(C) PLOS One
This story was originally published by PLOS One and is unaltered.
. . . . . . . . . .



A method to estimate the cellular composition of the mouse brain from heterogeneous datasets [1]

['Dimitri Rodarie', 'Blue Brain Project', 'École Polytechnique Fédérale De Lausanne', 'Epfl', 'Geneva', 'Csaba Verasztó', 'Yann Roussel', 'Michael Reimann', 'Daniel Keller', 'Srikanth Ramaswamy']

Date: 2023-01

The mouse brain contains a rich diversity of inhibitory neuron types that have been characterized by their patterns of gene expression. However, it is still unclear how these cell types are distributed across the mouse brain. We developed a computational method to estimate the densities of different inhibitory neuron types across the mouse brain. Our method allows the unbiased integration of diverse and disparate datasets into one framework to predict inhibitory neuron densities for uncharted brain regions. We constrained our estimates based on previously computed brain-wide neuron densities, gene expression data from in situ hybridization image stacks together with a wide range of values reported in the literature. Using constrained optimization, we derived coherent estimates of cell densities for the different inhibitory neuron types. We estimate that 20.3% of all neurons in the mouse brain are inhibitory. Among all inhibitory neurons, 18% predominantly express parvalbumin (PV), 16% express somatostatin (SST), 3% express vasoactive intestinal peptide (VIP), and the remainder 63% belong to the residual GABAergic population. We find that our density estimations improve as more literature values are integrated. Our pipeline is extensible, allowing new cell types or data to be integrated as they become available. The data, algorithms, software, and results of our pipeline are publicly available and update the Blue Brain Cell Atlas. This work therefore leverages the research community to collectively converge on the numbers of each cell type in each brain region.

Previously, we presented a model of a cell atlas, which provided an estimate of the densities of neurons, glia and their subtypes for each region in the mouse brain. Here, we describe an extension of this model to include more inhibitory neuron types. We collected estimates of inhibitory neuron counts from literature and built a framework to combine them into a consistent cell atlas. Using brain slice images, we also estimated inhibitory neuron density in regions where no literature data are available. We estimated that in the mouse brain 20.3% of all neurons are inhibitory. Among all inhibitory neurons, 18% predominantly express parvalbumin (PV), 16% express somatostatin (SST), 3% express vasoactive intestinal peptide (VIP), and the remainder 63% belong to the residual GABAergic population. Our approach can be further extended to any other cell type and provides a resource to build tissue-level models of the rodent brain.

Obtaining a global understanding of the cellular composition of the brain is a very complex task, not only because of the great variability that exists between reports of similar counts but also because of the numerous brain regions and cell types that make up the brain.

Funding: This study was supported by funding to the Blue Brain Project, a research center of the École polytechnique fédérale de Lausanne (EPFL), from the Swiss government’s ETH Board of the Swiss Federal Institutes of Technology. The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.

Data Availability: All relevant data are within the manuscript and its Supporting Information files. The code is available at: https://github.com/BlueBrain/atlas-densities Results can be visualized at: https://bbp.epfl.ch/nexus/cell-atlas/ The data downloaded from the Allen Institute website is available at the following links (see also Acknowledgements Section in the main text): Nissl CCFv2: http://download.alleninstitute.org/informatics-archive/current-release/mouse_ccf/ara_nissl/ Annotation atlas CCFv2: http://download.alleninstitute.org/informatics-archive/current-release/mouse_ccf/annotation/mouse_2011/ Annotation atlas CCFv3: http://download.alleninstitute.org/informatics-archive/current-release/mouse_ccf/annotation/ccf_2017/ The AIBS ISH dataset can be found at http://mouse.brain-map.org/search/index The corresponding experiment identifiers are listed below: • Parvalbumin: Experiment id #868 used for the BBCAv2 and for S8 Fig • Somatostatin: #1001 for the BBCAv2 • Vasoactive Intestinal Peptide: #77371835 for the BBCAv2 • GAD1 (equivalent of GAD67): #479 for the BBCAv2, #79556706 for Fig 3 , #480 for Fig 10 , #75457536 for S7 Fig and #79556706 used in Section 3.4 • GAD2 (equivalent of GAD65): #79903740 for S7 Fig .

Copyright: © 2022 Rodarie et al. This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.

