AnimalsAll experiments and techniques involving live animals have been approved and conform to the regulatory standards set by the Institutional Animal Care and Use Committee (IACUC) at the Pennsylvania State University College of Medicine. We used in-house bred C57bl/6 J mice (originally purchased from the Jackson Laboratory, Strain #:000664) or transgenic animals with C57bl/6 J background to create MRI and LSFM templates. For brain-wide labeling of pan-GABAergic cell types during embryonic and early postnatal development, we used Gad2-IRES-Cre mice (The Jackson Laboratory, stock 028867) crossed with Ai14 mice which express a Cre-dependent tdTomato fluorescent reporter (The Jackson Laboratory, stock 007908). We used tail samples with PCR for genotyping including Rbm31-based sex genotyping for mice younger than P6. All mice were housed under a 12-hour light/12-hour dark cycle at 22–25 °C with access to food and water ad libitum.Timed pregnancies and brain sample collectionTimed pregnancies followed recommendations from The Jackson Laboratory (https://www.jax.org/news-and-insights/jax-blog/2014/september/six-steps-for-setting-up-timed-pregnant-mice). Adult breeder males were singly housed for 1 week prior to pairing. We paired a male and a female breeder in the evening, removed the male breeder the following morning, and checked for the presence of a vaginal plug in the females. After separation, we measured the baseline body weight of all females on what we considered E0.5 and added extra enrichment to the cages for female breeders. Subsequently, we weighed the females again on E7.5 and E14.5 to assess weight gain, expecting a 1 g or 2–8 g gain at these respective timepoints. Pregnancy was confirmed either by the presence of fetuses during euthanasia of the dams and fetal tissue collection (on E11.5, E13.5, E15.5, and E18.5) or by the birth of a litter.On target collection dates for embryonic samples (E11.5, E13.5, E15.5, and E18.5), pregnant dams were placed in an isoflurane chamber until deeply anesthetized at which point a mixture of ketamine and xylazine was administered via intraperitoneal injection. The dams were subjected to cervical decapitation once fully anesthetized. Then, the uterine horns were immediately removed and placed in ice-cold petri dishes filled with 0.05 M PBS for careful removal of embryos from the uterine casing. E18.5 brains were dissected out at this point while E11.5, E13.5, E15.5 were processed as whole embryos. Embryonic samples were incubated in a 4% paraformaldehyde (PFA) in 0.05 M PBS solution for two days at 4 °C before storing in 0.05 M PBS at 4 °C until use. All embryos were characterized according to the Theiler Staging (TS) Criteria for Mouse Embryo Development86. We used E11.5 for TS19, E13.5 for TS21, E15.5 for TS24, and E18.5 for TS26. After PFA fixation was complete, dissection of the embryos from the yolk sac (E11.5) and eye/ eye pigment removal (E11.5, E13.5, E15.5) were performed under a dissection microscope (M165 FC Stereomicroscope, Leica), followed by storage in 0.05 M PBS at 4 °C until use. Whole embryos were used for E11.5, E13.5, and E15.5, and dissected brains were used for E18.5 for tissue clearing and LSFM imaging. All embryonic samples were tailed for sex genotyping.For postnatal brains (P4, P14, and P56), we defined pups at birth as P0. For collection, mice were anesthetized by a mixture of ketamine and xylazine via intraperitoneal injection. Anesthetized animals were subsequently perfused by saline (0.9% NaCl) and freshly made 4% PFA. Decapitated heads were fixed in 4% PFA at 4 °C overnight, followed by brain dissection and storage in 0.05 M PBS at 4 °C until use.The Kim Lab at Penn State University prepared all animal samples and performed tissue clearing with LSFM imaging. Whole embryos (E11.5, E13.5, E15.5) or decapitated samples (E18.5, P4, P14, and P56) were sent to Dr. Jiangyang Zhang’s lab at NYU for high resolution MRI.Magnetic resonance imagingAll imaging was performed on a horizontal 7 Tesla MRI system (Bruker Biospin, Billerica, MA, USA) equipped with a high-performance gradient system (maximum gradient strength of 670 mT/m). We used a transmit volume coil (72 mm inner diameter) together with a 4-channel receive-only phased array cryoprobe with high sensitivity. Hair and scalp