Main

To understand neural circuit function, it is critical to determine both the cell types within the circuit and how they interconnect. Inhibitory cells regulate when and how information travels within a neural circuit20 and, in the mouse primary visual cortex (VISp), the landscape of inhibitory neurons is complex. At least 28 inhibitory cell types (MET-types, that is, cell types defined by concordant morphology (M), electrophysiology (E) and transcriptomic expression (T)) have been identified by Patch-seq21. Patch-seq enables the morphology, transcriptomic profile and electrophysical signature to be determined for individual cells from in vitro slices13,14,15,22. However, this technique cannot resolve how these cells connect with each other. Such connectivity patterns can be identified with large-scale electron microscopy (EM)23. However, EM lacks the molecular information that could be used to relate connectivity data to transcriptomically defined cell types and their gene expression. We therefore performed an integrative morphological analysis to understand how these multi-modal cell types connect within the local circuit.

We leverage the morphological features of inhibitory MET-types to predict the MET-type or subclass identity of inhibitory cells in a large-scale EM dataset from mouse VISp. We then focused our study on Martinotti cells (MCs) in layers 4 and 5 (L4, L5) because they are well-accepted to be somatostatin (Sst) positive17,24, are morphologically distinct from other subclasses (for example, vasoactive intestinal peptide (Vip) and parvalbumin (Pvalb)) and are morphologically diverse16. Previous functional studies have also shown that MCs can connect broadly in a ‘blanket of inhibition’25,26,27 or inhibit a subset of excitatory neuron targets within and across cortical layers28,29,30,31,32. We sought to determine whether differences in synaptic connectivity of individual cells were aligned to differences in MET-types.

We reconstructed MCs from the EM, measured their synaptic connectivity and predicted their MET-type identity using a classifier trained on the morphology of inhibitory Patch-seq cells. We find that all of the EM MCs (n = 16) are predicted to belong to Sst MET-types. We also find that the predicted MET-types have distinct connectivity patterns. Individual cells within a predicted MET-type had largely consistent biases in connectivity onto excitatory targets. We also find that Sst MET-types can differ in the number and size of output synapses as well as axonal myelination patterns. These differences in synaptic features, connectivity, and myelination probably support distinct functional roles for inhibitory Sst MET-types.

By linking Patch-seq and EM data through neuron morphology, we have developed a method to explore the relationships between transcriptomically defined cell types, morphology, electrophysiology and synaptic connectivity. Using this approach, we find evidence that inhibitory MET-types participate in distinct cortical circuits and employ unique connectivity rules.

EM morphology is comparable with Patch-seq

A large-scale serial-section EM dataset, as described in the MICrONS consortium paper23, was collected from an adult mouse VISp and some higher-order visual cortical areas (Fig. 1a). In brief, the largest portion of the structural dataset contains 82,247 neurons, which have been segmented initially using a machine-learning model to allow synapse-level characterization of the connectivity and ultrastructure. However, these neurons require further extensive proofreading to extract their complete morphological information. This manuscript and ref. 33 focus on proofread subsets of these neurons to enable this characterization. Schneider-Mizell et al.33 proofread neurons in a 100 µm × 100 µm columnar sample, which contains 163 inhibitory cells. These neurons have morphological features that are consistent with most of the main inhibitory subclasses (Lamp5, Sst, Pvalb, Vip).

Fig. 1: Comparison of EM and Patch-seq pipelines and morphologies.
figure 1

a, Subset of morphologies available from large-scale millimetre-squared EM dataset volume. Each cell is in a different colour. b, Schematic representation of the Patch-seq and EM pipelines for generating morphological reconstructions and comparison of features across pipelines. Ephys, electrophysiology; swc, file format of skeletonized neuronal reconstructions. c, Confusion matrix of RFC MET-type predictions showing the frequency with which the RFC correctly predicted the MET-type of Patch-seq cells (normalized by row). The RFC was trained on morphological features of inhibitory Patch-seq data (n = 477). We used this classifier to predict MET-type or subclass identity of EM cells (n = 173). d, Top, example Patch-seq morphology and average axon/dendrite histograms of MET-types. Bottom, example EM cells morphology and average axon/dendrite histograms grouped by predicted MET-type (MET-8, n = 3; MET-4, n = 6; MET-6, n = 5). Illustration in a adapted from refs. 33,60, Springer Nature Limited.

Full size image

In this study, we focus on a dataset of curated MCs (n = 16) that are in the column and across the VISp region of the volume. MCs were identified by their sparsely spiny dendrites, an axon emerging from the pia side of the soma and a primary axon branch that reached L1 (ref. 19). The MCs were proofread using the same approach as the cells in the columnar sample, but every algorithmically identified branch and end point was checked manually and extended to create a comprehensive reconstruction. To date, these are some of the largest EM reconstructions of single neurons, with more than 5,000 μm of dendrites and more than 23,000 μm of axon reconstructed, and extending more than 600 μm vertically and horizontally across the cortex. The largest MC, in terms of output, makes 12,540 total output synapses across 3,937 unique postsynaptic targets.

To determine the feasibility of using a classifier trained on the morphological features from Patch-seq data to predict the identity of EM cells, we compared the mean and distribution of features of inhibitory neurons from both datasets (Extended Data Fig. 1). The Patch-seq data span all main inhibitory cell subclasses, but, unlike the EM dataset, cells were sampled using transgenic lines and targeted for electrophysiological recordings16.

To compare the morphology of EM and Patch-seq cells, we aligned EM reconstructions to an average cortical layer space and calculated 43 morphological features using a pipeline developed for Patch-seq data12,16 (Fig. 1b). Despite the different sampling methods, we find that the values from the inhibitory cells are similar between datasets. Comparing z scores, the mean of the Patch-seq and EM distributions are less than 1 s.d. apart for most features (42 of 43, axon_depth_pc_04) (Extended Data Figs. 1 and 2a). All but one feature (axon_depth_pc_03) of the EM MCs (presumed Sst-positive) also fall within the range of Sst Patch-seq values (Extended Data Fig. 3). We therefore proceeded to use the morphological features to predict inhibitory subclass identity for all inhibitory neurons in the columnar sample and MET-type identity for EM MCs (Extended Data Fig. 4).

