Whole human-brain mapping of single cortical neurons for profiling morphological diversity and stereotypy

Abstract

Quantifying neuron morphology and distribution at the whole-brain scale is essential to understand the structure and diversity of cell types. It is exceedingly challenging to reuse recent technologies of single-cell labeling and whole-brain imaging to study human brains. We propose adaptive cell tomography (ACTomography), a low-cost, high-throughput, and high-efficacy tomography approach, based on adaptive targeting of individual cells. We established a platform to inject dyes into cortical neurons in surgical tissues of 18 patients with brain tumors or other conditions and one donated fresh postmortem brain. We collected three-dimensional images of 1746 cortical neurons, of which 852 neurons were reconstructed to quantify local dendritic morphology, and mapped to standard atlases. In our data, human neurons are more diverse across brain regions than by subject age or gender. The strong stereotypy within cohorts of brain regions allows generating a statistical tensor field of neuron morphology to characterize anatomical modularity of a human brain.

ACTomography acquires and profiles morphological diversity and stereotypy of single neurons mapped in whole human brains.

INTRODUCTION

Study of the human brain has obtained great attention from modern brain science (1–3). Despite the existence of many animal models to understand brains (4–12), it is still of utmost and profound importance to examine surrogates of human brain directly (13, 14). Acquiring neuron data from postmortem brains has been a powerful way to produce human brain map models (15, 16) and to discover previously unknown cell types (17). Alternatively, screening of necessarily extracted brain tissues along a surgical path to remove brain tumors and/or other disease sites offers another opportunity to profile states of human brains. Previous efforts such as in vitro single-cell characterization of human brains have started to generate three-dimensional (3D) morphology (18, 19), electrophysiology, and transcriptional profiles of single cortical L5 neurons from surgically extracted brain tissues (20). Over the years, large datasets of single–human neuron morphologies consisting of several hundred injected and traced human neurons were documented in frontal, temporal, and occipital cortex (21–23); anterior temporal lobe (24); and CA1 (25). Other relatively large-scale studies of human neuron morphology, until now, have been conducted using bulk staining such as Golgi (26–29). The field calls for a scalable platform for routine production of human neuron morphology with sampling and target-controlling capability.

Human ex vivo brain slices are often produced from surgically extracted brain tissues for collecting various physiological, morphology, and molecular attributes of individual cells. The ex vivo slices may also be sectioned from freshly dissected tissues in donated postmortem brains. Because of the requirement of minimal invasive surgical procedures, truncation of neurite structures is almost inevitable especially for long projecting excitatory neurons. Even for postmortem brains, damage of neurites is also hard to avoid during tissue extraction. On the other hand, in mammalian model animals such as mouse, it is known that quantification of individual cells’ structures at the whole-brain scale is essential to study the complexity and diversity of cell types (11). For human brains, although biopsy and autopsy tissues were also used (25), it is an open challenge how to collect a sufficient amount of neuron morphology data to map and profile whole brains of individual neurons at the single-cell resolution. The limited availability of genetic and viral labeling methods applicable to ex vivo human brain samples and the lack of computational analysis methods specifically developed for human neurons combine to form further challenges for the field.

We note that much of the existing effort in single-cell characterization of human neurons is related to molecular, e.g., transcriptomic, analyses of cells (17, 20, 30, 31). In general, the limited platform technologies to screen human neuron morphology also make it hard to connect single-cell analyses of morphological and molecular attributes. Nonetheless, the recent advancement of transcriptomic classification (20) and Patch-sequencing (32, 33) characterization of single cells shows the possibility to produce morphological classifier that could be used in a comparative analysis of these datasets. In addition, the increasingly available detailed maps of cell-type distributions produced using spatial methods such as multiplexed error-robust fluorescence in situ hybridization (MERFISH) (31, 34) may help to make connections to neuron morphologies because there are very tight laminar distributions in different cortical areas. Therefore, it is becoming desirable to establish a scalable technical platform to routinely produce and analyze single–human neuron morphometry, which would be iteratively refined to eventually facilitate classification of human neurons using large-scale–conjugated analyses of molecular, physiological, and morphological features.

As one of the first steps to address this massive challenge, this study proposes a low-cost and high-throughput method to screen human neurons at the human brain scale, based on ex vivo brain slices. Our method focuses on generating morphological profiles of individual neurons as a framework with potential extension to other cellular attributes. Here, we describe this methodology and apply it to study both the diversity and stereotypy of human neurons in the morphological space. Our data suggest a hypothesis of morpho-gapping modularity of human brains at the single-neuron resolution.

RESULTS

Adaptive cell tomography and human ex vivo cell injection at whole-brain scale

We developed an approach called adaptive cell tomography (ACTomography) to build 3D neuron and distribution models from locally observed attributes of individual neurons (Fig. 1). Here, we limit the discussion to morphological attributes, or 3D morphology, of human neurons. Our method begins with collection of small trunks of freshly dissected human brain tissues in craniotomy path to remove tumors/lesions in deeper brain regions (Fig. 1, A to C) or tissues extracted from a postmortem brain. We produced thin, fixed brain slices (Fig. 1D), which allowed clear identification of neuron somas using differential interference contrast (DIC) imaging (Fig. 1E and Materials and Methods). We also used DIC imaging to guide the injection of dye [Lucifer yellow (LY)] into each of the selected somas (Fig. 1E). After the dye diffuses from a soma saturating the neurite region (fig. S1), we acquired a full 3D image stack of an individual neuron using two-photon laser scanning microscopy (Fig. 1F and Materials and Methods). We then traced a neuron’s morphology in 3D to reconstruct its shape digitally (Fig. 1G). This method was applied to as many neurons as possible on ex vivo brain slices, which were also mapped to the same spatial coordinate framework to produce a large pool of digital morphology representations of individual neurons. Next, for every neuron, we identified its clique of similar neurons (Fig. 1H), building a 3D statistical neuron model based on its clique (Fig. 1I) and retaining the single-cell resolution at the whole-brain scale. Statistical modeling of neurons, as shown later here, also makes it easier to cluster brain regions.

