Visual recognition of social signals by a tectothalamic neural circuit – Nature

Animal care and transgenic zebrafish

Adult, juvenile and larval zebrafish (Danio rerio) were housed and handled according to standard procedures. All animal experiments were performed under the regulations of the Max Planck Society and the regional government of Upper Bavaria (Regierung von Oberbayern), approved protocols: ROB-55.2Vet-2532.Vet 03-15-16, ROB-55.2Vet-2532.Vet 02-16-31, and ROB55.2Vet-2532.Vet 02-16-122. Experimental animals were outcrosses to TL or TLN (nacre) unless otherwise noted. The following transgenic lines were used: Tg(elavl3:H2B-GCaMP6s)jf5⁴⁵, SAGFF(lf)81c (TeO Gal4 line)⁴⁶, Tg(UAS-E1B:NTR-mCherry)c264⁴⁷, Et(fos:Gal4-VP16)s1026t (DT Gal4 line)⁴⁸, Tg(UAS:BGi-epNTR-TagRFPT-utr.zb3)mpn420 (this study).

Larvae were raised in Danieau solution under a 14–10 h light–dark cycle at 28.5 °C until 6 d.p.f. For experiments in juveniles, animals were then raised under standard facility conditions at 28.5 °C in groups of 20–25 individuals. The fish were fed by feeding robots once a day with artemia and 2–3 times a day with dry food.

Shoaling assay and behaviour quantification

Shoaling with real and virtual conspecifics was assayed as previously described². In brief, 15 or 35 individual animals were transferred individually into shallow watch glass dishes of 10 cm or 7 cm diameter, respectively, separated by a grid of visual barriers and resting on a projection screen. Custom-written Bonsai⁴⁹ workflows were used to project stimuli to each animal and to track animal location at 30 fps. Stimuli were black dots on a white background moving along a predefined, synthetic-trefoil shaped trajectory at an average speed of 5 mm s⁻¹. For continuous motion, the stimulus position was updated 30 times per second. For bout-like motion, the stimulus position was updated once every 666 ms. Dot diameter was 2 mm unless noted otherwise, and 0 mm in the no-stimulus condition.

To assay shoaling of pairs of real conspecifics, we introduced a second animal in the same dish and did not show any projected stimuli. FastTrack⁵⁰ was used for post hoc tracking of real pair shoaling.

Attraction and neighbourhood maps were quantified as previously described² using custom-written Python software. We calculate the ‘real’ average interanimal distance or animal dot distance for each animal in 5 min chunks (IADr). Next, we generate 10 time-shifted trajectories and recalculate the shifted average inter animal or animal dot distance (IADs) for each time shift. Mean IADs for all time shifts are used to compute attraction as (IADs – IADr)/IADs.

For neighbourhood maps, neighbour position time series were transformed into the focal animal’s reference frame to compute a binned 2D histogram.

Repulsion was quantified as the reduction in attraction at the centre of each animal’s neighbour density map. Neighbour density maps were gaussian-filtered (sigma = 3 mm) before obtaining 24 radial line scans (width of 5 mm) starting from the centre of the map. Repulsion was the area above the average line scan, at radii less than the radius at which maximum neighbour density occurred (Extended Data Fig. 7a), divided by the full length of the scan (29 mm).

Looming stimuli were presented in the virtual shoaling setup⁴¹. Looming discs appeared once every minute at a defined offset of 5 mm to the left or the right from the current centre of mass of each animal. Looming discs expanded within 500 ms to the indicated final size and followed the animal. To compute an escape fraction, we defined an escape response as a trial in which the animal moved more than twice as far in a time window of 1 s immediately following the loom compared to the 1.3 s before. Bout duration was computed using peak detection on the velocity time series of each animal.

c-fos activity mapping

Shoaling assay for c-fos

For c-fos labelling, we used nacre;elavl3:H2B-GCaMP6s fish at 21 d.p.f. Thirty five fish were transferred into individual dishes and left without stimulation in the presence of white projector illumination from below for acclimatization and to establish a low, non-social c-fos baseline. Each animal was assigned randomly to one of the four stimulus groups. After 2 h, continuous or bout-like motion were shown to groups 1 and 2, respectively, whereas groups 3 and 4 continued to see no stimulus. After 45 min, groups 1, 2 and 3 were quickly euthanized and fixed. Four animals of group 4 were then transferred into the dishes of four other animals of this group for shoaling. After 45 min, these eight animals were euthanized and fixed as well.

