Dopamine promotes head direction plasticity during orienting movements – Nature

Cell types and synonyms

We have adopted the cell type names used in the recent comprehensive anatomical survey of the central complex²⁵. ExR2 neurons are the dopamine neurons of the PPM3 cluster that innervate the ellipsoid body⁵² (that is, PPM3-EB neurons⁵³). EPG neurons have also been termed E-PG or PBG1-8.b-EBw.s-D/Vgall.b neurons^54,55, EB-IDFP D/VSB-PB neurons⁵⁶ or compass neurons³⁶. ER neurons⁵⁷ have been termed ring neurons or R neurons⁵⁸, as well as LTR-EB neurons⁵⁹. PEN_a neurons are also referred to as P-EN1 (ref. ³⁵) or PEN1 (ref. ⁵⁷).

Fly husbandry and genotypes

Unless otherwise stated, flies were raised on standard cornmeal–molasses food (New Brown 19L, Archon Scientific) in an incubator on a 12/12-h light/dark cycle at 25 °C with humidity about 50–70%. All experiments used flies with at least one wild-type copy of the white gene, and many experiments used flies with two wild-type copies of the white gene (as detailed below). The experimenters were not blind to genotype. In Fig. 2, flies of the appropriate genotype and age were selected randomly and then alternately assigned to the ‘ExR2 activation’ condition or the ‘ExR2 activation, no visual cue’ condition, so that the two conditions were interleaved in the data acquisition sequence; these assignments were made prior to the beginning of each experiment.

The genotypes of the fly stocks used in Fig. 1 and Extended Data Figs. 2–4 are as follows: +/+ or +/w*; P{75B10-LexA}attP40/P{R75B10-LexA}attP40; PBac{13×LexAop-IVS-jGCaMP7f}VK00005/PBac{13×LexAop-IVS-jGCaMP7f}VK00005.

The genotype of the fly stock used in Fig. 2a is as follows: +/+; P{R38A11-LexA}attP40/P{13×LexAop2-mCD8::GFP}attP40; P{13×LexAop2-mCD8::GFP}attP2/P{LexAop-P2X2.Y}3.

The genotypes of the fly stocks used in Fig. 2b–f and Extended Data Figs. 6 and 7 are as follows: for control, +/+ or +/w*; R38A11-LexA/UAS-mCD8::GFP; R60D05-Gal4/+; for ExR2 activation, +/+ or +/w*; R38A11-LexA/UAS-mCD8::GFP; LexAop-P2X2.Y/R60D05-Gal4; for dopamine, +/+; UAS-mCD8::GFP/UAS-mCD8::GFP; R60D05-Gal4/R60D05-Gal4 (n = 9) or +/+ or +/w*; R38A11-LexA/UAS-mCD8::GFP; R60D05-Gal4/+ (n = 3).

The genotypes of the fly stocks used in Extended Data Fig. 8 are as follows: for no Chrimson control, +/+; P{20XUAS-IVS-mCD8::GFP}attP40, PBac{13×LexAop2-IVS-Syn21-Chrimson::tdT-3.1-p10} su(Hw)attP5/+; P{R60D05-Gal4}attP2/+; for ExR2 activation and ‘no cue’ conditions, +/+; P{20XUAS-IVS-mCD8::GFP}attP40, PBac{13×LexAop2-IVS-Syn21-Chrimson::tdT-3.1-p10} su(Hw)attP5/P{R38All-LexA}attP40; P{R60D05-Gal4}attP2/+.

The genotypes of the fly stocks used in Fig. 3 are as follows: for control, +/+ or +/w*; P{R38A11-LexA}attP40/P{20XUAS-IVS-jGCaMP7f}su(Hw)attP5; P{R60D05-Gal4}attP2/+; for ExR2 activation, +/+ or +/w*; P{R38A11-LexA}attP40/P{20XUAS-IVS-jGCaMP7f}su(Hw)attP5; P{LexAop-P2X2.Y}3/P{R60D05-Gal4}attP2.

The genotypes of the fly stocks used in Fig. 4 and Extended Data Fig. 9 are as follows: for control, w[1118], P{13×LexAop-IVS-jGCaMP7f}su(Hw)attP8/+; P{R60D05-LexA}attP40/+; P{UAS-HsapKCNJ2.eGFP}7/+; for Kir2.1, w[1118], P{13×LexAop-IVS-jGCaMP7f}su(Hw)attP8/+; P{R60D05-LexA}attP40/+; P{UAS-HsapKCNJ2.eGFP}7/P{R75B10-Gal4}attP2.

Characterization and description of driver line expression

We used P{R60D05-Gal4}attP2 or P{R60D05-LexA}attP40 to target EPG neurons for calcium imaging or electrophysiology recording as established in previous studies^27,28.

We used P{R38A11-LexA}attP40 to drive expression of mCD8::GFP and P2X2 receptors in ExR2 neurons. Immunostaining of these brains showed that GFP expression was isolated to four ExR2 neurons, a bilateral pair of unidentified ascending neurons that arborize within the antennal lobe, and 1–2 pars intercerebralis neurons (data not shown). The database of the FlyLight Project Team at Janelia Research Campus (https://flweb.janelia.org/cgi-bin/flew.cgi) showed a similar expression pattern for this line.

