A worm has 302 neurons, and every wire between them has been traced. The map still can't tell you what the worm will do next.
This page follows 150 years of people learning to see, record and simulate nervous systems. Then it lays out what we have to do next to make a map behave like the animal it came from, what stands in the way, and the road from the worm to larger brains.
Hover a neuron. You'll see its wires, and none of its parameters.
Before the history, the finish line
The goal is to run a nervous system on a computer, starting from its map, well enough that the simulated animal behaves like the real one, including in situations the model was never fitted to.
How well is well enough? No two animals behave alike, so the fair ruler is a sibling raised the same way. A model has reached the goal when it predicts an animal as well as that animal's sibling does. Slide the score and watch.
No model reaches this yet, for any animal, not even the worm. To see why, go back to 1873.
Era I
1873 – 1906
Before anyone could ask how a nervous system computes, someone had to see that it is made of separate cells.
Silver chromate blackens a few nerve cells completely and leaves the thousands around them untouched, seemingly at random. For the first time a whole neuron stands out from the tangle, from the finest twig of its dendrites to its axon.
Cajal takes up Golgi's stain and draws what he sees, cell after cell. He argues that nerve tissue is made of separate cells that touch without fusing. From their shapes he guesses which way signals travel: in through the dendrites, out along the axon.
Golgi and Cajal share the prize while disagreeing about what they saw. Golgi still believes nerve tissue is one continuous net. The first maps of the nervous system arrive together with an argument about what such a map is.
A drawing shows where cells are and which way they point. It can't show anything move.
Era II
1939 – 1967
The next generation put electrodes inside nerve fibres and wrote down equations for what they heard.
The squid's giant axon is wide enough to thread a wire into. With an electrode inside it, they record the action potential from within a nerve fibre for the first time.
They explain the spike with two currents: sodium rushing in, then potassium flowing out, through channels that open and close with voltage. The equations won them a Nobel in 1963. Seventy years later they still sit at the core of every detailed neuron simulator.
Turn the sodium down and the spike dies partway. Change the potassium and the same fibre fires differently. The wire is identical every time. Only its channel numbers change. Call numbers like these θ: the parameters a wiring diagram never records. Most of this story is about where θ comes from.
A dendrite is a leaky cable. Where a synapse lands on the tree changes how much of its signal reaches the cell body, and fast signals fade more than slow ones. File that away for later.
A meteorologist's toy model of convection. Two runs from almost the same starting point drift apart until they have nothing in common. Prediction has a horizon.
Every tenfold improvement in the starting estimate buys the same fixed extra stretch of forecast. Past that horizon, a nervous system model should be scored the way weather models are: on distributions of activity, not exact trajectories. Whether a nervous system is chaotic like this is something to measure, not assume.
Era III
1963 – 1986
To understand a whole nervous system, you need one small enough to map completely.
Molecular biology, Brenner writes, has answered its classic questions. The next frontier is how animals develop and how nervous systems work, studied in a simple animal. He settles on Caenorhabditis elegans: a millimetre long, transparent, and quick to breed. His founding paper on its genetics appears in 1974.
On a temperature gradient, worms crawl to the temperature they were raised at, then follow it around the plate like a contour line. Later work traced the memory to one pair of sensory neurons, AFD.
Keep this worm in mind. Its preference is a set point that a neuron adjusts from experience, over hours. Any machinery that claims neurons tune themselves has to reproduce exactly this first. It is the positive control for everything that follows.
1986
The structure of the nervous system of the nematode C. elegans maps all 302 neurons and about 7,000 connections. It was traced by hand through thousands of electron micrographs, over more than a decade.
With the map finished, the hard question came into focus. The wiring was known. What the worm would do with it was not.
Era IV
1998 – 2014
In a ganglion of about thirty neurons that runs a crab's stomach, Eve Marder's lab found that nature doesn't pick one right answer.
A model neuron measures its own calcium and adjusts its channel densities until its activity is right. Scramble its channels and it finds its way back.
They simulated about 20 million versions of the crab's pyloric circuit. Wildly different parameter sets produced the same rhythm.
