Plastic landmark anchoring in zebrafish compass neurons

Animal husbandry

The animal handling and experiments were performed according to protocols approved by the animal welfare officer at Institüt für Neurowissenschaften, Technische Universität München (TUM) and the relevant department at the regional government (Regierung von Oberbayern, Sachgebiet 55.2; animal protocol number 55-2-1-54-2532 10112 and 55.2-2532.Vet_02-24-5). Adult zebrafish (Danio rerio) were housed in the facility at the Institute for Neuronal Cell Biology at TUM. The adult fish were maintained in water temperature of 27.5–28.0 °C on the 14–10 h light–dark cycle. All experiments were performed on 6–9 days post-fertilization larvae of undetermined sex. The eggs were kept in 0.3× Danieau solution, and in the water from the fish facility upon hatching. The larvae were maintained at 28.0 °C and under the 14–10 h light–dark cycle.

Animal strains

All imaging experiments of the HD neurons were performed on fish carrying Tg(gad1b–Gal4)mpn155 (ref. 18) and Tg(UAS–GCaMP6s)mpn101 (ref. 46). To record the activity of habenula neurons (Extended Data Fig. 8), either Tg(vglut2a–Gal4)nns20 (ref. 26) (n = 6 fish) or a previously uncharacterized enhancer trap line Tg(18107–Gal4) was used (n = 2 fish) with UAS–GCaMP6s. The expression pattern of the 18107–Gal4 line can be browed on Z Brain Atlas (https://zebrafishexplorer.zib.de/home/). For labelling habenula for the ablation (Fig. 5 and Extended Data Fig. 9), Tg(18107–Gal4) was used. A subset of fish in the ablation experiment possessed Tg(UAS–nfsB–mCherry)47 for a logistical reason. This does not affect the results of the ablation experiments as they were evenly distributed across the conditions, and the fish were not treated with relevant chemogenetic reagents.

The experiment to search for putative AHV cells (Extended Data Fig. 10e–j) was performed on fish carrying Tg(HuC–H2B–jGCaMP7c). All fish were mitfa−/− (that is, nacre) mutants lacking melanophores to allow optical access to the brain.

Two-photon microscopy experiments

Animal preparation and the stimulus presentation setup

Animals were embedded in 2% low-melting point agarose in 30-mm petri dishes. The agarose around the tail was carefully removed with a scalpel to allow tail movements. The dish was mounted on a 3D-printed pedestal and placed in a cube-shaped acrylic tank with the outer edge length of 51 mm. The height of the pedestal was designed so that the head of the animal came to the centre of the tank, taking the thickness of the dish and the typical amount of agarose into account. The tank was then filled with fish facility water to minimize the refraction due to the petri dish wall. The three sides of the tank (except for the one facing the back of the fish) were made of single-side frosted acrylic (PLEXIGLAS Satinice 0M033 SC), which functioned as projection screens. The diffusive side faced inwards to minimize the reflections between the walls.

The visual stimuli were projected onto the three walls through two sets of mirrors with a previously described geometry16 (Supplementary Video 1), subtending 270° horizontally and 90° vertically. The larvae were lit with an infrared LED array through the transparent back wall of the tank. Their tail movements were monitored from below with a high-speed camera (Allied Vision Pike F032) at 200 Hz, through a hot mirror and a short pass filter to reject the excitation beam.

Microscope

Functional imaging was performed with a custom-built two-photon microscope. The excitation was provided by a femtosecond pulsed laser with 920-nm wavelength, the repetition rate of 80 MHz and the average source power of 1.8 W (Spark ALCOR 920-2). The average power at the sample was approximately 10 mW. The scan head consisted of a horizontally scanning 12-kHz resonant mirror and a vertically scanning galvo mirror, controlled by a FPGA running a custom LabView code (LabView 2015)⁴⁸. Pixels were acquired at 20 MHz and averaged eightfold, resulting in the frame rate of 5 Hz. The typical dimension of the image was about 100 µm × 100 µm, with the resolution of about 0.2 µm per pixel. Only pixels corresponding to the middle 80% of the horizontal scanning range were acquired to avoid image distortion, and the area outside was not excited to minimize photo-damage. The fast power modulation was achieved with the acousto-optic modulator built in to the laser.