Classifier predicts Sst subclass of EM MCs

We trained a random forest classifier (RFC) on the morphological features of inhibitory Patch-seq cells to predict inhibitory MET-types. We achieved an estimated classifier accuracy of 59%, which is greater than expected if cells were assigned randomly to one of the 22 MET-types (around 4.5%) (Fig. 1c). At the subclass level, Sst Patch-seq cells were predicted correctly to be in the Sst subclass 87% of the time (Extended Data Fig. 5a).

To assess the relationship between the frequency and accuracy of a predicted label, we predicted a MET-type label for each Patch-seq cell over 500 iterations and compared the frequency versus the accuracy of that prediction (Extended Data Fig. 2b; Methods). We find that MET-type labels are more likely to be accurate if the label was predicted at least 54% of the time (Extended Data Fig. 2b). We then used this frequency as a threshold to select EM cells for connectivity analysis. All EM MCs were predicted to belong to an Sst MET-type (Fig. 1d and Extended Data Fig. 4) with a frequency of 79% or greater (‘Reliability’; Extended Data Table 1).

EM MCs are assigned to several MET-types

MET-types16 represent cell types with concordant morphology, electrophysiology and transcriptome. Some MET-types contain cells from a single transcriptomic type (t-type) whereas other MET-types contain cells across several transcriptomically similar t-types6. The L4 and L5 EM MCs are predicted to belong to one of five Sst MET-types described below.

Three EM cells are predicted to belong to the Sst MET-8 type (Extended Data Fig. 6). Sst MET-8 consists of cells from a single t-type, Hpse Cbln4, which have somas located in L4 and upper L5. Hpse Cbln4 cells are described as non-MCs in somatosensory cortex22,31, but have some L1 projection in primary visual cortex and hence are considered MCs16,22.

Six EM cells are predicted to be Sst MET-4 type (Extended Data Fig. 6), which consists predominantly of cells from the Sst Calb2 Pdlim5 t-type. The Sst Calb2 Pdlim5 t-type splits across layers with some in L2 and L3 (Sst MET-3) and some in upper L5 (Sst MET-4)16, but all Sst Calb2 Pdlim5 cells are characterized by a L1-dominant axon lamination (‘Martinotti’) pattern19,34, consistent with T-shaped MCs, although Sst MET-4 cells have the overall shape of a fanning-out MC28,30.

Five EM cells are predicted to be Sst MET-6 (Extended Data Fig. 6), which are described as having ‘T-shaped’ morphology with an axon that reaches L1 but has dominant L5 innervation30. The Sst MET-6 type is composed of several t-types that are in proximity along the transcriptomic dendrogram, including Sst Myh8 Etv1 and Sst Chrna2 Glra3.

Finally, one cell is predicted to be in the Sst MET-5 type and one is predicted to be Sst MET-9 (Extended Data Fig. 6). Sst MET-5 is morphologically similar to Sst MET-6 (ref. 16), with both Sst Myh8 Etv1 and Sst Nr2f2 Necab1 cells (similar to fanning-out MCs)16,28,30. The Sst MET-9 type is made up primarily of the Sst Tac2 Tacstd t-type and has an axon that reaches L1 but has a peak in L5.

These results demonstrate that we can reliably assign a morphological, electrophysiological and transcriptomic identity to neurons characterized in an EM volume using local dendritic and axonal morphology. In the remainder of the paper, we will refer to these ‘predicted MET-types’ as MET-types and focus analysis on those with at least three cells (MET-8, MET-4 and MET-6).

To determine what proportion of the Sst subclass is represented by each of these MET-types, we referenced a published whole mouse brain MERFISH dataset (Methods), which facilitates quantification of the MET- and t-type spatial distribution across the mouse brain. We found that Sst MET-4, MET-6 and MET-8 represent 3.4%, 16.5% and 18.5% of the Sst subclass, respectively, in VISp (Extended Data Table 2). Combined with three other Sst MET-types, the L2 or L3 Sst MET-types MET-2 (16.9%) and MET-3 (11.8%) and the L6 Sst MET-12 (9.9%), these six MET-types constitute most cell types in the Sst subclass in VISp. Thus, with our MET-type classifier, we are able to describe the connectivity of the most abundant L4 and L5 MET-types in the Sst population.

Synapses and myelin vary by MET-type

From EM, we can measure a rich set of ultrastructural features, including synapse distributions onto postsynaptic cell types, what fraction of each cell type receives synapses, and myelination patterns. There is striking and strong diversity in each of these properties across the population and our classifier results facilitate examination of how these differences align with Sst MET-type predictions.

Output synapse features

In the EM volume, most output synapses (average 82.5% of all output synapses; Extended Data Table 1) along the reconstructed axons can be mapped onto a single postsynaptic target cell in the dataset. Each target cell was assigned a cell type based on somatic features using automated methods (‘Identifying synapses’; Methods). This high rate of detection of individual target cells allowed us to confidently investigate the relationship between MET-type identity and connectivity statistics (Fig. 2a–c,f).

Fig. 2: Output synapses and myelination of MET-types.
figure 2

a, Example cell from each predicted MET-type showing output synapses (cyan dots) and myelination (magenta). Insets, myelination of a predicted MET-8 cell and myelinated axon in EM. b, Same example cells (plus further examples) from each predicted MET-type with output synapses colour-coded by synapse target (as identified by a classifier trained on somatic features60). c, Average histogram of synapses onto targets by predicted MET-type. d, Quantification of number and size of output synapses by MET-type. e, Quantification of myelination features by MET-type. f, Quantification of the total number and percentage of output synapses onto each target cell type by MET-type. Bars indicate significant differences (P < 0.05 Kruskall–Wallis for group and Conover post hoc with Bonferroni correction, *P < 0.05, **P < 0.01; MET-8, n = 3; MET-4, n = 6; MET-6, n = 5). Scale bars, 10 μm (first inset), 1 μm (second inset) (a), 100 μm (b).

