TOPIC 42 · PHASE E COMPUTATIONAL

Connectomics

What do you get when you draw every wire of a brain?

2026-08-18 · BigCat

We've had the whole wiring diagram of a worm for 40 years — every one of its 302 neurons on paper — and we still can't quite tell you which way it will crawl next.

"Draw the wiring diagram of the brain" — you have probably heard some version of that ambition. Picture every one of 86 billion neurons and every one of a few hundred trillion synapses labelled and connected into a single graph; intuition says once we have that, not much mystery can be left. In 1986 a group in the UK actually finished the job — on a worm with only 302 neurons. Four decades later that diagram is still tacked up on lab walls, and we still can't read the worm's mind off of it alone. This issue is about that gap: why "we know every wire" turned out to be so much further from "we understand the brain" than anyone expected.

// 01

A wiring diagram nobody could read for forty years

In the late 1960s Sydney Brenner picked a species he could see through — the nematode C. elegans, about 1 mm long, body transparent, and every adult hermaphrodite carrying exactly the same number of neurons in exactly the same places. His student John White spent more than a decade cutting it into slices under the electron microscope, and in 1986 they published: 302 neurons, roughly 7,000 chemical synapses, roughly 900 electrical synapses. The first complete connectome in history.

302 neurons · ~7000 chemical synapses · 40 years to slowly start "reading" C. elegans · what a complete connectome looks like
Schematic: each dot a neuron, each line a connection (the real map is denser)

Then something strange happened. For forty years that diagram has been printed on walls and loaded into databases, yet a plain question — "which way will this worm crawl in the next second?" — still won't come out of it. Getting even basic behaviours (avoidance, foraging, mating) predicted has taken the past decade of stacking gene-expression atlases, neuromodulator maps and live calcium imaging on top of the wiring, and even then it's partial.

Why? Because a wiring diagram gives you the skeleton of the network — whether two neurons have a physical cable between them. Whether they actually send a signal, how strong that signal is, and when it goes through, depends on a pile of things the diagram doesn't show: the current strength of the synapse (synaptic plasticity is editing it constantly), what neuromodulator the tissue is bathed in (a dose of dopamine can swap out the entire gain of a circuit, Topic 26), and whether the downstream cell is opening its gates in phase with the current oscillation (Topic 38). The connectomics field has an honest sentence for this: the wiring diagram is necessary; it is nowhere near sufficient.

// 02

From millimetres to nanometres: how you turn a brain into a graph

To map a brain at the resolution where individual synapses are visible — a few nanometres — there is currently one path: serial-section electron microscopy. The flow sounds like taking apart a very thick book.

① a small block fixed in resin ② slice at tens of nm 1/2000th of a hair ③ image every slice stitch back to 3D ④ trace every neuron mostly AI now Serial-section EM connectomics · four steps
Cut, image, stitch, trace — each step is terrifying once you scale it to a brain

The scary part is scale. The data volume from this pipeline is measured in exabytes (1018 bytes). Sort the milestones by year and the slope tells the story:

· 1986: C. elegans, 302 neurons (White et al.). · 2019: the male version and an updated hermaphrodite (Cook et al.) — both sexes finally done. · 2020: fly "hemibrain", ~25k neurons and ~20M synapses (Scheffer et al.). · 2024: FlyWire whole fly brain, ~130k neurons and ~50M synapses (Dorkenwald & Matsliah et al.) — the first complete brain of a complex animal. · 2024: a Harvard/Google sample of 1 mm³ of human temporal cortex (Shapson-Coe et al.) — 1.4 PB of data, ~50k neurons, ~150M synapses, for a piece the size of a fingernail. · 2025: MICrONS mapped 1 mm³ of mouse visual cortex — ~200k neurons, ~500M synapses — and, uniquely, recorded live functional data from the same tissue first.

Extrapolate that curve to the whole human brain — 1500 cm³, ~86 billion neurons — and the same method costs decades and electricity by the gigawatt-hour. So human connectomes take a different, lower-resolution road that can be done in a living person: diffusion MRI follows water molecules along the direction of white-matter tracts and traces the fibres out, producing a map of "which big regions are connected by thick cables" rather than a synapse-by-synapse graph. That's what the Human Connectome Project (from 2010, Van Essen and colleagues) has been doing — a map coarse where it needs to be fine. (The six layers of cortex, close up: cortical layers & microcircuit; one path from a sense organ: visual pathway)

// 03

Wiring isn't the circuit: structural vs functional

Even if you had a full human connectome on your desk, you still need to hold two things apart:

Structural connectivity: is there a physical cable between these two regions? Functional connectivity: are the signals actually flowing between them right now, and how much (usually estimated from fMRI activity correlations or MEG oscillation coherence)?

