AZBA: A 3D Adult Zebrafish Brain Atlas for the Digital Age

Zebrafish have made significant contributions to our understanding of the vertebrate brain and the neural basis of behavior, earning a place as one of the most widely used model organisms in neuroscience. Their appeal arises from the marriage of low cost, early life transparency, and ease of genetic manipulation with a behavioral repertoire that becomes more sophisticated as animals transition from larvae to adults. To further enhance the use of adult zebrafish, we created the first fully segmented three-dimensional digital adult zebrafish brain atlas (AZBA). AZBA was built by combining tissue clearing, light-sheet fluorescence microscopy, and three-dimensional image registration of nuclear and antibody stains. These images were used to guide segmentation of the atlas into over 200 neuroanatomical regions comprising the entirety of the adult zebrafish brain. As an open source, online azba.wayne.edu), updatable digital resource, AZBA will significantly enhance the use of adult zebrafish in furthering our understanding of vertebrate brain function in both health and disease.


Introduction 57
Uncovering general principles of neuroanatomical function and brain-behavior 58 relationships requires the integration of findings across model organisms that range in 59 complexity, organization, and accessibility (Brenowitz and Zakon, 2015;Marder, 2002;Yartsev, 60 2017). Amongst vertebrate model organisms in neuroscience, zebrafish are relative newcomers 61 that have grown in popularity in recent years (Kenney, 2020;Orger and de Polavieja, 2017). 62 Originally established as a model organism for developmental biology due to ease of 63 domestication, high fecundity, and early life transparency (Parichy, 2015), the increased 64 popularity of zebrafish is driven by recent advancements in brain imaging, molecular genetic 65 manipulation, and behavior. To further enhance the use of zebrafish as an animal model in 66 neuroscience, we created a digital three-dimensional brain atlas (AZBA: adult zebrafish brain 67 atlas). Although several digital atlases exist for larval zebrafish (Kunst et  repertoire that includes long-term associative memory, complex social interactions, and goal-72 driven behaviors (Gerlai, 2016;Kalueff et al., 2013;Kenney et al., 2017;Nakajo et al., 2020). 73 Three-dimensional digital brain atlases are essential tools for modern neuroscience 74 because they facilitate lines of inquiry that are not possible with two-dimensional book-based 75 atlases. For example, visualization of the three-dimensional structure of the brain and 76 incorporation of new data or discoveries are difficult, if not impossible, with a print atlas. In 77 contrast, digital atlases enable exploration of brain structures in any arbitrary three-dimensional 78 perspective and can be readily updated to incorporate new information such as patterns of gene 79 expression and anatomical connectivity, as has been done for the mouse (Wang et al., 2020). 80 Such features are important for fields like neurodevelopment and comparative neuroanatomy 81 neurochemical organization of the brain, images from ten different antibody stains were also 107 registered into the same anatomical space. Finally, we performed manual segmentation, 108 delineating the atlas into over 200 neuroanatomical regions, including nuclei and white matter 109

tracts. 110
Taken together, AZBA is the most comprehensive, detailed, and up to date atlas of the 111 adult zebrafish brain. We have made all averaged images freely available 112 (https://doi.org/10.5061/dryad.dfn2z351g; azba.wayne.edu) to enable their use in exploring the 113 organization of the zebrafish brain and automated segmentation for activity mapping. By 114 generating this resource using readily available techniques, AZBA can be continuously updated 115 to reflect the latest findings in zebrafish neuroanatomy. We anticipate this becoming an 116 indispensable resource as adult zebrafish continue to gain traction as a model organism in 117 understanding the intricacies of the vertebrate brain. To create an averaged three-dimensional atlas, we developed a sample preparation and 131 analysis pipeline for whole-mount 3D image acquisition and registration (Figure 1). We used a 132 whole-mount preparation to avoid issues with slice-based techniques such as tissue loss, 133 tearing, and distortion. To circumvent the challenge of tissue opacity, we used a rapid organic 134 solvent based tissue clearing technique, iDISCO+ (Renier et al., 2016), that renders brains 135 optically transparent. Because conventional microscopic techniques are not suitable for large 136 volume imaging, we used light-sheet microscopy. Image stacks from individual fish brains were 137 converted to 3D volumes and registered into the same anatomical space prior to averaging. 138 Finally, average 3D images were manually segmented into their constituent brain regions. 139