Full size image

We observed that MET-8 cells form significantly more output synapses (9,621 ± 1,460; mean ± s.e.m.) than MET-4 (2,280 ± 250) or MET-6 (1,499 ± 121) cells (Fig. 2a–c; for quantification see Fig. 2d,f). Whereas MET-8 cells have more total axon, the MET-8 synapse density is nearly twice the density of MET-4 and MET-6, approximately 0.54 versus 0.25 synapses per micrometre (Extended Data Fig. 5c,d).

We also find significant differences in the size of output synapses between MET-types. The output synapses of MET-8 cells are significantly smaller than those from MET-4 cells. MET-8 cells form synapses that are 70% the size of MET-4 synapses and 78% of the size of MET-6 synapses (Fig. 2d). MET-8 cells also form significantly more synapses per connection than other MET-types (Fig. 2d). This finding is discussed in greater detail in ‘Single versus multi-synaptic connections’.

Myelination patterns

EM data includes information about myelination, which can influence the biophysics of a cell35. Although inhibitory neurons36,37, including MCs38, have previously been shown to be myelinated, little is known so far regarding the relationship between myelination pattern, Sst cell types and biophysical properties. We found distinct myelination patterns across the EM MCs that varied with the three MET-types (Fig. 2a,e and Extended Data Fig. 6).

MET-4 cells (5/5 cells) are myelinated along their main ascending axon collateral. MET-8 cells (3/3) are myelinated near the soma along a short stretch of the primary axon branch, which rarely extends into L2/3. MET-6 cells are either not myelinated (1/6) or sparsely myelinated with a less clear pattern (5/6) (Fig. 2a,e and Extended Data Fig. 6). Most of the reconstructed cells had some portion of their axon myelinated, but the number of myelinated segments and total path length of the myelination varied by MET-type (Fig. 2e).

The ascending axon stalk of MET-4 cells is ensheathed by several segments (average 6.7 segments) separated by nodes of Ranvier. MET-4 cells have two-and-a-half times as many segments as MET-8 (2.7) and five times as many as MET-6 (1.4). We find no significant difference in the length of individual myelin segments between cell types. MET-4 cells have approximately 220 µm of total length of myelination, which is nearly three times the length of myelin of MET-8 (around 69 µm) and six times the length of myelin of MET-6 (36 µm) (Fig. 2e).

We hypothesize that such differences in output synapse number, size, density and myelination between these cells could underlie distinct signalling properties, both in terms of output strength and timing20,39,40. Further work to characterize the electrophysiological output of these neurons could shed light on the effects of these ultrastructural differences.

MET-types have distinct connection rules

Synaptic connectivity

After noting the distinct ultrastructural features between the inhibitory cells, we next examined whether MCs synapse onto similar cell types. Excitatory cortical cell types (pyramidal cells (PCs)) are named for both the layer in which they reside (for example, L2/3) and where they project their axons. Intra-telencephalic (IT) neurons project within the cortex-dominated anterior forebrain and extra-telencephalic (ET) neurons also project beyond the forebrain41. IT and ET (sometimes called PT) neurons are functionally distinct and differentially implicated in several diseases42. We determined the synaptic connectivity pattern of each Sst MET cell using automated methods to detect synapses and assign target type identity mentioned above (see ‘Identifying synapses and postsynaptic targets’ in Methods).

Diverse connectivity patterns amongst the inhibitory cells largely align with predicted MET-type (Fig. 2b,c,f and Extended Data Fig. 7). MET-8 cells preferentially target L4 IT PCs (65.7% of the total number of output synapses ± 4.0; mean ± s.e.m.). L5 IT and L2/3 PCs (12.5 ± 3.5% and 11.5 ± 1.1%, respectively) are the next main targets.

MET-4 cells target predominantly L2/3 PCs (41.0 ± 3.3%) followed by L5 IT PCs (19.6 ± 5.1%). L4 IT and L5 ET cells receive fewer synapses (14.7 ± 3.5% and 10.5 ± 4.3%, respectively). Four of the six predicted MET-4 followed this average pattern, although two MET-4 cells preferentially synapse with L5 ET rather than L5 IT cells, which may indicate variability either in targeting for this MET-type or misassignment of MET-type for these two cells (indicated by asterisks in Extended Data Fig. 7). Further characterization of the connectivity of MET-types will help to elucidate their intrinsic variability.

MET-6 cells form most synapses onto L2/3 PCs (35.0 ± 2.0%) followed by L5 ET neurons and L6 PCs (29.4 ± 2.8% and 11.2 ± 2.1%, respectively). Comparing the connectivity of MET-4 and MET-6 cells is of particular interest, because both groups have cells with classic Martinotti morphologies, yet differences in branching properties predict that they belong to distinct Sst MET-types and they have distinct connectivity profiles.

Despite having somas in L5, we find that the main targets for both Sst MET-4 and MET-6 are L2/3 PCs. This contrasts with the view that the primary role of L5 MCs is to inhibit the apical tufts of L5 excitatory neurons43,44, although these synapses are present on the apical tufts of the preferred L5 target type of each type (Fig. 2c). MC axons have previously been described as overlapping with the basal dendrites of their L5 targets19,28,45,46; here we show that more synapses onto the preferred L5 targets are formed within L5 than L1 (peaks at 0–100 µm from pia; Fig. 2c), as opposed to purely targeting apical tufts.

Connectivity rates

The distinct output synapse distributions of the Sst MET-types suggests that Sst cells are not synapsing indiscriminately, but it does not measure directly whether all available cells are innervated, as predicted by a ‘blanket’ inhibition model25. Therefore, we calculated the percentage of neighbouring cells (within a given target cell type) that receive synapses from a presynaptic MC. We find that MCs form the most connections with neighbours within 100 µm. However, the percentages vary by MET-type and target cell type and rarely reach 100% (Fig. 3a).