were removed and heads with intact skulls were imaged to prevent deformations. As the embryonic mouse has immature soft skulls, embryonic mouse heads were embedded in 5% agarose gel (Sigma Aldrich, St Louis, MO, USA) for additional support. Specimens were placed in 5 mL syringes filled with Fomblin (Solvay Solexis, Thorofare, NJ, USA) to prevent dehydration and susceptibility to artifacts.High-resolution diffusion MRI was acquired using an in-house 3D diffusion-weighted gradient and spin-echo (DW-GRASE) sequence87 with the following parameters: echo time (TE)/repetition time (TR) = 30/400 ms, two signal averages, diffusion gradient duration/separation = 4/12 ms, 60 diffusion directions with a b-value of 1.0 ms/µm2 for E11.5, 2.0 ms/µm2 for E13.5-E17.5, and 5.0 ms/µm2 for P4, P14, and P56 brains. The increase in b-values with age was necessary as the diffusivity of brain tissues decreases with development88. Co-registered T2-weighted data were acquired using the same sequence but with TE/TR = 50/1000 ms. The native and interpolated spatial resolutions of the MRI data were 0.063/0.0315 mm for E11.5, 0.068/0.034 mm for E13.5, 0.075/0.037.5 mm for E15.5, 0.08/0.04 mm for E18.5, and 0.1/0.05 mm isotropic for P4-P56.The 3D MRI data were reconstructed from k-space to images and zero-padded to twice the raw image resolution in each dimension in MATLAB R2022a (Mathworks, Natick, MA, USA). Diffusion tensor images89 were constructed using the log-linear fitting method in DTI Studio v1.8 (http://www.mristudio.org), and the tensor-based scalar metrics were generated, including the mean diffusivity (MD) and fractional anisotropy (FA).Tissue clearingWe mainly used SHIELD (Stabilization under Harsh conditions via Intramolecular Epoxide Linkages to prevent Degradation) tissue clearing to ensure minimal tissue volume changes while preserving endogenous fluorescence signals when available90. Commercially available SHIELD preservation, passive clearing reagents (see Supplementary Data 4), and detailed protocols were obtained from LifeCanvas Technologies (https://sites.google.com/lifecanvastech.com/protocol/outline). For P56 brains, PFA-fixed samples were incubated in SHIELD OFF solution for 4 days at 4 °C on an orbital shaker. Subsequently, the SHIELD OFF solution was replaced with SHIELD ON buffer and incubated for 24 hours at 37 °C with shaking. Tissues were incubated in 20 mL of delipidation buffer at 37 °C for 4-6 days followed by an overnight wash in 1x PBS at 37 °C with gentle shaking. If sample imaging did not occur within a week after the delipidation step, tissue samples were stored in 1x PBS containing 0.02% sodium azide at 4 °C before continuing to the next step. To match the refractive index (RI) of the delipidated tissues (RI = 1.52) and obtain optical clearing, samples were incubated in 20 mL of 50% EasyIndex (LifeCanvas Technologies, Cat. no.: EI-Z1001) + 50% distilled water for 1 day, then switched to 100% EasyIndex solution for another day at 37 °C with gentle shaking. For samples in earlier time points, we used whole embryos (E11.5, E13.5, and E15.5) or dissected brains (E18.5, P4 and P14) with the same protocol but with smaller reagent quantities and shorter incubation times. Once cleared, embryonic samples were stored in tightly sealed containers with 100% EasyIndex at room temperature (20-22 °C). For 3D immunolabeling and histological staining, we used electrophoresis based active clearing using the SmartClear II Pro (LifeCanvas Technologies) and active labeling using SmartLabel (LifeCanvas Technologies) based on LifeCanvas protocols. Briefly, after the SHIELD OFF step described above, samples were incubated in 100 mL of delipidation buffer overnight at room temperature (RT). Samples were inserted in a mesh bag, placed in the SmartClear II Pro chamber, and delipidated overnight. Samples were transferred to 20 mL of primary sample buffer and incubated overnight at RT. Samples were placed in a sample cup in the SmartLabel device with an antibody cocktail to perform active labeling91. We used Mouse Monoclonal Antibody Neurofilament NF-H (Encor, cat. no. MCA-9B12, Lot no. 012022, 10 µl diluted 1:100 per hemisphere) with Alexa Fluor® 647 AffiniPure™ Fab Fragment Donkey Anti-Mouse IgG (H + L) secondary antibody (Jackson Immuno Research, cat. no.: 