Their relationship is not trivial. The same wiring supports radically different functional states. Asleep, doing arithmetic, dreaming, anaesthetised — the same cables are carrying wildly different things (Topics 16, 22). Why? Whether a signal gets through depends, besides the presence of the cable, on: the current synaptic strength (synaptic plasticity is editing it constantly), what neuromodulator the tissue is bathed in (dopamine, serotonin and friends turning gain knobs), and whether the downstream cell hits the current oscillatory beat — miss the beat and it might as well be unconnected (Topic 38).

same wiring structural asleep slow waves integrate long-range solving frontoparietal circuits lit up anaesthetised local fragments only Same skeleton, radically different functional maps
The wires are fixed; function is a state-dependent pattern of "selective activation" running on them

The 2025 MICrONS release gave a particularly clean example: 1 mm³ of mouse visual cortex, nanoscale connectome and live functional recordings from the same tissue, published together. The verdict was underwhelming — how many synapses two neurons share is only a weak predictor of their functional correlation. Put differently: a brain isn't the sum of its telephone lines; it's a multi-layered network that selectively connects depending on the current state. The static diagram is the skeleton. Everything that makes the skeleton alive — transmitters, plasticity, rhythms — is not written on it.

// 04

What a brain looks like as a graph

Shift levels. Treat neurons as nodes and synapses as edges, run the usual graph-theoretic quantities, and a small number of features come back — from worm to fly to human brain, they all rhyme.

local cluster (module) another local cluster hubs · rich club modularity + a few super-connected hubs = small world Three things brain networks keep growing
Lots of local clusters · a few super-connected hubs · hubs preferentially wired to one another ("rich club")

· Modularity: neurons cluster into communities, dense inside, sparse between. · Small-worldness: any two nodes reach each other in a few hops — dense local clusters plus a few long shortcuts. · Rich club: a small set of super-connected hubs are wired to each other more than chance would predict — the ones that can afford long-distance are largely the same "aristocrats" (van den Heuvel & Sporns 2011).

Once Bullmore and Sporns brought this graph-theoretic view into neuroscience, whole-brain properties became quantitatively discussable for the first time: the wiring is a negotiated solution between two goals — being able to integrate (signals must reach across to meet somewhere) and being able to run cheaply (Topic 41's budget). These properties weren't hand-picked by a designer; they fall out of that negotiation. So whether connectomics can "read" a brain may depend less on tracing every cable than on reading the shape of the whole graph — which is where the field really starts to talk.

AI cross-read

Artificial networks are literally graphs — units as nodes, weights as edges. But the architectures are regular: fully-connected layers, convolutional layers, attention layers, all templated wiring rules. A biologically-inspired branch now taking off, graph neural networks (GNNs), starts from the opposite premise — the graph is arbitrary: molecules, social networks, protein interactomes; nodes and edges can carry information; the shape is whatever it is. And there is a closer parallel — if you had every weight of a trained large model, could you tell what it computes? That's what mechanistic interpretability is asking, and the current answer is: nowhere close. The two fields have walked into the same wall: a static graph isn't enough; you need to see how the current actually flows through it.

🌀 CROSSING OVER · interdisciplinary echoes

"A thing is defined by what it connects to" sounds modern, but several old disciplines got there first; connectomics just adds physical coordinates:

// GOING DEEPER

If you scan and "upload" my connectome, is that still me?
Today the question splits in two. Physically: even at synaptic resolution, you'd be missing a lot. A meaningful share of your memory and habit doesn't live in "which wire goes where" but in the current strength of every wire (LTP/STDP), the local chemical context (which neuromodulator is high right now), and real-time feedback from the body (Topic 8's interoception). None of that is on a static wiring diagram. Philosophically: even if you copied all of it and started it running, you'd have made "another you" — from his first-person view, "me"; from the original you's view, still "him". The instant two "me"s exist at once, personal identity in the everyday sense has already broken. So don't rush to believe the "upload equals immortality" line.
Forty years and we still can't read a worm — is this road dead?
Not dead, but the naive version is. Naive version: get the connectome → plug into a model → out comes behaviour. What people actually do now is much more layered — overlay the wiring with gene expression (which receptors sit on which wire), neuromodulator maps (what the patch is bathed in right now), and live imaging (who fires when the animal actually moves), and part of the worm's behaviour finally comes out. Projects like OpenWorm and NemaNode have been accumulating exactly this multi-layer stack. The lesson: the connectome is one necessary layer, not the only layer. Expecting it to speak alone was a mistake about what it is.
Is spending billions to draw these maps worth it?
Honestly, the short-term theoretical return isn't high — most of the output is tools, datasets, methods. But infrastructure-type science tends to pay off across generations: the Human Genome Project was called wildly expensive at the time and nobody says that now. Realistic payoffs from connectomics probably show up in three places: giving circuit-level disease hypotheses an anchor (in autism or schizophrenia, which wire actually goes wrong); giving AI architecture some less-regular biological inspiration; and giving consciousness theories (like IIT vs GNW in Topic 11) the first anatomical data they can be directly falsified against — not answers, but the chance for their answers to be falsified.
Would a perfect connectome tell us where consciousness is?
No — at least not on its own. Consciousness isn't "which wires connect", it's "how these wires are being used right now" — a functional property, not a structural one. That's exactly why the adversarial experiments in Topic 11 (the Cogitate collaboration) had to be run in living brains. Structure sets the space of the possible; consciousness happens as some pattern of activity within that space. Which is also to say: without the connectome as boundary condition, every sentence a consciousness theory writes is free improvisation. So the connectome isn't the answer — it's the precondition for answers to be answerable at all.

// FURTHER READING