Fig. 1. Schematic illustration of ACTomography for building neuron models in a number of brain regions at single-cell resolution.

Method is shown with surgically extracted brain tissues and is also applicable to postmortem brains. (A) Magnetic resonance imaging (MRI) T1 scan of a patient and schematic display of a small region (magenta) from which normal brain tissues would be extracted during surgery to remove the deep tumor/lesion along the “path” (green). (B) Example of a tumor in surgery. (C) Extracted human brain tissue, ~1 cm³ in this case. (D) Ex vivo slice (200 μm thick). (E) Differential interference contrast (DIC) imaging of the brain slice. Scale bar, 20 μm. (F) Two-photon 3D imaging of an injected neuron (Maximum intensity projection image is showed). Scale bar, 20 μm. (G) 3D traced morphology of the neuron in (F). (H) The clique of neurons that share similar features with (G). (I) Statistical analysis of a neuron clique to build a 3D neuron model that is also mapped to the approximate standard brain region.

ACTomography has two major components: adaptive cell targeting (ACT) and single-cell tomography. Because the choice of somas for dye injection can be both selective (i.e., applied only to specific neurons or neuron types) and nonspecific (i.e., injection into every possible type of neurons visible in DIC imaging), our approach is capable for targeting and profiling a wide range of neurons adaptively, limited only by the availability of human ex vivo brain slices. Previously, we succeeded in developing ACT for both in vivo single-cell electrophysiology (35) and 3D morphology reconstruction of large human neurons of ex vivo brain slices (18) as well as electroporation-based 3D tracing of in vivo mouse neurons (35). Therefore, a key step to make ACT work for ACTomography is to sample a large enough number of neurons independently to increase the success rate in reconstructing 3D models of neurons, based on limited and often truncated neurites in human brain slices. As our previous experimental protocol for ACT-based dye injection into human neurons was time-consuming, here, we developed an inexpensive and high-throughput method to increase the sampling rate of neurons to improve the usage efficiency of human tissue while, at the same time, increase the availability of human brain specimens especially in different brain regions.

We thus established a whole-brain–scale workflow to collect human ex vivo cell injection images. Human ex vivo brain slices (thickness of 200 μm in most cases) that contain “normal” neurons were produced for 8 males and 10 females, who suffered from severe tumors such as glioblastoma or other neurological conditions but survived after brain surgeries, and the donated postmortem brain of 1 male died of glioblastoma (table S1 and Materials and Methods). To maximize the potential to use brain tissues from subjects available only at different geographical regions, we introduced a fixation-and-injection approach so that ex vivo tissues could be prepared and shipped out to our processing center 1000 km away without severe restriction of the freshness of brain tissues at the time of injection (Materials and Methods). For each neuron, we controlled the total dye injection and diffusion time to be around 18.6 ± 3.4 min (fig. S1), which could be further parallelized, followed by two-photon imaging to acquire a 3D image stack of the neuron quickly (t = 98.1 ± 23.7 s). In this way, we acquired totally 3D image stacks of 1746 neurons (figs. S2 and S3) of the 3231 injected neurons in eight major nonoverlapping brain regions and two overlapping regions (semantically in Fig. 2A and more precisely in Fig. 3). Now, this dataset is among the largest human single-neuron archives that has been produced. The aggregated production rate of both efficacy in image acquisition over cell injection (54.0% on average, within the range of 25 to 80% depending on the actual human tissue samples) (Materials and Methods) and neuron image production speed (~20 min per neuron) in ex vivo slices was at least 20 times more efficient than our previous cell-injection attempt (e.g., 18).

Fig. 2. Semantic mapping of 1746 3D injected, imaged, and 852 traced human neurons in eight exclusive brain regions (irregular shape) and two overlapping regions (oval) at whole-brain scale (abbreviations in table S1), along with feature analysis of traced neurons.

(A) Sematic mapping scheme. Each brain region is shown in a different color along with the number of imaged neurons in it. Images (maximum intensity projection) and morphology (green) of one or two representative neurons are also shown for each region (some images are zoomed in for clear visualization; scale bars, 20 μm). Note that the actual locations of brain tissues extracted in this study as more precisely illustrated in Fig. 3 were much smaller than the color-labeled areas here. The brain template is based on (58). (B) 4D feature relationship among neuron bifurcation, length, number of stems, and volume of 852 neurons. (C) 4D feature relationship among neuron branching number, average fragmentation score, Hausdorff dimension, and volume. (D) 4D feature relationship among neuron maximal path distance, surface, average bifurcation angle (local), and volume. Optimal polynomial fitting curves (red) are also shown for (B) to (D).

Fig. 3. Whole-brain computational mapping of tissue extraction regions of human ex vivo brain slices, based on registration of postoperational MRI scans (for surgical tissues) or the last MRI scans (for postmortem brains).