HCR staining and imaging

Animals were euthanized and fixed on 4% ice cold paraformaldehyde (PFA). The PFA was washed out after 24 h with 1× PBS and the samples were gradually dehydrated and permeabilized with methanol and stored in −20 °C for several days until the HCR in situ labelling was performed. All of the HCR reagents were purchased from Molecular Instruments and the staining was performed according to the manufacturer’s protocol for whole-mount zebrafish larvae. In brief, the samples were separated into 2 juvenile fish per single 1.5 ml Eppendorf tube. Rehydration steps were performed by washing for 5 min each in 75% methanol/PBST (1× PBS + 0.1% Tween-20), 50% methanol/PBST, 25% methanol/PBST and finally five times with 100% PBST. The samples were permeabilized with 30 µg ml⁻¹ proteinase K for 45 min at room temperature, followed by postfix with 4% PFA for 20 min at room temperature and 5 washes in PBST for 5 min each. The samples were prehybridized in a 500 µl probe hybridization buffer (Molecular Instruments) for 30 min at 37 °C. Hybridization was performed by adding 2 pmol of each probe set to the hybridization buffer and incubating for 16 h at 37 °C. Probe sets for c-fos (B5 initiator), cort (B3 initiator) and elavl3 (B2 initiator) were purchased from and designed by Molecular Instruments. To remove the excess probes, the samples were then washed 4 times for 15 min each with a wash buffer (Molecular Instruments) at 37 °C, followed by 2 washes of 5 min each with 5× SSCT (5× SSC + 0.1% Tween-20) at room temperature. Pre-amplification was performed by incubating the samples for 30 min in an amplification buffer (Molecular Instruments) at room temperature. The fluorescently labelled hairpins (B2-488, B3-647, B5-546) were prepared by snap cooling: heating at 95 °C for 90 s and then cooling to room temperature for 30 min. Hairpin solution was prepared by adding 10 µl of the snapped-cooled hairpins (3 µM stock concentration) to a 500 µl amplification buffer. The pre-amplification buffer was removed, and the samples were incubated in the hairpin solution for 16 h at room temperature. The excess hairpins were washed three times with 5× SSCT for 20 min each wash, and the samples were stored in 5× SSCT in the dark at 4 °C until imaging.

For dorsal imaging, the samples were embedded in 2.5% low melting agarose in 1× PBS. Imaging was performed with a Leica SP8 confocal microscope equipped with a ×20 water-immersion objective. z-Stacks, comprising four tiles, covering of the entire brain were taken (final stitched image size: 1,950 px × 1,950 px, 1,406 µm × 1,406 µm, 3 µm in z). All 32 samples were imaged with the exact same laser power, gain, zoom, averaging and speed to faithfully quantify and compare the fluorescent signal between the samples. For ventral imaging, the samples were removed from the agarose and dissected to remove the jaw and the gills. After the dissections, the samples were embedded upside down and imaged in the same manner. Four brains were lost during ventral imaging and were therefore excluded entirely from the subsequent analysis.

Image registration

Image registration was performed using Advanced Normalization Tools (ANTs⁵¹) running on the MPCDF Draco/Raven Garching computing cluster. Before registration, stacks were batch-processed in ImageJ. Each stack was downsampled to 512 px width at the original aspect ratio using bilinear interpolation, split into individual channels and saved as .nrrd files. For ventral stacks, artefacts of the dissection such as left-over autofluorescent muscle fibres and skin were masked before registration. Initial attempts to register the elavl3 HCR channel of dorsal or ventral HCR confocal stacks to a live-imaged two-photon reference of elavl3-H2B-GCaMP6s expression were not successful, probably due to deformations resulting from the HCR protocol and diverse qualitative differences in image features between the imaging modalities. Instead, separate dorsal and ventral HCR registration templates were generated from scratch by running antsMultivariateTemplateConstruction2.sh on three manually selected stacks. Next, all dorsal and ventral stacks were registered to their respective templates using antsRegistration. Finally, the ventral template was registered to the dorsal template using affine + b-spline transformations via antsLandmarkBasedTransformInitializer with the help of 25 manually curated landmarks in each stack before applying standard antsRegistration. The resulting ventral-to-dorsal transform was then applied to re-register all ventral stacks into one common (dorsal) reference frame.