We used P{R75B10-LexA}attP40 to carry out calcium imaging from ExR2 neurons because R38A11-LexA produces relatively weak jGCaMP7f expression, and we obtained a better signal-to-noise ratio in two-photon imaging when we used R75B10-LexA to drive jGCaMP7f expression. Images from the FlyLight database (https://flweb.janelia.org/cgi-bin/flew.cgi) show that this line has strong expression in a subset of dopamine neurons from the PPM3 cluster, namely ExR2 FB4M and/or FB4L. This line also has weaker expression in a lobula cell type, the SEZ, and the SMP. During calcium imaging, the only detectable jGCaMP7f expression in the ellipsoid body and bulb regions was from the ExR2 neurons. We observed that jGCaMP7f signals were highly correlated in the ellipsoid body and bulb, as we would expect if these signals arose from ExR2 neurons exclusively.

We used P{R75B10-Gal4}attP2 to drive expression of Kir2.1 in ExR2 neurons. Images from the FlyLight database (https://flweb.janelia.org/cgi-bin/flew.cgi) show that this line has strong expression in ExR2 dopamine neurons in the ellipsoid body, as well as some labelling in a middle layer of the fan-shaped body. This line also has weaker expression in a lobula cell type, the SEZ, and the SMP. Immunostaining of GFP protein in flies with the genotype w[1118], P{13×LexAop-IVS-jGCaMP7f}su(Hw)attP8 / +; P{R60D05-LexA}attP40 / +; P{UAS-HsapKCNJ2.EGFP}7 / P{R75B10-Gal4}attP2 revealed the clearest Kir2.1::eGFP expression in the ExR2 neuron and middle fan-shaped body layers.

Origins of transgenic stocks

P{20XUAS-IVS-mCD8::GFP}attP40 and PBac{13×LexAop2-IVS-Syn21-Chrimson::tdT-3.1-p10}su(Hw)attP5 were gifts from B. Pfeiffer and G. Rubin and have previously been published^60,61. Rubin Gal4 and LexA lines (P{R60D05-Gal4}attP2, P{R38A11-LexA}attP40, P{R75B10-LexA}attP40 and P{R75B10-Gal4}attP2) were obtained from the Bloomington Drosophila Stock Center (BDSC); the general methods for constructing these lines have previously been published^60,62,63. Lines for expressing jGCaMP7f under LexAop or UAS control (PBac{13×LexAop-IVS-jGCaMP7f}VK00005 and P{20XUAS-IVS-jGCaMP7f }su(Hw)attP5) were obtained from the BDSC and have previously been published³⁷. Lines for expressing GFP under LexAop control (P{13×LexAop2-mCD8::GFP}attP40 and P{13×LexAop2-mCD8::GFP}attP2) were obtained from the BDSC and have previously been published⁶⁰. The line used for expressing P2X2 under LexAop control (P{LexAop-P2X2.Y}3) was obtained from the BDSC and has previously been published⁶⁴. The line for expressing Kir2.1::eGFP under UAS control (P{UAS-HsapKCNJ2.eGFP}7) was obtained from the BDSC and has previously been published⁶⁵.

Fly preparation and dissection

Newly eclosed female Drosophila melanogaster were anaesthetized on ice and were collected about 3–10 h (electrophysiology) or 1–4 days (imaging) before the experiment. For the ExR2 imaging experiments in Fig. 1, we deprived the flies of food (but not water) and kept them in isolation for approximately 24 h before the experiment to promote walking behaviour.

For electrophysiology experiments, the fly holder consisted of flat titanium foil secured within an acrylic platform. The fly’s head was pitched forwards so that the posterior surface was parallel to the foil and most of each eye was under the foil. For imaging experiments, a holder was custom designed using CAD software (OnShape) and created in-house using a three-dimensional (3D) printer (Form 2 with Grey Pro and Rigid resins, Formlabs) to expose a larger surface area of the fly’s eye below the holder, and the fly’s head was pitched only slightly forwards (<30°).

The fly was secured in the holder using ultraviolet-curable adhesive (Loctite AA 3972) cured by a brief (<1 s) pulse of ultraviolet light (LED-200, Electro-Lite Co.). After the dorsal portion of the head (above the holder surface) was covered in saline solution, a hole was cut in the head capsule and trachea were removed to expose the posterior surface of the brain. To reduce brain movement, muscle 16 and proboscis muscles were clipped with forceps. For electrophysiology, an aperture was made in the perineural sheath by pulling gently with fine forceps or by using suction from a patch pipette containing external solution.

The external saline solution contained (in mM): 103 NaCl, 3 KCl, 5 N-tris(hydroxymethyl) methyl-2-aminoethane-sulfonic acid, 8 trehalose, 10 glucose, 26 NaHCO₃, 1 NaH₂PO₄, 1.5 CaCl₂ and 4 MgCl₂, with osmolarity adjusted to 270–273 mOsm. External solution was bubbled with 95% O₂ and 5% CO₂ and reached a final pH of 7.3. The external saline solution was continuously perfused over the brain during experiments.