In models across biology, behaviour depends on a few stiff combinations of parameters and barely notices the rest.
Self-tuning rules explain why neurons of one type show correlated channel levels, and how a cell type keeps its identity while its parts vary.
Two things follow. Recordings of activity can't pin down every parameter. And the cell doesn't need them pinned down: it keeps adjusting until its activity is right.
Back to the worm. An electron microscope sees structure: membranes, vesicles, the places where two cells press together.
It can't see a neuron fire. For that, you make the cells glow.
Era V
2013 – 2023
Proteins that glow when calcium rises made neurons visible as they fired. For the first time, you could watch most of a nervous system at once.
Nearly every head neuron recorded at once. Activity moves through a few coordinated global states that track crawling forward, reversing and turning.
Complete connectomes of the hermaphrodite and the male, nerve ring to tail.
Witvliet and colleagues map eight connectomes from birth to adulthood. The CeNGEN project lists the genes every neuron class expresses, and NeuroPAL gives each neuron a colour so a recording can name it.
Stimulate one neuron with light and record the rest, then repeat, for about 23,000 measured pairs. Anatomy mispredicts many of them, partly because neuropeptides carry signals off the wiring diagram.
By 2023 the worm had a map, a parts list, a name for every neuron and a way to watch them all. It still had no model that could take the map and predict the atlas.
Era VI
2023 – 2026
Then came the fly. Connectomes grew by three orders of magnitude, and the race to turn them into working models began.
The larval fly brain: about 3,000 neurons and 548,000 synapses.
About 140,000 neurons, segmented by machine learning and proofread by a community of scientists and volunteers.
Shiu and colleagues run the whole fly connectome as simple spiking neurons and recover some sensorimotor pathways. Lappalainen and colleagues train a connectome-shaped visual system on a task and predict single-neuron responses. BAAIWorm closes the loop between a detailed worm brain and a simulated body.
Jaxley makes detailed biophysical networks trainable by gradient descent. The BANC dataset adds the fly's nerve cord. A foundation model trained on activity alone predicts responses to new stimuli in mouse visual cortex. The State of Brain Emulation report notes that the field has no agreed way to measure progress.
A complete connectome of the male fly's central nervous system, with 166,700 neurons and about 125 million synapses. FlyGM drives a simulated fly body with the connectome as a controller trained by reinforcement learning.
How big the maps got
Each map swallows the last. Drag through the years: the worm becomes a speck inside the larva, the larva inside the fly. The two largest brains on the scale have not been mapped at all.
Where we stand
Every one of these models has to decide where its neuron parameters come from. Some set them by hand, some fit them globally, some learn them for a task, and some skip biophysics altogether. Each group is placed below by how much of a nervous system it simulates and how much biophysical detail it keeps.
Nobody yet has a principled way to set the parameters of a whole nervous system. That is the next thing to build.
What we have to do next
The crab's lesson points to a way through. The parameters a connectome leaves out are not free numbers waiting to be fitted. Each cell sets its own, and a simulation can let it.
To fit θ directly you would have to fit every channel density in every compartment of every neuron, from activity recordings that can't pin them down. Here is the size of that problem.
H* Per-neuron biophysical parameters are not free parameters to be fitted. They are the steady state of a regulatory process that each cell runs, and that process can be simulated.
Instead of a billion conductances, fit a regulatory program for each of about 10⁴ cell types, with 10 to 50 parameters each: sensors, set points, gains, leak and a spread of starting states. Every neuron's parameters then grow out of simulated development, inside the wired network, inside a body moving through its environment.
This is one neuron with two channels, g and h. The landscape is its activity: valleys are quiet, ridges are busy. Along the blue contour, activity sits exactly at the cell's target. Drop neurons and watch each move its two channels in a fixed ratio, its gain direction, until it reaches the contour.
Each neuron moves along its own straight track, and the tracks never cross. Along a track, c = log h − ρ log g never changes, so where a cell stops depends on where it started.
The variability that made fitting hopeless becomes a prediction: a crowd lands spread along the contour and only along it. Test F4 checks that shape against single-cell expression data.