We also highlighted the fact that the alignment of multiple datasets to a common coordinate system is essential to obtain the best results in terms of cell density estimations. We therefore used Krepl et al.’s algorithm [ 56 ], based on deep learning techniques to automatically realign ISH datasets to our reference system. This tool replaces the manual realignment step in Erö et al.’s pipeline [ 17 ]. We combined all these new tools and methods to produce a second version of the Blue Brain Mouse Cell Atlas pipeline (BBCAv2) and model. We further extended the BBCAv2 in a companion paper [ 57 ], mapping well identified morpho-electrical types to our GABAergic neuron subclasses. These refinements to the model will help neuroscientists to get a better understanding of cellular composition in the brain and will pave the way for more accurate in silico reconstructions of brain tissues.

Steps of the pipeline are displayed in rows with the data consumed in the left column, the method used in the middle column, and the produced results on the right. Each step of this pipeline builds on the results of the previous steps.

We chose to include these subclasses in the new pipeline because they have been extensively studied in the literature [ 7 , 8 , 22 – 52 ], which means we could find data to build and validate our model. Moreover, the associations of these neuron types with brain health or disease make them an important refinement to our model. PV+ neurons are fast spiking, resulting in fast and efficient suppression of their surrounding neurons’ activity [ 53 , 54 ]. A deficit of PV+ neurons seems to be linked to neural diseases such as schizophrenia [ 42 ], Alzheimer’s [ 26 ] and autism [ 32 , 36 ]. SST+ neurons contribute to long range inhibition [ 54 ]. Their population decreases with age and seems to contribute to the development of Alzheimer’s disease [ 39 , 41 , 47 ]. VIP+ neurons, on the other hand, tend to target other inhibitory neurons in the isocortex, playing the role of local disinhibitors [ 38 ]. SST+ and VIP+ neurons are also involved in the circadian clock which regulates daily brain activity [ 55 ]. As in Erö et al. [ 17 ], we did not use counting algorithms to extract counts of cells from the AIBS ISH datasets. Indeed, the authors pointed out that counting techniques were not always producing satisfactory results especially in dense regions because of cell overlapping (see their Figs 1C and 7A–7D ). Using counting techniques at the whole brain scale is therefore out of the scope of this study.

Specifically, using additional ISH experiment datasets from the AIBS, we were able to further refine the GABAergic neuron densities from Erö et al. [ 17 ] and estimate the densities of GABAergic subclasses. This includes parvalbumin (PV+), somatostatin (SST+), vasoactive intestinal peptide (VIP+) immunoreactive neurons, and the remaining inhibitory neurons.

We extended the workflow from Erö et al. [ 17 ], using estimates of cell densities for individual brain regions from the literature instead of whole-brain (global) estimates for specific cell types.

Erö et al. argued that literature values are often inconsistent and are therefore unreliable [ 12 , 17 ]. Hence, the authors used only whole-brain values from literature (e.g., the total number of cells and neurons in the mouse brain from Herculano-Houzel et al. [ 20 ]) to constrain their estimates of cell densities in every brain region, avoiding any bias toward a particular region. However, there are also several regions for which a consensus on cell composition has been reached (e.g., for the cerebellar Purkinje layer, see review from Keller et al. [ 12 ]). For instance, in all regions of the isocortex (also called neocortex), it is well-known that layer 1 regions should only be composed of inhibitory neurons [ 5 ]. To take this aspect into account, the BBCAv1 had to manually integrate those regional constraints a posteriori (e.g., force each neuron in isocortex L1 to be inhibitory) which in turn contradicted the original global constraints. Moreover, the BBCAv1 strategy yielded regional estimates that sometimes did not match their respective literature values (e.g. cell estimates for cortical subregions compared to Herculano-Houzel et al. [ 21 ]). In order to obtain a more fine-grained classification, the BBCAv1 strategy required brain-wide estimates for every new cell type to be integrated. These constraints may not yet exist. We therefore decided to refine the BBCAv1 strategy to obtain inhibitory neuron subtypes densities.

Erö and colleagues [ 17 ] were able to estimate the cellular composition of each brain region defined by the Allen Mouse Brain Atlas [ 18 ] with respect to excitatory (glutamatergic) and inhibitory (GABAergic) neurons, astrocytes, oligodendrocytes, and microglia. The authors aligned multiple datasets, including genetic marker expression from in situ hybridization (ISH, see complete list of abbreviations in Table A in S1 Document ) experiments from the Allen Institute for Brain Science (AIBS), to the Allen Reference Nissl volume from Lein et al. [ 19 ] to determine the positions of all cells within the annotation volume. This resulted in the first version of the Blue Brain Mouse Cell Atlas pipeline (BBCAv1) and the associated model as described in Erö et al. [ 17 ].