Stimulus protocols

All visual stimulus presentation and behavioural tracking were performed using Stytra package49 (v0.8). The panoramic virtual reality environments were created and rendered using OpenGL through a Python wrapper (ModernGL). In each frame of the stimuli, three views of the virtual environment corresponding to the three screen walls were rendered, which were arranged on the projector window to fit the screens. Behavioural tracking was performed as previously reported49. In brief, seven to nine linear segments were fit to the tail of the larva, and the ‘tail angle’ was calculated at each camera frame as the cumulative sum of the angular offsets between the neighbouring pairs of segments. To detect swimming bouts, a running standard deviation of the tail angle within a 50-ms window was calculated (‘vigour’). A swimming bout is defined as a contiguous period during which the vigour surpassed 0.1 rad.

For each bout, the average tail angle within 70 ms after the onset was calculated, with a subtraction of the baseline angle 50 ms before the bout onset. This average angle (‘bout bias’) captures the first cycle of the tail oscillation in a bout and correlates well with the heading change in the freely swimming larvae50. We estimated the head direction of the fish in the virtual world as the cumulative sum of bout bias. The time trace of the head direction was also smoothed with a decaying exponential with the time constant of 50 ms, such that swim bouts result in smooth rotations of the scene (as opposed to instantaneous jumps).

For the recordings from the HD neurons, we simulated a virtual cylinder around the fish, whose height was determined so that the gaze angles to the top and bottom would be respectively ±60°. Various textures representing the visual scene, generated with the dimension of 720 × 240 pixel, were mapped onto this virtual cylinder. Dynamic aspects of the stimuli (that is, scene rotations and movements of the dots) were achieved by updating the textures on the cylinder. The visual scenes used were as follows:

• Flash: uniform fields of black or white.

• Sun-and-bars: three black vertical bars on a single radial gradient of luminance, ranging from white at the centre and black at the periphery. The bars were 15° wide and respectively centred at −90°, +75°,and +105° azimuths (0° is to the front and positive angles to the right). The centre of the gradient was in front and 45° above the horizon, and the radius was 135°.

• Translating dots: dots randomly distributed in a virtual 3D space at the density of 7.2 cm⁻³ moved at 10 mm s⁻¹ sideways. The dots within the 40-mm cubic region around the observer were projected as 3 × 3-pixel white squares against a black background (in a texture bitmap on the cylinder), regardless of the distance.

• Stonehenge: four white vertical bars on a black background. The bars were positioned at −120°, −90°, 0° and +135° azimuths, respectively. The rightmost bar was broken in the vertical direction with the periodicity of 20° elevation and the duty cycle of 50%.

• Cue-cards: a white rectangle with the 90° centred about the 0° azimuth, which spanned the elevation ranges of either above +20° (top cue) or below −20° (bottom cue). The background was black.

• Noise: a 2D array of uniform random numbers within (0, 1), smoothed with a 5 × 5 pixel 2D boxcar kernel and binarized into black and white at 0.5.

• Single sun: a radial gradient of luminance from white at the centre and black at the periphery, centred at −90° azimuth and 35° above the horizon, with the radius of 60°.

• Double sun: the same as the single-sun scene, but symmetrized around the vertical meridian.