c-fos signal intensity quantification

Image analysis was performed using custom scripts in Python. Registered dorsal and ventral stacks were merged as the arithmetic mean intensity for each animal. To normalize to a drop in signal intensity with tissue depth, the c-fos signal was divided voxel-wise by the elavl3 HCR signal. For visualizations of imaging planes, the elavl3 signal used in normalization was filtered by a 3D gaussian (filter width: 55 µm, 55 µm, 15 µm x,y,z). For area-wise c-fos quantification, unfiltered elavl3 signal was used in normalization. To identify activity clusters, merged stacks from all animals per condition were generated by finding the maximum intensity at each voxel across animals. A combined RGB hyperstack was generated that showed c-fos signal for each condition, cort HCR and elavl3 HCR for reference in different colours for visual inspection. Activity clusters were manually drawn as 3D masks on the hyperstack using the ImageJ segmentation editor on orthogonal overlay views. Masks were drawn with the intent to outline prominent, distinct clusters of c-fos signal, irrespective of their modulation by social condition. The full hyperstack, including cluster masks is available. Brain areas housing the activity clusters were identified by comparison of the elavl3 reference to the mapzebrain atlas³⁰ and additional resources^13,52,53.

Individual cort– and c-fos-positive cells in DT were counted manually using the ImageJ cell counter plugin. For statistical analysis across activity clusters and conditions, bulk normalized c-fos signal was computed as the average intensity of all voxels belonging to a given cluster. Effect size was determined in each cluster for each condition versus the no-stimulus condition by pairwise computation of Cohen’s d defined as the difference of the means divided by the pooled standard deviation. To determine significant activity modulation compared to the no-stimulus condition, we performed repeated two-tailed t-tests and corrected for multiple comparisons in each family of tests (each activity cluster) using the Bonferroni correction. Hierarchical clustering of the activity clusters was performed on the effect sizes using the seaborn method clustermap with the default parameters for average Euclidean clustering.

Functional two-photon calcium imaging

Two-photon functional calcium imaging was performed on 6–8 d.p.f. larvae and 17–22 d.p.f. juvenile elavl3:H2B-GCaMP6s transgenic fish without paralysis or anaesthesia. The 6–8 d.p.f. larvae were embedded in 2% agarose with the tail freed as previously described³⁹. Juveniles (17–22 d.p.f.) were embedded in 3% agarose. As juvenile zebrafish are prone to hypoxia in this preparation, several precautions were taken. A drop of low-melting agarose was placed onto a petri dish and allowed to cool before a fish was introduced and oriented with a pipette tip. Once solidified, agarose was removed from the mouth, gills and tail using scalpels to restore active and passive breathing (Extended Data Fig. 3). Additional oxygen was supplied by continuously perfusing the dish. The perfusion medium consisted of fish water freshly oxygenated to saturation at the start of the experiment and diluted 1:1 with demineralized water to support ionoregulation and buffered with 1.2 mM NaH₂PO₄ and 23 mM NaHCO₃⁵⁴. To monitor health, we checked heartbeat and breathing movements of gills and mouth before and after an experiment. Only fish that were breathing and moving after the end of the experiment were included in the analysis. The embedded fish were mounted onto the stage of a modified two-photon moveable objective microscope (MOM, Sutter Instrument, with resonant-galvo scanhead) with a ×20 objective (Olympus XLUMPLFLN, NA 1.0) and imaged for at least 60 min. Typically, fish resumed swimming immediately after release from embedding. Only fish that did not drift up or down in their preparation were used for analysis. Fish in which no tectal responses could be observed were eliminated from the analysis. Fast volumetric imaging of the tectum and/or thalamus was performed using a custom-built remote focusing arm added before the microscope (Extended Data Fig. 3b). The remote focusing path was constructed using the 30 mm and 60 mm Thorlabs cage system, and consists of the following parts (in order of forwards traversal): a half-wave plate (Thorlabs, AHWP05M-980), a polarizing beam splitter (Thorlabs, PBS102), two lenses (Thorlabs, AC254-100-B-ML and AC508-200-B-ML), a quarter-wave plate (Thorlabs, AQWP10M-980), remote objective (Nikon, CFI ×16 0.8 NA), and a gold mirror (Thorlabs, PF05-03-M01) mounted onto a custom piezo (PPS-D08300-001 nanoFaktur, 300 µm closed loop range, with a nPoint LC.402 controller). The piezo was mounted on a xyz translation stage with tip-tilt control. Changing the mirror position is rapid (for the step sizes used for imaging, 1–2 ms) and results in a change of focus of the excitation beam exiting the main objective. Refocusing through the remote arm enabled rapid sequential imaging of 6 planes with a nonlinear step size ranging from 6–24 µm at 5 volumes per second. Remote focusing was not used for the high-resolution single-plane imaging in Fig. 2h,i. The plane size ranged from 370 µm × 370 µm for larvae to 1,075 µm × 1,075 µm for juveniles. Laser power ranged from 12.3 mW to 15.4 mW. The spatial sampling (0.7–2.1 µm px⁻¹) and optical resolution enabled discrimination of single cells with cell body diameters typically in the range of 5 µm to 8 µm.

