Inhibitory specificity from a connectomic census of mouse visual cortex

Abstract

Mammalian cortex features a vast diversity of neuronal cell types, each with characteristic anatomical, molecular and functional properties¹. Synaptic connectivity shapes how each cell type participates in the cortical circuit, but mapping connectivity rules at the resolution of distinct cell types remains difficult. Here we used millimetre-scale volumetric electron microscopy² to investigate the connectivity of all inhibitory neurons across a densely segmented neuronal population of 1,352 cells spanning all layers of mouse visual cortex, producing a wiring diagram of inhibition with more than 70,000 synapses. Inspired by classical neuroanatomy, we classified inhibitory neurons based on targeting of dendritic compartments and developed an excitatory neuron classification based on dendritic reconstructions with whole-cell maps of synaptic input. Single-cell connectivity showed a class of disinhibitory specialist that targets basket cells. Analysis of inhibitory connectivity onto excitatory neurons found widespread specificity, with many interneurons exhibiting differential targeting of spatially intermingled subpopulations. Inhibitory targeting was organized into 'motif groups', diverse sets of cells that collectively target both perisomatic and dendritic compartments of the same excitatory targets. Collectively, our analysis identified new organizing principles for cortical inhibition and will serve as a foundation for linking contemporary multimodal neuronal atlases with the cortical wiring diagram.

Fig. 1. A columnar reconstruction of mouse visual cortex.

a, The millimetre-scale EM volume is large enough to capture complete dendrites of cells across all layers. Neurons shown are a random subset of the volume, with a single example at right for clarity. b, The autosegmented EM data show ultrastructural features such as membranes, synapses and mitochondria. Scale bar, 500 nm. c, Top view of EM data with approximate regional boundaries indicated. The yellow box indicates the 100 µm × 100 µm column of interest. Scale bar, 200 µm. d, All soma locations in the column coloured by cell class. Scale bar, 100 µm. e, Example neurons from along the column. Note that anatomical continuity required adding a bend in deeper layers. f, Proofreading workflow by cell class. g, Cell density for column cells along cortical depth by cell class. Scale bar, 200 µm. h, Input synapse count per micrometre of depth across all excitatory (purple) and inhibitory (green) column cells along cortical depth by target neuronal cell class. Scale bar, 200 µm. i, All excitatory dendrites, with arbours of cells with deeper somata coloured darker. Same orientation as in d. Scale bar, 200 µm. j, Number of input synapses for each excitatory neuron as a function of soma depth. k, All inhibitory dendrites, as in j. l, Number of input synapses for inhibitory neurons, as in k. m, Axons of inhibitory neurons, as in j. n, Number of output synapses for inhibitory neurons, as in k. VISrl, rostrolateral visual area; VISal, anterolateral visual area; Exc, excitatory; inh, inhibitory; non, non-neuronal; syn, synapses; WM, white matter. Source Data

Fig. 2. Inhibitory subclasses and the inhibition of inhibition.