In this paper, we provide a method to generate a cell atlas model of the whole mouse brain. This method integrates disparate datasets from literature for multiple cell types, using a constrained optimization approach. The final cell densities in our model are constrained by one another to maintain a coherent framework.

Over the past century, numerous studies have reported a great variety of cells in the mouse brain according to their morphological, electrical, and molecular features. Several groups such as the US BRAIN Initiative have launched ambitious projects to undertake a comprehensive census of all brain cells, determining their molecular, structural and functional properties [ 1 – 4 ]. Building a catalog of the vast diversity of neuron types in the brain is a challenge for modern neuroscience. Characterizing this variety is exacerbated by the fact that distinct cell types are localized to specific brain regions—for example, pyramidal cells in the cerebral cortex or Purkinje cells in the cerebellum [ 5 , 6 ]. Therefore, studies tend to simplify this task by focusing on specific brain regions or a small subset of known cell types. For example, there is a small number of studies that attempt a brain wide estimation of cell densities, using molecular and microscopic techniques to label cells across all brain areas, however, these studies are always limited to only a few cell classes [ 7 – 10 ]. However, these brain-wide datasets cover only a small proportion of the mouse brain cell classifications. Moreover, the variety of methods used has resulted in considerable variability in the outcomes, even when the same region is considered [ 7 , 11 , 12 ]. As a result, it is difficult for neuroscientists to evaluate the quality of published data sets and to combine different data sets (even brain-wide mappings) to estimate the cellular composition of the entire brain. Compiling a complete and comprehensive cell atlas of the mouse brain is, therefore, a monumental task, which needs to be tackled to enable the in silico reconstruction of multiscale neural circuits [ 6 , 13 – 16 ].

2. Materials and methods

2.1. Overview of the pipeline 2.1.1. Original pipeline. In this paper we refined several steps of the BBCAv1 to facilitate the integration of more and more specific cell types without having to rely on the total number or ratio of these cell types. We started with the original Blue Brain Mouse Cell Atlas workflow (BBCAv1) from Erö et al. [17] which consisted of six steps: The Nissl-stained brain slices from Dong [18], and the corresponding mouse brain Annotation Volume (AV) were manually realigned along the sagittal axis. Every ISH dataset [19] used in the following steps was also manually realigned to match the Nissl volume. The annotated Nissl volume from the AIBS was combined with a total number of cells in the mouse brain from Herculano-Houzel et al. [20] to estimate the number of cells in each brain region. A combination of genetic marker datasets and a global ratio of glial cells in the brain from Herculano-Houzel et al. [20] was used to distinguish neurons from glial cells. Glial cells were labeled as astrocytes, oligodendrocytes or microglia based on whole brain ratios together with their respective ISH datasets. GAD67 (glutamic acid decarboxylase) marker experiment from Lein et al. [19], associated with a global number of inhibitory neurons from Kim et al. [7] was used to distinguish the excitatory from the inhibitory neurons. This number stands for the sum of PV+, SST+ and VIP+ reacting cells in the brain and was assumed in this pipeline to represent the entire inhibitory population. Manual correction of purely inhibitory regions: molecular layer regions of the cerebellum, layer 1 of the isocortex, the reticular nucleus of the thalamus, and the striatum were considered to be fully inhibitory, which means that all their neurons are labeled as GABAergic. 2.1.2. Adjustments for improved pipeline. Step 1 of the BBCAv1 involved manual selection of landmark points on each slice of the AV, and ISH data to realign to the Nissl volume. This step was not only labor intensive but also error prone. In our updated workflow, we replaced the manual realignment by a novel deep learning based alignment algorithm by Krepl et al. [56]. In step 5, the total number of inhibitory neurons in the mouse brain was a global constraint. Its biological value was taken from Kim et al. [7] and represents the sum of SST+, PV+, and VIP+ neurons. However, these three types do not represent all inhibitory neurons (e.g., LAMP5 cells in the isocortex [58]). This means that the total number of inhibitory cells in BBCAv1 was underestimated, despite the corrections applied in step 6. We therefore reworked steps 5 and 6 by integrating more literature estimates of neuron type densities as well as by using additional ISH datasets. This not only improved our estimates of the excitatory to inhibitory ratios in the brain, but also provided more precise estimates of GABAergic neuron types, including PV+, SST+, and VIP+ cells. However, as we will discuss later, the spatial resolution (200 μm between each coronal slice) and the whole-brain coverage of these datasets are not sufficient to estimate the density of all inhibitory subtypes in each voxel of the mouse brain, as it was done in BBCAv1. Therefore, we computed the mean density of each inhibitory type for each region of the AV, instead of estimating the density per voxel. Since several AV and Nissl volumes have been released [59,60], we also developed a method to choose the best combination to estimate cell densities (see Section 2.2). 2.1.3. New pipeline. Fig 1 shows the new pipeline for the Blue Brain Mouse Cell Atlas version 2 (BBCAv2). It consumes data from the AIBS in the form of image stacks of stained coronal brain slices, as well as a combination of annotation and Nissl reference atlases. The BBCAv2 has four main steps. In the first step, the different image stacks from ISH experiments from the AIBS are automatically aligned and registered to a reference volume. We describe this step in Section 2.3. The second step generates cell, neuron, and glial densities for each region of the AV. It corresponds to steps 2 and 3 of the BBCAv1 pipeline. We devise assumptions to estimate densities of different inhibitory neuron types in a coherent framework (see Section 2.4) and apply them for step 3 and 4. In step 3, we assume a correlation between genetic marker expression and cell type density gathered from literature (see Section 2.5) and the previously realigned and filtered ISH datasets. The details of this step are described in Section 2.6. Step 4 integrates the results of step 3 into a consistent estimate of the cell densities for each inhibitory subtype, according to our assumptions (see Section 2.7). Finally, the cells are placed in the brain volume to produce the updated BBCAv2 model (see Section 2.8).