Fig. 3: Exploring pairwise connectivity of MET-types and postsynaptic targets.
figure 3

a, Average histogram of percentage of each cell type that receives synapses as a function of radial (xz) distance between presynaptic EM cell and postsynaptic target. All distributions were compared pairwise by a Kolmogorov–Smirnov test with a false discovery rate correction. *P < 0.05 for all comparisons of that distribution to others. b, Percentage of connections that contain a single versus multiple synapses (shown up to five) for each target cell type across predicted MET-types (for example, predicted MET-6 cells form more multi-synaptic contacts onto L5 ET targets than onto other cell types). c, Example cell from each predicted MET-type showing soma locations of postsynaptic targets. Somas are colour-coded to indicate the number of synapses that cell receives from the presynaptic cell (soma in cyan). All examples use the same scale. d, Histograms of inter-synaptic distances onto target types (distances calculated per postsynaptic target). MET-8, -4 and -6 (n = 3, 6 and 5, respectively).

Full size image

MET-8, MET-4 and MET-6 cells synapse onto about 40% or fewer of the neighbouring cells from each target cell type within 50 µm, with a few notable exceptions as mentioned below. This result demonstrates that these inhibitory cells rarely synapse onto every available neuron. However, MET-8 cells do synapse onto nearly 100% of L4 IT cells within a 50-µm radius and 80% within a 100-µm radius, producing a ‘blanket of inhibition’ for this specific cell type. Similarly, MET-6 cells target nearly 60% of their preferred L5 postsynaptic target (L5 ET neurons within 100 µm), but not other cell types (they target fewer than 31% of the available L2/3 cells within 50-µm radial distance). MET-4 cells have a different targeting pattern. They target less than 40% of all available cells across types within 50 µm), including their preferred target cell types. Thus, each Sst MET-type uses a distinct inhibition pattern.

Single versus multi-synapse connections

Given that individual MCs do not connect to all neighbours, preferences in output synapse distribution could be due to differences in the number of synapses onto each target cell (number of synapses per connection). Therefore, we calculated the average number of synapses from each MC onto each target cell across MET-types (Fig. 3b). We found that most connections contain a single synapse; however, the fraction of single-synapse and multi-synapse connections differs by postsynaptic target cell type (Fig. 3b).

MET-8 cells make multi-synaptic contacts onto their preferred targets: L4, L5 IT and inhibitory cells. Specifically, more than 60% of all connections onto L4 cells are multi-synaptic and approximately 50% of connections onto L5 IT and inhibitory targets are multi-synaptic. Some targets receive more than 20 synapses from a single MET-8 cell (Fig. 3c). MET-4 cells form the fewest multi-synaptic connections and show the least modulation with respect to target cell subclass. MET-6 cells form predominantly single-synapse connections, including onto their preferred targets (L2/3 PCs), but do form highly multi-synaptic connections onto around 50% of targeted L5 ET PCs. In summary, MET-8 forms many multi-synaptic connections onto its preferred postsynaptic targets, MET-4 forms predominantly single-synapse connections and MET-6 forms multi-synaptic connections onto one specific target cell type.

Given the total number of output synapses from an individual neuron onto a limited number of targets of that type, a model that distributes those synapses randomly amongst the targets would produce a Poisson distribution of multi-synaptic connections. We find that we can reject the hypothesis that synapses from individual presynaptic cells onto a given target cell type are Poisson distributed, with only a few exceptions. For example, most (4/5 cells) MET-6 cells synapse onto L4 IT cells in a Poisson distribution (where P > 0.05) (Extended Data Fig. 5e).

In Fig. 3c, we provide examples of the soma locations of postsynaptic targets (colour-coded by the number of synapses received) from each MET-type. These plots illustrate the differences in the target location and range in synapse number per connection for each MET-type. MET-8 cells form multi-synaptic connections with L4 and L5 IT PCs, visualized by the cluster of orange–yellow somas directly above the presynaptic (cyan) cell soma. MET-4 cells target diffusely across most targets but form a few multi-synapse connections in L5. Finally, MET-6 cells target diffusely across most targets, but form multi-synaptic connections with L5 ET cells located just above the presynaptic cell soma (Fig. 3c).

Dendritic targeting of MCs onto L5 PCs

To determine whether Sst MET-types form spatially clustered synapses onto a postsynaptic target, we quantified the inter-synapse distances of a given pre–postsynaptic pair. We find that most synapses are formed within 150 µm of each other (Euclidean distance) (Fig. 3d). However, both MET-4 and MET-6 cells have many synapses that are more than 300 µm apart. These distances may be due to synapses formed onto both the apical and basal dendrites of a target cell or across a wide lateral extent of basal dendrites.

One hypothesis regarding L5 MC connectivity is that they form synapses onto both the apical and basal dendrites of L5 excitatory neurons (see histograms in Fig. 2c and more than 300-µm distances in Fig. 3d) to coordinate inhibition across compartments of individual cells. We determined the location of synapses from MET-4 and MET-6 cells onto L5 IT and L5 ET targets (Extended Data Table 3). Both MET-4 and MET-6 cells form approximately half of their synapses onto apical dendrites (around 51% and around 47%, respectively). Approximately 42% of MET-4 connections are apical only, around 46% are non-apical and around 11% are onto both apical and non-apical (probably basal) dendrites. We find about 34% of MET-6 connections are apical only, around 50% are non-apical and around 16% are onto both apical and basal dendrites. Thus, the coordinated inhibition across compartments of L5 PCs by individual MCs occurs in only a small fraction of connections to postsynaptic targets; however, most connections are made onto a single compartment, and both apical and non-apical dendrites receive similar numbers of synapses.

Generalization of the morpho-classifier

After characterizing the ultrastructural differences that align with the predicted MC MET-types, we tested the generalizability of the classifier. We compared the predicted MET-subclass for inhibitory cells within the columnar sample to the ‘targeting type’ of the cells defined by Schneider-Mizell et al.33. The predicted subclasses largely align with the targeting type (Extended Data Fig. 2c). Most predicted Sst cells (64%) are distal dendrite targeting cells, most predicted Pvalb cells (78%) are perisomatic targeting cells, and Vip cells (91%) are inhibitory targeting cells. These patterns match with published connectivity patterns for these subclasses9. We observe some off-diagonal mapping, although further work is needed to determine whether this is driven by methodology or biology. Patch-seq and EM cells grouped by predicted subclass and targeting type are shown in Extended Data Fig. 4.