z-Stack acquisition and image registration

For each functionally imaged fish, a z-stack of the entire brain was taken (512 × 512 or 1,024 × 1,024 pixels, 2 µm in z, 835–920 nm laser wavelength, plane averaging 50–100×) with the two-photon microscope. Larval data were registered to the mapzebrain atlas³⁰ using the elavl3:H2B-GCaMP6s reference. For juvenile data, a standard brain was generated from three high-quality z-stacks (150× frame averaging) as described in the ‘c-fos activity mapping’ section and each juvenile brain was registered to it. The generation of a standard brain and the parameters used for ANTs registration have been described in detail previously³⁰.

To align functional regions of interest (ROIs) from 2P data to a common reference frame, a two-step strategy was used. First, average frames of all imaging planes were registered to individual z-stacks using template matching. Converted ROI locations in z-stack coordinates were then transformed to the larval and juvenile common reference frames by running the ANTs command antsApplyTransformsToPoints with the matrices from the z-stack registrations.

Visual stimuli

Visual stimuli were designed using PsychoPy and projected by an LED projector (Texas Instruments, DLP Lightcrafter 4500, with 561 nm long-pass filter) on Rosco tough rolux 3000 diffusive paper placed into a petri dish filled with fish water.

Frequency tuning

A black dot moving on a circular trajectory (radius, 18 mm) with the fish head in the centre was shown starting either perpendicular to the fish at the left, or in front of the fish. The dot was moved in discrete jumps at 0.75, 1.5, 3.0, 6.0 or 60.0 Hz at an overall speed of 5 mm s⁻¹ (15.9 degrees (deg) s⁻¹). Each frequency was presented using a dot diameter of 4 mm (12.7 deg). Moreover, 1.5 Hz and 60.0 Hz stimuli were also presented with dot diameters of 2 mm (6.4 deg) and 8 mm (25.1 deg). Both clockwise and counter-clockwise presentations were shown. The frequency, direction and, if applicable, size were randomly drawn at each stimulus instance. Each stimulus had a duration of 22.6 s and was followed by a 20 s break. A total of 13 stimuli were shown per 10 min recording. Five to nine of these recordings were performed in each fish, leading to an average of four to six presentations of each stimulus. For Fig. 2b, only responses to dots (4 mm diameter) with 1.5 Hz and 60 Hz bout frequency were analysed. For Fig. 2g, again only responses to 4 mm dots were analysed.