a, Example of an inhibitory axon making synaptic outputs (green dots) onto specific locations on a target pyramidal cell (purple). b, Dendritic compartment definitions for excitatory neurons. c, Cartoon definition for a multisynaptic connection (left) and the synapses in the multisynaptic connection considered ‘clumped’ along the presynaptic axon (right). d, Targeting features for all inhibitory neurons, measured as fraction of synapses onto column cells (for fraction clumped only: synapses in multisynaptic connections). e, Relationship between anatomical connectivity categories (top), typical associated classical cell categories (middle) and anatomical examples (bottom) of the inhibitory subclasses. Dendrite is darker, axon lighter. f, Adjacency matrix for inhibitory neurons. Each dot represents a connection from a presynaptic to a postsynaptic cell, with dot size proportional to synapse count. Dots are coloured by presynaptic subclass and ordered by subclass, connectivity group (Fig. 5) and soma depth. g, Standard model of inhibition of inhibition between molecular subclasses. h, Mean number of synaptic inputs a postsynaptic cell received from all cells of a given presynaptic subclass. i, Potential InhTC targets. j, Synaptic output fraction each InhTC (columns) places onto target subclasses (rows). InhTCs are clustered into two subtypes: one targeting DistTCs (InhTC^dist) and another targeting PeriTCs (InhTC^peri). k, Connectivity diagram for InhTC^peri suggested by data. l, Morphology of example InhTC^dist. m, Morphology of all InhTC^peri. n, Median synapse size (arbitrary units measuring voxels in segmented cleft) from InhTC^dist (left) and InhTC^peri (right) onto inhibitory subclasses. Error bars indicate 95% confidence interval. T-test P-values indicated: *, P < 0.05; ***, P < 0.005 after Holm–Sidak correction. o, Distribution of synapses per connection for InhTC^peri and InhTC^dist onto their preferred and non-preferred targets. Scale bars, 500 µm. CCK, cholecystokinin; frac, fraction; multisyn, multisynaptic; no, number of. Source Data

Fig. 3. Characterization of excitatory neuron M-types.

a, Morphology (black) and synapse (cyan dots) properties were used to extract features for each excitatory neuron, such as this layer 2/3 pyramidal cell. b, Heatmap of Z-scored feature values for all excitatory neurons, ordered by anatomical cluster (Fig. 5) and soma depth. Anatomical properties were tip length, tortuosity, dendritic and somatic synapse counts, total path length, radial extent, median synapse distance from soma, somatic and dendritic synapse sizes, dynamic range of synapse size, shallowest and deepest ranges of synapse depth, range of synapse depths, linear synapse density and dendritic radius. All synapse measures use synaptic inputs only. See Methods for detailed feature descriptions. c, Uniform manifold approximation and projection (UMAP) of neuron features coloured by anatomical cluster. Inset shows number of cells per cluster. d, Example morphologies for each cluster. Scale bar, 500 µm. e, Soma depth of cells in each anatomical cluster. f, Median linear density of input synapses across dendrites by M-type. g, Median synapse size (Methods). In f and g, coloured dots indicate single cells; black dots and error bars indicate a bootstrapped (n = 1,000) estimate of the median and 95% confidence interval. a.u., arbitrary units. Source Data

Fig. 4. Inhibition of excitatory neurons.

a, Connectivity from all inhibitory neurons (columns) onto all excitatory neurons (rows), sorted by M-type and soma depth. Dot size indicates net number of synapses observed. b, Net synapses onto column cells for each inhibitory subclass. Black dots indicate median; bars show 5% confidence interval. c, Mean net synapses per target cell from each inhibitory subclass onto each excitatory M-type. d, Spearman correlation of PeriTC and DistTC net input onto individual cells, measured in each M-type. Bars indicate 95% confidence interval based on bootstrapping (n = 2,000). Stars indicate M-types significantly different from zero with a P value < 0.05 after Holm–Sidak multiple test correction. e,f, Pearson correlation of connectivity density between excitatory M-types on the basis of PeriTCs (e) and DistTCs (f). Dotted lines indicate groups of cells roughly in a layer. Source Data

Fig. 5. Inhibitory motif groups organize inhibitory connectivity.

a, Distribution of synaptic output for all interneurons, clustered into motif groups with common target distributions. Each row is an excitatory target M-type, each column is an interneuron, and colour indicates fraction of observed synapses from the interneuron onto the target M-type. Only synapses onto excitatory neurons are used to compute the fraction. Neurons are ordered by motif group and soma depth. Bar plots along top indicate number of synapses onto column cells, with colour showing subclass (as in d). Bar plots along right indicate number of cells in target M-type. b,c, Morphology of all cells in group 4 (b) and group 13 (c), with colours as in a. Scale bar, 500 µm. d, Soma depth and subclass for cells in each motif group. Scale bar, 200 µm. e, Net synaptic output distribution across M-types for each motif group. f, Synaptic input for each M-type from each motif group as a fraction of all within-column inhibition. g, Schematic of motif group connectivity in upper layers. h, Schematic of motif group connectivity in Layer 5. Source Data

