Perisomatic ultrastructure efficiently classifies cells in mouse cortex

Abstract

Mammalian neocortex contains a highly diverse set of cell types. These cell types have been mapped systematically using a variety of molecular, electrophysiological and morphological approaches^1-4. Each modality offers new perspectives on the variation of biological processes underlying cell-type specialization. Cellular-scale electron microscopy provides dense ultrastructural examination and an unbiased perspective on the subcellular organization of brain cells, including their synaptic connectivity and nanometre-scale morphology. In data that contain tens of thousands of neurons, most of which have incomplete reconstructions, identifying cell types becomes a clear challenge for analysis⁵. Here, to address this challenge, we present a systematic survey of the somatic region of all cells in a cubic millimetre of cortex using quantitative features obtained from electron microscopy. This analysis demonstrates that the perisomatic region is sufficient to identify cell types, including types defined primarily on the basis of their connectivity patterns. We then describe how this classification facilitates cell-type-specific connectivity characterization and locating cells with rare connectivity patterns in the dataset.

Fig. 1. Large-scale automated segmentations necessitate proofreading insensitive cell classifications.

a, Rendering of a small fraction of neurons from the MICrONS dataset (1.1 mm × 800 μm × 600 μm) covering all layers of cortex and several visual areas, with 1,207 rendered and then cut away to reveal the full morphology of 2 selected neurons on the right portion of the dataset. b, Example neuronal morphologies before and after proofreading. Left, excitatory neuron; right, inhibitory neuron. c, Fraction of input and output synapses removed (left), added (middle) and maintained (right) after proofreading for 1,347 neurons. For all box plots: centre line, median; box limits, upper and lower quartiles; whiskers, 1.5× interquartile range; outliers not shown (visible in the adjacent scatter plots). d, Neurons near the volume borders will have truncated morphologies. e, Top: histogram of the radial extent of dendrites from a sample of 1,347 proofread neurons (left) and the cumulative distribution of those cells (right). Bottom: histogram of the minimum distance from a volume border for all high-quality nuclear detections (n = 94,010; left) and the cumulative distribution of those distances (right). The grey shading indicates the portion of cells less than the median radial extent (33% of cells), shown in teal.

Fig. 2. Perisomatic region of cortical cells.

a, A measure of the distance from the soma for each edit that was made to the segmentation during proofreading of 1,347 cells. The teal line denotes the average and the shading marks the 10th–90th percentile across all cells. The arrow marks 15 µm. b, Example cell demonstrating the extent of mesh information used to extract somatic, nuclear and synapse features. All cell meshes were restricted to 15 µm from the centre of the nucleus. c, Representative example of nuclear infolding in a single electron microscopy image. The soma is highlighted in grey, black represents the nuclear envelope and orange marks the areas classified as infolded on the basis of the shrink-wrap method (Methods). d, Example cell demonstrating the extent of mesh information used to extract postsynaptic features (left) and three example PSSs (right). All synapses included in the PSS analyses were within 60 µm from the centre of the nucleus. e, A 3D rendering of the somatic cutouts from all the cells from a 100-µm column that was densely reconstructed for which manual labels were given (n = 1,619). Cells rendered are organized by their cell class and coloured by their cell subclass according to the colour scheme displayed. Scale bars, 5 µm (b), 2 µm (c) and 20 µm (d).

Fig. 3. Variations of nucleus and somatic features show stark laminar and cell-class-based distinctions.

a, Nuclear volume (μm³; left), fraction of nuclear membrane infolded (middle) and somatic synapse density (μm²; right) against distance from the pial surface (μm). Cortical layer boundaries are marked by dashed lines. b, Somatic surface cutout area (μm²; within 15 µm from the nuclear centre) against nuclear volume (μm³). c, 2D UMAP embedding of all neuronal and non-neuronal cells inferred from somatic features, nuclear features and cortical depth. d, z-scored feature matrix representing all the somatic and nuclear features on the manually labelled cells from the cortical column. Cells are organized by their annotated subclass. Tick marks along the x axis denote segments of 100 cells (1,115 excitatory neurons, 143 inhibitory neurons, 361 non-neurons). For all plots, manually labelled cell classes are represented in colour (1,619) and unlabelled examples are shown in light grey (n = 92,391). e, 3D mesh renderings of representative examples of different neuronal and non-neuronal cell classes. In the top row, nuclei are shown with the folded surface area highlighted in orange. Corresponding cell bodies are shown in the bottom row with somatic synapses in orange. Sphere size corresponds to predicted synapse size from the synapse detection model. Exc., excitatory. Scale bars, 5 μm (top row) and 15 μm (bottom row).