Patch-clamp recordings

Patch pipettes were made from borosilicate glass (Sutter, 1.5 mm o.d., 86 mm i.d.) using a Sutter P-97 puller. Pipettes were fire polished down after pulling⁶⁶ using a microforge (ALA Scientific Instruments) to a final resistance of 8–15 MΩ. The internal pipette solution contained (in mM): 140 potassium aspartate, 10 4-(2-hydroxyethyl)-1-piperazineethanesulfonic acid, 4 MgATP, 0.5 Na₃GTP, 1 ethylene glycol tetraacetic acid, 1 KCl and 13 biocytin hydrazide. The pH was 7.3, and the osmolarity was adjusted to about 268 mOsm. For experiments that did not require rapid pharmacology, the saline solution was perfused at about 2 ml min⁻¹; higher flow rates were used for rapid drug delivery (see below). Recordings were carried out at room temperature.

To visualize the brain, we used a custom-modified Olympus BX51WI microscope with a 40× water-immersion objective. We removed the light source and condenser below the preparation, and we instead illuminated the brain with far-red light delivered by a fibre-coupled light-emitting diode (LED; 740 nm, M740F2, Thorlabs) through a ferrule patch cable (200 µm Core, Thorlabs) plugged into a fibre optic cannula (1.25-mm SS ferrule 200-µm core, 0.22 NA, Thorlabs) glued to the recording platform, with the tip of the cannula about 1 cm behind the fly. GFP fluorescence was visualized using a mercury arc lamp (U-LH100HG, Olympus) with an eGFP-long-pass filter (U-N41012, Chroma).

Somatic recordings were obtained in current-clamp mode using an Axopatch 200B amplifier and a CV-203BU headstage (Molecular Devices). Voltage signals were low-pass filtered at 5 kHz before digitization and then acquired with a NiDAQ PCI-6251 (National Instruments) at 20 kHz. To counteract the depolarizing leak current caused by the finite resistance of the patch electrode seal⁶⁷, we applied a constant hyperpolarizing current throughout the experiment that lowered the somatic membrane potential by about 5 mV. For all electrophysiological recordings, liquid junction potential correction was carried out post hoc by subtracting 13 mV from recorded voltages⁶⁷.

Two-photon calcium imaging

Imaging experiments were carried out as previously described²⁸ using a two-photon microscope with a movable stage (Thorlabs Bergamo II) and a fast piezoelectric objective scanner (Physik Instrumente P725) for volumetric imaging. We used a Chameleon Vision-S Ti-sapphire femtosecond laser tuned to 940 nm for two-photon excitation. Images were collected using a 20× 0.95-NA objective (Olympus). Emission fluorescence was filtered with a 525-nm bandpass filter (Thorlabs) and collected using a GaAsP photomultiplier tube (Hamamatsu).

For EPG neurons, the imaging region was centred on the protocerebral bridge, where EPG axons terminate. The imaging view was 256 × 128 pixels, and 8–12 slices deep in the z axis (4-6 µm per slice), resulting in a 6–9 Hz volumetric scanning rate. For ExR2 neurons, the imaging region was centred on the bulb and ellipsoid body. The imaging view was 256 × 128 pixels, and 12 slices deep in the z axis (6-8 µm per slice), resulting in a 6–7 Hz volumetric scanning rate.

Volumetric z-scanning signals from the piezoelectric objective scanner were acquired simultaneously with analog output signals from the visual panorama and/or analog outputs from FicTrac 2.1 through a NiDAQ PCI-6341 at 40 kHz. Data were acquired using ScanImage 2018 (Vidrio Technologies) with National Instruments hardware from Vidrio (NI PXIe-6341).

Measuring locomotor behaviour

For the experiments in Figs. 1 and 4, the fly stood on a spherical treadmill, which was a 9-mm-diameter ball made of white foam (FR-4615, General Plastics) painted with black shapes. The ball floated above a 3D printed plenum made of clear acrylic (Autotiv). Medical-grade breathing air flowed into the base of the plenum and flowed out into a hemispherical depression that cradled the ball to allow it to rotate freely. The ball was illuminated from below by two round boards of 36 IR red LED lamps (SODIAL). The movement of the ball was tracked at about 60 Hz using a video camera (CM-3-U3-13S2M-CS, FLIR) fitted with a Tamron 23FM08L 8-mm 1:1.4 macro zoom lens. Machine vision software (FicTrac 2.1) was used to convert the image of the ball to a running estimate of the ball’s position in all three axes of rotation⁶⁸. The fly’s rotational velocity was inferred from the yaw velocity of the ball. The fly’s forward velocity was inferred from the pitch velocity of the ball. The fly’s fictive heading direction (head direction) was inferred from the temporal integral of the fly’s rotational velocity.