Fig. 6. Synaptic selectivity and cell connectivity cards.

a, Example inhibitory neuron (cell ID 303085). Axon in blue, dendrite in red. Scale bar, 200 µm. b, Distribution of synaptic outputs across target compartments for the cell in a. c, Distribution of synaptic outputs across M-types (bar length) and compartments (bar colours) for the cell in a. d, Selectivity index values for the cell in a, measured as the ratio of observed synapse count to median shuffled synapse count for a null model as described below. Error bars indicate 95th percentile interval. Coloured dots (blue, low; orange, high) indicate significant differences (two-sided shuffle, P < 0.05) relative to the shuffle distribution after Holm–Sidak multiple test correction. e, As a baseline synapse distribution for null models, all synaptic inputs onto all cells in the column were binned by compartment, depth and M-type. (See Extended Data Fig. 10 for more details.) f, Shuffled connectivity for the cell in a was computed by sampling from the baseline synapse distribution with the observed depth and compartment bins and counting synapses onto each M-type across all bins (n = 1,000 shuffles). Example shuffle values for L3a (top) and L4a (bottom) M-types versus observed synapses are shown. g, Selectivity index for all cells in motif group 5. Non-significant values are assigned a value of 1. The cell in a is highlighted by a black box. h, Direction of the median cell’s selectivity index from each motif group onto each M-type. Orange indicates more connected, blue less connected. Connections where the median selectivity index was non-significant are indicated with a dot. i–l, Compact cell connectivity cards encapsulating anatomy (left), M-type target distribution (middle, bar length), compartment targeting (middle, bar colours as in d) and selectivity index (right, as in g) for four example neurons: an L5ET-specific basket cell (i), a deep-layer-specific upper layer neuron (j), a translaminar basket cell (k) and a translaminar layer 6 interneuron (l). Full connectivity cards for all cells can be found in Supplementary File 1. Scale bars, 200 µm. SI, selectivity index. Source Data

Extended Data Fig. 1. Compartment classification pipeline.

a) Description of the compartment classification pipeline. b–d) Pipeline applied to an example layer 3 pyramidal cell. b) Apical probability per vertex. c) Branch-level apical classification. d) Final organization into four dendritic compartments based on apical classification and distance rules. e) Quantification of quality of apical branch classification based on leave-one-out classification with a training set based on 50 randomly selected cells and 23 cells chosen to improve difficult classifications. Each dot is a branch of a test pyramidal cell, colored red if apical and blue if not apical. X-axis is the net log-odds of the branch being apical (capped at ±200) and the y-axis is the relative apical quality based on a soft-max operation (see Methods for details). Branches in the upper right quadrant were classified as apical. The method was able to correctly classify at least one apical branch for all cells, and “false positives” were often associated with borderline cases. f) Distribution of synaptic inputs onto excitatory neurons with depth by dendritic compartment. Values are based on counting synapses in bins at a given depth, but at any location laterally.

Extended Data Fig. 2. Closest distances between synapses in multisynaptic connections.

a) Cumulative distributions of the closest synapse onto the same target along the axonal arbor per manually labeled inhibitory neuron subclass. Excitatory (left) and inhibitory (right) targets shown separately. Vertical gray line indicates the value used for “clumpiness” in the main text. b) Same as a, but for the cluster-based labels and with log scale to highlight shorter distances. c) The “clumpiness” metric using different distance thresholds. The qualitative relationships are extremely robust to distance thresholds.