140
Light-sheet Imaging 141 Tissue clearing using iDISCO+ resulted in transparent brains (Figure 2A). iDISCO+ is 142 compatible with a variety of stains, such as a nuclear stain (TO-PRO), that allowed us to 143 approximate the Nissl stain in the print atlas (Wulliman et al., 1996). Cleared and stained brains 144 were imaged in the horizontal plane with an in-plane resolution of 3.25 μm and an axial step-145 size of 4 μm ( Figure 2B) yielding near-isotropic signals at sufficient resolution to clearly 146 distinguish regional boundaries. From this collection of images, we generated three-dimensional 147 volumes ( Figure 2C) that enabled viewing at any arbitrary angle including the coronal and 148 sagittal planes ( Figure 2D). Images were subject to quality-control so that those damaged 149 during dissection were discarded. We retained seventeen sets of nuclear stained and 150 associated autofluorescence images from both male and female fish (8 female), each of which 151 were transformed into 3D volumes for registration. 152

3D Image Registration 154
The atlas was generated by registering images from individual animals into the same 155 space, thereby creating an anatomical average. This approach has been previously used in and an image registration pipeline toolkit (Friedel et al., 2014) to perform iterative registration to 160 generate a consensus image (Figure 3). This method begins with a 6-parameter linear 161 registration to rotate and translate the initial image dataset followed by a 12-parameter affine 162 registration to scale, translate, rotate, and shear the dataset with a pair-wise approach to avoid 163 bias by outlier images ( Figure 3A). Lastly, an iterative non-linear registration with 6 iterations at 164 subsequently higher resolutions was performed using minctracc (Collins and Evans, 1997). This 165 resulted in a set of linear and non-linear transformations for each TO-PRO image in our dataset 166 from native space to a consensus space and orientation. These transformations were then 167 applied to corresponding autofluorescence images, thereby creating an atlas with averaged 168 images containing TO-PRO and autofluorescence signals ( Figure 3B). 169 170

Antibody stains 171
To provide additional guidance for segmentation and generate insight into the 172 neurochemical organization of the adult zebrafish brain, we also acquired images using ten 173 different antibody stains ( Figure 4A and Table S1). We sought stains that would identify different 174 cell types in the brain, such as neurons (HuC/D), radial glial cells (glial fibrillary associated 175 protein; GFAP), and proliferating cells (proliferating cell nuclear antigen; PCNA), markers for different neurotransmitters (tyrosine hydroxylase (TH), 5-hydroxytryptamine (5-HT), and choline 177 acetyltransferase (ChAT)) and calcium binding proteins (parvalbumin (PV), calbindin, and 178 calretinin). Some of these stains, such as TH, 5-HT, ChAT, and calretinin have already been 179 subject to brain-wide analysis, making them useful for guiding segmentation. 180 Fully realizing the utility of different stains requires images to be brought into the same 181 anatomical space as the previously generated TO-PRO average. To achieve this, during 182 imaging of antibody stains we also acquired autofluorescence images, thereby providing a 183 bridge between the antibody images and the TO-PRO images. Autofluorescence images from 184 antibody stains were registered with the autofluorescence channel of the TO-PRO images, 185 yielding a set of transformations that were used to bring antibody stains into the same 186 anatomical space as the TO-PRO stain ( Figure 4B). To generate a representative image for 187 each antibody, we averaged together at least five independent brains. Our approach resulted in 188 strong correspondence between antibody images and the TO-PRO stain ( Figure 4C). The utility 189 of this approach is apparent from examining structures known to express high levels of specific 190 enzymes, like TH in the locus coeruleus ( Figure 4D; green arrows). 191