Specificity

Naturalistic stimulus trajectories consisted of a 4 mm diameter dot (12.7 deg) moving along real trajectories from one of two interacting juvenile zebrafish that were previously recorded². The trajectory was computed as a fish-centric view of the conspecific with respect to a focal fish. To avoid noise in the heading calculation due to tracking jitter, the trajectory was convolved with a normalized hamming kernel (mode: valid, window length: 20). The naturalistic motion sequences were shown for 1 min each (Extended Data Fig. 5b). For the whole-field motion stimulus an image was created by combining random intensities and restricted spatial distributions in Fourier space, matching the size of the moving dot (Extended Data Fig. 5c). The computed image either rotated in discrete jumps of 1.5 Hz or continuously at 60 Hz. In both cases the stimulus took 22.6 s to finish a complete round. All stimuli, 1.5 Hz dot, 60 Hz dot, 1.5 Hz whole-field, 60.0 Hz whole-field and naturalistic dot motion were shown in a pseudo-random order during 6×10 min recordings.

Kinetic parameters

Presented 4 mm diameter (12.7 deg) dots moved clockwise on a circular trajectory (18 mm radius). Five speeds were tested using a continuously moving dot: 2.5, 5, 15, 50 and 150 mm s⁻¹ (8, 15.9, 47.7, 159.2 and 477.5 deg s⁻¹). Five speeds at a bout frequency of 1.5 Hz were tested by increasing the distance the dot moved during each bout. This increased both the average speed and the acceleration during bouts. The following parameters were tested: 1.25 mm s⁻¹; 3 m s⁻², 2.5 mm s⁻¹; 6 m s⁻², 5 mm s⁻¹; 12 m s⁻², 10 mm s⁻¹; 24 m s⁻², and 20 mm s⁻¹; 48 m s⁻² (4 deg s⁻¹; 9.5 × 10³ deg s⁻², 8 deg s⁻¹; 19.1 × 10³ deg s⁻², 15.9 deg s⁻¹; 38.2 × 10³ deg s⁻², 31.8 deg s⁻¹; 76.4 × 10³ deg s⁻² and 63.7 deg s⁻¹; 152.8 × 10³ deg s⁻²). Finally, for changing acceleration during each bout, we modelled each bout as a gaussian speed profile and changed the width of the curve. Each stimulus still had an average speed of 5 mm s⁻¹ (15.9 deg s⁻¹) through a normalization factor. The following peak accelerations were tested: 0.0, 0.02, 0.5, 2.0 and 12.0 m s⁻² (0, 0.06, 1.6, 6.3 and 38.2 × 10³ deg s⁻²).

Control stimuli after tectal ablation

Control stimuli consisted of translational gratings moving caudorostrally with respect to the fish (width, 20 mm; frequency, 0.12 Hz; duration, 20 s) and a looming stimulus (expansion from 0.6 deg to 110 deg in 83 ms, delay 10 s with disk and 20 s without stimulus) centred below the fish. One grating was shown at the beginning, followed by the dot stimuli, another grating and finally the looming stimulus. These recording sessions took 10 min each and were separated by a 1 min break to avoid potential habituation or response suppression due to the looming stimulus.

Data analysis for two-photon imaging

Suite2P⁵⁵ was used for motion correction, ROI detection, cell classification and signal extraction. For the entire analysis, a GCaMP6s time-constant of 7 s was used to accommodate the slow kinetics partially due to the nuclear localization of this sensor. On the basis of a visual inspection of the raw data, a cell diameter of 4–6 px was used. In detail, raw recording files were deinterleaved into separate time series for each plane. An extra motion-correction step was required owing to ripple noise stemming from the resonant mirror: to avoid spurious alignment to the noise pattern, rigid and non-rigid motion correction was performed on a spatially low-pass filtered time series (Gaussian, sigma = 4). The resulting motion-correction parameters were applied to the raw data. Next, the time series were downsampled fivefold to one volume per second. On the downsampled data, ROIs were detected and fluorescent traces were extracted.

Thresholding

Neuron ROIs were thresholded in a two-step process. First, the built-in Suite2p classification algorithm iscell was applied using the default parameters. Second, iscell⁺ ROIs that showed a mean ∆F/F response to any stimulus above the 95th percentile (Fig. 2)/90th percentile (Fig. 4).

Mean ∆F/F responses

For each functional ROI, the fluorescent trace was normalized and split into stimulus episodes. ∆F/F was computed by using the 5 s before stimulus onset as the baseline. ∆F/F temporal responses were averaged across stimulus presentations per stimulus and then averaged over time to receive one value per stimulus.