2.3. Alignment of in situ hybridization coronal brain slices For step 3 of the BBCAv2, we consume ISH datasets from the AIBS [19] which are sectioned in the coronal plane (one 25 μm thick section every 200 μm, see Fig 3D). However, the individual images are not aligned to the same Nissl volume that we use to estimate neuron densities. In the BBCAv1, ISH images slices were manually realigned to the Nissl volume with a non-rigid landmark-based algorithm (see S8 Fig). This task was very tedious and time consuming because it required an expert to manually choose common landmarks between adjacent sections. For our new pipeline, we therefore replace this step with a fully automatic machine learning based method. The results of the automated realignment on a single slice are shown in S8 Fig. PPT PowerPoint slide

PNG larger image

TIFF original image Download: Fig 3. AIBS GAD67 ISH experiment data. (A) raw image, (B) filtered image and (C) realigned binarized image. This AIBS ISH experiment highlights inhibitory cells of the mouse brain (experiment No 79556706—slice 22). The raw slice image (A) contains artifacts and a minimum expression value offsets the dataset. The filtered image (B) is derived from the raw image (A) by the AIBS. Somas reacting to GAD67 are detected in the raw image and isolated from the background. The different colors of the cells in (B) represent the different levels of expression (AIBS 2015 white paper). (A) and (B) have been downloaded on the AIBS website [68] with permission of the AIBS. The filtered images are realigned and thresholded to obtain the binarized image shown in (C). (D) Positions of the manually realigned coronal stained slices, shown in blue on a sagittal slice of the CCFbbp brain volume. https://doi.org/10.1371/journal.pcbi.1010739.g003 The details of this automatic alignment method are described in a separate paper [56]. The different steps of the automatic realignment of every ISH dataset from the AIBS to the reference atlas are as follows: Original ISH images (or raw images—see Fig 3A) and their filtered versions, later called filtered images (see Fig 3B), are downloaded from the AIBS website. The filtered images isolate the somas reacting to their specific marker from the brain background in the raw image.

Each image is scaled, positioned, and rotated according to the reference brain volume, using the provided metadata. The metadata includes the position of the corners of the images because the slices are not perfectly vertical. Here we assumed the slices to be vertical as the algorithm was trained with this assumption and it makes the interpolation between slices easier. Hence, we used the mean position of the coronal ISH images in the AIBS average brain along the rostral-caudal axis also from the metadata. The ISH brain slices, the AV2 and the Nissl2 volume have been all registered within the average brain [62]. So, the ISH coronal section ids should correspond to the Nissl2’s. For the Missl1a, used here, for each ISH coronal slice to realign, we manually selected the best corresponding reference Nissl slice.