For each panel H001 to H019, the cross-sectional view shows the envelope of brain tissue extraction region (red), 3D registered to a standard brain coordinate template MNI152. Scale bar, 20 mm and the same for all cross-sectional panels. The 3D views show both disease sites (green) and brain tissue extraction regions (red). Bottom right shows all brain tissue extraction regions, each in a different color, mapped and overlaid on MNI152. L, left; R, right.

In the dye-injection step, as we have not restricted to specific neuron types except the soma visibility and the soma volume in DIC imaging, our current collection more likely contains large cells like pyramidal neurons in L3 and L5. We can also observe subneuronal structures, such as spines, in neurons (see fig. S4 for examples). We could also determine the layer of neurons via prestaining of Nissl or 4′,6-diamidino-2-phenylindole (DAPI) (fig. S5). To validate that the acquired images recapitulate the “true” neuron morphology faithfully, we also used our approach to inject 491 mouse neurons in freshly dissected whole mouse brains. The efficacy of image acquisition over cell injection was 83.1%, reasonably higher than the upper range of efficacy in our human cell injection for good ex vivo tissues, due to the overall better condition to extract and use mouse brain tissues. The resultant 3D image datasets of 408 mouse neurons were also compared to Cre-labeled and fluorescence micro-optical sectioning tomography imaged mouse neurons (11) of the same types, as exemplified in caudoputamen neurons (fig. S6).

Whole-brain mapping of 3D imaged and traced human neurons

We mapped the locations of extracted brain tissues and the accordingly acquired neuronal images to standard brain atlases, both computationally (Fig. 3 and Materials and Methods) and semantically (Fig. 2A). Of note, the actual brain tissues extracted for this study occupied only much smaller regions (Fig. 3) than the semantically depicted cortical regions (Fig. 2A). The estimated Brodmann areas of the extracted tissues from each patient are summarized in table S1. This whole-brain mapping paradigm can assemble precious morphological, distributional, and associational information of single neurons.

We traced the 3D morphology of 852 neurons that arborize heavily in dendrites (Materials and Methods). The traceable neurites are abundant during careful visual inspection with adjusted contrast of the neuronal images (Fig. 2). The five largest brain regions containing these traced neurons are middle frontal gyrus (MFG) (n = 304), temporal pole (TP) (n = 174), superior temporal gyrus (STG) (n = 131), inferior parietal lobule (IPL) (n = 112), and parietal lobe (PL) (n = 91). We calculated 22 global morphological features (Fig. 2, B to D, and Materials and Methods), which are independent of the orientations of neurons. Distributions of exemplar global morpho-features are also shown in figs. S7 and S8. Complex relationships among these features not only could be modeled using various data fitting methods but also exhibit clear redundancy (Fig. 2, B to D). Therefore, the feature space was further reduced to six dimensions {#Stems, #Branches, Length, Average Bifurcation Angle - Local, Average Bifurcation Angle - Remote, Hausdorff Dimension}. The two pairs of features, {#Branches, Length} and {Average Bifurcation Angle - Local, Average Bifurcation Angle - Remote}, show well-expected correlation among themselves (Fig. 4, pair plot) that also indicates that neuron morphology in this study is consistent as previously documented human neuron morphology at NeuroMorpho.Org (36). The distribution of #Stems and Hausdorff Dimension features is also consistent with previous documented human neurons. Overall, these two pairs of features and two remaining features, #Stems and Hausdorff Dimension, form a compact subspace to characterize the feature space of traced neurons in these five brains regions (Fig. 4), also in all 10 brain regions (fig. S9).

Fig. 4. Analysis of global morphological features indicates greater diversity of human neurons across brain regions than age or gender of subjects and the stereotypy of neuron distribution patterns.

(A) Analysis of the five brain regions with most reconstructed neurons. Pair plot of feature distributions (left): Six key global morphological features reduced from the original 22D space. Uniform Manifold Approximation and Projection (UMAP) plots (right): The feature space was further reduced using UMAP along with color codes indicating three different groupings by brain region, age, and gender. (B) Detailed analysis of two brain regions, PL and IPL, similar to (A). Two examples of 3D traced neurons for each region are shown, along with patient ID and neuron Identity document (ID).

In addition, we compared our traced neurons (n = 18) with three previous datasets available from Neuromorpho.org, including human neurons via Golgi staining (n = 100) (37), vervet (n = 12), and baboon (n = 12) neurons stained with LY via cell injection (4). The comparison was focused on brain regions of Brodmann area of 6, which was shared by all our and the other three datasets (figs. S10 and S11). We calculated the L-measure features of all neurons in two cases: (i) using all data as it is or (ii) using morphologies no further than (120, 120, 70) μm in xyz from the somas. The second case allows to compare the local fiber structures of different datasets more objectively. Despite the limitations of our approach to access distal neuron structures (lower maximum path distance), our human data contain richer local morphologies than those from Golgi staining, as it is seen in relatively higher numbers of stems, tips, bifurcations, and other features. The cross-species comparison also indicated that human neurons are more complex than those of vervet and baboons, with higher number of stems and larger angles of the bifurcations and parent-daughter ratio.