BPI

On the basis of the behavioural tuning curves to bout frequency², stimuli were split into bout-like (0.75–3 Hz) and continuous (6–60 Hz) categories, regardless of stimulus size or directionality. BPI was defined as the difference in mean over mean ∆F/F to bout-like stimuli and mean over mean ∆F/F to continuous stimuli divided by their sum (equation (1)). BPNs were considered all ROIs that scored BPI >0.5, which equates to a threefold higher bout response.

Tuning peaks

For computing peaks in the tuning of neurons to a variable, mean ∆F/F responses were interpolated with a one-dimensional spline (scipy.interpolate.InterpolatedUnivariateSpline, k = 2, second degree) and the location of the maximum was computed.

Gaussian KDE

To generate a kernel density estimate of BPNs in anatomical space, BPN coordinates were used to fit a Gaussian Kernel (sklearn.neighbors.KernelDensity(*, bandwidth = 10 (14 for 7 d.p.f.), algorithm=‘auto’, kernel=‘gaussian’, metric=‘euclidean’). In detail, the brain was divided along the rostrocaudal axis and, for each hemisphere, a separate kernel was fitted with the contained BPNs. The resulting two kernels were used to generate probability density fields of each hemisphere, which were then merged again. The resulting density was thresholded so that only voxels within the brain itself had values > 0 and all voxels in the volume surrounding the brain equalled 0. Probability values were then normalized so that the sum would result in the total number of BPNs. To draw contours of areas with certain threshold BPN density, the KDE volume was binarized so that all voxels above threshold equalled 1. Of the resulting binarized volume a two-dimensional maximum intensity projection was computed for each orthogonal anatomical axis and a contour-finding algorithm (skimage.measure.find_contours) was applied to the two-dimensional projection.

Definition of the larval DT

The outline of the larval thalamus proper was refined with expert help of M. Wullimann (LMU). The refinement was based on extensive analysis of gene expression⁵². The elavl3 reference stain was used to identify the diencephalic regions. Proliferative cells, however, which are abundant in the anterior DT at the larval stage, are not labelled by elavl3. The neurogenin line was used to indicate the early glutamatergic cells belonging to DT. Neurogenin is absent in the prethalamus (VT). The VT/DT boundary was further defined using gad1b and dlx4, which label late and early GABAergic cells, respectively. GABAergic cells are mainly found in VT, although the intercalated nucleus and the anterior nucleus of DT may contain some gad1b positive cells. The pretectum/DT boundary was defined using gad1b and th. The latter marks dopamine cells present in the pretectum.

EM and segmentation of mapzebrain regions

A detailed description of the EM dataset and region mapping is published elsewhere³⁵. In brief, the Serial Block Face Scanning EM dataset was of a 5 d.p.f. larval zebrafish imaged at a resolution of 14 × 14 × 25 nm. A diffeomorphic mapping between the mapzebrain light-microscopy brain reference coordinate system and the EM coordinate system generated by the dipy (https://dipy.org/) Python library was used to overlay mapzebrain (https://mapzebrain.org) region annotations over the EM data. Registration accuracy was reviewed for different brain regions with an alignment error of maximal ~5 µm (midbrain) to ~20 µm (hindbrain). We applied flood-filling networks for an automated reconstruction of all neurons⁵⁶ within the whole-brain EM dataset³⁵. To correct for split and merge errors of the segmentation, we used the Knossos application (www.knossos.app). Proof-reading of single pBPN-DT cells started at the cell body location and ended when all branches were completely traced. Growth cones defined premature neurons. Proof-reading of partner cells started at the synapse and was again performed until the whole cell was completed. Synapses have been automatically segmented using the SyConn v2 pipeline³⁵. Input to pBPN-DT cells and tectal PVPNs was quantified by randomly selecting ten incoming synapses per cell and tracing input partner cells, until their cell bodies were identified.

Nitroreductase ablations

To chemogenetically ablate neurons, fish expressing nitroreductase (NTR) were treated with 7.5 mM metronidazole (MTZ) in Danieau’s solution with 1 ml l⁻¹ dimethylsulfoxid (DMSO) for larvae, or 7.5 mM MTZ in fish facility water with 2 ml l⁻¹ DMSO for juvenile fish, respectively. Control fish were only treated with the respective DMSO concentration in the absence of MTZ. For SAGFF(lf)81c ablations, the canonical UAS–E1B:NTR-mCherry(c264) nitroreductase was used⁴⁷.