Limitations

There are some technical artifacts that will have minor effects on the quantifications of connectivity we report. First, we focused our analysis on a small number of reconstructed EM cells. These cells are highlighted because they are comprehensively proofread and are within the same cortical region (VISp) as the Patch-seq data. It is possible that we have only sub-sampled MET-types and our findings are an incomplete description of the diversity within the cell type. Another possibility is that some MET-types are more intrinsically variable. The consistency of MET-type connectivity will need to be evaluated with further studies. The top 10 µm of the cortical surface is also not included in these reconstructions owing to segmentation errors. However, we find that synapses onto L5 targets peak at a lower depth compared with those onto L2/3 targets. We therefore do not expect either of these effects to be large enough to change the overall conclusions reported here.

There are potential differences in the morphological feature values that are driven by differences in the Patch-seq and EM experimental procedures. Theoretically, the classifier trained on Patch-seq data could be less accurate for EM data than it appears to be in cross-validation experiments applied to Patch-Seq. However, the differences in morphology we observe between MET-types, which we focus on here, are clear to a trained anatomist, and differences in connectivity are largely consistent from cell to cell. We take the consistency between the MET predictions and stark shifts in connectivity as support for true biological differences between these groups. It is also possible that closely related MET-types (for example, Sst MET-5 and Sst MET-6) can have similar targeting and that this modality can be used to further refine or collapse cell type(s). Thus, we believe that exploring the synaptic connectivity across all cortical MET-types is an exciting future application of this type of work.

Discussion

By establishing a morphological feature set aligned across EM and a multi-modal Patch-seq dataset, we can map cells between the two datasets to predict the connectivity of different types of inhibitory neuron. We can also predict the molecular identity and electrophysiological profile of neurons sampled in the large volume EM data. These predictions reveal that the MET-types analysed consistently differ in their myelination, synaptic features and target cell subclass connectivity profiles, which suggests that they play unique roles in cortical circuitry (Fig. 4). We believe this approach could be a powerful tool to leverage existing and future Patch-seq and EM datasets to predict cellular properties across methodologies.

Fig. 4: Integrated view of MET-types including modalities from Patch-seq and EM.
figure 4

a, First column, average percentage of synapses onto each postsynaptic target group. Second column, schematic summarizing the connectivity motifs observed from EM data. Third column, example cells from the EM dataset. Fourth column, average action potential (AP) traces per MET-type. Fifth column, transcriptomic cell types that comprise previously defined MET-types. b, This integrated view of MET-types now allows us to generate hypotheses such as the role of highly expressed genes in a given transcriptomic type on connectivity patterns. c, Dot plot showing the fraction of cells within the MET-type (circle size) and expression level (red to blue scale bar) of the genes that are differentially expressed across the transcriptomic types in MET-8, 4 and 6 (n = 18, 15 and 18, respectively) and present in at least 50% of one MET-type. Genes listed are the top five upregulated or downregulated genes by pairwise comparison (minus duplicates). aData previously published16. Credits: plots in a reproduced with permission from ref. 16, Cell Press.

Full size image

Integration of MET-type and connectivity

Previously defined MET-types16 have been shown to have distinct morphological, electrical and transcriptomic properties (Fig. 4a). We can now see how Sst MET-8, MET-4 and MET-6 target distinct excitatory neuron subclasses using different connectivity motifs (Fig. 4a). We can also generate hypotheses about what features measured in Patch-seq (for example, electrophysical properties, transcriptomic expression patterns, morphology) may correlate with the connectivity patterns observed for cells with similar morphology in EM (Fig. 4b). We show expression patterns of the highest differentially expressed genes between the principal Sst t-types that comprise the MET-types (Fig. 4c). These genes could play a role in setting up and/or maintaining the distinct morphology and connectivity of these MET-types. Several of the genes identified above are involved in cell adhesion (Cdh12 (ref. 47), Pcdh8 (ref. 48)) or both adhesion and inhibitory synapse formation (Ptprd)49,50. Calsyntenin-2 (encoded by Clstn2) is also found at GABAergic synapses, and knocking-out the gene leads to fewer parvalbumin neurons and inhibitory synapses in the hippocampus51 and loss of inhibitory synapses in cortex52. Identifying subcellular protein expression patterns and loss-of-function studies of these proteins during development would shed light on their role in synaptic patterning in these MET-types.

Potential role of diverse myelination

Although myelination is frequently described for excitatory neurons, previous studies have shown that inhibitory neurons36,53, including MCs38, can be myelinated. Here we find significant differences in the number and total length of myelinated segments across Sst MET-types (Fig. 2e). Reduced myelination onto Pvalb neurons can reduce firing rates and conduction velocity36,54. Differences in the myelination of Sst cells may similarly influence their firing properties. For example, the myelination of the ascending axon stalk of MET-4 cells could help to synchronize inhibition of L2/3 and L5 IT PCs and/or increase the speed of inhibition onto L2/3 targets of MET-4 versus MET-6 cells (Fig. 2a,e). Oligodendrocytes—the glial cells that myelinate axons—have also been shown to metabolically support the axons they ensheath55,56. Thus, myelination onto MET-4 cells may indicate glial support of their larger output synapses (Fig. 2d). Our observations suggest that myelination may vary in a cell-type-specific manner and could provide another marker of cell-type identity in EM datasets.

Potential circuit roles of synapse size

We find significant differences in output synapse sizes across the observed MET-types (Fig. 2d). Correlated slice electrophysiology and EM performed on excitatory cells demonstrated a linear relationship between chemical synapse size and strength39. Thus, an individual MET-8 synapse could be weaker than an individual MET-4 or MET-6 synapse. However, we also find that MET-8 cells form significantly more output synapses, so a single synapse analysis might be under-counting their inhibitory influence on the circuit.