To identify and exclude naively visual neurons, each HD cell recording started with the alternating presentation of white and black flashes (8 s long each, five repetitions). In the experiment in Extended Data Fig. 3 the translating dots moving leftwards and rightwards alternatingly were also presented (8 s long, five repetitions). Afterwards, epochs of closed-loop scene presentations started. At the beginning of each epoch, the scene orientation was reset to 0°. On top of the closed-loop control, episodes of exogenous slow rotation (18° s⁻¹) were superimposed intermittently (5 s every 30 s (Figs. 1 and 2 and Extended Data Figs. 3 and 4a–g) or 20 s (Figs. 3–5 and Extended Data Fig. 4h–m)). The directions of the rotations flipped after every four rotational episodes. The structures of the virtual cylinder experiments were as follows:

• Sun-and-bars experiment (introduced in Fig. 1): in the first epoch (10 min), the sun-and-bars scene was presented. In the second epoch (10 min), fish received no visual stimuli (that is, darkness).

• Translating dots experiment (introduced in Extended Data Fig. 3): in the first epoch (8 min), the sun-and-bars (n = 10 fish) or Stonehenge (n = 15 fish) scene was presented. In the second epoch (15 min), the fish observed translating dots moving either left or right. The dots disappeared and the screen turned uniform white if the fish performed a bout or 10 s passed without a bout (that is, no rotational visual feedback). The dots reappeared after waiting for 10 s.

• Stonehenge experiment (introduced in Extended Data Fig. 4a–g): in the first and second epochs (8 min each), the sun-and-bars and the Stonehenge scenes were presented, respectively.

• Cue-card experiment (introduced in Extended Data Fig. 4h–m): the sun-and-bars, bottom cue-card and top cue-card scenes were presented for 4 min each.

• Jump and noise experiment (introduced in Fig. 2): in the first and second epochs (6 min each), the sun-and-bars scene was presented. In the second epoch, the superimposed exogenous rotations were swapped with abrupt 90° jumps. In the third epoch (6 min), the noise scene was presented.

• Symmetry experiment (introduced in Fig. 3): in the first and third epochs (12 min each), the single-sun scene was presented. In the second epoch (12 min), either the double-sun scene (n = 25 fish; Fig. 3) or single-sun scene (n = 20 fish; Extended Data Fig. 7) was presented.

• Ablation experiment (introduced in Fig. 5): in all epochs, the sun-and bars scene was presented. The first epochs were 12 min and 6 min long in the pre-ablation and post-ablation recordings, respectively. In the second epochs (6 min), the superimposed exogenous rotations were swapped with abrupt 90° jumps.

The experiment to characterize habenula visual responses (Extended Data Fig. 8) started with alternating black and white flash presentations (6 s each, five repetitions), which were used to select visually responsive cells. Next, white vertical or horizontal bars against a dark (25% luminance) background (which respectively subtended the entire height or circumference of the cylinder) were presented at 16 different azimuths and 5 different elevations, respectively. The width of the bar was 14°, and their azimuths and elevations were evenly spaced in the range of (−112.5° to 112.5°) and (−30° to 30°), respectively. Each presentation of bars lasted 4 s, with the interleave of 6 s. Each orientation and position combination were repeated three times, and the presentation order was randomized. Finally, the sun-and-bars scene rotating about the fish for the full 360° at 9° s−1 in an open loop was presented four times, in alternating directions.

In the habenula ablation experiment, we recorded the responses of the habenula neurons to alternating black and white flash stimuli (8 s each, ten repetitions) to determine the visually responsive side (Extended Data Fig. 9b,c).

For the experiment to look for putative AHV cells (Extended Data Fig. 10e–j), we simulated a uniformly distributed point cloud in the 3D virtual reality environment (instead of simulating dots as a texture on the cylinder). Specifically, we simulated 2,000 dots within a cubic area with side length of 40 mm, and dots within the 20-mm radius from the observer were rendered as bright spots on a dark background with a diameter of about 1.2° (regardless of distance). Translational and rotational optic flow was simulated by moving the camera in the virtual environment. The experiment started with alternating presentations of short (5 s) yaw rotational optic flow and translational optic flow sideways, interleaved with 5 s of blank, dark screens, repeated five times, which were intended for characterizing the sensory responses of the cells.