715-607-003, Lot no. 153490, 2.7 µl diluted 1:500 per hemisphere), and propidium iodide (Thermo Fisher, cat. no.: P1304MP, 12 µl /hemisphere) for pan-cellular labeling. For the limited dataset at P56, we used the iDISCO (immunolabeling-enabled three-dimensional imaging of solvent-cleared organs) tissue clearing protocol92,93.Light sheet fluorescence microscopy imaging and reconstructionFor LSFM imaging, all samples were embedded in an agarose solution containing 2% low-melting agarose (Millipore Sigma, cat. no.: A6013, CAS Number: 9012-36-6) in EasyIndex using a custom sample holder. Embedded samples were then incubated in EasyIndex at room temperature (20-22 °C) for at least 12 hours before imaging using the SmartSPIM light sheet fluorescence microscope (LifeCanvas Technologies, Cambridge, MA, USA). During the imaging process, the sample holder arm securing the embedded sample was immersed in 100% EasyIndex. Our imaging setup consisted of a 3.6X objective lens (LifeCanvas Technologies, 0.2 NA, 12 mm working distance, 1.8 μm lateral resolution), three lasers with wavelengths of 488 nm, 561 nm, and 642 nm, and a 5 μm z step size. After imaging, all samples were stored in 100% EasyIndex at room temperature (20–22 °C). For our iDISCO cleared P56 brain samples, we did not use agarose embedding and directly mounted samples in a custom-built holder. LSFM imaging was performed in ethyl cinnamate for index matching (Millipore Sigma, cat.no.: 112372, CAS number: 103-36-6) using the same imaging parameters.For 3D reconstruction, we developed a parallelized stitching algorithm optimized for conserving hard drive space and memory consumption initially based on Wobbly Stitcher17. The algorithm started by collecting 10% outer edges of each image tile and making a maximum intensity projection (MIP) of outer edges in the axial (z) direction for every set of 32 slices of the entire stack. The algorithm then aligned z coordinates of MIP images across image columns, followed by the x and y coordinate alignment. Finally, 32 slices within each MIP were adjusted based on curve fitting to reach final coordinates of each tile. This algorithm only reads the raw images two times (at the beginning and the final writing), which significantly reduced the bottleneck of reading large files in a storage drive. See Code Availability section for algorithm access.Symmetric template constructionEach symmetric template is an intensity and morphological average of multiple male and female samples with a sample size ranging from 6 to 14 (Table 2). After stitching, images were preprocessed for template construction. For postnatal ages, MRI data preprocessing began with digital brain extraction by hand drawing brain masks using ITK-SNAP v494 or Avizo 2021.1 (Thermo Fisher Scientific). Brain tissue was assigned a value of 1 and non-brain was assigned 0. The mask was multiplied by MRI data to extract the brain. Next, sample orientation was viewed in ITK-SNAP and corrected using ANTs v2.3.538. MRI data was normalized using N4 bias field correction95. LSFM data preprocessing began with image resampling to 3 sizes: 50 µm, 20 µm, and 10 µm isotropic voxel resolution, then orientation correction. To ensure template symmetry, each preprocessed image was duplicated and reflected across the sagittal midline, doubling the number of input datasets used in the template construction pipeline and allowing final parcellations to be reflected over the midline, reducing annotation efforts. Template construction, ANTs call ‘antsMultivariateTempalteConstruction2.sh’38,96, was employed on Penn State’s High-Performance Computing system (HPC). Briefly, starting from an initial template estimate derived as the average image of the input cohort, this function iteratively performed three steps: (1) non-linearly registered each input image to the current estimate of the template, (2) voxel-wise averaged the warped images, and (3) applied the average transform to the resulting image from step 2 to update the morphology of the current estimate of the template. Iterations continued until the template shape and intensity values stabilized. MRI templates were constructed at their imaged resolution using ADC MRI contrasts for initial postnatal templates and diffusion weighted imaging (DWI) contrasts for embryonic templates. Once