Diversity and stereotypy of human neuron dendritic arbors

We analyzed dendritic arbors of all traced neurons to study neuron stereotypy, which is closely related to the diversity of neurons’ morphology with respect to different groupings of our data. We first considered the orientation-independent morphological features. A Uniform Manifold Approximation and Projection (UMAP) analysis indicates that IPL, STG, PL, and MFG neurons form a visible gradient of their feature distribution, while TP neurons spread out more in the feature space (Fig. 4A, UMAP plots). In clear contrast, the age and gender of the human subjects do not differentiate these neurons (Fig. 4A). Thus, these neurons are more diverse across brain regions than the age or gender. UMAP analysis also indicates that neurons’ 3D morphology is not correlated with human subjects’ age and gender and is more stereotyped within a specific brain region than across different brain regions (Fig. 4A). In particular, the two neighboring brain regions, IPL and PL, have distinctively conserved neuron morpho-features, as indicated by all individual key features (Fig. 4B, pair plot) and the UMAP analysis (Fig. 4B, UMAP).

Although the 22D morpho-features did not yet produce substantial overall differences between neurons with varying orientations, we further validated the interareal diversity of neurons by examining neurons that share similar orientations in the raw images. We classified 3D traced neurons into three pools: pool-A, pool-B, and pool-T. Pool-A contains neurons that have well-aligned and visible apical dendrites in the xy imaging plane (figs. S12A, S13, and S2). Neurons in pool-B have their apical dendrite oriented axially along the z axis, i.e., orthogonal to the xy imaging plane, so that basal dendrites are mostly visible in the raw images (figs. S12B, S13, and S3). Cortical neurons in pool-A and pool-B also correspond to the conventionally named “coronal” and “horizontal” views in previous studies. Pool-T contains all neurons with a tilted orientation in raw images. When subgroups of neurons sharing similar orientations (fig. S12) were compared with the situation that all neurons were analyzed together (Fig. 4 and fig. S9), we observed similar continuum distribution and/or separation patterns, providing additional data to show that human neurons have greater diversity across brain regions than across gender or age.

We also considered other types of neuronal features in the orientation-classified analysis. We clustered the Sholl analysis (38) profiles of neurons that share similar orientations (Fig. 5). It is clear that most neurons innervate heavily around 50 μm from somas, while some subgroups arborize more dominantly around 150 μm from somas. Pool-A and pool-B neurons form two and three major subgroups, respectively. They have evidently lower entropy values than brain region–based groupings (using IPL, PL, STG, TP, and MFG that contain most traced neurons) or random subgroupings (Fig. 5, bar chart), implying the neuron arbors do form clusters in the space of Sholl analysis profiles. However, the gain of the normalized mutual information (NMI) between these subgroups and the five brain regions as examined (Figs. 2 and 3 and table S1) is three to four times greater than the NMI scores between random subgrouping and the brain region groupings (Fig. 5, bottom left, bar chart). This shows that the neuron classes that arborize in varying density correlate with the cortical regions that we examined, consistent with observation in Fig. 4 based on global morphological features. The joint pattern of (i) clear clustering in the feature space and (ii) the correlation with brain region distribution indicates that different types of neuron dendritic arbors in our data may spatially interlace with brain regions and potentially form a manifold across human cortical regions.

Fig. 5. Orientation-classified analysis of human neuron distribution.

The top panel shows the Sholl analysis profiles, i.e., the number of branches of a neuron at specific distance to its soma. Neurons in pool-A and pool-B are separately clustered on the basis of cosine similarity. For clear visualization of the generalizable similarity matrices, the sorted Pearson correlation (similarity) matrices matching the orders of clustered neurons at the top panel are also aligned at the bottom panel. Pool-A and pool-B neurons can be subgrouped into two and three clusters, respectively, as indicated by red double arrows. The entropy values of such subgroupings [H(sg)], random subgrouping [H(rand)], and brain region–based subgrouping [H(rgn)] are shown in the left-side bar chart. The gains of the normalized mutual information [Gain(NMI)] scores are also displayed.

3D neuron models and distribution models at the human brain scale

The strong morphological stereotypy of neurons allows us to use feature search, similar to BlastNeuron (39), to detect N-nearest neighbors of any target neuron, ideally also within each pool, to form a clique of similar neurons (Fig. 1H). This method filters out morphologically dissimilar neurons that may be close in global feature space due to orientation. For each neuron clique, it is reasonable to assume that individual neurons are more likely from the same neuron type and share similar neuronal features. It then allows building a statistical neuron feature model that could generate new simulated neuron structures in 3D (Fig. 1I). Specifically, using the 22D morphological features as an example, we detected the top five nearest neighbors in the feature space and computed the mean and median feature vectors, as well as their variation feature vectors. This led to a 66D mean feature vector and a 66D variation vector of each neuron. We also estimated the overall mean and variation feature vectors for all traced neurons as the global background features, which were used to standardize the individual mean and variation features of single neurons. Optionally, we further normalized the mean vector using the respective variation, so that individual feature dimensions could also be compared with each other.

We examined the distribution and their associations among morphological features by removing global background. We found that morphological features in different brain regions could be well modeled by covarying patterns. For instance, for three anatomically nearby brain regions, inferior/middle/superior frontal gyri, the mean features in SFG tend to covary linearly with those of MFG (negative correlation) and IFG (positive correlation) but are relatively independent of the variation features of these regions in a 4D correlation analysis (Fig. 6A, top left panel). However, the variation features of these brain regions do not exhibit clear correlation among themselves and are also independent of the mean features (Fig. 6A, middle left panel). When the variation features were used to further normalize the mean features, we observed a strong condensation of features to the range of 0.75 to 1.5, while the entire distribution could be modeled by a second-order polynomial (Fig. 6A, bottom left panel). Very differently, for another pair of nearby brain regions, PL and IPL, their mean features cannot be modeled linearly (Fig. 6A, top right panel), while their variation features exhibit weak negative correlation (Fig. 6A, middle right panel). However, the variation-normalized mean features are also condensed to a narrow window of 0.8 and 2, while the entire normalized distribution can be approximated by a second-order polynomial (Fig. 6A, bottom right panel).

