durations.

Our design for a chemigenetic voltage indicator combines a voltage-sensitive microbial rhodopsin domain (6,7,11) with a dye-capture protein domain (Fig. 1A) that irreversibly binds a synthetic fluorophore dye ligand (14,15) (Fig. 1B), analogous to previously reported voltage indicators that use fluorescent proteins (10,11,18). Transmembranevoltage-dependent changes in the absorption spectrum (6,19) of the rhodopsin domain of Voltron reversibly modulate the degree of fluorescence quenching of the nearby bound dye through Förster resonance energy transfer (FRET). We investigated the modularity of this approach, finding that three different rhodopsin domains, QuasAr1 (7), QuasAr2 (7), and Ace2N (11,20), modulated the fluorescence of the rhodamine dye Janelia Fluor ® 549 (JF 549 ) after binding to either HaloTag (15) or SNAP-tag (21) dye-capture protein domains (figs. S1 to S8). Removing a small number of amino acid residues at the junction of the rhodopsin and selflabeling tag domains increased the amplitude of fluorescent voltage signals (fig. S1), presumably by decreasing average distance and thus increasing FRET efficiency between the dye and rhodopsin retinal cofactor. The configuration providing the best signal-to-noise ratio for spikes was Ace2N fused to HaloTag with five amino acids removed at their junction (Fig. 1, A and B, and fig. S2), hereafter referred to as Voltron.

We tested several Voltron-dye combinations in cultured rat neurons and acute mouse brain slices with high-speed imaging and simultaneous whole-cell patch clamp electrophysiology (Fig. 1C, figs. S6 and S9 to S13, and tables S1 and S2). Voltron could detect neuronal action potentials and sub-threshold potential changes with a variety of JF dye ligands with emission maxima between 520 nm and 660 nm using fluorescence imaging with one-photon excitation (Fig. 1, C to E, and fig. S6), but was not compatible with twophoton imaging, as described previously for rhodopsincontaining GEVIs (22,23). Voltron bound to JF525 (Voltron 525 ) exhibited the highest sensitivity, giving a fluorescence change of -23 ± 1% ΔF/F 0 for a voltage step from -70 mV to +30 mV (Fig. 1E and fig. S9); Voltron 549 showed similar sensitivity. Voltron 525 responded to voltage steps with submillisecond on and off time constants (table S3 and fig. S10). We compared the brightness and photostability of Voltron in neuronal cultures with those of two other fluorescent protein-based GEVIs: Ace2N-mNeon (11) and ASAP2f (13). Both Voltron 525 and Voltron 549 were brighter than Ace2N-mNeon (3-4-fold) and ASAP2f (16-18-fold) (Fig. 1F) in cell culture. This difference did not result from differences in expression; we compared the brightness of Voltron 549 and Ace2N-mNeon at the single-molecule level and observed a similar 3-4-fold brightness difference (Fig. 1G). Voltron 525 and Voltron 549 were also more photostable in ensemble measurements (Fig. 1H, tables S4 and S5, and figs. S14 and S15) as well as in single-molecule assays, in which photobleaching times were 8-fold longer for Voltron 549 than those of Ace2N-mNeon (Fig. 1I). Overall, the improved brightness and photostability of Voltron increased the photon yield by at least 10-fold in neurons over existing GEVIs that rely on fluorescence from FPs.

In vivo, Voltron could be reliably expressed and labeled with dye in mice, larval zebrafish, and adult fruit flies (Figs.

1 to 4 and figs. S16 to S19 and S21 to S45). Simultaneous in vivo electrophysiology and Voltron imaging in each of these organisms confirmed the detection of individual action potentials (Fig. 1, J and K, and figs. S17 to S19). For imaging in the mouse brain, we used a variant of Voltron appended with a 63 amino acid sequence from the rat potassium channel Kv2.1 that restricts expression to the membrane of the cell body and proximal dendrites (24, 25) (Voltron-ST, fig. S20). The rapid kinetics of Voltron525-ST allowed clear observation of action potentials in fast-spiking parvalbuminexpressing interneurons in the CA1 region of mouse hippocampus (Fig. 2, A to G, and fig. S21). We measured the orientation tuning of the spiking and subthreshold responses of cortical layer 2/3 pyramidal neurons in mouse primary visual cortex in response to the mouse observing directional movement of light and dark stripes, a benchmark for new indicators (1, 11) (Fig. 2, H to L, and figs. S22 to S24), and confirmed that spiking activity showed sharper orientation selectivity than did subthreshold voltage signals (26). We extended the imaging period over several consecutive weeks by injecting additional JF525 HaloTag ligand prior to each imaging session (Fig. 2, J to L, and fig. S24).

Next, we attempted to image larger areas containing more neurons for longer times in vivo in mouse cortex (Fig. 3). By widefield microscopy at illumination intensities between 3 and 20 mW/mm 2 , we could clearly identify and distinguish action potentials from nearby neurons throughout 15 min of continuous imaging (SNR = 5.3 during the first minute, 4.4 during final minute); (Fig. 3,B to E). We expanded the field-of-view to include dozens of cortical interneurons labeled with Voltron 525 -ST in a transgenic mouse line (NDNF-Cre) (27), while imaging at 400 Hz (Fig. 3, F andG, and figs. S25 to S42). Overall, we imaged a total of 449 neurons (12 fields of view in 3 mice), demonstrating routine voltage imaging of populations of neurons in superficial mouse cortex (Fig. 3G and figs. S25 to S42). This scale of in vivo voltage imaging enabled analysis of membrane potential correlations between many neuron pairs (fig. S26).