Pharmacology

For pharmacology experiments, external saline solution was perfused quickly using a Watson Marlow Pump (120U) set at 60 r.p.m. (about 5 ml min⁻¹). At the start of a treatment trial, the intake tube was moved for exactly 30 s (electrophysiology) or 60 s (imaging) from normal saline solution into saline containing dopamine (200 µM) or ATP (5 mM). On each rig, we carried out measurements of the time it takes a new solution to flow through the tubing and reach the recording chamber. For electrophysiology, we validated this tubing delay time estimate by placing a recording electrode in the bath in voltage-clamp mode and then perfusing in external solution with a much higher salt content (>1 M NaCl) in the exact same manner that ATP or dopamine solutions were normally delivered. The first deviation in current signal measured by the electrode was used to estimate the entry of the new solution. For data display, measurements are plotted relative to the earliest time when the drug (ATP or dopamine) entered the recording chamber (t = 0).

We prepared the ATP solution (5 mM in bubbled saline) immediately before application from a frozen stock of 500 mM ATP in water. We prepared the dopamine solution (200 µM dopamine + 100 µM sodium metabisulfite to minimize oxidation⁶⁹, prepared in saline) on a daily basis, using a frozen stock of 100 mM dopamine in 50 mM sodium metabisulfite in water. ATP solutions and dopamine solutions were bubbled with 95% O₂ and 5% CO₂ before application.

To verify the timing of ExR2 activation using ATP, we carried out whole-cell current-clamp recordings from ExR2 neurons (Fig. 2a), using the same ATP protocol, the same ExR2 LexA driver line and the same LexAop-P2X2 effector transgenes that we used in our EPG electrophysiology experiments. This comparison was only possible in electrophysiology experiments because the highly specific ExR2 driver line used to drive P2X2 expression did not drive enough jGCaMP7f expression for high-quality two-photon imaging from ExR2 neurons.

Chrimson stimulation

Flies expressing Chrimson⁷⁰ were raised for all of development on Nutri-Fly “German Food” Sick Fly Formulation no. 66-115 (Genesee) containing about 0.6 mM all-trans retinal (Sigma) and tegosept anti-fungal agent. Fly vials were wrapped in foil to prevent photo-conversion of the all-trans retinal. For optogenetic stimulation, we used the Hg-lamp source (U-LH100HG) to deliver pulses of orange light (590–650 nm, 2–3 mW, Cy5 long-pass filter cube (49019, Chroma) through the objective. A shutter (Uniblitz Electronic) controlled the pulse pattern (0.5 s on, 1 s off) that was paired with the visual presentation of a bright vertical bar that rotated around the fly in the same manner as described below for all other electrophysiology experiments.

Visual panorama

Visual stimuli were presented to the fly using a circular panorama (IORodeo) made of modular square panels⁷¹ as previously described²⁸. Each square panel was made up of an 8 × 8 array of LEDs (8 × 8 ‘pixels’) that refresh at 372 Hz or faster⁷¹. For electrophysiology experiments, these LEDs were green (LED peak = 525 nm). In imaging experiments, these LEDs were blue (LED peak = 470 nm) to minimize overlap with GCaMP emission spectra.

For electrophysiology experiments, we used a panorama that spanned 270° in azimuth; the panorama covered the azimuthal range from about 124° left of the midline to about 147° right of the midline (that is, it was slightly asymmetric). For imaging experiments, we used a 360° panorama, with one square panel behind the fly removed for camera positioning. The visual panoramas spanned about 43° vertically of the fly’s visual field and a single LED pixel along the top of the arena subtended about 3.6–3.7° of the fly’s visual field.

In electrophysiology experiments, to reduce electrical noise, the panorama was wrapped with grounded copper mesh. To reduce reflections, the mesh was covered with black ink, and the front surface of each panel was covered with a diffuser (SXF-0600 Snow White Light Diffuser, Decorative Films). In imaging experiments, five layers of filters (Rosco R381, bandpass centre 440, full-width at half-maximum 40 nm) were placed in front of the panels to minimize detection of the visual stimulus by the GCaMP emission collection channel. Analog output signals from the visual panel system were digitized with a NiDAQ PCI-6251 (National Instruments) at 20 kHz (electrophysiology) or with a NiDAQ PCI-6341 (National Instruments) at 40 kHz (calcium imaging).

Visual stimuli

The visual cue was a bright vertical bar (2 pixels wide, 7°) that spanned the full height of the panorama (about 43°). For open-loop trials, the bar was rotated continuously around the fly at about 18° s⁻¹ in the rightward direction. In imaging experiments with the 360° arena, the top half of the bar was ‘jumped’ over the missing panel directly behind the fly to maintain a constant total luminance within the arena at all times. For the closed-loop training period described in Fig. 4, the angular position voltage signal from FicTrac 2.1 was used to continuously update the azimuthal position of the visual cue displayed on the panorama. Thus, when the ball rotated rightwards (indicating an attempted leftward rotational manoeuvre by the fly), the visual cue rotated rightwards at the same velocity. For the optic flow stimulus in Fig. 1f, we presented a 360° vertical grating consisting of alternating 7° dark and light stripes. The grating appeared on the screen and remained static for 4 s before starting a about 18° s⁻¹ leftward or rightward rotation, to isolate responses to the optic flow from any responses to the appearance of the grating. On average, the appearance of the vertical grating produced a relatively small transient ExR2 response (ΔF/F ≈ 0.4, not shown). The grating rotated for 4 s before disappearing, and ΔF/F responses were measured in the last 2 s of this stimulus period. The analog output signals from the visual panel system and from FicTrac 2.1 were digitized with a NiDAQ PCI-6341 (National Instruments) at 40 kHz. A visual cue position of 0° means the cue is directly in front of the fly.

