Cut a flatworm in half, and the tail end grows a new head. That part biologists have known for over a century. What a lab at Tufts University showed in the early 2010s is stranger: if you briefly change the voltage across the membranes of the cells at the cut site, using a drug that blocks a specific ion channel rather than any genetic edit, you can get a worm that regrows two heads instead of one. No new genes. No CRISPR. Just a different electrical state held for a few hours, and the tissue reads a different anatomical instruction from it.
That result came out of Michael Levin’s lab, and it is the experiment most people cite when they want to argue that cells are not just chemical machines, they are informational ones, and the information in question is voltage. The claim that follows, and the one worth taking seriously without swallowing whole, is that networks of non-neural cells can perform something close to computation: taking in distributed signals, integrating them, and committing to a discrete output. Call it a decision. Whether “logic” is the right word for it is where this gets contested.
Voltage Is a Variable, Not Just a Trigger
Every cell in the body maintains a voltage difference across its outer membrane, usually somewhere between negative ten and negative ninety millivolts, negative meaning the inside is more electrically negative than the outside. In neurons, that voltage is famous for one trick: it spikes, fires an action potential, and resets. Textbooks teach that story so thoroughly that it is easy to assume voltage is a neural phenomenon on loan to the rest of biology.
It isn’t. Skin cells, gut cells, developing embryonic tissue, even bacteria maintain and modulate membrane voltage without ever firing anything resembling a spike. The ion channels and pumps that set this voltage, the same families of proteins that Roderick MacKinnon’s structural work in the late 1990s and early 2000s helped the field finally see atom by atom, are present in essentially every cell type. What differs is how that voltage gets used. In a neuron, it is a fast, transient trigger. In a developing embryo or a healing wound, it behaves more like a slow-changing variable, a value the tissue can hold stable for hours or days, and that value carries information about what the tissue should become.
Voltage is doing more in that worm than keeping its neurons firing.
Levin’s group demonstrated this most vividly in Xenopus tadpoles. In a 2012 study led by Vaibhav Pai, forcing a particular voltage state in cells on the tadpole’s flank, far from the head, caused those cells to grow a functioning eye, complete with retina and optic nerve connections, in the wrong part of the body. The genome in those cells was identical to every other cell’s genome. The instruction to build an eye came from the electrical state, not from new DNA. That is a hard result to explain away as a curiosity. It means voltage state can override or redirect positional identity on its own.
Gap Junctions Turn a Tissue Into a Circuit
None of this works cell by cell. The reason voltage can carry a patterning instruction across a chunk of tissue, rather than staying locked inside one cell, is a structure called the gap junction: a channel made of proteins called connexins that physically links the interior of one cell to the interior of its neighbor, letting ions and small molecules pass directly between them without ever touching the extracellular space.
A sheet of cells wired together by gap junctions stops behaving like a collection of independent chemical reactions and starts behaving like an electrical network. Change the voltage in one region and that change propagates, is filtered, and sometimes amplified across the whole connected sheet, the way current flows through resistors wired in series and parallel. This is the physical basis for calling these systems circuits rather than metaphors for circuits. The wiring is real tissue, the current is real ion flow, and the propagation follows real, measurable electrical rules.
A tissue wired by gap junctions behaves like a circuit, not a crowd of separate chemical reactions.
This is also why bioelectric effects can be regional and bounded rather than spreading uncontrolled through the whole organism. Gap junction connectivity itself is regulated: cells can open or close these channels, effectively rewiring the circuit’s topology in response to injury or developmental stage. The planarian head-versus-tail decision, the tadpole eye experiment, both depend on a patch of tissue being electrically connected enough to share a voltage state, and electrically isolated enough that the effect stays local instead of reprogramming the entire animal.
The Case For Calling It Computation
Here is the argument Levin and collaborators have made across more than a decade of papers: a bioelectric network takes in multiple, distributed inputs (local injury signals, neighboring cell states, developmental history), integrates them over a connected tissue, and produces a small number of discrete, stable outputs (grow a head, grow a tail, grow an eye, grow nothing). That input-integration-discrete-output structure is exactly the structure engineers use to describe a logic circuit. The output is not a smooth, proportional response to a single input. It is a committed decision among a limited set of anatomical outcomes, the kind of behavior you’d expect from a system running something like a lookup table rather than a thermostat.