the initial MRI template was constructed, the sample to template warp fields generated were applied to all MRI contrasts for each sample. Warped samples were averaged to construct templates for each contrast. LSFM templates were constructed from autofluorescence data collected from C57bl/6 J mice and transgenic mice with a C57bl/6 J background. To save memory and improve speed, LSFM templates were initially constructed at 50 µm isotropic resolution. This template was resampled for template construction initialization at 20 µm isotropic resolution, a process repeated to construct the final LSFM template with 10 µm isotropic resolution input images. The in-house template construction script is referenced in the Code Availability section.Multimodal registration of 3D imaging to the DevCCFWe aligned the CCFv3 to the P56 DevCCF and each LSFM template to the corresponding age matched DevCCF MRI template to enable data integration across multiple modalities with undistorted morphology. Our protocol aims to address multimodal registration challenges due to differences in brain and ventricle volume that often result in internal structure misalignment97. We performed initial non-linear registration of the 3D datasets (CCFv3 and LSFM templates) to the age-matched DevCCF MRI template using ANTs with the mutual information similarity metric38. We then visually compared the warped 3D dataset with the DevCCF template in ITK-SNAP98 to identify landmark brain regions that remained misaligned after the initial registration. Whole brain masks and misaligned regions were segmented for the 3D dataset and DevCCF template in 3D using Avizo (Thermo Fisher Scientific). The segmented regions were subtracted from the brain masks, creating modified brain masks with identifiable boundaries around misaligned brain regions. This provided a map of regions that needed correction. Next, linear registration was performed of the modified brain masks, followed by equally weighted non-linear alignment of both the 3D data images and their modified brain masks. Landmark-assisted multimodal registration warp fields were resampled and applied to transform CCFv3 and LSFM templates to MRI template morphology at 20 μm isotropic voxel resolution, which is sufficient resolution for mapping cell-type data with a reasonable compute time, yet small enough file size to be downloadable to a local computer. CCFv3 annotations were also transformed to MRI template morphology at 20 μm isotropic voxel resolution for comparison with the DevCCF. We elected not to register 10 μm isotropic voxel resolution images to MRI for several reasons Increasing resolution does not have sufficient resolution in MRI with which to be paired, therefore is unlikely to improve resolution. Additionally, computation time and resources to register the 10 μm isotropic voxel resolution was not reliably available. Lastly, this helps keep the DevCCF package small enough to easily download and work with on any computer. Because 10 μm isotropic voxel resolution LSFM templates were generated, we have made them available unwarped alongside the DevCCF.We also registered LSFM whole brain data featuring immunostaining (e.g. Nissl) or Cre-dependent fluorescence (e.g. Gad2) to the DevCCF to assist with annotation, segmentation, and validation. Each LSFM dataset collected up to three channels simultaneously, always including one autofluorescence channel. We used non-linear registration to align LSFM autofluorescence channel data to the warped DevCCF LSFM template. The LSFM template was chosen to achieve optimal registration quality due to matching the autofluorescence contrasts. Forward transforms were applied to LSFM immunostaining and Cre-dependent fluorescence channel data to align them to the DevCCF template.2D gene expression mapping onto DevCCFWe utilized the Allen Brain Atlas API to download in situ hybridization (ISH) data from the ADMBA28 (E11.5, E13.5, E15.5, E18.5, P4, P14, P28) and Allen Brain Atlas50 (P56) as both imaged histological sections and analyzed gene expression. Except for P28, this dataset is age-matched to the DevCCF, which allows reliable validation of DevCCF annotations. Each dataset consists of 2D images of coronal or sagittal sliced brain sections depicting expression of a single gene. Each brain sample