit that fires based on the weighted sum of its inputs — has the same abstract structure as a CA cell.
The second is CA-style abstraction: continuous or stochastic biological systems whose behavior can be well-approximated by a CA model. Turing reaction-diffusion systems are continuous partial differential equations, not cellular automata. But their core logic — local activation, long-range inhibition, no global coordinator — is structurally the same as a CA update rule, and CA models of reaction-diffusion dynamics capture the qualitative behavior accurately.
The third is shared mechanism class: the claim that both CA and many biological systems belong to the same broad family of processes — distributed local interactions generating large-scale structure without central control. This is the most defensible of the three connections, and the most interesting. It is also a philosophical claim about what kind of thing a CA is, not an empirical claim about how any specific biological system works.
The pages in this cluster move across all three types of connection. Keeping them distinct prevents overclaiming without sacrificing the genuine insight: the mathematical tools developed for CA apply directly to biological systems, and the biological phenomena studied in developmental biology and ecology have illuminated what kinds of behavior CA can and cannot produce.
Turing and the Activator-Inhibitor Framework
Turing’s 1952 insight began with a puzzle. A fertilized egg is, to a first approximation, a uniform sphere. Yet it develops into an organism with a front and back, a head and tail, and a surface patterned in ways specific to its species. How does spatial structure arise from spatial uniformity?
Turing’s answer was that you don’t need an encoding. Two chemicals suffice. An activator promotes its own production (positive feedback) and stimulates production of an inhibitor. The inhibitor diffuses faster than the activator and suppresses activator production at a distance. Given the right diffusion coefficients and reaction rates, the uniform mixture becomes unstable: small random fluctuations are amplified locally by the activator’s positive feedback but damped at larger scales by the inhibitor’s faster diffusion. The result is self-organized pattern — peaks of activator at a characteristic spacing, determined by the ratio of diffusion speeds, not by any external template.
In 1972, Alfred Gierer and Hans Meinhardt gave this mechanism a more concrete biological interpretation in Kybernetik (volume 12, pages 30–39), making the activator-inhibitor framework directly applicable to developmental biology. J.D. Murray’s subsequent work, collected in his 1989 textbook Mathematical Biology, showed that a single reaction-diffusion system applied to different domain geometries — the elongated tail of a cheetah versus the broad flank of a jaguar — produces the qualitative differences between spotted and striped coat patterns. The specific pattern depends on the geometry of the domain, not on the rule. Thin regions produce stripes; broad regions produce spots. Animals with thin tails are always striped at the tip. This is a quantitative prediction, not a post-hoc explanation.
Experimental Confirmation: The Zebrafish
For decades after Turing’s paper, the reaction-diffusion mechanism remained theoretically compelling but experimentally unverified in any specific organism. The breakthrough came from an unexpected direction: pigment-cell dynamics in zebrafish.
In 1995, Shigeru Kondo and Rihito Asai published “A Reaction-Diffusion Wave on the Skin of the Marine Angelfish Pomacanthus” in Nature (volume 376, pages 765–768). Angelfish stripes are not static — they shift, split, and merge as the fish grows. Kondo and Asai showed that these dynamics matched the predictions of a Turing model: stripes split when the domain exceeds a critical size and shift at speeds consistent with the predicted wave velocity. The movement of the pattern in real time was the signature.
In zebrafish (Danio rerio), the stripe-forming mechanism turned out to operate through cell-cell signaling rather than molecular diffusion. Dark melanophores and light xanthophores interact at close range by repelling each other, and at longer range each cell type depends on the other for survival and differentiation. Kondo’s 2009 paper in PNAS, “Interactions Between Zebrafish Pigment Cells Responsible for the Generation of Turing Patterns,” demonstrated that this interaction network has exactly the mathematical properties required for Turing instability.
The connection to CA here is structural, not literal. These are living cells, not CA grid squares; their interactions are continuous and stochastic, not discrete update functions. What the zebrafish system shares with a CA is the logical organization of the interactions: local activation, longer-range inhibition, no global pattern director. It instantiates the same class of local-interaction dynamics that CA models formalize — which is why CA-style models of zebrafish pigmentation reproduce the observed patterns with quantitative accuracy. That is a meaningful scientific connection, and a weaker claim than identity.
Morphogenesis as Local Computation
The question of how a single fertilized cell produces a geometrically precise body plan is, in formal terms, a question about how local rules generate global structure. That framing makes CA a natural modeling language.