The 2020 xenobot experiments, run by Levin’s lab with computer scientist Josh Bongard at the University of Vermont, pushed this further. Clusters of frog skin and muscle cells, with no neurons and no added genes, were shown to self-organize into small mobile constructs capable of coordinated movement, and in later work, of locating and gathering loose cells into piles resembling themselves. Nothing in that behavior was explicitly programmed cell by cell. It emerged from the same kind of voltage-mediated, gap-junction-coupled collective dynamics that build a head where a head belongs. If a swarm of unremarkable cells can produce goal-directed behavior nobody wired in directly, “the tissue is computing an outcome” starts to feel like the most economical description available, even if it is an uncomfortable one.
The Case Against the Word Logic
Here is where the metaphor should be pushed back on, hard, before it hardens into fact. Synthetic biologists already build literal digital logic into cells, and the contrast is instructive. Groups including Timothy Lu’s lab at MIT have constructed genetic circuits using transcription factors and recombinases that function as AND gates, OR gates, and even flip-flop memory elements, with discrete on and off states encoded in gene expression levels that can be read out unambiguously. That is computation in something close to the engineering sense: digital-ish, modular, designed input-output mapping.
Bioelectric networks are not that. Membrane voltage is a continuous, analog variable, not a clean binary state, and it decays, leaks, and drifts in ways that genetic circuits mostly don’t. The “decision” a bioelectric network makes, head versus tail, eye versus no eye, is real, but it looks more like a dynamical system settling into one of several stable attractor states than like a transistor flipping. Dynamical systems theory already has language for this: bistability, hysteresis, attractor landscapes. None of that language requires importing the word “logic” from computer science, and doing so risks implying a level of discrete, symbolic processing that the biophysics doesn’t actually show.
Calling it logic flatters the metaphor. Calling it control doesn’t flatter it enough.
The honest middle position is that bioelectric networks perform analog computation in the control-theory sense, continuous signal processing that integrates inputs and produces stable, categorizable outputs, which is a real and useful thing to call computation, just not the same thing a logic gate does. The table below makes the contrast concrete.
| Feature | Genetic logic circuits (e.g. Lu lab, MIT) | Bioelectric networks (e.g. Levin lab, Tufts) |
|---|---|---|
| Signal carrier | Transcription factor concentration | Membrane voltage, ion flux |
| Output character | Discrete, gene expression on/off | Continuous, settles into stable states |
| Wiring | Designed DNA sequence, promoters | Gap junctions, native ion channels |
| Timescale | Minutes to hours (transcription, translation) | Milliseconds to days depending on context |
| Best-known demonstration | Engineered toggle switches, recombinase memory | Two-headed planaria, ectopic tadpole eyes |
Where the Analogy Earns Its Keep Anyway
Even with that caveat, the computation framing has done something genuinely useful: it has reoriented regenerative medicine away from asking only “which gene do we need to turn on” and toward asking “which electrical state does this tissue need to be told it’s in.” That reframing matters because it opens a route to influencing anatomy without gene therapy’s delivery problems. A drug that shifts ion channel activity, something closer to existing pharmacology than to viral vectors, can in principle nudge a tissue toward a different bioelectric attractor and a different anatomical outcome. Levin’s group has pursued this explicitly for limb regeneration and wound healing, treating the bioelectric state as a lever that is pharmacologically reachable in a way the genome often isn’t.
It also reframes what a “decision” means at the tissue level, which has downstream relevance for cancer biology. Several of Levin’s papers describe tumor-like, disorganized growth appearing in tissue that has been electrically decoupled from its neighbors, as though losing bioelectric connectivity to the surrounding network is sufficient to let a cell’s growth escape the collective’s constraint. If that holds up under more scrutiny, it suggests cancer is, in part, a communication failure in a voltage-coordinated network rather than purely a genetic one, which is a genuinely different place to look for interventions.
The counterargument deserves one more pass before closing: skeptics are right that none of this is “computation” in any sense an engineer building a CPU would recognize, and that the field should resist importing computer-science vocabulary wholesale just because it sounds rigorous. But the skeptics are usually arguing against a strawman version of the claim. Nobody serious is saying cells run Python. The actual claim, that distributed voltage networks integrate signals and settle into a limited set of stable anatomical outputs, is empirically supported, mechanistically grounded in real ion channels and real gap junctions, and different enough from ordinary single-cell signaling to deserve its own vocabulary. Whether that vocabulary should borrow the word “logic” or stick with “control” and “dynamical systems” is a semantic fight, not a factual one.
The more interesting open question isn’t whether cells compute. It’s whether bioelectric states can be read and written with enough precision, with a drug, a device, or an electrode array, to direct a specific anatomical outcome on demand, the way a pacemaker directs a specific cardiac rhythm on demand. Right now the field can do that reliably in flatworms and tadpoles. Whether that precision scales to a human limb or a human tumor is the actual frontier, and it is a wet, slow, electrochemical one, not a digital one.