was used for 4 to 7 ISH gene expression experiments, alternating slices for each.We used the python v3.8.16 package ANTsPy v0.3.838 to register the ISH data to the age-matched DevCCF. P28 data was matched to the P56 DevCCF as mouse brain volume is stable after 3 weeks of age27. To prepare for registration, sample level data was compiled from each of 4-7 ISH experiments, then reconstructed to a 3D volume containing all sample sections and 3D volumes of each individual analyzed gene expression experiment. Missing sections were included in the volumes as empty slices. Reconstructed volumes were resampled to 512×512 pixels for computational tractability and histological reconstruction intensity was inverted to set the background to zero. Empty histological analyzed sections were filled using B-spline scattered data approximation99 to generate a full intact volumes.To begin alignment, we ran an initial linear multi-metric registration from DevCCF MRI templates (FA and ADC contrasts) to the reconstructed histological sample volume. This aided the next step of slice-wise correction of the 3D sample reconstruction. Here, we ran nonlinear multi-metric registration of each sample section to the neighboring experimental section and the aligned MRI template section mask. Next, the template to subject alignment was refined by nonlinearly registering the age-matched MRI templates to the slice-wise corrected sample reconstruction. Finally, the B-spline approximation filled analyzed gene expression volumes for each individual experiment were warped to DevCCF morphology using the transform parameters saved during sample slice-wise correction and inverse transform parameters saved during refined template to subject alignment. All registrations are used the ANTsPy ‘antsRegistrationSyNQuick’ transform type with the mutual information similarity metric, with the exception of slice-wise correction using the MRI mask, which used the mean squared difference similarity metric38.Integrating CCFv3 to DevCCFTo compute the voxel-to-voxel mapping between CCFv3 and DevCCF anatomical label correspondence, we performed voxel-wise comparison in the P56 aligned annotations. Each voxel in the P56 brain was labeled by one DevCCF and one CCFv3 annotation and grouped into a parent structure. Unique combinations of DevCCF-CCFv3 mappings were summed to generate voxel counts. Voxel counts meeting a threshold of 250 voxels were then plotted as a Sankey flow diagram using the Plotly v5.10 library in python 3.11.To quantify cell type distribution between CCFv3 and P56 DevCCF labels for spatial transcriptomic data, we obtained soma coordinates for each cell in the MERFISH dataset in CCFv3 morphology51. We used the CCFv3 to DevCCF inverse transforms to warp the P56 DevCCF annotations to the CCFv3 template. For each coordinate, we compared the DevCCF and CCFv3 labels and grouped cells by parent division and subclass label. We then normalized the data by the total cell count per subclass, resulting in a proportional representation of cell subclasses by parent division. Heatmaps were plotted using python’s seaborn 0.12 library. Code associated with these methods are referenced in the Code Availability section.Molecular Atlas8 3D meshes for each region were downloaded from molecularatlas.org and pixelized using the Dragonfly 2022.2 (Comet Technologies Canada Inc., Montreal, Canada) multi ROI toolbox, giving the Molecular Atlas in CCFv3 space. The Molecular Atlas was transformed to DevCCF morphology using CCFv3 to DevCCF inverse transforms for annotation comparison.Anatomical segmentations for the DevCCFWe performed theory and data-driven anatomical segmentation of the multimodal templates at each age using Avizo (Thermo Fisher Scientific). We manually drew contours on coronal, horizontal, and sagittal slices of templates to generate 3D segmentations. To assist in the process, we used various interpolation, thresholding, and smoothing tools. We assigned unique labels and colors to each region and further developed the hierarchical nomenclatures following standards in the ADMBA. Annotations were delineated on one hemisphere, then copied and reflected over the midline to the empty hemisphere to make them bilateral.We followed the principles of the prosomeric model to define brain regions based on