Fig. 6. Generation of the brain region–wise neuron feature tensor field and the respective variation tensor field at the single-cell resolution.

(A) Brain region–wise multidimensional correlation analysis of mean feature vectors (top panels), variation features (middle panels), and the variation-normalized features (bottom panels). “-std,” SD; “-n,” normalized by SD. Note both color and size of the data points are used to indicate additional dimensions other than the x and y axes in the analysis. Optimal regression curves are also shown, with translucent regions indicating 95% confidence intervals of regressions. (B) Variation-normalized 66D mean feature map of all brain regions. Feature value range was clipped at 2 for clear visualization. (C) Biclustering map of mean features; the cosine similarity is shown in each bin. (D) Biclustering map of variation-normalized mean features in (B); the cosine similarity is shown in each bin. (E) The sorted mean feature tensor; sorting order is the same as in (C). (F) The sorted variation feature tensor; sorting order is the same as in (C). (G) 3D visualized ex vivo brain regions of this study; color scheme, same as in Figs. 2 and 3. (H) Spatial adjacency map of brain regions in this study. Euclidean distances, d, between pair-wise representative medoids of extracted and 3D registered brain tissues were computed; spatial adjacency = exp(−d). (I) Spatial adjacency–augmented mean feature tensor field. (J) Spatial adjacency–augmented variation feature tensor field.

We used mean features and variation features to investigate the distribution of single neurons in different brain regions. The variation-normalized mean features characterize different brain regions clearly (Fig. 6B). Biclustering the similarity among these region-wise feature vectors shows details of neuron distribution patterns in brain anatomical regions. We identified two clear cohorts of brain regions [IPL, MTG, SFG, and IFG] and [PL, S(M,I)FG, STG, S(M)FG, MFG, and TP] that have different morphological mean feature vectors (Fig. 6C), while the variation-normalized mean features are biclustered in a greater hierarchy across brain regions (Fig. 6D). In both cases, PL versus IPL features are evidently far away from each other. We also noticed that, in both cases, S(M,I)FG and S(M)FG features are close to MFG, which indicates that the respective traced neurons might be actually more adjacent to or in the MFG region. In addition, in both cases, SFG/IFG features always differ from MFG features, consistent with the negative correlation observed in Fig. 6A (top left panel). Together with the observed contrast between the PL/IPL pair, our data suggest a hypothesis of gapping modularity of neuron morphology distributed across nearby brain regions.

Therefore, we were able to model the brain-scale interareal association of individual neurons’ morphology using tensors, which describe multiregion multidimensional statistics of neurons across the brain (Fig. 6, E to J). In particular, we generated the mean tensor (Fig. 6E) and variation tensor (Fig. 6H) of the brain regions involved in this study, based on the two cohorts of anatomical regions identified in Fig. 6C. These tensors demonstrate clear modules of the organization of brain anatomy, indicating that the dendritic features of neurons are strongly conserved within each cohort but interlace across brain areas. Within each cohort, the dendritic features have different levels of variations, which could be further modeled by using additional combinations of features in future work. Considering the 3D spatial locations of the extracted brain tissues (Fig. 6F), we also computed a spatial adjacency map of these brain regions (Fig. 6I). We aggregated the spatial closeness of brain regions and the feature tensors and generated the spatially modulated morphological feature tensor field (Fig. 6G) and the respective variation tensor field (Fig. 6J). Both augmented tensor fields uncover the rich structures of dendrite profiles of individual human neurons that are embedded deeply in the neuron-wise distributions shown in Figs. 4 and 5. The strong stereotypy as exhibited by localized morpho-features also indicates that statistical modeling of neurons’ 3D shape and neuron distribution can be reasonably achieved by applying adaptive cell tomography at the human brain scale.

DISCUSSION

In this work, we proposed ACTomography to acquire one of the largest known single-neuron morphology datasets for human brain. Different from several recent successful single-cell morphology studies on model animals such as mouse (11, 40) and monkey (41), for human brains, there lacked a suitable large-scale technical platform to conduct similar studies. The pioneering efforts of cell injection, imaging, and morpho-analysis were promising (18, 21, 24, 25, 35, 42–45) yet also had a limited scale for human brains. In comparison with previous approaches, the primary goal of this paper is to provide a generalizable framework for setting up a high-throughput platform of human neuron studies including the modules of cell injection, imaging, neuron tracing, brain registration, pattern analysis, and modeling of neuron morphology and brain anatomy. Our approach is suitable for both brain tissues necessarily extracted in surgeries to remove deep brain tumors and lesions or in donated postmortem brains. Despite the limitation to access the complete structure of human neurons, which is the main drawback by using brain slices, our data do show superiority in capturing more local structures in comparison with previous Golgi-stained human data. While we restricted the comparison within the same Brodmann area but not the same layers, this finding does generally implicate that the studies using Golgi stains of the brain neurons may require revision. The rich information produced here shows that, with this low-cost platform, we were able to achieve high efficacy of successful neuron targeting and cell injection to produce reasonable 3D dendritic morphology in fixed ex vivo brain tissues. The number of human neurons and mouse neurons in our experiments was large enough to provide statistical validation of our approach in comparison with previous studies of neuron labeling, imaging, and morphology analysis. The efficacy ratio in harvesting neuron data was also sufficiently high, which makes it desirable to boost the throughput further by automation and parallelizing soma detection, neuron-type identification, cell injection, and imaging. Optimization in the system is also desired to manage thicker and larger tissue slices to access more complete neuron structures. Despite that the neurons were acquired from specimens remote to the disease loci, we cautiously note that the neurons may still bear certain morphological changes due to pathological changes.