MCs can form several synapses onto a single target, so we calculated the average connection size of a cell (Extended Data Fig. 5c). We find that MET-8 cell connections are larger than those of MET-6 (not MET-4). It is possible that the greater number of synapses onto individual targets may be related, in part, to the smaller synapse size. The number and size of inhibitory synapses onto a target may balance the excitation impinging onto that cell and shape its response. In the chicken auditory nucleus, inhibition varies along the tonotopic axis, which shapes the timing and dynamic range of postsynaptic responses57. Thus, inhibitory synapse size may be a function of homeostatic plasticity and/or be suited to the features encoded by the postsynaptic targets of each MET-type.

Inhibitory connectivity rules

Each Sst MET-type uses a distinct inhibition pattern: inhibit cell types in proportion to preference, culminating in targeting nearly 100% of the neighbouring preferred cell type (MET-8); inhibit a small fraction of all target cell types (MET-4); or inhibit a small fraction of most target cell types except the preferred ‘local’ target (MET-6). These findings align well with paired recordings showing high connection probability from L4 Sst cells to L4 PCs and L5 Sst cells to L5 ET cells58. Our radial distance findings also align with studies from brain slices showing that most connections from Sst cells are onto excitatory targets within a 200-µm lateral distance58,59.

Previous studies found dense connectivity from L2/3 Sst cells onto L2/3 PC targets, creating ‘blanket inhibition’ in the cortex25, and predicted ‘non-specific’ connectivity onto most cell types10. We observe ‘blanket inhibition’ from the MET-8 cells onto L4 PCs, but not from MET-8 onto other target subclasses or from other Sst MET-types. Thus, ‘blanket inhibition’ is present in specific contexts, but it is not a universal connectivity pattern for Sst cells, highlighting the diversity of connectivity motifs within this subclass.

Alignment with other studies

Previous studies have shown that Hpse Cbln4 cells in mouse primary somatosensory and visual cortex synapse predominantly onto L4 and not L5 PCs22,31. Our EM results not only recapitulate the connectivity bias observed in mouse sensory cortices (L4 > L5), but also reveal the degree to which MET-8 cells synapse across cell types.

Previous studies also found that Chrna2+ MCs in L5 of auditory cortex preferentially inhibit thick-tufted (probably L5 ET cells) but not thin-tufted (probably L5/L6 IT) cells29, and optogenetic mapping of Sst Myh8 cells (genetically targeted using Chrna2-cre) shows that they more strongly inhibit L5 ET than IT cells32. MET-6 cells, which include Chrna2+ cells16, form more synapses onto L5 ET than IT cells. Thus, the MET-6-type structural connectivity we characterized aligns with the probed physiology of Chrna2+ MCs29.

Recent work in mouse VISp finds that bulk optogenetic activation of virally labelled Sst Calb2 cells (Sst MET-3 and Sst MET-4) produces larger amplitude iPSCs onto L5 ET than L5 IT cells32, whereas we find that the MET-4 group forms more output synapses onto L5 IT cells and has similar connection probabilities to both populations. Thus differences in connection strength could be driven by differences in the convergence of connections or the output connectivity of MET-3 and MET-4 types, or could reflect that MET-4/Sst Calb2 cells have functionally stronger individual synapses onto ET than IT cells. Predicting a molecular cell type for EM data enables the comparison of circuit and synaptic features to physiological features. Future targeted studies to examine the functional dynamics of individual inhibitory synapses onto distinct excitatory target populations would help to resolve observed differences such as this.

Finally, we have demonstrated that local morphological features can be used to link cell types across datasets. Linking these cell type identities enables the investigation of synaptic connectivity with respect to morphology, electrophysiology and transcriptomic expression in advance of the availability of technologies that allow the direct measurement of these properties from the same cell. As larger electron microscopy datasets are generated, this approach could be extended to cell types across the whole mouse brain and other species. Measuring the synaptic connectivity of transcriptomic cell types will facilitate future work to characterize the behaviour of these types, in local and brain-wide circuits, using genetic tools.

Methods

EM dataset generation and image alignment

The EM dataset described above is from tissue collected from the visual cortex (including primary and higher-order visual areas) dissected from a P87 (postnatal day) male SLC17a7-CrexAi162 heterozygous mouse23. In brief, a large-scale serial-section EM dataset was collected and imaged using automated transmission electron microscopes61. The data above are from a sub-volume representing 65% of the original EM volume with images of approximately 4 × 4 × 40 nm per pixel resolution. These images were segmented into meshes using convolutional neural networks and subsequent agglomeration62. The EM images and meshes are visualized in Neuroglancer. These meshes can be proofread (merged or split) within the ChunkedGraph system33,63 Neuroglancer framework to facilitate proofreading of cells.

Correcting and generating representations of cells

Meshes underwent skeletonization (skeleton originated from a defined soma point) to generate a list of branch and end points for each mesh, visible in Neuroglancer33. Each branch point was inspected manually. True branch points were left alone and false branch points (often due to overlapping processes from distinct cells) were split using Neuroglancer tools. Subsequently, each end point was inspected manually. True endpoints were left alone and false endpoints (premature end of a process) were extended by an expert annotator, who would follow the process along the EM imagery to a natural ending (bouton, tapered end) or until the process could no longer be extended reliably (for example, edge of block). The total number of EM reconstructions of inhibitory cells used in this study is 173 (including clean and comprehensive reconstructions).

Morphological analysis and MET-type prediction

Determining cortical layers

Soma position, pia, white matter and laminar borders were drawn manually. For Patch-seq cells (data previously collected) a ×20 brightfield and fluorescent image of 4′,6-diamidino-2-phenylindole-stained tissue was used16 and for EM cells, an expert annotator drew layer boundaries based on soma density near a z plane that contained the soma of the cell of interest. Drawings were made in Neuroglancer on a single EM z plane for the 16 MCs. Cortical layers were drawn at a central z plane of the cortical column and were applied to all cells from the column that were not the curated MCs. These polygons were then exported to be used for feature calculation.

