A spinal sensorimotor circuit and what it asks of MEA interfaces

What the work claims

Picton, Madrid, Pazzaglia, Wang, Bertuzzi, Ferrario, Anastasiadis, Arreguit, Fontanel, Huang, Mulleners, Song, Ijspeert, and El Manira report that a low-order spinal sensorimotor circuit is necessary for real-time social coordination during zebrafish schooling.¹ The core of the circuit is a class of intraspinal proprioceptive neurons that sense local body bending and deliver direct, curvature-based inhibition to V2a excitatory interneurons in the locomotor network. The authors claim this single feedback loop encodes both self-generated (egocentric) and neighbor-induced (allocentric) bending, allowing fish to match the phase of their swimming to the wakes of neighbors. Disrupting the circuit by blocking tail movement, mutating the mechanosensor Piezo2, or genetically ablating the proprioceptors abolishes schooling.

This is a primary experimental paper. It combines ex vivo patch-clamp electrophysiology, calcium imaging, optogenetics, mechanical tail pacing, a neuromechanical model, and a physical robotic fish.

How it works

The circuit is built around three cell classes. V2a interneurons form the excitatory core of the locomotor pattern generator. Motor neurons drive the muscles. Intraspinal proprioceptors sit in the spinal cord and respond to body curvature. When the tail bends, these proprioceptors activate and inhibit contralateral V2a interneurons, closing a sensorimotor loop.

The authors first showed, using Tg(glyt2:GCaMP6s) fish, that imposed tail bending increased calcium activity in intraspinal proprioceptors (p < 0.0001, n = 5 cells from 2 fish) and decreased activity in contralateral V2a interneurons (p = 0.0005, n = 6 cells from 3 fish).¹ Whole-cell recordings from V2a neurons confirmed larger outward inhibitory currents during bending.

Next they removed egocentric feedback by blocking tail movement in head-fixed preparations. Without proprioception, the swim frequency dropped (p = 0.0013, 9 cells from 5 fish), the excitation window of V2a neurons lengthened (p < 0.0001), and inhibitory currents grew (p = 0.0041). The rhythm lost its normal curvature-gated timing.

To test allocentric signaling, they paced the tail ex vivo at 2, 4, 6, 8, and 10 Hz and found a near-perfect linear relationship between pacing frequency and swim frequency (R² = 0.9964, n = 10 fish). A schooling-like sequence of leader-tail kinematics induced time-locked inhibition in silent preparations and synchronized V2a firing and EMG bursts during ongoing swimming (p < 0.0001 for pre-versus-pacing and pacing-versus-post, n = 8 fish).

The group then translated the circuit into a neuromechanical model and a physical robot. With proprioception enabled, a follower fish or robot matched its phase to a leader-generated wake. With proprioception disabled, phase matching collapsed. In simulation, swimming behind a leader reduced power consumption by about 17% (p < 0.0001, n = 50 networks); the robot saved about 7% (p < 0.0001, n = 6 trials). In vivo, Piezo2 mutants and DTA-ablated fish showed loss of group cohesion and lower polarization than wild-type fish.

Where a skeptic should push

The biggest limitation is translation. A zebrafish spinal cord is not a mammalian cortex, a human brain slice, or a cortical organoid. The V2a-proprioceptor motif is conserved in vertebrates, but the synaptic organization, cell types, and timescales differ. The authors are careful not to overclaim generalizability, and neither should a reader.

The electrophysiology sample sizes are small: 5 to 9 cells across a few fish. The effects are large and statistically significant, but the numbers are typical of difficult whole-cell recordings rather than a population-scale survey. Piezo2 mutation and DTA ablation are broad manipulations; either could affect sensory systems beyond the proprioceptors studied here. The robotic validation is elegant, yet a robot's hydrodynamics are not a live fish's.

Finally, the in vivo schooling analysis is correlational at the level of individual fish pairs. The causal argument rests on the ex vivo and genetic manipulations. That chain is reasonable, but a skeptic would want to see closed-loop optogenetic activation or silencing of proprioceptors during free swimming to prove the circuit is necessary and sufficient in the behaving animal.

Why MEA interfaces may need a mechanical feedback channel

MEA hardware has spent decades making electrodes smaller, denser, and lower noise. The implicit model is that tissue generates voltage, the array reads voltage, and the array injects current. The zebrafish paper says that for this kind of coordinated rhythm, voltage is only half the control loop. The other half is mechanical state: curvature, strain, and the flow field created by neighbors.

The non-obvious implication is that the next generation of neural interfaces should be multimodal. Electrodes would record and stimulate, while integrated strain gauges, optical shape sensors, or flow transducers report the mechanical state of the tissue and its surroundings. The controller would then close a sensorimotor loop: sense deformation, compute a phase correction, and inject current at the right point in the cycle. A voltage-only MEA is like recording from the motor neurons of a fish whose tail has been blocked; you can see the rhythm, but you cannot see why it drifts.

The opportunity is to make organoid and slice models more biologically realistic. Cortical organoids on MEAs often produce spontaneous bursts, but they lack the sensory feedback that shapes real neural dynamics. Adding a simulated mechanical feedback pathway could entrain organoids into more stable, coordinated states and make them better models of sensorimotor circuits. The paper's circuit motif is a blueprint: a deformation sensor, a sign-inverting synapse, and a phase-locked oscillator.

The threat is added complexity. Mechanical sensors need wiring, power, calibration, and artifact rejection. A strain gauge on a flexible MEA will pick up stimulation artifacts and motion noise. If the sensor fails or is misregistered, the controller will misinterpret tissue state and inject current at the wrong phase. There is also a dual-use risk: a system that can entrain a culture's phase can manipulate it, not just observe it.

There is a subtler instrumentation angle. The paper shows that inhibition from proprioceptors is curvature-based, not rate-based. The timing of inhibition matters more than its average amplitude. That places a hard requirement on the acquisition chain: the mechanical signal must be sampled at the same temporal resolution as the electrophysiology, and the latency from mechanical sensing to stimulation must be small compared with the rhythm period. For a 5 to 10 Hz locomotor-like rhythm, that means millisecond-scale latency budgets.

The bottom line

The study identifies a specific spinal circuit that uses curvature-based proprioceptive feedback to synchronize swimming. The evidence is strongest for ex vivo phase locking and weakest for the full in vivo schooling behavior. What would confirm the model is closed-loop optogenetic control of proprioceptors during free swimming. What would break it is evidence that Piezo2 mutants lose schooling for reasons unrelated to spinal proprioception, or that the V2a-proprioceptor circuit is not present in the species one wants to interface.

For MEA engineering, the lesson is not to copy a fish spinal cord. It is that effective neural interfaces may need to close sensorimotor loops, not just open them. Voltage is the signal we know how to record; curvature and force may be the signals the tissue actually uses to coordinate itself.