Connectome analysis

To analyse the proximity of ExR2 output sites to ER-to-EPG neuron synapses, we analysed a partial connectome of the dorsal part of the right central brain of an adult female fly obtained by the FlyEM project at Janelia Research Campus (https://janelia.org/project-team/flyem/hemibrain)⁵⁷. Analysis was carried out in R using the neuprintr extension⁷² of neuprint⁷³. ExR2, ER and EPG annotations were taken from the hemibrain v1.2.1 release. We calculated the Euclidean distance between each ER-to-EPG neuron pre-synapse and the ExR2 release site that is closest to that pre-synapse. Analysis was restricted to synapses within the ellipsoid body neuropil region. The boundaries of the ellipsoid body and other neuropil regions (Extended Data Fig. 1) were extracted from the hemibrain dataset.

Our analysis of mushroom body synapses (Extended Data Fig. 1) focused on Kenyon cells (KCs), mushroom body output neurons (MBONs) and mushroom body dopamine neurons (MB-DANs). We calculated the Euclidean distance between each KC-to-MBON synapse and its closest MB-DAN output synapse. This analysis was restricted to neurons on the right side of the brain with clear compartmentalization in the gamma lobe of the mushroom body. Our results of mushroom body synapses are consistent with those of a previous analysis⁷⁴.

Data analysis

Data analysis was carried out using Matlab R2016b, R2017a, R2017b, 2019b, R2020b (MathWorks), Python 3.9.5, R 4.1.0 and RStudio 1.4.1717. Throughout the figures, statistical tests are summarized as follows: NS, P > 0.05; *P < 0.05; **P < 0.01; ***P < 0.0005.

Calcium imaging alignment and data processing

Rigid motion correction in the x and y axes was carried out for each acquisition using the NoRMCorre algorithm⁷⁵. Volumetric regions of interest (ROIs) were defined by combining 2D ROIs drawn in multiple imaging planes. Fluorescence values for each ROI were determined by averaging all pixels in that volumetric ROI. All fluorescence data were smoothed with a Gaussian kernel 600 ms in width before analysis. For ExR2 imaging (Fig. 1), ROIs were drawn throughout the ellipsoid body (EB) and the left and right bulbs. ROIs from the EB and the left and right bulbs were combined for analysis to improve the signal-to-noise ratio. For EPG imaging, 16 volumetric ROIs corresponding to the 16 glomeruli in the protocerebral bridge were defined on the basis of visible anatomical boundaries. ROIs of EPG neurons from the left and right hemispheres that occupy the same part of the EB were then combined to create one ROI for each of the eight EB wedges. To calculate the time-dependent change in fluorescence (ΔF/F) for each ROI, we used a baseline fluorescence (F) defined as the fifth percentile of raw fluorescence values for that ROI throughout the entire experiment. For the optic flow analysis in Fig. 1f, fluorescence in an empty background ROI was subtracted from fluorescence in the ExR2 ROI, to adjust for any light from the visual panels that was picked up by the photomultiplier tube during imaging. For that analysis, ΔF/F was calculated using the mean fluorescence from the 2 s before the stimulus onset as the baseline F.

ExR2 locomotor correlations

The displacement of the spherical treadmill was computed by FicTrac 2.1 (ref. ⁶⁸) in each of the three axes of rotation at about 60 Hz. This was then used to calculate forward and rotational speeds of the fly at each time point. Speed data were downsampled to match the volume rate of the imaging data and smoothed 10 times with a Gaussian kernel 60 ms in width. Finally, the speed data were shifted back in time by two imaging volumes (about 300 ms), because this maximized the correlation between the two signals. In the example traces shown in Fig. 1b, the speed traces were smoothed one additional time with a Gaussian kernel 600 ms in width after downsampling and were not shifted in time. For the binned rotational speed analysis (Fig. 1c), all of the ΔF/F data for a given experiment were divided into 45 speed bins, and the mean ΔF/F within each bin was calculated. Bin widths and edges were chosen so that each bin contained the same number of data points. For the binned 2D speed analysis (Fig. 1d), speed and ΔF/F from all experiments were combined and then divided into bins on the basis of 2D speed. The bins in this analysis had a uniform size in each speed axis, meaning that the number of data points per bin was not uniform across bins. The grey shading in Fig. 1d shows the range of very low rotational speeds that occur during transitions between resting and moving; all of these speeds (including rest periods and rest–move transitions) were included in our analyses. In a subset of experiments, the same analysis was carried out on trials in which the visual cue was rotated around the fly in an open loop (Extended Data Fig. 4b), and for separate experiments the equivalent analysis was carried out with the same cue in a closed loop (Extended Data Fig. 4a). A line was fitted to the portion of each of the resulting curves with rotational speeds greater than 30° s⁻¹ using Matlab’s fitlm() function. To generate the plots in Extended Data Fig. 3a, ROIs were drawn around ExR2 neurites in the left and right lateral accessory lobes, and ΔF/F was calculated as described above. Correlation coefficients between the fly’s rotational speed and ExR2 activity in each ROI were calculated using Matlab’s corrcoef() function. Only epochs in which the fly was moving (rotational speed >15° s⁻¹ or forward speed >2 mm s⁻¹) were included in the analysis. The plots in Extended Data Fig. 3b were generated using data from two different 5-min epochs during a single recording, with a 5-min gap between them. The correlation between rotational speed and ExR2 activity was calculated as described above, and the resulting coefficients were smoothed with a 2D Gaussian kernel (σ = 1.5) and manually thresholded to show only the pixels with the strongest positive and negative correlations. Background (greyscale) images in Extended Data Fig. 3b show trial-averaged fluorescence from these same five imaging planes.