Fig. 4. PSS features.

a, Inhibitory neurons exhibit great variability in ultrastructural morphology. b, Procedure for building a PSS dictionary model. The set of shapes is used to train a PointNet autoencoder that learns a latent feature vector of a fixed size (1,024). This autoencoder is then applied to all shapes in the dictionary to generate a set of latent feature vectors. k-means with k = 30 is applied to this to obtain a set of cluster centres for binning the shapes. c, For each cell, the PSSs are binned by shape type and distance from the soma (4 bins) from 0 to 60 µm with 15 µm bin sizes. The resulting histogram is a 2D histogram shown above with the shapes in the x direction and distances in the y direction. d, Examples of 60-µm cutouts of the four predicted inhibitory subclasses with their spatial histograms shown as heat maps. The top row shows the shape of the cluster centre of each of the 30 clusters. In each heat map, darker rectangles indicate higher values. e, z-scored feature matrix representing the distance-binned PSS features on the manually labelled inhibitory cells from the cortical column (n = 143). Cells are organized by their annotated subclass. Tick marks along the x axis denote segments of 100 cells. f, 2D UMAP of all the inhibitory neurons (n = 6,805) inferred after concatenating nucleus, soma and PSS features, with cortical column cells in colour and dataset-wide inhibitory neurons in grey.

Fig. 5. Hierarchical predictions enable dataset-wide circuit analyses.

a, Diagram of the hierarchical model framework used to predict neuronal and non-neuronal subclasses using a set of five classifiers. Nucleus and soma features alone were used for models 1–4. PSS features were added to predict inhibitory subclasses in model 5 (Extended Data Table 1). Oligo, oligodendrocyte. b, Confusion matrix of the cross-validation performance for all cells within the manually labelled column. Note that classifiers for excitatory neurons, inhibitory neurons and non-neurons were trained separately (models 2, 4 and 5 in a). The confusion rate between these classes can be seen in Extended Data Fig. 4. c, 2D UMAP embedding inferred from depth, nucleus and soma features of all cells in the dataset coloured by the hierarchical model predictions (n = 94,010). d, Left: 2D rendering of a representative 23P cell morphology, with dendrite in black and axon in grey. Points represent the somatic position of all downstream target cells coloured by the hierarchical model subclass prediction. Right: synapse count (top), total synapse area (middle; voxels are 4 × 4 × 40 nm) and number of synapses per connection (bottom) displayed by the model-predicted subclasses illustrating the local targeting profile of this individual cell. e, Similar information as in d but for an inhibitory bipolar cell that is predicted to preferentially target basket cells. This unique population of bipolar cells has been further characterized. For all box plots: centre line, median; box limits, upper and lower quartiles; whiskers, 1.5× interquartile range; outliers shown.

Fig. 6. Perisomatic feature space enables more efficient search for unique cells.

a, 2D UMAP embedding highlighting a chandelier cell (orange dot), a 5P-NP-targeting cell (blue dot) and their respective 20 nearest neighbours in the high-dimensional perisomatic feature space. Note, UMAP nonlinearly distorts feature space, so not all nearest neighbours appear closest in the plot. b, Example proofread chandelier cell in L2/3 (dendrite in black, axon in grey). Output synapses are marked along the axon and coloured by subclass prediction. Note the characteristic vertical chains of synapses onto 23P cells. c, Chandelier cells are characterized by their preference to synapse onto the AIS of target cells^, (quantified here by measuring the angle between the target soma and the synapse (ϕ) and the distance from the soma (r)). d, Heat map illustrating the angle and distance distribution of the chandelier cell shown in b as well as two non-Chandelier inhibitory examples. Colour denotes the normalized synapse density for each cell. Synapses located at an angle >160° were considered onto the AIS of the target cell (shown by the dashed line). e, Angular distribution histogram of the chandelier cell (top row), 20 nearest neighbours in the perisomatic feature space, and 20 random inhibitory cells (P < 0.00001). f, Example cell that preferentially targets the rare 5P-NP subclass (dendrite in black, axon in grey); points represent target cell soma locations coloured by predicted subclass. Output synapse counts reflect strong preference to 5P-NP cells. For all box plots: centre line, median; box limits, upper and lower quartiles; whiskers, 1.5× interquartile range; outliers shown. g, Fraction of output connectivity onto neuronal subclasses of the 5P-NP-targeting cell (top row), 20 nearest neighbours in the perisomatic feature space, and 20 random inhibitory cells (P < 0.00001). All P values reported and asterisks represent significance by two-tailed Fisher exact test.

Extended Data Fig. 1. Neuronal and non-neuronal subclass distribution of individual soma and nucleus features.

a) 2D UMAP embedding of all neuronal and non-neuronal cells inferred from somatic features, nuclear features and cortical depth. Manually labeled cellular subclasses are represented in color (1,619) and unlabeled examples in light gray (n = 92,391). b) Distribution and variation of cortical depth of all cells from the manually labeled column dataset. c) Distribution and variation of nucleus and somatic features of all cells from the column dataset. For all box plots, center line, median; box limits, upper and lower quartiles; whiskers, 1.5x interquartile range, outliers not shown. Individual cells, including outliers, are shown in the overlaid swarm plots. Color denotes human assigned subclass labels.