Next, leftwards and rightwards translational optic flow was continuously presented for 5 min each, which was intended to facilitate fish to turn, so that we could analyse the bout-triggered activity of the cells. The whole experiment was in open loop. The data acquired with two different sets of velocity parameters were merged together: in five fish, the speed of the rotational optic flow was 18° s−1, and the translational optic flow moved 90° to the side at 5.0 mm s−1. For the rest (n = 16 fish), the rotational optic flow was at 6° s−1, and the translational optic flow moved 45° to the side-front at 3.0 mm s−1.

Laser ablations

The habenula axons (that is, fasciculus retroflexus) were ablated unilaterally either by scanning a laser within a small region of interest (ROI) on the fasciculus retroflexus (setup A: Spectra Physics MaiTai, 830 nm, source power 1.5 W) or by pointing a laser on the fasciculus retroflexus (setup B: Spark ALCOR 920-2, 920 nm, source power 1.8 W). The pulsing characteristics of the two lasers were comparable (repetition rate of 80 MHz, pulse width of less than 100 fs), but ALCOR was group delay dispersion corrected and thus more efficient for ablations. On setup A, scanning with the duration of 200 ms was repeated 10 times with an interval of 300 ms. On setup B, a couple of approximately 100-ms-long pulses were delivered. These procedures were repeated until the successful ablation was confirmed either by a spot of increased fluorescence due to the photo-damage or a cavitation bubble.

Ablations were repeated at two to three locations around the midbrain or pretectum levels to ensure the complete cut of the fasciculus retroflexus. The numbers of the fish treated on setups A and B were n = 8 and 8 (visual side ablated and control side, respectively) and 5 and 7 (visual side ablated and control side, respectively) fish, respectively. We waited at least 1 h after the ablation before making the post-ablation recordings.

Data analysis

Behavioural data analysis

The swim bouts estimated online during the experiments were analysed without additional preprocessing. In particular, we calculated trial-averaged cumulative turns around the exogenous scene rotations (Extended Data Fig. 1a,c,e–g), as well as comparing the biases of individual swim bouts with the scene orientation (Extended Data Fig. 1b,d).

Imaging data preprocessing

All imaging data were pre-processed using the suite2p package⁵¹. In brief, frames were iteratively aligned to reference frames randomly picked from the movie, using phase correlation. To detect ROIs representing cellular somata, a singular-value decomposition was performed on the aligned movie, and the ROIs were seeded from the peaks of the spatial singular-value vectors. All ROIs were used without morphological classifiers for cells, because the functional ROI selection procedure described below rejected spurious non-cell ROIs. In the ablation experiments, ROIs were manually defined, as described below.

ROI selection

For the HD neuron recordings, fluorescent time traces were first normalized into Z-scores for each ROI by subtracting the mean and dividing by the standard deviation. The normalized traces were then smoothed with a box-car kernel with the width of 1 s. Next, a scaled shifted sinusoid (atimes cos (theta -b)+c) was fit to the smoothed traces of each ROI, where (theta ) is the orientation of the visual scene relative to the fish. The fitting was performed with the ‘curve_fit’ function from the scipy package, and parameters were bounded in the range of (age 0,{b}in (,-,pi ,pi )). Only fractions of the data were used for this fitting to allow cross-validated quantifications of bump-to-scene alignments on the held-out fraction (described below). Specifically, either the second half of the first epoch (Figs. 1 and 3–5) or the entire first epoch (Fig. 2 and Extended Data Figs.

3 and 4) were used for fitting. ROIs with R2 larger than 0.15 were considered to be sufficiently modulated by the scene orientation and included. In addition, pairwise Pearson correlations of response time traces to repeatedly presented flashes (and translating dots in Extended Data Fig. 3) at the beginning of the experiments were calculated for each ROI, and averaged across all pairs of repetitions. Naively visual ROIs with mean pairwise correlations above 0.1 were excluded from further analyses. Finally, rectangular masks were manually drawn around rhombomere 1 of the aHB, and the ROIs outside the mask were excluded.