Linear model analysis

Forward and rotational speed data were processed as described above, and then z-scored and used as predictor variables. Fluorescence data were z-scored and used as the dependent variable. For each experiment, two regression models were fitted using Matlab’s fitlm() function, one using only rotational speed as a predictor variable, and one using both forward and rotational speed. Adjusted R² values were obtained from the output of the fitlm() function; note that all of these R² values are adjusted for the degrees of freedom in the model fit, which allows us to compare the explanatory powers of models having different numbers of free parameters.

Visual learning network model

Our model was a modified version of a published model²⁹ (https://research.janelia.org/jayaraman/Kim_etal_Nature2019_Downloads). In this model, the dynamics of the EPG population are given by

in which [𝑥]₊ denotes that we are taking the non-negative values of 𝑥, f_n is the activity of EPG neuron n, τ is a decay time constant (50 ms), (a) is the strength of self-excitation, D is the strength of excitation from neighbouring neurons, β is the strength of global inhibition, and v is a rotational velocity signal that can be positive (rightward rotation) or negative (leftward rotation). ({I}_{n}) is the inhibitory visual input to EPG neuron n:

in which w_n,m is the strength of the synapse from visual neuron m onto EPG neuron n, and g_m is the activity of visual neuron m. Weights are constrained to the range [0 0.33] and visual neuron activity is constrained to the range of [0 0.35]. Note that ({I}_{n}le 0), as the vectors w and g contain only non-negative values. Ensemble visual neuron activity g was modelled as a von Mises function ((kappa =15)) over 1D azimuthal space, with the minimal value of g_m set to zero. To simulate visual noise, we generate random samples from a uniform distribution over the range [0, 0.5g_max] and then apply temporal smoothing (box-car averaging over 80 ms); this noisy fluctuation was then added to I_n. The Hebbian learning rule at visual synapses onto EPG neurons was assumed to be postsynaptically gated, meaning that learning can occur only at synapses onto active EPG neurons:

in which w_max = 0.33, g₀ = 0.33 and η is the learning rate (see below for more details). The first term (η[f_n]₊ (w_max − w_n,m)) can be interpreted as nonassociative long-term potentiation that depends on postsynaptic activity, whereas the second term (η[f_n]₊ w_maxg_m/g₀) can be interpreted as associative long-term depression that depends on the conjunction of presynaptic and postsynaptic activity. All elements of the model noted thus far are identical to one of the model variants in ref. ²⁹.

Taking this framework as a starting point, we then modified the model of ref. ²⁹ in several ways. First, as the input to the model, we used rotational velocity data that we recorded from flies walking on a spherical treadmill in a virtual environment with a single visual cue (that is, the same closed-loop visual cue that we use in the training period in Fig. 4). For each model run, we combined nine different 5-min epochs of rotational velocity data in a random order. These rotational velocity values from our data were taken as the time-varying parameter (v); they were also used to shift the bump of activity in the visual neuron population (that is, they were used to shift the von Mises function across azimuthal space). We used a time step length of 16.1 ms for our simulation to match the sampling rate of the rotational velocity data.

Next, we scaled the synaptic learning rate (eta ) at each time point so that it was linearly dependent on the fly’s rotational speed; this choice was motivated by our finding that the activity of ExR2 dopamine neurons is quasi-linearly dependent on the fly’s rotational speed (Fig. 1c). We refer to this as an adaptive learning rate. Note that our formulation (η = |0.5v|) is different from that of ref. ²⁹ (η = 0.5v²).

Finally, we re-ran all of the simulations with the same behavioural data sequences as inputs, but now setting the learning rate η to a fixed value, obtained by taking the mean value of η throughout the training period in the adaptive learning models. Choosing this fixed value for η matched the total amount of learning across the two conditions. After the training was complete, we calculated the circular correlation between the population vector averages (PVAs) of the synaptic weights of each EPG neuron and each ER neuron at each time step, using a published method⁷⁶. Circular correlation was computed over 3,000 s of simulation time; the mean value reported is the mean of 117 simulations (trained on shuffled blocks of behavioural data) ± 95% confidence interval (1,000 bootstrap resamples).