We have used ex vivo brain tissues from 19 subjects covering eight exclusive and two overlapping cortical regions while, at the same time, used registration methods to spatially map these regions at the whole-brain scale. This approach sets up a scalable framework to add more brain regions and cell types into this whole-brain map of single-cell morphologies. Now, ACTomography allows us to adaptively select neurons whose somas are visible for injection. In cell selection, although, here, we focus on reporting injection and analysis of pyramidal neurons using DIC imaging, ACTomography is not limited by the imaging modality. First, various wide-field costaining methods, such as DAPI labeling or specific sparse-labeling techniques, could be applied at the same time so that we may also inject dyes into a variety of potential cells under fluorescent microscopes. Second, we could deploy an injection array or parallelize the injection to many loci simultaneously and identify the successfully injected neurons in a post hoc way to triage specific neuron types. Our fixation-and-injection approach lowers the requirement of the freshness of brain tissues for injection. In turn, it maximizes the possibility to screen more neurons from different sources and conditions.

Our analyses provide a general assessment of the “background” dendritic model of neurons and their anatomical distribution, based on multidimensional correlation of morphological features and brain regions, of normal brain tissues. To map the diversity and stereotypy of neurons, we have used both semantic and computational brain mapping to pinpoint neurons’ anatomical locations. We found that dendritic features of human neurons in our data correlate with major cohorts of brain regions [cf., e.g., (46) for similar rodent finding and (37) for a human Golgi study]. Such dendritic features do not exhibit strong correlation with gender or age of the subjects. This means that the differences between cells in the same areas of different individuals (despite age and sex) are probably small in comparison with the differences between cortical areas. It could be a point of validation of previous work, which typically relied on small samples and, in many cases, with each individual contributing cells from one area. The spatial distribution of brain regions can also be fused with the single-neuron–level dendritic feature maps as tensor fields to model and visualize the anatomical modularity and complexity implied by the diversity and stereotypy of morpho-features of neurons. This study provides an orthogonal, complementary paradigm to previous work on analyzing morphological property and distribution of mammalian neurons at the whole-brain scale [e.g., (11)]. Rich information can be well anticipated when more data, especially metadata, become increasingly available in continuing efforts, such as the cortical-layer, spiny, left and right hemisphere information of these neurons that we would report in a sister piece of this study, functional and behavioral data of the subjects, comparative data of “abnormal” cells in the near-lesion or in-lesion brain tissues, and many other brain regions not reported in this study.

In contrast to other model animals, one of the key challenges to study human neurons’ morphology at whole-brain scale is about the completeness and fine details of neuron shape, especially arborizing patterns of long projecting axons and apical dendrites. An intrinsic limitation is due to the ex vivo nature of human brain research in this direction. For single-cell injection technologies, an obvious improvement is to optimize tracer (e.g., dye) diffusion along neurites. While we may switch tracers and also keep upgrading tracer injection techniques, we note that one strength of ACTomography is the fast and effective accumulation of a number of locally visualized human neurites in precisely pinpointed anatomical regions. The strong stereotypy of intraregion neuron features suggests two approaches to obtain more complete shape information of individual neurons. First, for an experimentally observed target neuron, we may aggregate local information of its neurites in a mean model along with estimated variations of morphological dimensions, as shown in Fig. 6, and therefore synthesize the missing structures by learning from existing features of nearby neurons. Second, we may generate new neurons by integrating morphology tensor models over anatomical regions. When neurons’ orientations are aligned (e.g., Fig. 5), a silhouette of neurons’ shapes may be turned into a statistically sound representation to model fine details of neuron arbors. By putting together numerous pieces of neurites, a “shotgun fragment assembly” approach similar to whole-genome sequencing algorithms (47) might be possible to reconstruct long projecting axons in human brains.

Last, previous studies [e.g., (20)] and current large-scale brain science initiatives strongly suggest that the morphological profiling of human neurons at multiple brain regions should be combined with molecular and physiological profiling of individual neurons. ACTomography is generalizable to incorporate some of those experimental components. We envision that a possible combination could be spatial genomics or in situ transcriptomics (48) for the ex vivo brain slices. As we have successfully mapped various morphological tensors extracted from ex vivo samples over the standardized brain anatomy, the same brain mapping strategy is highly applicable to fuse such multimodality single-neuron resolution data at the whole-brain scale. In addition, one may also use genetic tools such as enhancer adeno-associated viruses (49) to prelabel specific cell types genetically in cultured human tissues (50, 51). Incorporating such a prelabeling method could be highly beneficial to the ACT step of ACTomography or the post hoc triage of neuron types as discussed above.

MATERIALS AND METHODS

Human specimens

A total of 19 human cases were enrolled in this study, with 18 cases obtained from surgical procedure in Beijing Tiantan Hospital and 1 case obtained from donated brain from Human Brain and Tissue Bank of National Clinical Research Centre for Neurological Disorders. All the cases were informed for consent, and all procedures involving human tissue for research purposes were approved by the ethics committee of the Beijing Tiantan Hospital.