Transgenic animals (RFP⁺) and control sibling fish (RFP⁻) were incubated in MTZ + DMSO solution for 16–24 h overnight. For the experiment shown in Extended Data Fig. 7b,e, we additionally incubated RFP⁺ and RFP⁻ animals in DMSO only. Animals recovered for 16–24 h in system water before starting imaging or behaviour experiments.

In larvae, s1026tEt drives strong expression in the DT, dorsal VT, ventral pretectum and ventral telencephalon, and additional background expression in the heart and trunk musculature^31,48. We found that MTZ mediated ablation of s1026tEt, UAS-E1B:NTR-mCherry(c264) double transgenic fish was lethal in 100% of animals. We therefore used a nitroreductase transgene of which the background muscle expression was suppressed by a 3′UTR: UAS:BGi-epNTR-TagRFPT-utr.zb3⁵⁷. To overcome strong variegation of the existing allele (y362) in s1026tEt cells, we created new alleles via Tol2 mediated transgenesis. We injected UAS:BGi-epNTR-TagRFPT-utr.zb3 together with Tol2 mRNA into the TLN s1026tEt background. We outcrossed individual founders to TLN and raised transgenic RFP⁺ offspring and control RFP⁻ siblings. At 19 d.p.f., we selected transgenic offspring of one founder for homogeneous RFP signal in the DT. This founder established the allele mpn420 of which the expression is largely confined to the DT (Extended Data Fig. 8). Half of transgenic and control animals were randomly assigned to MTZ + DMSO versus DMSO-only treatment. We observed no lethality after ablation in this line. Analogous, non-transgenic siblings were also split into two groups for MTZ + DMSO versus DMSO-only treatment. MTZ had no detectable effect on non-transgenic animals and we subsequently pooled all three control conditions.

Laser ablations

Juvenile (20–23 d.p.f.) elavl3:H2B-GCaMP6s transgenic fish were embedded as described above, anesthetized with Tricaine and placed under a two-photon laser scanning microscope (Femtonics). The DT-BPN region within each brain was visually identified by the experimenter. A 20 μm ROI was specified on the DT-BPN region of each brain hemisphere and scanned with a 800 nm/400 mW laser beam for 30 ms. After each scan, one image was captured to observe the resulting damage and potential off-target effects. This procedure was repeated 7 ± 3 times until no more nuclei could be observed in the target region that could be targeted without blood vessel damage. Fish were removed from the embedding and anaesthesia immediately after ablation. One fish that did not start swimming within a few minutes was excluded from subsequent behaviour testing and analysis. Fish were allowed to recover for at least 16 h before they were tested in a shoaling assay as described above.

Social isolation

Individual eggs from elav:H2B-GCaMP6s incrosses were placed into small petri dishes at 0 d.p.f. The side walls of the dishes were taped to prevent visual contact between dishes. For controls, 15 eggs were placed in one small petri dish. Larvae were imaged at 7 d.p.f. As no brain stacks were acquired for these animals, anatomical DT masks were drawn for each fish manually. The response threshold for data analysis was adjusted to 50% due to a lower number of recorded neurons in single-plane imaging.

Statistical analysis

All analyses were performed in Python, using NumPy, Scipy, MatplotLib, Suite2p, Pandas and Scikit-learn. All statistical details are described in the figure legends and the Methods. All tests were two-tailed, unless noted otherwise. Error bars represent 1 s.d., unless noted otherwise. N denotes the number of animals, unless noted otherwise.

Data collection software

The following data collection software were used: Bonsai (v.2.4.1); Leica LAS X (v.3.5.7); and ScanImage (v.5.6).

Data analysis software

The following data analysis software was used: Python (v.3.9) with NumPy (v.1.21.0), Scipy (v.1.7.0), MatplotLib (v.3.4.2), Pandas (v.1.3.0) and additional packages (full python environments are available with our code on bitbucket); Ants (v.1.9); Suite2p (v.0.9.3); and ImageJ (v.1.53c).

Reporting summary

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