Registration Precision 193
To compare registration precision using TO-PRO and autofluorescence signals we 194 labelled six different landmarks in the atlas images and corresponding points in acquired 195 images. Transforms derived from the registration process were then applied to the acquired 196 image's labelled landmarks. We then measured the Euclidean distance between the 197 transformed points and the points in the atlas. Using a mixed model ANOVA (2 × 16; signal 198 (between subject) × landmark (within subject)), we found a main effect of signal (F(1,13) = 1084, 199 P = 6.6 × 10 -14 ) with the TO-PRO signal having greater mean precision (15 ± 10 µm vs 99 ± 53 200 µm; Figure S1). However, there was also a significant effect of landmark (F(5,65) = 98, P < 2 × 201 10 -16 ), and an interaction between signal and landmark (F (5, 65) = 120, P < 2 × 10 -16 ). A closer 202 examination of the data revealed that in the autofluorescence images the average precision of 203 each landmark covers a much wider range (17 to 180 µm) than TO-PRO (9 to 23 µm). In the 204 autofluorescence image, the landmark with the highest precision (point 5; 17 ± 4 µm) has 205 precision on par with the TO-PRO average. This suggests that the larger error measured using 206 autofluorescence images is most likely due to experimenter error in selecting points, reflecting 207 the paucity of well-defined landmarks in this signal compared to the richer, high contrast TO-208 PRO images. Registered images were used to segment the brain into its constituent parts (Figure 5; 212 see Table S2 for anatomical abbreviations and colors). Segmentation was primarily guided by 213 the seminal atlas of Wullimann and colleagues (1996). Regional boundaries and terminology 214 were updated for parts of the brain that have been subject to more recent analysis such as the 215 telencephalon, hypothalamic regions, and motor nuclei (Mueller et al., 2004;Porter and Mueller, 216 2020; Rink and Wullimann, 2001). Segmenting large, clearly delineated regions, such as the 217 optic tectum and the cerebellum, was straightforward ( Figure 5A). Small nuclei that only appear 218 in one or two slices in the atlas or in images from only one axis proved more challenging. For 219 such regions, we primarily made use of the coronal axis due to it being the most extensively 220 represented in both the atlas and the literature ( Figure 5A, bottom). The horizontal and sagittal 221 planes enabled us to identify the anterior-posterior boundaries ( Figure 5A, top and middle). To 222 ensure we captured as many neuronal structures as possible, we also made extensive use of a 223 neuronal marker (HuC/D) in conjunction with the nuclear stain, which allowed us to safely 224 identify many boundaries ( Figure S2A). Other challenges included the fact that the original brain 225 atlas contains a significant amount of unsegmented space. We labelled these regions as 226 "unknown" and according to their lowest anatomical level (e.g. unknown ventral telencephalon 227 (UnkVT), unknown diencephalon (UnkD), etc.). Explicitly labelling these regions distinguishes 228 them from the "clear" label that is used for areas outside the brain to facilitate computational 229 analysis using the atlas. Identification of tracts was largely based on a combination of 230 autofluorescence and lack nuclear and neuronal staining since we were unsuccessful in finding 231 a white matter stain compatible with iDISCO+ (e.g. the MLF: Figure 4D, pink arrowheads). We 232 anticipate future work will aid in filling these unsegmented regions with the potential to discover 233 new neuronal circuits and anatomical structures. Our segmentation resulted in a three-dimensional model of the zebrafish brain that can 255 be viewed from any arbitrary angle ( Figure 5D). Each nucleus, white matter tract, ventricle, and 256 anatomical space was given a unique abbreviation and color, totaling 204 regions (Table S2) 257 and associated with an anatomical hierarchy (Table S3, File S1). The full extent of the atlas can

Sex and brain volume 265
Because the goal of the atlas was to generate a representative brain, we combined 266 images from both male and female fish. To determine if there was an effect of sex on brain 267 volumes in our TO-PRO image set, we used a mixed model ANOVA (2 × 204; sex (between 268 subjects) × brain region (within subjects)). We found neither an effect of sex (F(1,15) = 0.11, P = 269 0.75; Figure S3) nor an interaction between sex and brain region (F(199, 2985) = 0.14, P = 1; 270 Figure S4) suggesting that sex did not affect the overall size of the brain or individual regions. 271 We did find a main effect of region (F(199, 2985) = 497, P < 2 x 10 -16 ), consistent with the wide 272 range of region sizes observed across the brain (0.000026 to 0.53 mm 3 ; Table S3). 273