morphological features, such as sulci, fissures, ventricles, commissures, tracts, cytoarchitecture, and gene expression32. Segmentations started with large regions defined early in development, such as the neural plate, and were progressively subdivided into smaller regions as defined by the prosomeric model, such as the forebrain, hindbrain, and spinal cord until reaching a level of detail comparable to at least level 5 (fundamental caudo-rostral and dorsoventral partitions) in all brain regions (Fig. 4a). Additionally, we updated the telencephalon to reflect the cortical concentric ring topology (Supplementary Fig. 3)39. Anatomical divisions and subdivisions were drawn across the whole brain in multiple segments and combined. For example, floor, basal, alar, and roof plate segmentations in the ventral-dorsal direction were drawn separately from rostro-caudal neuromere segmentations and later overlayed with one another (Table 3). Similarly, cortical region segmentations (e.g., hippocampal cortex and olfactory cortex) were drawn separately from cortical layer segmentations (e.g., ventricular zone, mantle zone), and later overlayed to combine. Likewise, the subpallium was segmented by the primary domains (striatum, pallidum, diagonal, and preoptic) and secondary divisions (septum, paraseptum, central, and amygdalar), then combined. This technique allows efficient segmentation correction upon validation. Additional brain regions were modified from the ontology described in the ADMBA to meet the current anatomical understanding (Supplementary Data 2)39,42,100.Neuromeres and pallial cortex division boundaries follow smooth trajectories that end perpendicular to the brain surface and the ventricles. Each neuromere stretches from floor to roof plate and is in contact with only one neuromere caudally and one neuromere rostrally32. Pallial neocortex (NeoCx), allocortex (AlloCx), and mesocortex (MesoCx) are segmented according to the concentric ring topology as 6-layer structures, 3-layer structures, and transitional structures, respectively, where layers are overlayed from ventricular to superficial (Supplementary Fig. 3)39. Choroidal tissue is within the roof plate. The medullary hindbrain (MH) neuromeres are divided evenly into 5 segments (r7-r11). Parcellations are based on previously defined landmarks28,32,39,42,100,101,102 primarily visualized with template contrast features. For example, the lateral and medial ganglionic eminences mark the early subpallium (Supplementary Fig. 4a). The facial motor nerve identifies the ventral surface and alar-basal boundary of rhombomere 6 (r6) (Supplementary Fig. 4e). In late postnatal ages, the neocortex, allocortex, and mesocortex were segmented by separating regions by thick cortical layer areas, 3-layer areas, and transitional zones, respectively, visible in MRI templates (Supplementary Fig. 4f, h). Regions without easily visible landmarks are delineated based on neighboring regions. Mesomere 2 (m2), for example is a thin neuromere between the caudal most landmarks of mesomere 1 (m1) and the rostral most landmarks of isthmus (r0) (Table 3).The morphological foundation of the DevCCF was primarily guided by MRI templates, additional details from warped LSFM autofluorescence templates and registered LSFM cell-type data (e.g. Neurofilament, Nissl, Gad2). Here we provide a few examples of each data type (Figs. 3, 4d–f, Table 3). The MRI T2-weighted templates, in conjunction with LSFM templates, facilitated delineation of brain tissue from non-brain structures, as well as ventricle identification in embryonic DevCCF templates (Fig. 3). MRI FA templates highlight white matter tract landmarks, which serve as markers for many boundaries (e.g., the boundary of p1 and p2 is immediately caudal to the retroflex tract; (Fig. 3p, Table 3). LSFM templates were instrumental in segmenting features such as the choroid plexus in the developmental roof plate (Fig. 3m, q). We imaged and aligned additional 3D fluorescent cell-type specific datasets to provide reference data for DevCCF segmentations. For example, we used LSFM images of embryonic Gad2-Cre;Ai14 samples to delineate the subpallium (SPall), prosomere 1 (p1), prosomere 3 (p3), and the cerebellum (Fig. 5). Additionally, LSFM imaging of SYTOX stained mouse brains samples provided