A model fitted to a healthy animal can't say what happens hours after you remove one of its channels. A regulating model has to say. The experiment that decides it: auxin degrades a tagged channel in an adult worm while imaging follows activity for hours.
Three accounts predict three curves. Under H*, regulation integrates the error and the compensation fraction κ climbs toward 1. Under R1, genetic hardwiring, nothing moves and κ stays at 0. A leaky rule stops partway. A third rival, R2, makes every neuron idiosyncratic. All three are limits of one equation, so the data can place each cell type somewhere in the family instead of only naming a winner.
Remember the worm that remembered its temperature. Raise two animals differently, then put them in the same dish. Draw a rearing history on the amber strip.
With a constant-gain rule, every history ends in the same adult. With a small squared-error term, history leaves a lasting mark in c. In a first synthetic demonstrator, A0, two rearing protocols produced adult conductances that differed by up to 13.2% while adult activity agreed to 6×10⁻¹¹.
Which kind of rule real neurons run is an experimental question. Test F7 asks it by silencing identified neurons during development.
H2, electrotonic anchoring. Rall's cables return. Instead of weighting a connection by its synapse count, compute each synapse's effect from where it lands on the EM-traced dendrite.
The B1 prototype ran this on real FlyWire neurons. For slow signals, synapse counts turned out to be a fine proxy almost everywhere: no super-class moved more than 0.074 of its input weight. The headroom is in fast signalling by central-brain and visual centrifugal neurons, and in huge graded neurons like APL, which moves 0.40 of its weight for slow signals and 0.58 at 100 Hz. Headroom shows that H2 could matter. Only test F5, on fast functional data, can show that it does.
H3, residual cartography. The 2023 atlas showed the wiring diagram misses signals. Fit the wired model, then read its errors: each missing mechanism leaves its own shape in the table of errors. Switch on explanations until the residual is gone.
A hypothesis this large is only worth something if it can lose. The plan fixes seven tests, F1 to F7, before anything is built. Each has a margin and a sample size written down in advance, and each can end in pass, fail or inconclusive.
Underpowered data can only come out inconclusive. Before anyone claims evidence against regulation, they must rule out a missing mechanism, an intervention too weak, too little measurement and a failed optimisation. Otherwise the verdict reads, more modestly, "the tested formulation does not account for the data".
First results
Synthetic: three neurons with body feedback, 18 of 18 numerical gates passed. The endpoint solver matches direct development to 2×10⁻¹⁰. Recovery after perturbation pins the regulation rate to within 0.7%.
not an organism result, not a verdict on H*
Passive-cable transfer for about 4,000 FlyWire neurons. Counts suffice for slow signals. The headroom is concentrated in fast centrifugal signalling and large graded neurons.
default cable parameters, rankings firmer than absolute values
Open problems
known obstacles and proposed crossings
Nothing on this page shows that H* is true. These are the obstacles already known, and the ideas for getting past them. Pick one on the map.
Some crossings will come from this plan and some from the groups mapped in Era VI: faster differentiable simulators, cheaper connectomes, voltage imaging, same-animal EM and imaging. The work is organised to absorb them, like an Earth system model: physical components joined by a coupler, organisms as configurations, and an open benchmark with blind prediction rounds.
Ahead
2027 onward: the worm, then the fly, then larger brains
The worm comes first, because there every approximation can be checked against an unapproximated reference. The fly follows once the decision rule has been applied to H*. Each step can end in pass, fail or inconclusive, and every outcome leads somewhere.
Public data can't settle it
Whether cells really set their own parameters can't be decided from the data that exists today. These experiments would decide it, in priority order. Each one changes a neuron's history and watches what follows.
The finish line, again
Progress is scored against the ceiling from the start of this page: how well one animal predicts another. A score of 1 means the model predicts an animal as well as a sibling does.
"Solved" for an organism means levels E1 to E5 passed, with E6 attempted. E6 is the hardest: given one animal's connectome and a few recordings, predict that particular animal.
You reared this page. Everything you stained, fired, dropped, depleted and drew was activity, and the page integrated it.
Read the full plan Hypotheses and tests Roadmap Proposals T1–T9 A0 prototype B1 prototype