Neurochemical organization of the adult zebrafish brain 275
We used AZBA to generate insight into the neurochemical organization of the adult 276 zebrafish brain using antibody stains that have not previously been subject to brain-wide 277 examination. Parvalbumin (PV) is a calcium binding protein that labels a class of inhibitory 278 interneurons (Celio, 1986). We found several highly concentrated areas of PV staining such as  Calbindin is a calcium binding protein important for regulating intracellular signaling that 301 is often used in comparative neurological studies (Schmidt, 2012). We found calbindin to be 302 concentrated in the fiber layers of the olfactory bulbs ( Figure S2D-a), the Dm in the 303 telencephalon ( Figure S2D   In the present article, we introduce a new resource for the zebrafish community: AZBA, a 326 three-dimensional adult zebrafish brain atlas that can be downloaded 327 (https://doi.org/10.5061/dryad.dfn2z351g) or explored on the web (azba.wayne.edu). This 328 resource will facilitate a wide variety of neurobiological studies using adult zebrafish aimed at 329 dissecting neural circuits of behavior, understanding brain pathology, and discovering novel and 330 conserved neuroanatomy. We created AZBA by leveraging advances in tissue clearing, light-331 sheet fluorescent microscopy, and image registration, resulting in the most detailed atlas for 332 adult zebrafish to date. Tissue clearing allowed us to take a whole-mount approach, overcoming 333 the natural opacity of the adult brain and issues associated with slice-based techniques such as 334 tissue loss, tearing, and distortion. Laser fluorescence light-sheet microscopy was used to 335 image the large volume of the zebrafish brain with high resolution and minimal photobleaching. 336 Finally, we used three-dimensional image registration to create images derived from multiple 337 animals and inclusion of ten antibody stains into the same anatomical space. These were then 338 used to guide segmentation of the atlas into over 200 different neuroanatomical regions. 339 AZBA represents a significant departure from two prior adult zebrafish brain atlases. The 340 seminal book atlas from Wullimann and colleagues (1996) is exceptionally detailed and has 341 guided zebrafish neuroscience research for over two decades. However, being in print, it has 342 not been updated with the latest findings and its two-dimensional visual presentation and lack of 343 chemoarchitectural markers makes identifying regional boundaries across anatomical planes  We view the current segmentation and image collection that comprises AZBA as a first 397 version that will be continually updated. To facilitate updating, and encourage input from the 398 scientific community, we have created a website (azba.wayne.edu) where we invite comments 399 and suggestions for updates. In addition, we expect future work will incorporate more antibody 400 images and in situ hybridization probes for understanding how patterns of protein and gene 401 expression vary across the brain. Through collaboration with the zebrafish community, we plan 402 to incorporate the wealth of Gal4 and Cre/loxP lines that have been generated to characterize 403 expression patterns in the adult brain. We envision a process where scientists send fixed brain 404 samples to a central lab for tissue clearing, imaging, and registration to the atlas for 405 incorporation into our online resource. A similar approach has been taken with larval fish 406

Sample preparation 480
Zebrafish were euthanized by anesthetizing in 4% tricaine followed by immersion in ice 481 cold water for five minutes. Animals were then decapitated using a razor blade and heads were 482 placed in ice cold PBS for five minutes to let blood drain. Heads were then fixed in 4% PFA 483 overnight after which brains were then carefully dissected into cold PBS and stored at 4 C until 484 processing for iDISCO+. Brains that were damaged during the dissection process were not 485 used for generating the atlas. 486 487

Tissue staining 488
Tissue staining and clearing was performed using iDISCO+ (Renier et al., 2016). 489 Samples were first washed three times in PBS at room temperature, followed by dehydration in 490 a series of methanol/water mixtures (an hour each in 20%, 40%, 60%, 80%, 100% methanol). 491 Samples were further washed in 100% methanol, chilled on ice, and then incubated in 5% 492 hydrogen peroxide in methanol overnight at 4 C. The next day, samples were rehydrated in a 493 methanol/water series at room temperature (80%, 60%, 40%, 20% methanol) followed by a PBS 494 wash and two one-hour washes in PTx.2 (PBS with 0.2% TritonX-100). Samples were then washed overnight at 37 C in permeabilization solution (PBS with 0.2% TritonX-100, 0.3 M 496 glycine, 20% DMSO) followed by an overnight incubation at 37 C in blocking solution (PBS with 497 0.2% TritionX-100, 6% normal donkey serum, and 10% DMSO). Samples were then labelled 498 with TO-PRO3 iodide (TO-PRO) (1 night) or primary antibodies (2-3 nights) via incubation at 37 499 C in PTwH (PTx.2 with 10 µg/mL heparin) with 3% donkey serum and 5% DMSO. Samples 500 were then washed at 37 C for one day with five changes of PTwH. Antibody stained samples 501 were followed by incubation with secondary antibodies at 37 C for 2-3 days in PTwH with 3% 502 donkey serum. For samples labelled with TO-PRO, the secondary antibody labelling step was 503 omitted. Following secondary antibody labelling, samples were again washed at 37 C in PTwH 504 for one day with five solution changes. 505 506

