WHAT THE FLY'S WIRING REALLY DECIDES

Scientists now hold complete wiring diagrams, synapse by synapse, of the adult fruit fly’s nerve cord: one from a male (MANC), one from a female, and one of a male’s entire central nervous system (MaleCNS). Simulations built on these connectomes already reproduce sensory responses and rhythmic leg activity that looks like walking.

But a model that walks does not prove that its precise wiring matters. Networks of interconnected units oscillate easily, and coarse statistics — how many connections each neuron has — already fix much of their behaviour. Isabel Guan of the Hong Kong University of Science and Technology and colleagues at the Hong Kong Polytechnic University, Nanyang Technological University and the company ZENBOT set out to test which features of leg movement truly depend on the specific wiring.

Real wiring against scrambled copies

They built models of the leg motor system from two independent connectomes: MaleCNS, 10,173 neurons and 462,650 connections, and MANC, 10,440 neurons and 634,207 connections. Connection strengths were simply the synapse counts, never trained. Input came only through two descending neurons, DNg100, whose activation starts forward walking; output was read from motor neurons grouped into extensor and flexor pools at the four joints of each leg — 20 antagonistic pairs.

Each real network was compared with 30 scrambled ones in six families, each keeping more of the true structure: the number of connections, connections between cell types, each neuron’s number of inputs and outputs, and blocks by leg or by developmental lineage. The decision criteria were recorded before the results were seen, a practice the authors describe as rare in connectome modelling; and re-running the simulations reproduced all 4,692 records per connectome to rounding error.

Rhythm is cheap, coordination is not

• Rhythm is generic. Many scrambled networks oscillated as well as the real ones, and some better: up to 0.96 against 0.70 in the parameter scan.
• Alternation is specific. The real wiring drove opposing pools to take turns more strongly than every scrambled network: a score of 0.312 against at most 0.147 in MaleCNS, and 0.173 against at most 0.084 in MANC.
• The effect concentrates at the thorax–coxa joint, the first of the four leg joints studied. In MaleCNS, the coordinated state held for the full five minutes of the longest simulations.

A principle from more than a century ago

Why? More than a century ago, C. S. Sherrington proposed reciprocal innervation: pathways that excite one muscle inhibit its antagonist. The team measured how each neuron influences the two pools of every pair. In both real connectomes, the reciprocal-innervation index was strongly positive — 0.48 and 0.38 — and negative in every scrambled network. Of the influence that neurons exert on both pools, 88% (MaleCNS) and 83% (MANC) pushes them in opposite directions, against 16–49% after scrambling. It works two ways: excitatory neurons that excite one pool and, through an inhibitory interneuron, silence the other; and inhibitory neurons that silence one pool and release the other.

The decisive test moved premotor inputs from one pool to its antagonist while changing each motor neuron’s total input by only 2–4%. Coordination collapsed: the score fell to about −0.08 in MaleCNS, meaning opposing pools now fired together, while control shuffles within a pool kept nearly 90% of the score.

Limits

These are results on a model, not on the animal: no body, no sensory feedback, one neuron model, predicted transmitter signs, and glutamate treated as inhibitory — when it was treated as excitatory in earlier runs, the advantage of the real wiring vanished. About a quarter of motor-neuron joint assignments are approximate, and MANC reproduced the findings with smaller margins. The authors also note they used AI assistants, including Anthropic’s Claude, to write code and draft text.

Their lesson for anyone modelling a nervous system: test the wiring on coordination, not on rhythm, because a scrambled network can keep time just as well.