Extended Data Fig. 3. Inhibitory neuron properties.

a) Projections of all analyzed interneurons (n = 163) projected on a 3-d space based on linear discriminant analysis (LDA) using connectivity features (shown in c). Fully colored dots indicate manually classified cells used as training data for LDA, while dots with grey centers were labeled based on this classification. b) Matrix showing relationship between anatomical subclasses and manual classifications. c) Individual connectivity features, organized by subclass. Colored dots are individual cells, black dots indicate median with error bars showing a bootstrapped 95% confidence interval. d–g) Morphology of all PeriTCs (d), DistTCs (e), SparTCs (f), and InhTCs (g). Scale bars are 500 µm. Dark and thick lines are dendrite, thinner and lighter are axon. Cells are ordered by soma depth.

Extended Data Fig. 4. Inhibition of inhibition.

a) Connectivity dotplot between inhibitory neurons, organized by inhibitory subclasses, organized by soma depth. For each panel, the scatterplot reflects the connectivity from cells in the presynaptic subclass (x-axis) to cells in the postsynaptic subclass (y-axis). Each dot is a single connection, with larger dots having more synapses. The location of each dot corresponds to the depth of the pre- and post-synaptic cell bodies. Stem plots on top and side indicate the net synaptic inputs and net synaptic outputs of each cell in each subclass within the column sample. b) Same as a, but for InhTC^Peri and InhTC^Dist onto PeriTCs separately.

Extended Data Fig. 5. A laminar-specific circuit for InhTC^dist cells.

a) Morphology of all InhTC that preferentially target DistTCs. Cells are sorted by soma depth. b) Connectivity dotplot for synapses from InhTC^dist onto DistTCs. In the grid, each dot represents a connection from one InhTC onto one DistTC, with the number of synapses indicated by dot size. The location of the dot corresponds to the soma depth of the pre- and post-synaptic cells. Stem plots on top and side indicate the net synaptic inputs and net synaptic outputs of each InhTC^dist and DistTC. Note that DistTCs in layer 2/3 receive little input from InhTCs, compared to those in layer 4 and upper layer 5. c) Connectivity scatterplot for synapses from DistTCs onto InhTC^dist, as in b. Note that the DistTCs in layer 2/3 also form few synaptic outputs onto InhTC^dist. d) Distribution across M-types of synaptic outputs across low-connection DistTCs and high-connection DistTCs. e) Connectivity cartoon suggested by this data.

Extended Data Fig. 6. M-type clustering and manual labels.

a) Matrix of manual labels (x-axis) vs M-types. b) UMAP representation of features, colored by manually labeled cell types. c) Co-clustering matrix of excitatory cells, indicating the number of times a pair of cells was clustered together by iterations of the phonograph algorithm. Cells are ordered by subsequent agglomerative clustering on this matrix. d) Feature importance for each M-type, based on training binary random forest classifiers to predict each M-type separately and computing the mean decrease in impurity for each feature.

Extended Data Fig. 7. Somatic versus dendritic synapses across all excitatory M-types.

a) Median number of dendritic and somatic synapses for excitatory neurons of all M-types. Pearson r = 0.96, p = 5 × 10⁻¹⁰. b) Number of dendritic and somatic input synapses across all excitatory neurons, colored by M-type. Pearson r = 0.86, p < 1 × 10⁻¹⁰. Black line indicates linear fit with 95% confidence intervals from bootstrapping. c) Individual ordinary least square fits (with 95% confidence interval) for each M-type of z-scored dendritic synapses vs z-scored somatic synapses. With the exception of L5NP cells and deep layer 6b L6wm cells, all M-types have a positive relationship between predominantly inhibitory somatic synapses and predominantly excitatory dendritic synapses. d) Number of dendritic vs somatic input synapses for each M-type separately, linear fit line and 95% confidence interval.