For each dataset, every coronal section is automatically realigned to the corresponding anatomical section of the Nissl-stained mouse brain using the Krepl algorithm [56]. Here, the raw in situ images are used as they provide more landmarks for the algorithm.

Resulting displacement fields are then applied to the filtered images of the corresponding ISH experiment.

A 3D volume for each gene is created by downsampling the images to match the final voxel dimensions of the reference atlas. Missing sections between the 2D coronal slices are generated with a linear intensity-based interpolation. The resulting 3D voxel-based datasets are consumed by our pipeline to estimate inhibitory neuron subtype densities (see Section 2.6 and S8 Fig).

2.4. Assumptions underlying the estimation of inhibitory subtypes densities We wish to obtain the number of inhibitory neurons and their subtypes for each region of the AV. The regions are indexed by region ids R = {0, r 1 , r 2 , …, r n } ϵ |N861, with 0 denoting the outside of the mouse brain. We are particularly interested in inhibitory subtypes expressing the markers GAD67, PV, SST and VIP. First, the number of neurons can be expressed as the sum of the different subtypes in each region r, ∀ r ϵ R: (1) where nNeu, nInh, and nExc are respectively the total, the inhibitory, and the excitatory neuron numbers. As reported by Zeisel et al. [69], these populations seem to be mutually exclusive. The term nOther represents the number of neurons that are neither excitatory nor inhibitory, meaning they could be purely modulatory neurons. The number nInh can be estimated by the number of neurons, reacting to the markers GAD67 and GAD65 [69] which implies that nInh = nGAD. We further subdivide nGAD into inhibitory subtypes positively reacting to the markers PV, SST and VIP. This yields the following sum, ∀ r ϵ R: (2) Here nPV, nSST and nVIP are respectively the number of PV+, SST+ and VIP+ immunoreactive neurons in the brain region r. The term nRest corresponds to the number of inhibitory neurons which react neither to PV, SST nor to VIP (InhR neurons), including, for instance, the LAMP5 cells in the isocortex [58]. InhR neurons might define various populations of GABAergic cells, reacting to a great variety of markers [69], which makes it more difficult to estimate nRest. It is therefore easier to estimate nGAD from literature and the marker GAD67 (for regions not covered by literature). Similarly, nPV, nSST and nVIP can be estimated from literature and the markers PV, SST and VIP, respectively. Finally, we can estimate nRest by subtraction in Eq (2). The remaining neuronal populations (nExc + nOther) can be deduced from nGAD (Eq (1)) and the neuron distribution (nNeu) obtained by step 2 of our BBCAv2 pipeline (see Fig 1). In this estimation, we rely on the following assumptions: Every GABAergic neuron expresses GAD67 and every GAD67 reacting cell is a GABAergic neuron. This genetic marker is indeed responsible for over 90% of the synthesis of GABA [70,71]. Additionally, no cells expressing GAD65 without expressing GAD67 have been reported in the RNA-sequence study from Zeisel et al. [69]. GAD67, PV, SST and VIP are only expressed in neurons [55,69,72,73]. PV+, SST+ and VIP+ populations are non-overlapping i.e., there are no cells in the mouse brain that co-express a combination of these markers. This assumption is supported in the isocortex and hippocampus by transcriptomic studies such as Huang and Paul [58] and Zeisel et al. [69] and we extrapolate these findings to other areas. Every PV+, SST+ and VIP+ neuron also expresses GAD67 as observed by Celio [74] and Tasic et al. [75]. Neuronal composition is homogeneous within subregions at the lowest level of the AIBS region hierarchy. We will also consider the cell, glia and neuron density distributions obtained in step 2 of the pipeline (see Fig 1) to be correct. These numbers have been validated against literature by Erö et al. [17] and can be refined in the future. Based on these assumptions, BBCAv2 will provide estimates of the densities of GAD67+, PV+, SST+, and VIP+ neurons for each region of the brain. In contrast with BBCAv1, this new version will present estimates that are as close as possible to available literature values.