Extended Data Fig. 2. PSS embedding space organized by postsynaptic ultrastructural morphologies.

2D UMAP embedding of all shapes in the PSS Dictionary. The numbers indicate the bin centers mapped in this 2D space and the corresponding PSS meshes on the right show the shape associated with each bin center. Bins 1–8 range in spine shapes, Bins 9–23 are shaft shapes and Bins 24–29 are soma shapes.

Extended Data Fig. 3. Inhibitory neuron subclasses exhibit spatial patterns to PSS distributions.

The UMAP embedding of all the perisomatic features, including PSS features, across all inhibitory cells, colored with respect to what fraction of that cell’s input (within the 60 µm cutout) comes from what PSS/distance bin. PSS shape bins were simplified from 29 bins to 5 broad categories to simplify the visualization (bins 0-4: short spines, 5–8: long-spines, 9-18 + 23: smooth shafts, 19–22: spiny shafts, 24–29: soma). This visualization gives insight into how different cells in different parts of this embedding space receive varying amounts of input onto different shapes within different spatial zones of the perisomatic area. Cells on the far left hand side of the embedding, where in general bipolar type neurons were found, have larger fractions of their inputs near the soma, including dendritic shafts which are more irregular in shape (“spiny shafts”), and smooth shaft inputs farther away where the dendrites begin to elaborate. Basket cells on the right hand side of the side of the embedding are dominated by somatic inputs and smooth shaft inputs which are more evenly distributed spatially. The island at the bottom that is dominated by neurogliaform cells is characterized by having relatively fewer somatic inputs, but an increasing amount of shaft and spiny input at distal dendrites.

Extended Data Fig. 4. Classifier validation.

a) Confusion matrix of hierarchical model performance for all cells within the manually labeled column after training. b) Confusion matrix of hierarchical model performance on a dataset wide sample of 100 cell predictions from each subclass. c) Comparison of column cross validation vs. dataset wide model performance, asterisk notes significance by two-tailed Fisher Exact Test.

Extended Data Fig. 5. Cell densities across the dataset by cellular subclass.

Predicted cell densities per mm² for each subclass across the entire dataset in the XZ plane. Each square represents 50 micron and color denotes the density scaled per mm². Note due to the approximate 1 mm depth of cortex, these values are also roughly densities per mm³. They roughly agree with densities of cells estimated from light microscopy stereology of subclasses, usually utilizing histochemical markers or genetic tools. For some subclasses, there is not a 1-1 to alignment between the definitions of types in this study and the usual molecular markers used in those studies, as molecular markers are not directly measurable in this electron microscopy volume.

Extended Data Fig. 6. Perisomatic feature based classification utilized with different cell-type labels.

a) Alternative excitatory subclass labels in the column from Schneider-Mizell et al., based on unsupervised clustering of dendritic and synaptic features rather than manual human expert calls. Labels on the clusters were inferred based on the overlap with expert labels and cortical depth, with finer distinctions added when necessary (i.e. L4a, L4b, L4c). b) Alternative inhibitory subclass labels from Schneider-Mizell et al. in the column based on unsupervised clustering of their output connectivity statistics. These subclasses (Perisomatic Targeting, Distal Targeting, Sparsely Targeting and Inhibitory Targeting) likely largely but not completely align with broad molecular distinctions made amongst inhibitory cells, based on reviews of the literature where molecular and output connectivity has been measured in the same cells. c) A confusion matrix of a hierarchical model retrained to utilize these subclass labels for excitatory neurons vs inhibitory neurons rather than human expert labels. Cross validation performance on the excitatory (67%) and inhibitory (85%) subclass models was lower than the expert labels, due primarily to the fine grained distinctions made amongst layer 4 and 6 types. The confusion matrix shown here is the output of the final model trained on all samples from the column.

Extended Data Fig. 7. Basic perisomatic feature patterns maintained across a second dataset from a different animal.

a) A cutout of a second dataset, which covers layer 2/3 to 6 of cortex, but is only 50 µm thick. Somas contained within this volume (n = 1,944) were analyzed in a manner identical to the larger dataset and soma, nucleus and PSS features were extracted. Excitatory nuclei highlighted in light blue and inhibitory nuclei in magenta. b) Feature to feature Pearson correlations exhibit similar correlation structure between the two datasets. c) A joint UMAP of the perisomatic features with the MICrONS dataset data shown in gray, and the smaller dataset covered by manually identified cell classes overlaid. In general, the same overall patterns and degree of separation amongst layers and cell classes was observed. Note: pericytes were manually excluded from this dataset due to the lower quality of nucleus and somatic segmentations. Extensive detailed subclass cell type validation is not possible in this dataset due to the truncation of axons and dendrites.