The encoding of light increments and decrements by separate On- and Off- systems is a fundamental ingredient of vision, which supports edge detection and makes efficient use of the limited dynamic range of visual neurons. Theory predicts that the neural representation of On- and Off-signals should be balanced, including across an animals’ visible spectrum. Here we find that larval zebrafish violate this textbook expectation: in the zebrafish brain, UV-stimulation near exclusively gives On-responses, blue/green stimulation mostly Off-responses, and red-light alone elicits approximately balanced On- and Off-responses. We link these findings to zebrafish visual ecology, and suggest that the observed spectral tuning boosts the encoding of object “colourfulness”, which correlates with object proximity in their underwater world.

Animals. All procedures were performed in accordance with the UK Animals (Scientific Procedures) act 1986 and approved by the animal welfare committee of the University of Sussex. For all experiments, we used 6-7 days post fertilization (dpf) zebrafish (Danio rerio) larvae. The following previously published transgenic line was used: Tg(elavl3:H2B-GCaMP6f); ZFIN ZDB-ALT-150916-4. Animals were housed under a standard 14:10 day/night rhythm and fed three times a day. For 2-photon in-vivo imaging, zebrafish larvae were immobilised in 2% low melting point agarose (Fisher Scientific, BP1360-100), placed on a glass coverslip and submerged in fish water.

Light Stimulation. With fish mounted upright, light stimulation was delivered as wide-field flashes from a spectrally broad liquid waveguide with a low NA (0.59, 77555 Newport), positioned next to the objective at ~45˚. The other end of the waveguide collected light from 13 “spectrally narrowed” LEDs, as described in detail elsewhere. All stimuli were series of single LED flashes of light lasting 3 s, separated by gaps of 3 s (1 stimulus loop: 13 LEDs * (3+3) s = 78 s. 3-4 loops were presented and averaged for each recording.

2-photon calcium imaging. All 2-photon imaging was performed on a MOM-type 2-photon microscope (designed by W. Denk, MPI, Martinsried; purchased through Sutter Instruments/Science Products) equipped with a mode-locked Ti:Sapphire laser (Chameleon Vision-S, Coherent) tuned to 960 nm for SyGCaMP imaging. We used one fluorescence detection channel (F48x573, AHF/Chroma), and a water immersion objective (W Plan-Apochromat 20x/1,0 DIC M27, Zeiss). For image acquisition, we used custom-written software (ScanM, by M. Mueller, MPI, Martinsried and T. Euler, CIN, Tuebingen) running under IGOR pro 6.3 for Windows (Wavemetrics).

To expand the field of view to ~1.2 mm diameter, which allowed capturing the entire brain’s length in a single scan, we used a non-telecentric optical approach as described in detail elsewhere. The excitation spot (point spread function) in this configuration was ~0.7 µm (xy) and ~11 µm (z) at full width half maximum. This optical configuration can in principle capture the signals from individual larval zebrafish somata. However, in this work it was our intention to capture the bulk spectral responses across large fractions of the brain. Accordingly, we balanced recording area and spatial sampling such that individual somata effectively corresponded to single, or at most groups of 2-4 pixels (3 planes covering ~450x1,000 µm with a 160x350 px scan each to yield ~2.9 µm voxel xy-spacing, compared to average zebrafish neuronal soma diameter of ~7 µm; 1 ms per line, 2.08 Hz volume rate).

To follow the brain’s natural 3D curvature, we also systematically 3D-bent each scan-plane as a function of the slow scanning-mirror’s position to form a “half-pipe”. Curvature was achieved via rapid remote focussing synchronised with the scan pattern, as described in detail elsewhere. The degree of peak axial curvature was empirically adjusted between 0-150 µm between scans and planes to achieve best overall sampling of the entire brain.

Pre-processing and extraction of response amplitudes of 2-photon data. Recordings were linearly interpolated to 42 Hz and manually aligned between fish using a time-averaged brightness projection. Regions of interest (ROIs), corresponding to individual and/or small groups of neighbouring neuronal somata were defined automatically using custom Python scripts. In short, we used a “quality-index” (QI, described in detail elsewhere [S4]) to first identify individual pixels that exhibited reliable responses to repeated stimulation. For this, we computed a pixel-wise QI-projection of the deinterleaved recording, sorting QI-pixels in descending order. The resulting curve was differentiated using scipy.interpolate.splrep. Pixel indices between inflections of the differential were projected back into space. Contours were identified using dilation (3,3)-erosion(2,2) and contour finding of Python-OpenCV. Individual contours were taken as ROIs, discarding any ROIs with a diameter > 15 µm. QI per ROI was then recalculated and used for further thresholding at QI>0.5. From here, fluorescence traces were extracted and z-normalized based on the 6 s at the beginning of recording prior to stimulus presentation. Overall, this strategy served to balance the need to combine multiple pixels into ROIs to boost their signal-to-noise, with a goal of keeping ROIs as small and localised as possible to approximately report the signals single, or from at most very small groups of somata that responded in a similar manner. This compromise was necessary to accommodate the large size of the scan pattern capturing the entire length of the brain while also maintaining a reasonable imaging rate. A stimulus time marker embedded in the recording data served to align the traces relative to the visual stimulus with a temporal precision of 1 ms.

Separation of On- and Off responses. Calcium traces were deconvolved using ARMA(1) (caiman.source_extraction.cnmf.deconvolution). Inferred discrete events were partitioned into events occurring during stimulus presentation and the complement.

Computing the brain’s bulk spectral tuning functions. Inferred events were summed over respective stimulus time windows. Sums were averaged over all recorded traces. Contrast between On and Off portions of the response was calculated as their difference over their sum.

Natural Imaging Data Analysis. Hyperspectral data were obtained from Nevala and Baden (2018) and element-wise multiplied with a deuterium light source derived correction curve. The data were restricted to the domain of 360-650 nm. Here, the long-wavelength end of the domain was decided based on the long-wavelength opsin absorption curve; the short-wavelength end was dictated by the sensitivity of the spectrometer. Spectra were scaled by standard deviation within a given scene. Traces were multiplied with the respective On- and Off-filters. The responses were summed within spectrum to produce a single number per point spectrum (or 800-long vector per scan). These vectors were standard-deviation-scaled within a scene. Spatial projections of filter responses were Gaussian-smoothed in space (σ=2px).

Bartel_etal_2020.csv:
'Address',
'Filename',
'RoiNum',
'Contour', - contour of the ROI; in pixel indices
'SN', - the quality index of the ROI
'Zoom', - software zoom value
'CentreXY', - centre of the contour
'Trace', - calcium trace
'XYZ', - the objective position
'Subplane',
'Plane',
'Fish',
'UniqueID',
'Alignment',
'Xal',
'Yal',
'AdjustedXY',
'OnDec', - the On portion of the inferred event trace
'OffDec', - the Off portion of the inferred event trace
'Cluster' - clustering, to see major differences between ROIs