2.7. Combination of neuron type densities Since our unconstrained estimates of inhibitory neuron (η and σ) densities are derived from the transfer functions or independent literature sources, their combination can lead to incongruent results. For instance, the unconstrained estimates of PV+ neurons might be greater than the estimates of GAD67+ neurons, which contradicts our assumption 4 (see Section 2.4). We therefore want to ensure that η and σ match our assumptions from Section 2.4. To this end, we deduce a list of linear constraints based on Eqs (1) and (2), ∀ r ∈ R: (6) and similarly for nSST r , nVIP r and nGAD r . (7) We also ensure that the consistency of the region hierarchy is maintained for every parent region (R m ∈ R). The number of inhibitory neurons (and their subtypes) in R m is equal to the sum of the corresponding estimates of R m ’s direct child regions in the region hierarchy (children Rm ) plus the estimates in voxels of the AV labeled as belonging to R m but none of its children (R m \child): (8) and similarly for nSST Rm , nVIP Rm and nGAD Rm . Thus, whenever one of the constraints (6), (7) or (8) is violated, we rescale our unconstrained estimates (η). We minimize the amount of corrections required to find a solution for the BBCAv2 model through optimization. The amount of corrections is defined as the sum of the distances between the unconstrained (η) and corrected values counterparts (x) divided by the standard deviation of the unconstrained value (σ): (9) This nonlinear function is convex which guarantees a global minimum. We convert it to a linear problem, by introducing a slack variable z so that ∀ r ∈ R: (10) Hence, our problem becomes: (11) st. Eqs (6), (7), (8), (10) The solution x is a vector with 3444 values and matches a total of 8613 linear constraints. An initial solution can also be found for each region, correcting the unconstrained estimates to match our constraints starting from leaf regions in the region hierarchy to the top-level brain regions (see S1 Document, S2 and S3 Figs). We use the simplex algorithm from the scipy python library [93]. We find that for 16% of the regions of the brain (shown in colors in Fig 6A and 6B), their η estimates are incongruent with the rest of the brain. Their corrected values x are therefore not falling within the range η ± σ. In some regions of the striatum (including caudoputamen), the number of GAD67 neurons from η overshoots the estimated total number of neurons. Our optimization makes these regions fully inhibitory. This is in line with literature findings on the mouse striatum, which describe it as almost fully inhibitory [94]. Among the rest of the regions where significant corrections are needed, almost one-third are subregions of the hindbrain, for which we collected very few literature values. Conversely, few inconsistent first estimates are found in subregions of the isocortex, for which more literature data are available. The remaining inconsistencies for η may have different sources: They can be explained by a poor estimate of the densities from the transfer function.

Divergent estimates from the literature for the same region might violate Eq (7). These regions would appear in yellow on Fig 6A and 6B).

Our assumptions may not be accurate in some regions of the brain.

The densities of neurons from the step 2 of the pipeline (see Fig 1), used in Eq (6) and (7), diverge locally from the literature findings. PPT PowerPoint slide

PNG larger image

TIFF original image Download: Fig 6. Results of the density pipeline. (A) Sagittal and (B) coronal view of the Nissl reference atlas used for BBCAv2, showing cells of the mouse brain. Regions with high cell density appear in dark grey. Regions that required significant corrections (see Section 2.7) for their neuron subtype density estimates are colorized (Neu = neuron). The coronal slice chosen is displayed as an orange line on the sagittal slices. Arrow in black shows the location of the lateral ventricle of the mouse brain. The blue line behind the sagittal slice highlights the drop of Nissl expression coming from the original Nissl experiment from Dong [18]. (C). Sagittal and (D). coronal view of the BBCAv2, showing the positions of the different types of neurons. PV+, SST+ and VIP+ cells appear respectively in yellow, dark orange and blue. The rest of the GABAergic population is color-coded in green and the remaining neurons in gray. The variation of the distribution of neurons is following the original distribution of Nissl expression. https://doi.org/10.1371/journal.pcbi.1010739.g006 After our corrections, we obtain consistent density estimates for PV+, SST+, VIP+, and GAD67+ neurons for each region of the AV.

[END]
---
[1] Url: https://journals.plos.org/ploscompbiol/article?id=10.1371/journal.pcbi.1010739

Published and (C) by PLOS One
Content appears here under this condition or license: Creative Commons - Attribution BY 4.0.

via Magical.Fish Gopher News Feeds:
gopher://magical.fish/1/feeds/news/plosone/

You are viewing proxied material from magical.fish. The copyright of proxied material belongs to its original authors. Any comments or complaints in relation to proxied material should be directed to the original authors of the content concerned. Please see the disclaimer for more details.