We used Voltron to image behaving zebrafish larvae, which respond to visual input with fast, directed swim bouts that are tailored to the details of the stimulus (28). We sought to uncover how this sensory-to-motor transformation unfolds in neuronal populations at fine timescales that are inaccessible with calcium imaging. We verified that Voltron could detect action potentials and subthreshold voltage signals in live zebrafish after labeling with several different colors of dye ligands (figs. S17 and S43). We then used Voltron 525 to monitor neural activity during swim bouts induced by visual motion (Fig. 4A). We recorded Voltron signals from 179 neurons across 43 fish in a motor-sensory nucleus in the tegmental area of the midbrain (Fig. 4B and fig. S44), yielding data on subthreshold membrane voltage modulation as well as automatically-detected spike times (Fig. 4C and fig. S45). We found neuron populations with different temporal activity patterns, including neurons whose firing rate increased ~1 s before the fish started swimming (fig. S44, B andC, "Ramp"), neurons whose firing rate was suppressed each time the fish swam (Fig. 4D, "Off"), and neurons that fired each time the fish swam (Fig. 4D, "Onset" and "Late"). Of the latter types, some fired just before swimming (~20 ms before swim onset, "Onset") and others fired just after swimming (~10 ms after swim onset, "Late"). There was a change in subthreshold voltage that preceded these firing-rate changes by tens of milliseconds (Fig. 4D and fig. S44D). The neuron types were spatially intermingled within this midbrain nucleus (Fig. 4, E andF).

The existence of neurons that fired before swimming and neurons that fired after swimming may indicate that this nucleus both partakes in the generation of swim bouts and is influenced by the motor output (Fig. 4G). Thus, Voltron allows for the dissection of population motor coding and sensorimotor integration circuits in ways that neither single-cell electrophysiology nor population calcium imaging can.

We tested Voltron in adult Drosophila in vivo by expressing the protein in a pair of dopaminergic neurons, one in each brain hemisphere, which innervate a single compartment in the mushroom body. We detected strong spiking signals from axons and dendrites of these neurons with Voltron549 (Fig. 1K and fig. S18). The fluorescence signals matched action potentials detected using electrophysiology. In some neuronal cell types in Drosophila, calcium indicators located in the cell body have failed to exhibit fluorescence changes even under conditions where high spike rates are expected (29). However, spikes were clearly detectable when imaging from the soma of dopamine neurons with Voltron (fig. S18E). We could clearly distinguish spikes from the two neurons based on the amplitude of the spiking signals even when imaging from neuropil where their axons overlap extensively, likely because each bilaterallyprojecting cell contributes a denser innervation of the mushroom body in the ipsilateral hemisphere (fig. S18D).

Combining the molecular specificity of genetically encoded reagents with the superior photophysics of chemical dyes is an established path to improved imaging reagents (14). However, previous attempts to create hybrid proteinsmall molecule indicators by various approaches have not been successful for in vivo imaging (30). We engineered a modular sensor scaffold in which the targeting and sensor domains are genetically encoded and only the fluorophore and its protein-binding anchor are synthetic. The resulting chemigenetic indicator, Voltron, exhibits increased photon output, enabling in vivo voltage imaging of many more neurons over longer times-approximately 10 2 more neuronminutes than other sensors. This improvement enables imaging experiments that can help reveal how the precise electrical dynamics of neuronal populations orchestrate behavior over different time scales. Arrows below represent the direction of movement of the drifting grating. (J to L) Top left, images of a pyramidal cell at a depth of 148 µm, imaged three times over a period of four weeks on the indicated weeks after virus injection. Scale bar: 10 µm. Top right, average of all spikes in session (black) and standard deviation (grey). Middle, raw ∆F/F 0 trace for five repetitions in each session, showing two orthogonal orientations (indicated with arrows below) from the neuron pictured on the top left. Bottom, orientation tuning to full-frame drifting gratings of the neuron pictured on the top left, displayed from number of spikes during trials (blue), number of spikes during preceding intertrial intervals (grey), and subthreshold ∆F/F 0 (right y-axis) after low-pass filtering traces using a 10-point median filter. For each orientation, response is calculated by averaging the low-pass filtered trace between 100 -400 ms after trial onset, and baseline is calculated by averaging the low pass filtered trace from 80 ms preceding trial onset to 20 ms after trial onset. Displayed as response minus baseline. Error bars represent standard error of the mean (s.e.m.) (20 -22 repetitions per session). , mean frequency of action potentials (middle) and raster plots of action potentials (bottom) near the initiation of swim bouts from three representative neurons: "Off" (green), "Onset" (red) and "Late" (blue) neuron. Shadows in the top and middle panels represent s.e.m. across swim events. (E) Classification of recorded neurons by their mean subthreshold signals near the initiation of swim bouts. 179 neurons recorded from 43 fish were classified using nonnegative matrix factorization and colored according to the weights for three factors: "onset" (red), "off" (green) and "late" (blue). The details of this classification are described in the Methods. (F) Spatial organization of the same population of neurons as in (E). Neurons from multiple fish are superimposed to a single map based on the distance from the center of this midbrain nucleus. (G) Hypothetical model of neural activity modulation in this midbrain nucleus.