Ultrasensitive voltage imaging reveals distinct electrical microdomains in neurons

TL;DR: The authors built ASAP7y, a voltage sensor sensitive enough to read out small, sub-spike membrane potential changes in living tissue, and fast enough to watch them travel along a neuron's branches in real time. Using it in the fly visual system — where every cell type's exact wiring is known from EM — they show directly that some neurons are electrically "one blob" (voltage spreads uniformly) while others are carved into semi-isolated electrical compartments, each with its own receptive field. A single neuron, in other words, can run several independent computations in parallel. The compact-vs-compartmentalized distinction is measured; the explanation in terms of three geometric knobs is modeled.

Why measuring voltage inside a neuron is so hard

The textbook cartoon of a neuron is an integrator: dendrites collect synaptic inputs, the cell sums them, and if the total crosses threshold it spikes. That picture quietly assumes the whole cell shares one membrane potential — that a depolarization on one branch is felt everywhere.

Cable theory says this is only sometimes true. Voltage injected at a point decays as it spreads, roughly exponentially, with a characteristic length called the space constant \lambda. If a neuron's arbor is small compared to \lambda, it is electrically compact — isopotential, a single computational unit. If the arbor is large compared to \lambda, distant branches are electrically decoupled, and the cell can host multiple local computations at once. Which regime a given cell type lives in has been genuinely unclear in vivo, because you would need to measure membrane potential at several points along a thin neurite, at millisecond resolution, sensitive enough to catch subthreshold signals of a few millivolts — all in a living, moving animal.

Patch electrodes can't reach thin neurites. Calcium imaging is a slow, nonlinear proxy for voltage. Existing genetically-encoded voltage indicators (GEVIs) had the speed but not the signal-to-noise for sustained, single-trial, subthreshold recordings. That is the gap this paper fills first with a tool, then with a result.

The tool: a sensor tuned for whispers, not just shouts

ASAP7y is an engineered descendant of the ASAP family — a voltage-sensitive protein domain fused to a circularly-permuted GFP, so membrane potential modulates brightness. The key number is the slope of fluorescence vs. voltage near rest: 2.7% ΔF/F per mV, versus 1.3%/mV for its predecessor ASAP5. That factor-of-two matters enormously for subthreshold work, because a 3 mV EPSP now produces a −6% signal instead of −3%. Its full dynamic range grew from 3.5 to 18, and its onset time constant is 0.9 ms at body temperature — fast enough to track propagation.

Two other properties do real work. First, ASAP7y is red-shifted, with peak two-photon brightness at 1000 nm and >50% brightness out to 1038 nm. That makes it compatible with high-power ytterbium fiber lasers at 1030 nm, which drive the parallelized, high-speed two-photon methods (scanless holographic, SLAP2) the authors lean on. Second, it is photostable enough to record for 20 minutes continuously. The authors also validate it broadly in mice — matching local field potentials with r^2 = 0.8 through the skull, and reporting gamma oscillations with ~10× the power of ASAP3 — but the mammalian data is mostly a "this works everywhere" demonstration. The scientific payload is in the fly.

The result: two kinds of neuron, seen directly

The experimental core is beautifully simple. Sparsely label a single genetically-defined neuron with ASAP7y, use acousto-optic random-access two-photon microscopy to place several small ROIs along its arbor, find the cell's receptive field center, then flash a small dark spot there and watch the voltage at every ROI at ~1 kHz.

voltage spreads everywhere voltage stays local

The neurons split cleanly. LPLC2, LC11, and LC4 — loom- and small-object detectors — depolarize nearly uniformly across ROIs tens to hundreds of microns apart. Dm9, CT1, MeLo13, Y3, and LC14b show the opposite: strong response on the stimulated branch, little on its neighbor.

To quantify this they define an Electrical Uniformity Index (EUI): \text{EUI} = 1 - |x-y|/|x+y|, where x and y are the integrated responses in two regions. EUI = 1 means isopotential; EUI = 0 means fully compartmentalized. They also fit a zero-intercept regression (region B vs region A) per cell; slope 1 = compact, slope 0 = isolated.

This is the strongest, most direct part of the paper: a non-uniform voltage distribution across an arbor cannot be faked by an isopotential cell, so compartmentalization is measured, not inferred. The one subtlety is that a uniform reading could in principle mean either "electrically compact" or "identical input everywhere." The authors address this with modeling (Extended Data Fig. 11), showing the uniform cells stay compact across a plausible range of biophysical parameters — so that particular conclusion is model-supported, not purely observed.