In total, 8 male and 10 female patients between 18 and 71 years old underwent the surgical resection of primary tumor or similarly documented brain diseases for treatment purpose, with clear surgical indications. All the patients were confirmed by preoperative imaging that the brain tumor or lesion located in the deep brain structure or suspected for high-grade gliomas, which were indicated for surgical approaches path through the normal cortex following the neurosurgical principles. The surgical procedure cautiously preserved the normal cortex limited in the approached or in the edge of high-grade gliomas. Postoperative histological review confirmed that the tumor/lesion pathology and function status were assessed by follow-up after patient discharge.

In the meantime, the case H002 was obtained from a donated patient died of glioblastoma, which extensively infiltrated in the left brain. The normal frontal lobe tissue was acquired from the right side following the protocols approved. Specifically, this patient received tumor resection on 13 September 2020, due to space-occupying lesions in the left frontal lobe. Postoperative pathology reported glioblastoma, and postoperative chemoradiotherapy and electric field therapy were given. The patient underwent the last magnetic resonance imaging (MRI) examination on 23 February 2022 and eventually died due to disease progression on 22 April 2022. After death, the brain tissues were donated to the Human Brain and Tissue Bank of National Clinical Research Centre for Neurological Disorders and dissected and preserved by qualified personnel. In this study, the right frontal lobe tissues of patients were obtained in strict accordance with ethical requirements, and the corresponding anatomical sites were judged by professionals and photographed for preservation.

Brain tissue preparation and sectioning

The brain tissues were fixed, fluorescent-labeled, and imaged. About 5 to 10 min (except two cases of large tissues, it was within 20 min) after the surgery, the extracted specimens were immersed in 4% paraformaldehyde (PFA) phosphate buffer solution (0.1 M, pH 7.4) for fixation. Specimens larger than 1 cm³ were cut into such a unit volume before fixation. Before sectioning, the tissues in PFA solution were stored on a shaking table at 4°C for ~24 hours.

The fixed tissue blocks were then sectioned into slices before being labeled. The tissue blocks from the PFA solution were taken out to remove blood clots and meninges on the surface, followed by rinsing three times with double distilled water. It was then put in a dry petri dish, where water on the surface was removed using absorbent paper. Then, we poured 3% agarose gel (42°C) into the petri dish until the tissue is completely submerged while removing the bubbles around the tissue. The excess agar blocks were removed after the agar had solidified. Thereafter, we sectioned the tissue block into 200- to 300-μm-thick slices (~80%, 200 μm thick) with vibratome (Leica vt1200s). During this procedure, the tissue block was fixed on the vibratome with 404 glue (LOCTITE) and immersed within 1× phosphate-buffered saline (PBS). We cut the tissue longitudinally along the surface of the cerebral cortex, from gray matter to white matter. The slices were transferred to 5-ml centrifuge tubes and stored in 4% PFA solution on the shaking table at 4°C for 3 to 5 days to be further fixed. The fixing brain slices were still placed.

Fluorescent neuron labeling through cell injection

To distinguish individual neurons from surrounding cells, we labeled single neurons with YL CH, Lithium Salt (Thermo Fisher Scientific) through cell injection with micromanipulators. Injection was conducted under an upright differential interference (DIC) microscope (Scientifica 1000P) equipped with near infrared light (~900 nm) as light source and a water immersion objective. We found that LY could be effectively injected into neurons of fixed tissue slices and could label the neuron including dendritic tree and spines and, potentially, axons, under optimized conditions related to the specimen. The negatively charged YL was pumped into a neuron with the help of a microelectrophoretic current generator (WPI Microiontophoresis Current Programmer SYS-260).

The detailed procedure of cell injection is as follows.

1) Brain slices were carefully transferred onto the sample stage of the DIC microscope and held confined with a silk wound copper ring tablet above it. The sample chamber was filled with sufficient amount of 1× PBS to keep brain slices immersed in PBS.

2) Pipettes needed for the cell injection were produced by a micropipette puller (Sutter instrument P-2000). Borosilicate glass capillaries (WPI 1B100-4, 1.0-mm outer diameter by 0.58-mm inside diameter) were used to pull very fine pipette with a long needle tip (needle tip size of 1 to 2 μm and needle tip angle of ~10°). These pipettes also served as the microelectrodes of the microiontophoresis.

3) Pipettes filled with 4% LY PBS solution were mounted on a microelectrode holder (WPI, PE MEH710), which was fixed on a 3D adjustable micromanipulor (Scientifica, MicroStar S-MST-1000-X). In this process, it is critical to have sufficient solvent and without bubbles in the in the pipette and holder. The negative electrode of the microiontophoresis device (WPI SYS-260) was connected to the microelectrode holder filled with LY solution. The positive electrode was connected to the tablet iron wire and immersed in PBS solution.

4) In DIC microscope, we used a 40× water immersion objective and a charge-coupled device to acquire the image. After the sample was placed, needle tips were moved to the field of vision of the microscope. Then, the objective and needle tips were moved down synchronously until the objective is well focused on the brain slice. The neuronal somas about 30- to 50-μm deep within the brain slice were selected for labeling. For a given neuron soma, the needle tip was first moved to the position directly above the soma and then slowly moved downward to approach the soma. The cell is punctured after the tip of the needle contacted the cell membrane. We know that the needle penetrated the cell membrane if we could see an instantaneous rebounding of the cell membrane. The brain slices were imaged on live at DIC mode during this whole procedure.