Equation (3) represents the postsynaptically gated learning rule that is the focus of ref. ²⁹. In separate simulations, we also implemented the presynaptically gated learning rule of ref. ²⁹, and we confirmed that our conclusions are the same. Specifically, adding an adaptive learning rate can have the same effect. The only difference is that, for the presynaptically gated learning rule, the overall weight structure is slightly less uniform (that is, there is a slightly lower correlation between the PVA of ER output weights and EPG input weights).

For the analysis in Extended Data Fig. 5, the final weights of 16 simulations were taken as initial weights, and 3 replicates of these 16 simulations were continued for another 3,000 s of simulation time, but with Gaussian noise added to the synaptic weight matrices at each time step. The first two replicates used the synaptic learning rules described above, whereas the third had no synaptic learning at all.

EPG ensemble responses to ExR2 activation or hyperpolarization

In Figs. 3 and 4, ΔF/F data for each EB wedge were calculated as described above. The PVA was calculated by converting the ΔF/F data for each wedge into a vector with a direction based on the wedge’s position in the EB, and then adding those vectors to obtain the PVA for each time point. We then computed the circular distance between the cue and the PVA position for each time point; we refer to this as the ‘offset’ between cue position and the bump position⁷⁷. We computed bump amplitude for each time point by taking the difference between the minimum and maximum ΔF/F across all eight EB wedges.

Mutual information between the PVA and the visual cue was estimated using a published method⁷⁸ with k = 3 nearest neighbours, using the Python implementation from that reference; see also ref. ⁷⁹. For Fig. 3c, mutual information was estimated using the time points within each individual rotation of the visual cue around the fly. For Fig. 3d, the estimate used all time points in the baseline or post-ExR2 activation periods (as indicated in Fig. 3c). For Fig. 4c, mutual information was estimated for each fly using all time points within the open-loop test period.

Time points when the cue was behind the fly (at positions between 150° and 210°) were excluded from the calculations of offset and mutual information because the fly should have been unable to see the cue during these time points; including these data did not change the conclusion from either analysis.

Correlating bump amplitude and locomotor speed

In Fig. 4e and Extended Data Fig. 9c, speed data were collected through FicTrac 2.1 (ref. ⁶⁸) at about 60 Hz, downsampled to the imaging data volume rate, smoothed 10 times with a Gaussian kernel 60 ms in width, and shifted back 2 imaging volumes (about 300 ms) in time to maximize correlation between the signals. Bump amplitude was calculated as maximum − minimum ΔF/F across all eight EB wedges at each time point, and then smoothed with a Gaussian kernel 600 ms in width. Then, for each fly, bump amplitude was z-scored and binned according to the fly’s rotational speed at each time point; and the mean bump amplitude was then calculated within each bin. Bin edges were the same for each fly and were chosen so that the number of samples in each bin was the same after aggregating data across flies. Time points with rotational speeds exceeding 100° s⁻¹ (<3% of the total samples) were excluded from the analysis to ensure consistency across flies, because some individuals had higher walking speeds than others. For Fig. 4e and Extended Data Fig. 9c, the Pearson correlation coefficient between bump amplitude and rotational speed throughout the entire training and test period was calculated in Matlab using the corrcoef() function.

Whole-cell recording EPG neuron visual tuning

To describe the preferred cue position of an EPG neuron based on its membrane voltage, we computed its vector phase for each full cue rotation using the following equation⁸⁰:

in which the cue position (θ) ranges from 0 to 360° and V_θ is the membrane voltage at a given stimulus angle. To adapt this analysis to the 270° azimuthal extent of the visual panorama used in our electrophysiology experiments, θ was obtained by rescaling and shifting cue positions from −123.75°–146.25° to 0–360°. The cell’s preferred cue position was then obtained by rescaling and shifting the calculated vector phase to the range [−123.75, +146.25°]. To determine the preferred cue position based on EPG spiking (Extended Data Fig. 6e), we calculated the vector phase as follows:

in which θ is the list of cue positions that were present at the time of a spike.

In Fig. 2d, the circular standard deviation of the preferred cue position was calculated over a pre-stimulus-window (−11 min to −1.5 min) and a post-stimulus window (3 min to 12.5 min). Changes in preferred cue position tuning for each cell were assessed using the parametric Watson–Williams multi-sample test (implemented through circ_wwtest in Matlab⁷⁷), with Bonferroni correction. Specifically, we compared the preferred cue position values in a pre-stimulus window (−3.5 min to −1.5 min) to the preferred cue position values in a post-stimulus window (3 min to 5 min), as shown in Fig. 2c. We chose windows that are relatively narrow and closely spaced to minimize the contribution of the slow representational drift in the EPG ensemble that occurs even in control experiments over long time intervals. This drift is clearly distinct from the abrupt changes in preferred cue position that often accompany ExR2 stimulation or dopamine application, as shown in Fig. 2c. A small subset of cells were not significantly tuned to the visual cue position in the baseline period before ExR2 application or dopamine application (that is, they did not pass the Rayleigh test for uniformity with a threshold of P = 0.05). These cells were excluded from further statistical analysis of change in preferred cue position because the preferred cue position is not a meaningful value if there is no baseline tuning. For ExR2 activation, 11 cells were recorded, and of these, 2 cells were excluded for non-significant baseline tuning, and 6 out of the 9 remaining cells showed significant changes in preferred cue position after ExR2 activation (Bonferroni-corrected P = 1.9 × 10⁻¹⁴, 4.4 × 10⁻¹⁴, 1.3 × 10⁻⁹, 6.6 × 10⁻⁵, 0.0018, 0.011, 4.15, 4.51, 7.99). For the control genotype, 10 cells were recorded, 0 cells were excluded, and 0 cells showed significant changes in preferred cue position after ATP application (Bonferroni-corrected P = 0.064, 0.33, 0.56, 0.57, 1.95, 5.49, 6.26, 6.62, 6.99, 7.15). For dopamine application, 12 cells were recorded, and of these, 1 cell was excluded for non-significant baseline tuning, and 3 out of the remaining 11 cells showed significant changes in preferred cue position after dopamine application (Bonferroni-corrected P = 0.0066, 0.044, 0.045, 0.097, 1.17, 1.84, 1.86, 5.08, 5.70, 4.22, 6.89). Example cells 3, 4 and 5 that are shown in Fig. 2c showed significant changes in mean preferred cue position, whereas example cells 1, 2 and 6 did not.