The mechanism: three geometric knobs (this part is modeled)

Why do cell types differ? The authors take the male-fly optic-lobe connectome, import each neuron's EM skeleton into NEURON with passive membrane properties, inject current at the dominant input site, and read voltage at output synapses — for 717 cell types.

A nice piece of physical reasoning simplifies the parameter problem. Since passive cable theory is linear, the relative decay between two regions is invariant to injection amplitude and duration, and to membrane capacitance C_m. It depends only on \rho = 1/(4 R_a G_{pas}) — the combination of axial resistivity and leak conductance — times geometry. They verify this numerically on Y3. So the interesting variation is geometric, and it factors into three features:

• Path length from input to output (Pearson r=0.82 with voltage decay): longer neurites decay more. The dominant factor.
• Radial curvature — effectively how thin the neurite is (equivalent-cylinder radius; r=0.27): thinner cables have a shorter \lambda, since \lambda \propto \sqrt{d}.
• Branch number along the path (r=0.57): each branch point sheds current into side branches, sharpening decay.

The striking claim is that cell types don't mix these at their extremes. The joint distribution over the three axes is heavily skewed, with separate tails: H1 achieves compartmentalization through a long path, LPi21 through many branches, Cm17 through thin neurites — rarely by combining them. The authors read this as evidence that developmental and connectomic constraints push each cell type to a preferred strategy. That is a real, testable hypothesis, but it rests on the passive model and the assumption that all cell types share biophysical parameters \rho; active conductances, which the model omits, could change the picture for spiking cells.

Honestly, the authors flag the key caveat themselves: the model correctly classifies cells as compact vs compartmentalized, but "morphological features could not fully explain experimentally measured voltage distributions." So morphology sets the regime; the quantitative details need input patterns and possibly molecular specializations the model doesn't have.

The payoff: parallel computation inside one cell

The functional consequence is the part that should interest anyone thinking about neural computation. In Dm9, each branch has a distinct receptive field center — map the RF per ROI and neighboring branches point at different patches of visual space, while ROIs on the same branch agree. Stimulate one branch's RF center and that branch responds strongly while siblings barely move; the measured cross-branch "confusion matrix" matches a matrix simulated from pure electrotonic crosstalk. CT1 behaves the same way, consistent with prior calcium-compartmentalization results.

MeLo13 goes further: its medulla arbor shows classic center-surround organization while its lobula arbor has a large, non-opponent OFF field — and the two arbors can simultaneously hyperpolarize and depolarize. One cell, two qualitatively different visual computations at once, separated only by cable geometry. That is the concrete "aha": compartmentalization isn't just attenuation, it lets a neuron implement operations — parallel feature extraction, multi-scale integration, normalization — that were assumed to require a circuit.

The authors also note this could resolve a live puzzle: connectome-predicted receptive fields have sometimes been larger than measured ones. If a cell integrates only locally, its effective RF is a branch, not the whole arbor.

What to trust, and what to read

The tool is real and likely to be adopted — 2.7%/mV subthreshold sensitivity with 1030 nm compatibility and 20-minute recordings is a genuine capability jump, and it slots into existing high-speed two-photon rigs. The compact/compartmentalized dichotomy and the parallel-receptive-field demonstrations are direct, well-controlled measurements in a system where morphology is known to nanometer precision — an unusually clean setting.

Be appropriately skeptical of the modeling superstructure: the 717-cell-type continuum, the three-knob decomposition, and the "each cell type picks one strategy" claim all come from passive simulations with assumed, shared biophysical parameters, and the authors concede the model doesn't fully reproduce the measured voltage magnitudes. The scale of direct measurement is modest — eight cell types, single-digit-to-low-teens cells each — even if statistically clean. And everything is a non-spiking or lightly-spiking invertebrate visual system; whether mammalian dendrites with active conductances behave along the same continuum is untested here, though the mouse validation sets that up.

Most worth your time: Figure 5 (the EUI/slope quantification that carries the central empirical claim) and Figure 6 (the per-branch receptive fields and MeLo13's dual computation, where the "parallel processing in one neuron" idea becomes concrete). Figure 4 is the conceptual model — read it knowing it's inference, and note the Methods argument that decay depends only on \rho and geometry, which is the cleanest idea in the paper.