5) After the needle penetrated into the cell, the microiontophoresis device was turned on, the diffusion of the LY in a neuronal soma as well as along dendrites can be observed via the wide-field fluorescence imaging of the DIC microscope. The perfusion is continued with a negative current of 10 nA for 15 to 20 min until the branches were clearly visible. After the perfusion is finished and the microiontophoresis is turned off, we waited for another 5 min before pulling out the needle.

Two-photon fluorescence imaging of neurons

We used a two-photon fluorescence imaging system (Bruker, Ultima Investigator) to obtain the fluorescence microscopic imaging of the labeled neurons. The imaging system was setup on an upright microscope with a single set of galvanometers providing standard and resonant scanning mode, 8-kHz x-axis scanning and high speed of 30 frames per second (fps) for 2D 512 × 512 scan. For LY-stained neuron imaging, the fluorescence collection used was a filter module composed of 565-nm-long pass dichroic beam splitter and an emission filter ET525/70 m-2p. The objective used was an apochromatic water immersion 25× two-photon objective (Nikon, CFI75 Apochromat 25XW MP1300) with a numerical aperture of 1.1. Two-photon fluorescence excitation was provided by a Ti:sapphire femtosecond laser (Coherent Chameleon Discovery NX with TPC Laser System) generating widely tunable (680- to 1300-nm) femtosecond laser pulses (pulse duration of <120 fs). The repetition frequency of the laser output pulses is 80 MHz, with an average output power of 2 W at 900 nm (used for LY-labeled neuron imaging). The scan and image collection were controlled by the software of Prairie View. The brief two-photon imaging process is as follows.

1) The brain slices were transferred to the glass slide and confined by a silk wound tablet. An appropriate amount of 1× PBS was dropped on the brain slices ensuring that the objective lens can be submerged in PBS when taking images.

2) The slides were put on the stage of two-photon fluorescence microscope. The wide-field fluorescence imaging mode was used for the localization of target neurons.

3) The imaging system was switched to two-photon imaging mode after the cells were founded. Using Primitive view software, in the live scan mode, the x, y, and z axes could be adjusted to determine the imaging range of two-photon scan. In the stage control module, the upper and lower limits of the z-axis scanning of cells and the scanning step length could be set. The imaging zoom in/out was adjusted according to the extension of neurons in the xy plane. After setting, z-series scan could be started to get the image stack.

We defined the efficacy in image acquisition over cell injection as the ratio of the number of successfully imaged neurons with visible neuron arbors divided by the total number of attempted injections.

Brain registration

All subjects involved in this study had complete MRI data. On the basis of the operation records and the preoperation and postoperation MRI data, we applied 3D Slicer to T1 images to define the loci of the extracted brain tissue. For instance, for H003, we used 3D Slicer to load DICOM format T1 imaging data and then used the Segment Editor module of 3D Slicer to produce the outline of the extracted brain tissue. Then, both T1 image and the segmentation were output as files (in nrrd format).

We used the Brain Extraction Tool in FMRIB Software Library (FSL) to segment T1 images and exclude skull in images (fractional intensity threshold of 0.5). We used MNI152 template included in the ANTS software (52) as the target brain for registration, with the “SyN” elastic alignment method. The tumor and other lesion regions were manually defined in the T1-registrered images. When the automatic registration quality was not satisfactory, optionally, we used 3D Slicer to define the locations of extracted brain tissues on the MNI152 template directly.

Neuron tracing, neuron feature, and data analysis

All 3D neuron morphologies (in swc format) were traced and reconstructed using Vaa3D [v4.001, 2020-12-11, https://github.com/Vaa3D; (53)], an open-source and cross-platform system for visualization and annotation, by a group of annotators in a semiautomatic way. In particular, two modules, TeraFly (54) and TeraVR (55), which enable large-scale nonimmersive and immersive neuron reconstruction, were used to trace these neurons. Potential reconstruction errors were identified and removed through iterative quality control and correction. Thereafter, the morphology skeletons were refined to the center of the fluorescent signals via mean-shift algorithm, followed by pruning and sorting to generate neat neuron reconstructions, i.e., a single tree without breaks, loops, or unexpected branching. The root node of a neuron tree was set to be its soma. When needed, the contrast of the image was adjusted during the reconstruction to trace very weak neurites. It also helped to better visualize distal neuron fibers that might be less stained by fluorescent dyes during the cell injection. Neuron orientations were manually determined using Vaa3D based on clearly oriented apical dendritic stem(s) or the basal dendrite profiles.

The 22D morphological features are ('Nodes', 'SomaSurface', 'Stems', 'Bifurcations', 'Branches', 'Tips', 'OverallWidth', 'OverallHeight', 'OverallDepth', 'AverageDiameter', 'Length', 'Surface', 'Volume', 'MaxEuclideanDistance', 'MaxPathDistance', 'MaxBranchOrder', 'AverageContraction', 'AverageFragmentation', 'AverageParent-daughterRatio', 'AverageBifurcationAngleLocal', 'AverageBifurcationAngleRemote', and 'HausdorffDimension'), defined in L-measure (56). We used the implementation in Vaa3D’s global_neuron_features plugin.

The dendritic field area (57) was calculated to benchmark the dendritic spread of the neurons. Accordingly, we projected each neuron reconstruction onto the xy plane and calculated the area contained within the convex hull around the dendritic terminations. The calculation was performed separately for neuron reconstructions of pool-A and pool-B to take consideration of the neuron orientations. Data analysis was done mostly in Python, with the Python Data Analysis Library (pandas) and statistical data visualization library (seaborn).

Supplementary Materials

This PDF file includes:

Figs. S1 to S13

Table S1

References