Nissl-like 3D histology, allowing us to use classical neuron cell density patterns during segmentation (Fig. 4f).The ADMBA and associated ISH data were key resources to segment large rostro-caudal, and dorsoventral boundaries. While drawing segmentations, we viewed the atlas and gene expression data side-by-side via the web-portal (https://developingmouse.brain-map.org/) as supportive evidence for the prosomeric model of vertebrate brain development, aiding in validating relative relationships among neuromeres, dorsoventral plates, nuclei, white matter tracts, and cranial nerve roots. Moreover, we utilized registered ISH data associated with the ADMBA to validate DevCCF segmentations based on gene expression patterns known from existing literature (Supplementary Fig. 4).The CCFv315 provided delineations of the cortical layers and many nuclei in the P56 mouse brain. The CCFv3 template and annotations were aligned to the P56 MRI template using our landmark-assisted multimodal registration methods. Previously validated CCFv3 cortex layers (1, 2/3, 4, 5, 6a, and 6b) were imported to the P56 DevCCF. P56 DevCCF cortical regions (e.g., insular cortex, entorhinal cortex) were segmented manually and combined with imported CCFv3 layers to finalize cortical segmentations. Additional CCFv3 segmentations with corresponding prosomeric model regions (e.g., subdivisions of prosomere 2 and mesomere 1) were aligned to the DevCCF and used as primers to segment the DevCCF. CCFv3 cortical layers were also used for P14 and P56. CCFv3 cortical layers were similarly aligned from the P56 DevCCF to the P14 and P4 DevCCF templates via landmark-assisted multimodal registration methods and used to prime segmentation.Anatomical segmentations were paired with numerical identifiers (IDs). IDs below 18000 represent previously existing ADMBA labels with potential minor name, abbreviation, and ontology level updates. Newly added structures to the DevCCF ontology have IDs in the range 18000-19999. To ensure compatibility with viewing and analysis tools requiring 16-bit depth labels, existing 32-bit ADMBA IDs that fell outside the 16-bit depth range (0-65535) were fit into the 20000 to 29999 range. This was achieved through a formulaic approach where the last four digits of the original 32-bit IDs were extracted and given a prefix of 2. New 16-bit labels ensure the modified IDs retain their uniqueness and do not overlap with either original or newly generated labels. Annotation names, abbreviations, parent structures, associated colors, previous 32-bit IDs, and current 16-bit IDs are organized in the DevCCF ontology (Supplementary Data 2), where red text indicates DevCCF modifications or additions to the ADMBA ontology. Warped CCFv3 segmentation values outside the 16-bit range were also translated to 16-bit values for ease of use. Translated values can be viewed with original values in our modified CCFv3 ontology by viewing side-by-side columns ‘id‘ (original values), and ‘id_16bit‘ (new values) in the data download.GABAergic neuron quantificationWe used LSFM to image Gad2-Cre;Ai14 mouse embryos at E11.5 (3 male, 4 female), E13.5 (1 male, 3 female, 1 unknown), and E15.5 (1 male, 2 female). We used the interactive machine learning for (bio)image analysis (ilastik) v1.3.3post3103 to train a pixel classification-based machine learning model for each age to identify GAD2 positive (GAD2 + ) voxels in the whole embryo at 1.8 × 1.8 × 5 µm3 voxel resolution. We resampled the pixel classification images to 20 µm isotropic resolution for image registration to the DevCCF, where each voxel value represented the sum of GAD2+ voxels in the respective full resolution image. The GAD2+ voxel count was deformably registered to age matched DevCCF templates using ANTs call ‘antsRegistrationSyN.sh‘ as described in multimodal registration methods. GAD2+ voxel count relative occupancy per anatomical region was calculated as a ratio of the sum of positive voxels to the number of voxels per region.Interactive web visualization developmentThe DevCCF web app, built using Neuroglancer v3.5, provides an interactive way to explore the developing mouse brain atlas. A key feature is the “Ontology” button, which allows users to navigate the hierarchical structure of brain regions defined by the DevCCF ontology. This