The chemical basis of morphogenesis

TL;DR. In 1952, Alan Turing showed that two chemicals reacting and diffusing through tissue can spontaneously break symmetry and paint spots, stripes, and whorls onto an initially uniform blob — a mathematical mechanism by which genes could specify form without a blueprint. He could test it only by hand and on the Ferranti Mark 1, and biology had no way to find his hypothetical "morphogens," so the idea sat dormant for decades. It is now vindicated in real embryos, re-implemented in chemistry and robot swarms, and reborn as neural cellular automata that learn local rules to grow and regenerate target structures. If you care about growing materials, tissues, or robot bodies rather than assembling them, this is the founding document.

The idea Turing actually had

The 1952 paper asks the question developmental biology was avoiding: how does a symmetric ball of cells become an asymmetric organism? Every cell has the same genes; the fertilized egg is nearly homogeneous. Where does the pattern come from?

Turing's answer is deliberately minimal. Take two chemical species — he coined the term morphogens — that react with each other and diffuse through tissue:

where u, v are concentrations, f, g are reaction kinetics, and D_u, D_v are diffusion constants. The counterintuitive core result: a homogeneous steady state that is stable without diffusion can become unstable when you add it — even though diffusion is the archetypal smoothing, entropy-increasing process. Turing linearized around the uniform state, decomposed perturbations into spatial modes e^{ikx}, and showed that for the right kinetics and D_v \gg D_u, there is a band of wavenumbers k whose amplitude grows exponentially. Random molecular noise seeds every mode; the unstable band wins; the system locks into a periodic pattern with a wavelength set by the chemistry, roughly \lambda \sim 2\pi\sqrt[4]{D_u D_v / (\text{reaction rates})}, independent of initial conditions.

The intuition, crystallized later by Gierer and Meinhardt (1972), is local activation, long-range inhibition. An activator u makes more of itself and of an inhibitor v; the inhibitor suppresses the activator but diffuses much faster. A small bump of activator digs a wide moat of inhibition around itself, preventing neighbors from forming too close — and a periodic pattern crystallizes out of noise.

The paper works this out for a ring of cells (chosen for mathematical convenience — the modes are just Fourier modes on a circle), classifies six qualitatively different instabilities including stationary waves and traveling waves, treats a sphere as a model of gastrulation, and shows a hand-plus-machine simulation of 2D "dappling" reminiscent of animal coats. Turing suggested applications to Hydra tentacles, whorled leaves, and phyllotaxis. He was also disarmingly honest about the enterprise: the model, he wrote, "will be a simplification and an idealization, and consequently a falsification."

What makes this a computing paper and not just applied math: Turing is proposing that biological form is computed by a massively parallel, purely local, analog process — chemistry as the processor, the embryo as its own output device. It is arguably the first serious paper on self-organizing computation, written by the man who had just defined the other kind.

Why it stalled for forty years

Two independent walls, and Turing hit both.

Compute. Nonlinear reaction-diffusion has no closed-form solutions beyond the linear onset of instability; you must integrate numerically. Turing did his ring examples largely by hand and ran his dappling example on the Manchester/Ferranti Mark 1 — a machine executing very roughly 10^3 instructions per second with kilobytes of storage. A single coarse 2D pattern was days of work. Exploring parameter space, three or more species, growth of the domain, or 3D — the things that decide whether the mechanism explains real embryos — were unreachable. Turing said explicitly that the interesting nonlinear regime "cannot hope to have any very embracing theory" without machine computation, and he was drafting a follow-up on phyllotaxis when he died in 1954. Today a laptop integrates a 512^2 Gray-Scott grid in real time; a single GPU delivers roughly 10^{14}–10^{15} FLOP/s, some eleven-plus orders of magnitude beyond what he had — and, crucially, we can now differentiate through the simulation and optimize the rules.

Biology. In 1952 the morphogen was a hypothetical. DNA structure was published the following year; there was no molecular cloning, no in-situ hybridization, no GFP, no knockouts, no way to measure a diffusion constant in tissue. The theory made a sharp prediction — find two coupled diffusible species with the right sign structure and a large diffusivity ratio — that was untestable for decades. Meanwhile developmental biology organized itself around Wolpert's positional information paradigm (monotone gradients read out against thresholds, the "French flag"), which fit the emerging gene-expression data better and pushed self-organization to the margins. Even chemists couldn't build one: the first uncontested laboratory Turing pattern, in the CIMA (chlorite–iodide–malonic acid) reaction, arrived in 1990 — the trick being a gel that slowed the activator to achieve the diffusivity ratio.

What changed

The vindication came in waves.

Chemistry (1990–91): De Kepper's group and then Ouyang & Swinney produced sustained stationary Turing patterns in open gel reactors — the mechanism is physically real.