Extended Data Figure 7 shows the changes in preferred cue position for every cell, including the three with non-significant baseline tuning. The three cells that failed the Rayleigh test were not excluded from any other analyses, because no other analyses would be confounded by the lack of baseline tuning in these cells.

For Fig. 2e,f and Extended Data Fig. 6a,b,f, the cell’s visual response amplitude was computed by smoothing the membrane voltage using a third-order median filter with a 100-ms window (medfilt1() Matlab function), and then subtracting the minimum voltage from the maximum voltage on each cue rotation cycle. This voltage difference was reported directly (Fig. 2b and Extended Data Fig. 6a) or normalized to the mean response during a baseline period (Fig. 2e,f and Extended Data Fig. 6f; −3.5 to −1.5 min), or reported as post − pre (Extended Data Fig. 6b; pre is −3.5 min to −1.5 min; post is 4 min to 6 min). Mean responses across cells were calculated by taking the mean of all the single-cell measurements in bins of 40 s.

For Extended Data Fig. 8, the cycle-by-cycle circular standard deviation and visual response amplitude were calculated for the full 2-min trials that preceded or followed the optogenetic stimulus in the same manner as described above. Response amplitude was calculated by taking the mean voltage for each cue position across the eight cue rotations per trial and subtracting the maximum voltage from the minimum voltage.

ExR2 and EPG spiking

ExR2 and EPG spikes were detected from whole-cell recordings by a custom-written Matlab script. To detect ExR2 spikes, first the raw voltage was multiplied by −1 and low-pass filtered with a digital Butterworth filter using a cutoff frequency chosen by the user. Then, the derivative was taken of this filtered signal. We detected peaks in the derivative trace that passed above a user-chosen threshold, were wider than 2.5 ms, and followed the preceding peak by more than 2 ms. Values of the cutoff frequency for the low-pass filter and the differentiated peak threshold were chosen for each cell empirically. The peristimulus time histograms of spike rate were calculated using a bin size of 5 s. Mean baseline firing rates for each cell were as follows: 0.28, 0.26, 0.04, 0.52 and 0.02 Hz. Automatic EPG spike identification (Extended Data Fig. 6c,e) was based on the detection of transients in the current output from the Axopatch 200B amplifier⁸¹; these transients were counted as spikes if a fluctuation wider than 1.5 ms occurred in the voltage channel within 1.5 ms of that sample, and if the peak voltage during that fluctuation was in the top 15% of values for that trial.

Data inclusion

In one imaging experiment, an air bubble formed near the laser path soon after the beginning of the experiment (confirmed by eye with light microscopy at the end of the experiment), resulting in a marked reduction in brightness and signal-to-noise ratio, so this experiment was excluded from analysis.

For electrophysiology, cells were considered healthy and included in the analysis if their voltage was below −30 mV. For pharmacology experiments, the cell’s membrane voltage needed to remain healthy until 11 min into the treatment period to be analysed. If cells became more depolarized than −30 mV following this time point that data were excluded from analysis.

As previously reported²⁸, EPG neuron electrophysiological recordings occasionally exhibit large inhibitory postsynaptic potentials with a stereotyped sharp onset, a large amplitude (>10 mV) and a stereotyped time course. They are followed by a prolonged period of depolarization when the variance of the voltage trace is also diminished. These inhibitory events interfered with visual tuning measurements, and so for Fig. 2 and Extended Data Figs. 6–8, if an event occurred it was clipped out. Such clipping was required for 12% of trials. For one cell (fly 543, cell 1), the first 12 s of the first baseline trial had too much holding current applied. Those 12 s were also excluded from the analysis.

Statistics and sample sizes

Normality was evaluated with a Kolmogorov–Smirnov test or a Shapiro–Wilk test with α = 0.05. In cases when our data were not normally distributed, a nonparametric test was used. To analyse circular variables, we used statistical tests for circular statistics from the Matlab toolbox CircStat⁷⁷. Sample sizes were chosen on the basis of standard sample sizes in the field.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.