functionality is achieved through a combination of HTML, JavaScript, and the Neuroglancer API. First, the web app loads the necessary libraries and data, including the atlas template, annotations, and a JSON file representing the initial viewer state. A custom script then parses a CSV file containing the DevCCF ontology structure, which defines each brain region with its ID, name, acronym, parent region, and color. From this parsed data, the script builds a hierarchical JSON structure representing the parent-child relationships between brain regions. Next, the script creates HTML elements for each brain region, assigns unique IDs, and sets the text content and color based on the CSV data. These elements are then arranged hierarchically, with child regions nested within their parent regions. Initially, only the top-level regions are visible. The “Ontology” button is added to the interface, and when clicked, it reveals the ontology container with the hierarchical tree of brain regions. Users can expand or collapse parent regions using “+” and “-” icons and click on individual regions to highlight them in the Neuroglancer viewer. The script leverages the Neuroglancer API to control the visibility of brain regions based on user interactions with the ontology tree. Selecting a region in the tree updates the corresponding segmentation layer’s segment property in the Neuroglancer state, ensuring only the selected region and its descendants are visible. Conversely, changing the visible segments in the viewer updates the checked state of the corresponding checkboxes in the ontology tree, maintaining consistency between the two. Additional features include a search bar within the ontology container for finding specific regions and a hover function that displays detailed information about each region. This integration of HTML, JavaScript, and the Neuroglancer API provides a user-friendly and interactive way to explore the DevCCF ontology and visualize the corresponding brain regions.Aligning postnatal atlases to stereotaxic coordinatesPostnatal atlases were rotated to stereotaxic coordinates by aligning the estimated bregma and the anterior commissure in a common coronal plane. For each postnatal age, the MRI FA and T2w templates were opened in ITK-SNAP. Using 2D cross sections of the FA template and the 3D volume rendering of the T2w template, we annotated two midline points: the estimated bregma above the skull impression on the superior surface of the brain, and the center of the anterior commissure. Templates were rotated in the sagittal plane to place the estimated bregma directly superior to the anterior commissure.To display the coordinate system, we created 3D grid images for each postnatal atlas at native MRI resolution. The grids can be overlayed on the age-matching atlas. The central vertex is over the anterior commissure. Space between gridlines represents 1 mm. Datasets are coronally oriented. Grid labels parallel to the sagittal, horizontal, and coronal planes are labelled 1, 2, and 3 respectively, allowing differentiation when overlayed. Numeric spacing labels were created using FIJI text tools and given values of 4. These labels are visible from the coronal plane, displaying height and width in millimeters, with anterior-posterior slice distance from the anterior commissure on bottom left corner of each section. To avoid confusion across various 3D visualization and analysis software, image metadata origins are fixed at (0, 0, 0).Volumetric growth curvesWe calculated template average volume and standard deviation based on individual samples used to create each template. MRI DWI templates and annotations were aligned to each sample with ANTs ‘antsRegistrationSyNQuick.sh.‘ Sample volumes were calculated for whole brain and selected parent regions through development: pallium, subpallium, terminal hypothalamus, peduncular hypothalamus, diencephalon, midbrain, prepontine hindbrain, pontine hindbrain, pontomedullary hindbrain, and medullary hindbrain. Volume averages and standard deviations were calculated using NumPy v1.21.1 and Pandas v2.2.0 libraries in python. Data was plotted with Prism v10.2.0 (GraphPad).Reporting summaryFurther information on research design is available in the Nature Portfolio Reporting Summary linked to this article.