Biology (1995–2015): Kondo & Asai showed that stripes on the angelfish Pomacanthus move and rearrange as the fish grows — behavior a static gradient can't produce but a Turing system predicts. Then came molecular cases with named morphogens: WNT/DKK in hair-follicle spacing (Sick et al. 2006), FGF/SHH in the ridges of the mouse palate (Economou et al. 2012), Nodal/Lefty with directly measured differential diffusivity in zebrafish (Müller et al. 2012), and the flagship: digit patterning as a BMP–SOX9–WNT Turing network (Raspopovic, Marcon, Sharpe et al., Science 2014), with the earlier Sheth et al. 2012 result showing that deleting Hox genes shrinks the pattern wavelength — you get more, thinner digits, exactly the Turing signature. The modern synthesis (Green & Sharpe's framing) is that embryos use Turing self-organization to create periodic patterns and positional gradients to orient and modulate them.

Computation (2020–): Mordvintsev et al.'s "Growing Neural Cellular Automata" (Distill, 2020) is the direct descendant. Replace f, g with a small neural network shared by every cell, keep the strictly local update, and train by gradient descent through the unrolled dynamics so that the grown pattern matches a target image. Train with damage, and you get regeneration for free: cut the grown lizard in half and it regrows. This is Turing's mechanism turned inside out — instead of analyzing which patterns given chemistry produces, you compile a desired form into local rules.

Hardware (2018–): Slavkov et al. (Science Robotics, 2018) implemented Turing-style patterning in a swarm of ~300 Kilobots communicating only locally — emergent, self-repairing shape in physical hardware. Synthetic biologists have engineered pattern-forming gene circuits in bacteria (e.g., Karig et al.'s stochastic Turing patterns, PNAS 2018), though a clean designed deterministic Turing circuit in living cells remains hard.

A 2026 revival: morphogenesis as a compiler

The serious version of the revival treats Turing 1952 as the semantics of a new compilation target: specify a shape and a repair policy; compile to local rules; execute in a physical substrate.

The pipeline: (1) a differentiable simulator of the substrate — reaction-diffusion PDE, NCA, or a hybrid where an NCA's channels are constrained to realizable chemistry (mass-action kinetics, positivity, a small species count); (2) a loss combining target morphology, robustness under randomized damage and growth, and implementability — penalize diffusivity ratios and rate constants outside what the substrate offers; (3) a backend that maps learned rules onto DNA strand-displacement networks, synthetic gene circuits, swarm robot firmware, or bioelectric interventions in living tissue (Levin's program).

Keep from Turing: the locality discipline (no global coordinates, no addressing — this is what buys scalability and self-repair), linear stability analysis as a verification tool for the learned rules, and the insight that noise is a resource that seeds symmetry breaking. Replace: hand-chosen two-species kinetics with learned multi-channel dynamics; fixed domains with growth; and analysis with optimization. The interesting research question is the gap in between: learned NCA rules are wildly expressive but chemically unrealizable, while two-species RD is realizable but expressively poor. Characterizing the expressivity frontier — how many species and what diffusivity ratios you need to grow a given morphology robustly — is open and, I think, tractable.

Already tried, what's still open

Status: genuinely vindicated, which is rare for this series. Turing patterns are established in digits, skin appendages, palate ridges, fish skin, and left-right patterning; the CIMA reaction settled the chemistry; NCAs and Kilobot swarms settled the computation. Descendants include the NCA literature (texture synthesis, 3D structure growing, self-classifying and self-repairing grids), morphogenetic engineering, and Levin's bioelectric-morphogenesis program.

Open problems: a designed, deterministic Turing circuit running in living cells at chosen wavelengths; the compile-to-chemistry step for learned rules; scaling learned morphogenesis to 3D functional structures (vasculature-like transport networks, load-bearing geometry); and the theoretical question of which forms are growable at all under strictly local dynamics — a computability theory of morphogenesis that would close the loop back to Turing's other career.

Where to read it

The paper: doi.org/10.1098/rstb.1952.0012. It is long but astonishingly readable — Turing explains the biology and chemistry he needs from scratch. Read alongside: Gierer & Meinhardt (1972) for the activator-inhibitor formulation; Kondo & Miura, "Reaction-diffusion model as a framework for understanding biological pattern formation" (Science, 2010) for the biological synthesis; Raspopovic et al. (Science, 2014) for digits; Mordvintsev et al., "Growing Neural Cellular Automata" (Distill, 2020) for the learnable version; and Slavkov et al. (Science Robotics, 2018) for Turing patterns in a robot swarm. Bibliographic details of the main paper verified; citations to follow-up work are from memory of the literature and worth double-checking exact venues and years.