Tissue clearing 507
Labelled brains were first dehydrated in a series of methanol water mixtures at room 508 temperature (an hour each in 20%, 40%, 60%, 80%, 100% (x2) methanol) and then left 509 overnight in 100% methanol. Samples were then incubated at room temperature in 66% 510 dichloromethane in methanol for three hours followed by two 15-minute washes in 511 dicholormethane. After removal of dichloromethane, samples were incubated and stored in 512 dibenzyl ether until imaging. 513 514 Imaging 515 All imaging was done on a LaVision ultramicroscope I. Samples were mounted using an 516 ultraviolet curing resin (adhesive 61 from Norland Optical, Cranbury, NJ) that had a refractive 517 index (1.56) that matched the imaging solution, dibenzyl ether. Images were acquired in the 518 horizontal plane at 4X magnification. 519

Image processing 521
Data sets from light sheet imaging were stitched using Fiji's (NIH) extension for Grid 522 Stitching (Preibisch et al., 2009) and converted to a single stack, corresponding to the z-axis. All 523 image processing steps were run on a Linux-workstation with 64 GB of RAM and 12-core Intel 524

processor. 525
Each stack was converted to a 4 µm isotropic image using custom python code with 526 separate files for the autofluorescence channel and a second for the antibody or TO-PRO 527 channels. These images were resampled to 8 µm isotropic due to system constraints during the 528 image registration stages. 529 530

Registration 531
The TO-PRO and autofluorescence signals were acquired on an initial dataset of 17. To 532 create the initial average, we used image registration to align in a parallel group-wise fashion 533 the TO-PRO images. The variability was expected to be less in the TO-PRO because these 534 images contained more contrast than the autofluorescence images. 535 The creation of an initial average of the adult zebrafish brain was accomplished using 17 536 samples with the TO-PRO channel. The process was completed using a 3-step registration 537 process, similar to prior work (Lerch et al., 2011) using the pydpiper pipeline framework (Friedel 538 et al., 2014) and the minctracc registration tool (Collins and Evans, 1997). This involved taking a 539 single sample at random and registering the 17 samples to it using a 6-parameter linear 540 alignment process (LSQ6). This yielded 17 samples in similar orientation to allow a 12-541 parameter linear registration (LSQ12) to be performed in a pair-wise fashion (each sample is 542 paired with all the other samples, to avoid sample bias) and the final output of these 12-parameter registration was a group average. This represents a linearly registered average adult 544 zebrafish brain. This was then used as the target for non-linear registration with each of the 545 linearly registered 17 TO-PRO samples. This non-linear alignment was repeated successively 546 with smaller step sizes and blurring kernels to allow for an average with minimal bias from any 547 one sample brain. We then took this average and mirrored itself along the long axis of the brain 548 and repeated the registration process described above but instead of using a random brain as 549 the 6-parameter target, we used this mirrored brain. The result of this second pipeline was an 550 average brain where each plane of the brain (coronal, sagittal, horizontal) is parallel with the 551 imaging planes (x,y,z). This final average brain represented the starting point of the atlas. The 552 linear and non-linear transformations created in the registration pipeline were used to resample 553 the 4 µm isotropic TO-PRO and autofluorescence images to the atlas space, yielding an 554 average signal for each channel. The autofluorescence signal was used to register other sample 555 datasets with the atlas because it is common across all datasets. 556 To combine the additional cellular markers to better delineate structures and examine 557 their distribution across the brain, we converted all images and their channels to 4 µm isotropic 558 images as described above. We then converted them to 8 µm isotropic and used the 559 autofluorescence channel for each set to run the above registration pipeline (LSQ6, LSQ12 and 560 non-linear). The initial target was the autofluorescence average created with the TO-PRO 561 dataset described above. Following each registration pipeline, the transformations were used to 562 resample each autofluorescence and cellular marker channel to the atlas with a resolution of 4 563 µm isotropic. 564 To assess registration precision using TO-PRO or autofluorescence images, for each 565 signal we identified 6 landmarks in the atlas, and their corresponding location on 8 different 566 image sets. These points were then brought into atlas space using the transformations from the 567 registration process. We then computed the Euclidean distance between the points in the atlas 568 images from a single sample acquired in the horizontal plane during light-sheet imaging. C) 620 Three-dimensional volumes generated from a set of light-sheet images from an individual brain 621 visualized using a maximum intensity projection (left), and exterior volume (right). D) Coronal 622 (left) and sagittal (right) views of an individual brain generated from a single three-dimensional 623 volume. Autofluorescence images acquired during antibody staining (top) were registered into the same 668 space as autofluorescence images acquired during TO-PRO staining (bottom). C) 669 Transformations from autofluorescence registration were applied to antibody images to bring    Table S2.