Morphological feature calculation

Morphological features were calculated as previously described12, using features that were derived from previous studies11,64. Features were calculated using the skeleton keys Python package (https://github.com/AllenInstitute/skeleton_keys). First, neuron skeletons were extracted, resampled (at 1,550), and exported from the EM dataset using the pcg_skel python package (https://github.com/CAVEconnectome/pcg_skel). Features were extracted from neurons aligned in the direction perpendicular to pia and white matter. Laminar axon histograms (bin size of 5 µm) and earth movers’ distance features require a layer-aligned version of the morphology where node depths are registered to an average laminar depth template. These features include measurements such as total axon length, maximum path distance and total number of branches for both axon and dendrites from skeletonized representations.

RFC details

An RFC, support vector machine and logistic regression model were assessed for performance in predicting MET-type labels for Patch-seq cells using the morphological features of inhibitory cell types from a previously published Patch-seq dataset (Patch-seq n = 477, Sst n = 236, Pvalb n = 89, Vip n = 79, Lamp5 n = 44, Sncg n = 29)16. Specifically, we used a stratified fivefold cross-validation approach to assess classification accuracy. The data were split randomly into five partitions while maintaining the distribution of MET-type labels in each partition. This method iteratively rotates which partition is withheld from training and used to validate the model. MET-types with fewer than five morphological reconstructions (Lamp5 MET-2, Pvalb MET-5, Sncg MET-2, Sncg MET-3, Sst MET-11 and Vip MET-3) were omitted. An RFC with 250 estimators, a maximum depth of ten, balanced class weights, a minimum of ten samples per split and at least five samples per leaf node outperformed logistic regression and support vector machine models. Fivefold stratified cross-validation with shuffling was repeated 20 times and achieved a mean accuracy of 58.9 ± 4.1% (s.e.m.), far exceeding the expected chance accuracy for 22 categories (4.5%). Classifier accuracy was determined by how frequently the model correctly predicted the MET-type label of held out data (not used in training the model) (Fig. 1c and Extended Data Fig. 5a). We calculated an overall F1 score of 0.58, based on averaging F1 scores for each MET-type, based on classifying that type versus any other type. The cumulative confusion matrix for hold-out validation data was recorded in Fig. 1c.

EM MET-type inference

For 500 iterations, a random subsample (95%) of the Patch-seq data was selected with probabilities according to MET-type class size (a Patch-seq cell from a well-represented MET-type was more likely to be omitted). MET-types with five or fewer specimens were exempt from subsampling. In each iteration, a new RFC with the aforementioned parameters was fitted with sub-sampled Patch-seq data and MET labels were predicted for EM cells. The final MET assignment was given as the most frequently predicted MET label for each cell (Extended Data Table 1). We used these predicted MET-type labels to group cells for subsequent analysis.

Calculating a reliability threshold

A reliability metric was quantified as the fraction of iterations each sample was predicted as its final MET assignment out of all predictions (for example, a cell was predicted into this MET-type 80% of the time). To determine an appropriate threshold, we explored reliability scores in the Patch-seq data. We applied random subsampling iterations in a leave-one-out manner to the Patch-seq data. Here we set one Patch-seq sample aside and use the rest as training data. In each iteration, the training data are sub-sampled randomly as described above and used to fit a new RFC. The classifier predicts the MET-type label of the single left-out Patch-seq sample. This process was repeated 500 times for every Patch-seq cell until each had a predicted MET-type label and corresponding reliability metric. We then plotted a cumulative histogram of the correctly and incorrectly predicted labels versus the reliability score. We found that a reliability score of more than 0.54 is the most inclusive value at which Patch-seq samples are more frequently predicted correctly than incorrectly (Extended Data Fig. 2b).

Morphological feature z score analysis

The z scores for Patch-seq and EM morphological features were calculated; z scores for EM morphological features were derived using the mean and s.d. of the Patch-seq data.

Poisson distribution analysis

We fit a simple Poisson model to match the total number of connections and mean number of synapses per connection for a given cell type distribution. We used a chi-squared test to probe whether the measured distributions were significantly different from the Poisson model. Specifically, the total number and distribution of output synapses onto target cell types were calculated for each cell. We calculated the number of single-, double- and triple-synapse (up to 17-synapse) containing connections for each presynaptic MC onto each target cell type (L2/3, L4, L5 ET, and so on), then built a Poisson distribution for each presynaptic cell and postsynaptic target cell type using the total number of connections and the average number of synapses per connection. We then compared the real and Poisson distributions using a Chi-squared test (degrees of freedom = 0) to determine whether we could reject the null hypothesis (that the distributions were the same, P < 0.05).

Subclass prediction and cell targeting for EM column data

A MET-type prediction was made for each inhibitory neuron from a defined 100 µm × 100 µm cortical column (these cells are analysed in greater detail in ref. 33). Given the diverse population of interneurons captured in these data and the Sst-centric focus of this paper, the data are presented at the subclass level (apart from Extended Data Fig. 4, which shows Patch-seq and EM morphologies mapped to Sst MET-types or non-Sst subclasses). Subclass labels were assigned based on MET-type classifications (for example, all Pvalb MET-types are categorized under the Pvalb subclass).

Connectivity based inhibitory subclasses were defined by Schneider-Mizell et al.33. In brief, the dendritic arbour of excitatory neurons was assigned to different compartments: soma, proximal (less than 50 µm from the soma), apical and distal basal dendrite. Inhibitory cells were considered a fifth compartment. Synapses were then assigned a compartment based on where they were located on the postsynaptic target. The inhibitory subclass of a cell was then defined by where most of its synapses were located: perisomatic targeting cells that form synapses with the soma or proximal dendrites of postsynaptic cells, distal dendrite targeting cells synapse onto distal and/or apical dendrites, sparsely targeting cells do not form many multi-synaptic connections and inhibitory targeting cells predominantly synapse onto inhibitory cells.

Identifying synapses and postsynaptic targets

Synapses and their pre- and postsynaptic meshes in the EM dataset were previously algorithmically detected65. These data also included the automatically detected synapse size (number of voxels per synapse)66. In brief, a convolutional network was trained to identify synaptic clefts and assign voxels to each side of the cleft. Inference on the MICrONS volume was processed using the methods described in Wu et al.65 using 8 × 8 × 40 nm3 images. Subsequently, a second convolutional network was trained to perform pre- and postsynaptic partner assignment for the voxels for each detected synaptic cleft66. At the dataset level, manual inspection of the algorithmically identified synapses across subvolumes provides an estimated precision of 97% (correctly identifying a synapse) and recall of 89% (identifying all synapses)23.