Aristid Lindenmayer’s L-systems, introduced in 1968, are the most direct bridge. An L-system is a formal grammar that rewrites symbols in parallel — every symbol in a string is simultaneously replaced by one or more symbols according to a rule. This is structurally a one-dimensional CA applied to a developmental process. L-systems generate faithful representations of plant architecture: the branching patterns of ferns, the spiral leaf arrangements of sunflowers, the self-similar structure of trees. The mechanism is entirely local — each cell applies its rule without knowledge of the global structure — and the result is globally coherent form.
Lewis Wolpert’s “positional information” model (1969) and the Turing model are sometimes treated as alternatives, but they operate at different scales of the same process. Wolpert’s framework describes how cells interpret a pre-existing gradient; the Turing framework describes how the gradient is generated. The local computation that produces pattern happens in both layers.
Neural Models: The Original Cellular Automaton
The McCulloch-Pitts neuron, published in 1943 — three decades before the Game of Life — was effectively the first cellular automaton proposed for biological modeling.
McCulloch and Pitts described a neuron as a binary threshold unit: it fires if the weighted sum of its excitatory inputs exceeds a threshold, and stays silent otherwise. This is the CA update rule applied to biology. A network of such neurons, connected in some topology and updated synchronously, produces global behavior — perception, memory, decision — from local cell interactions.
The historical chain is explicit: von Neumann, designing his self-replicating CA in the late 1940s, was motivated directly by the McCulloch-Pitts model. His 29-state CA was an attempt to understand biological computation by building a formal system with similar architecture. Conway simplified von Neumann’s construction to two states. The intellectual sequence runs: real neurons → McCulloch-Pitts → von Neumann’s CA → Conway’s Life.
The neural CA connection has been renewed by recent work in machine learning. Mordvintsev and colleagues’ 2020 “Growing Neural Cellular Automata” (Distill) trained a CA rule encoded as a neural network to grow and regenerate a target pattern from a single cell — a model of morphogenesis in which the developmental program is local and the global form is emergent. The result regenerated after damage, as biological tissue does, without any explicit repair mechanism.
Ecology: Predators, Prey, and the Spatial Grid
The most influential models of population ecology — Lotka-Volterra equations, logistic growth — are mean-field models that treat every individual as equally likely to encounter every other. This is ecologically false. Predators find prey in their neighborhood.
CA models of population dynamics replace mean-field equations with spatial grids where individuals occupy cells and interact only with neighbors. The results differ qualitatively from mean-field predictions: predator-prey CA models produce traveling waves of population density that mean-field equations cannot generate. Local extinction is possible even when global populations are non-zero. Spatial refuges for prey can stabilize what the mean-field model predicts should collapse. Durrett and Levin’s 1994 analysis in Theoretical Population Biology, “The Importance of Being Discrete (and Spatial),” systematically showed how spatial structure changes the qualitative behavior of ecological models.
The forest fire model — a CA where trees grow, lightning strikes spark fires spreading to adjacent cells, and burned cells regenerate — exhibits self-organized criticality: fire sizes follow a power law, with catastrophic fires rare but non-negligible. This signature, poised at the boundary between order and chaos, appears in real forest fire data from California, Canadian boreal forest, and Australian eucalyptus. David Tilman and Peter Kareiva’s Spatial Ecology (1997) surveyed the emerging field systematically, case by case showing where spatial models produce empirically more accurate predictions than mean-field alternatives.
What Biology Teaches CA, and Vice Versa
The interaction between CA and biology has been generative in both directions. Biology has pushed CA researchers toward multi-state systems (real cells differentiate; they do not all run the same rule), three-dimensional lattices, and evolutionary rule-search. CA has given biology a precise vocabulary for articulating what “local rule, no central plan, emergent global pattern” actually means — and a formal language for testing whether a proposed mechanism can produce the observed behavior.
The parallel is productive because it is structural. Whether it points to something deeper — whether biological self-organization and CA dynamics are both expressions of a single, more general class of formal systems — is a philosophical inference that the empirical results are consistent with but do not establish. The most careful version of the claim is this: Conway’s Life and many biological systems belong to the same broad family of processes, one in which distributed local interactions generate large-scale structure without central control. The family is real. Whether its members share a common ancestor in some deeper mathematical sense is still an open question, and an interesting one.