Extended Data Fig. 8. Additional characterization of motif groups.

a–r) Morphology of all cells, organized by motif group. Within each group, cells are ordered by soma depth. Colors indicate M-type, darker lines indicate dendrites. s) The arbors of cells extend well beyond the columnar data. The scatterplot depicts a top-down view of soma locations of all synaptic targets of Cell ID 260622. Black dots are cells within the column, red dots are cells outside the column sample; dot size is proportional to number of synapses. t) The number of synapses from each interneuron onto target neurons within the column (black) and anywhere the dataset (red). Interneurons were ordered by within-column synapse count. The mean cell had 5.49 times more synapses across the dataset than onto column targets alone (black dashed line). Only targets passing basic quality control criteria were included. Note that while cells outside the sampled column are not necessarily proofread, synapses onto unproofread dendrites are nearly always correct (see Methods). u) Scatterplot of output synapse budget values within-column and dataset-wide (see v). The blue line indicates equality. The Pearson correlation between within-column measurements with the dataset-wide measurements was R = 0.9, not including trivial zeros (see Methods). v) Output synapse budget for each interneuron onto dataset-wide target M-types, using predictions from perisomatic features from Elabbady et al.. Note that the L6wm M-type was not included in predictions and is thus trivially zero for all interneurons.

Extended Data Fig. 9. Additional connectivity statistics within motif groups.

Connection density (left) measures the fraction of cells for a given M-type within the column targeted with at least 1 synapse. Synapses per connection (right) measures the average number of synapses in each observed connection. Single cell values are represented by dots, median values are shown with bars.

Extended Data Fig. 10. Selectivity and null models for inhibitory connectivity.

a) Number of synapses per M-type, compartment, and depth bin. These values were used as the baseline against which to compare synaptic output distributions for each inhibitory neuron. b) Expected value of each presynaptic inhibitory neuron according to an increasingly complex set of null models. Each row represents the fraction of synaptic outputs from a given inhibitory neuron (ordered as in Fig. 5a), distributed across excitatory M-types. From the left: 1) Synaptic outputs were proportional to the number of cells in each M-type, regardless of location in space. This approach accounts for the differing cell frequency for each M-type. 2) Synaptic outputs were proportional to the net number of input synapses for a given M-type, regardless of location in space. This approach accounts for the diversity in synaptic inputs for each M-type. 3) Synaptic outputs were distributed across compartments for each inhibitory cell as observed and distributed across M-types for each compartment separately. This approach accounts for the observed differences in compartment targeting for different interneurons. 4) Synaptic outputs were distributed across M-types within each of 50 depth bins, matching the observed depth distribution of synaptic outputs for each inhibitory neuron. This approach accounts for the spatial distribution of synapses, but not compartment targeting. 5) Synaptic outputs were distributed across M-types within both depth bins and compartments, matching the observed distribution of both. This approach accounts for both the spatial distribution of synapses and compartment targeting and is the most complete model considered here. At the far right, the observed distribution on the same scale, repeating the data in Fig. 5. c) Selectivity index (SI) for all cells, as described in the main text. Purple values have the observed number of output synapses significantly higher than a null model with matched compartment and depth targeting, while green are significantly less. Non-significant SI values are treated as 1. d) Difference between the observed distribution and the null model distribution for each cell as measured by the Kullback-Leibler divergence (from observed distribution to null distribution), by inhibitory subclass. Each colored dot is a cell, black dots are median with error bars indicating a 95% confidence interval based on a bootstrap. e) Comparison between the most complete null model across inhibitory subclasses. The PeriTCs have the lowest KL divergence of all types, indicating that the null model best predicts their connectivity. Note also that the individual cells exhibit a range of specificity relative to null models. f) Similarity of M-type synapse distributions in space, using the Bhattacharyya distance between the depth distribution of synaptic inputs onto soma and proximal dendrites (left) and distal and apical dendrite (right). Values closer to 1 indicate more similar distributions, values closer to 0 indicate more distinct distributions. g) All Bhattacharyya distance comparisons in e, with colored dots indicating pairs of distinct M-types, black dots indicating the median, and error bars showing a bootstrapped 95% confidence interval. Across all pairs, synaptic inputs onto the perisomatic and somatic compartments are more spatially segregated across different M-types than synaptic inputs onto distal and apical dendrites (p = 3.0 × 10⁻¹⁹, Mann-Whitney U test).