To confirm that these synapse detection rates apply for the cells in this study specifically, we manually inspected 5% of output synapses per cell (3,180 synapses across 11 of 16 MCs) to confirm the presence of true synapses and to determine the postsynaptic target cell identity. We found that overall 4% of synapses were false positives, which aligns with the estimate for the entire volume. We did observe some differences across individual cells (5.6 ± 4.3% s.d. false positives per cell). To determine a cell’s average connection size, we calculated the sum of synapse sizes onto each individual postsynaptic target and then divided by the number of postsynaptic targets.

Cell subclass identities were assigned to all meshes with single somas (individual cells) in the EM dataset using a support vector machine classifier trained on somatic and nuclear features60. This classifier was then applied across the EM dataset to generate predicted cell type identities for most cells. We use these assigned types in all plots shown above. There was broad agreement between the automated and manual cell typing except for a specific disagreement of L2/3 versus L4 identity for targets of the predicted MET-8 cell type due to differences in layer boundaries used by manual versus automated methods (Extended Data Fig. 5b). Labels shown in the figures are from automated assignment.

Most synapses from individual cells were onto postsynaptic targets that contained single somas in the reconstruction (Extended Data Table 1). Of the synapses that were not onto single-soma targets, most were ‘orphans’, referring to objects that cannot be connected to any somas, either because the soma of the cell is outside the volume of the dataset or because the dendrite or dendritic compartment was not connected to a soma.

Quantification of myelination

Using automatically detected synapses, annotators visualized all output synapses on a given presynaptic cell in Neuroglancer. Regions lacking synapses were manually inspected in the EM imagery. If myelination was seen, an annotator marked the start and end point of each myelinated segment in Neuroglancer to generate a line. The number of these annotations was summed to determine the number of myelinated segments per cell. The length of each annotation was summed to determine the distance of myelinated axon per cell.

Apical synapse location analysis

The details of defining an apical dendrite for excitatory neurons in the EM dataset are described in ref. 33. In brief, the apical dendrite of skeletonized excitatory neurons was identified algorithmically by an RFC trained on expert annotation data. The RFC predicts whether a skeletal vertex is apical from features of the vertex such as dendritic length (for example, distance of vertex from soma along the skeleton) and complexity (for example, number of branch points). The apical label is then associated with synapses near that vertex. For our analysis, we queried the output synapses of each MET-4 and MET-6 MC. We restricted our analysis to the subset of synapses that were made onto L5 excitatory targets (algorithmically defined as L5 ET or L5 IT by Elabbady et al.60) that had identified apical dendrites. We calculated the percentage of output synapses that each MET-4 and MET-6 cell formed onto the apical dendrites versus non-apical locations.

MET-type proportions derived from MERFISH data

Proportions of MET-types were estimated from a recently published MERFISH dataset67. Cell counts were calculated for each Sst t-type in VISp in the MERFISH dataset. As the cells in the MERFISH dataset were mapped to a whole-brain taxonomy67, whereas the MET-types were based on a VISp-specific taxonomy6, we identified correspondences between t-types across several taxonomies. Correspondences between the original taxonomy of Tasic et al.6 and the cortex or hippocampal formation (CTX/HPF) taxonomy were identified by finding the Tasic et al.6 t-types that had the highest number of shared cells for each CTX/HPF t-type68. Correspondences between the CTX/HPF and whole-brain taxonomy were taken directly from ref. 67. We assigned Tasic et al.6 t-types to the MET-types to which the largest number of cells belonged16, except for the t-type Sst Calb2 Pdlim5, for which corresponding cells were assigned to either Sst MET-3 (if located in L4 or above) or Sst MET-4 (if located in L2/3). Note that cells from the Sst Chodl subclass (Sst MET-1) were not analysed here. In addition, no clear t-type correspondences were identified for the MET-type Sst MET-11, which also contained a small number of cells in the original study16.

Gene expression analysis

We analysed differentially expressed genes using scrattch.hicat (https://github.com/AllenInstitute/scrattch.hicat/tree/master) on the reference fluorescence-activated cell sorting-collected single-cell RNA sequencing dataset6 to compare the main transcriptomic types that make up each MET-type (MET-8: Sst Hpse Cbln4; MET-4: Sst Calb2 Necab1and Sst Calb2 Pdlim5; MET-6: Sst Chrna2 Glra3, Sst Chrna2 Ptdg4 and Sst Myh8 Etv1). Pairwise differentially expressed genes were identified as previously described6 using the limma package69 and selecting genes with at least a twofold change in expression and an adjusted P value of less than 0.01. The top five upregulated and downregulated genes for each pairwise comparison were selected for visualization (ranked by adjusted P value). The average expression of these genes and the fraction of cells with non-zero expression was calculated for the neurons of the three Sst MET-types in the Patch-seq data16 and presented as a dot plot. Only genes expressed in at least 50% of cells in at least one MET-type were selected for visualization.

Statistics

Comparisons across several MET-types were performed using non-parametric Kruskal–Wallis tests followed by Conover post hoc tests with Bonferroni corrections for pairwise comparisons. P values are indicated on plots if the Kruskal–Wallis P values < 0.05 and post hoc tests for pairwise comparisons are P < 0.05 or P < 0.01. Errors reported are s.e.m. unless otherwise indicated. Comparisons of the fraction of cell types targeted by a MET-type (from 0 to 300 µm) (Fig. 3a) were performed using a non-parametric Kolmogorov–Smirnov test with P values adjusted by a false discovery rate (Benjamini–Hochberg) correction. Boxplot whiskers indicate the range (maximum/minimum) of data; n = 173 inhibitory neurons from the EM dataset; n = 16 extended MCs and, of those, n = 3 were predicted to be MET-8, n = 6 were predicted to be MET-4 and n = 5 were predicted to be MET-6. Measurements are taken repeatedly from these cells.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.