Morpho-electric and transcriptomic divergence of the layer 1 interneuron repertoire in human versus mouse neocortex

Abstract

Neocortical layer 1 (L1) is a site of convergence between pyramidal neuron dendrites and feedback axons where local inhibitory signaling can profoundly shape cortical processing. Evolutionary expansion of human neocortex is marked by distinctive pyramidal neurons with extensive L1 branching, but whether L1 interneurons are similarly diverse is underexplored. Using patch-seq recordings from human neurosurgical tissue, we identified four transcriptomic subclasses with mouse L1 homologues, along with unique subtypes and types unmatched in mouse L1. Subclass and subtype comparisons showed stronger transcriptomic differences in human and were correlated with strong morpho-electric variability along dimensions distinct from mouse L1 variability. Accompanied by greater layer thickness and other cytoarchitecture changes, these findings suggest L1 has diverged in evolution, reflecting demands of regulating the expanded human neocortical circuit.

One-Sentence Summary:

Linking cellular transcriptomic identity to intrinsic morpho-electric features, we describe innovations in human neocortical layer 1 interneurons.

Main Text:

Neocortical layer 1 (L1) is implicated in several higher order brain functions, including state modulation (1), learning (2–5), sensory perception (6), and consciousness (7). The neural circuitry that mediates these functions consists of converging pyramidal cell dendrites, long-range axons originating from thalamic, cortical and neuromodulatory regions and axons from local GABAergic interneurons (8). Although some of this inhibitory input arises from other layers (Martinotti cells for example), much of it arises from neurons with cell bodies in L1, an entirely GABAergic cell population with distinct developmental origins (9, 10). Emerging evidence suggests that these L1 interneurons profoundly shape cortical processing and that diversity within this population is linked to diversity of function (11, 12). As such, the L1 interneuron repertoire is a potential site of evolutionary divergence that could contribute to specialized cortical function in humans and other primates. In rodents, a progression of classification schemes for L1 neurons (13–18) has evolved towards a view of 4 canonical types based on molecular markers (19), but the robustness of this scheme, both across modalities and across species, remains unclear (particularly in human and non-human primates). Indeed, the observation of a ‘rosehip’ cell type found in human and not mouse neocortex (20) highlights the importance of studying human L1 to identify potential species specializations and to relate mouse literature to human L1 cell types and function.

Traditionally, L1 cell types have been defined by their morphology, sublaminar location, intrinsic membrane properties, and a handful of marker genes. However, applying distinctive features from rodent to define and study human cell types can be tenuous. Single-cell whole transcriptome data, on the other hand, can be leveraged to define cross-species cell type homologies (21, 22) and reveal genetic and phenotypic diversity obscured by the marker gene approach (23, 24), as observed in vivo in mouse L1 (11). The patch-seq technique (25, 26), combining patch-clamp electrophysiology, RNA sequencing, and morphological reconstruction from the same neuron, gives us unprecedented ability to reveal cell type diversity in human L1. We leveraged this multimodal data to provide a comprehensive view of cell-type distinctions previously proposed from a subset of modalities, make principled cross-species comparisons, and robustly identify distinct phenotypes found in human L1 across modalities.

Results

L1 patch-seq pipeline and transcriptomic references

To guide our analysis of L1 cell types, we used transcriptomic types (t-types) previously defined in reference datasets from human middle temporal gyrus (MTG) and mouse primary visual cortex (VISp), single-nucleus or snRNA-seq in human and single-cell or scRNA-seq in mouse (21, 27). With annotations from layer dissections as a guide, we identified 10 L1 t-types in human and 8 in mouse (Materials and Methods; Fig. S1A). In UMAP (Uniform Manifold Approximation and Projection) (28) projections of transcriptomic data (Fig. 1A), many human t-types formed separated clusters, with others clustered in groups of related t-types, whereas mouse L1 t-types showed more continuous variability. This contrast suggests stronger transcriptomic specialization in human L1, similar to supragranular excitatory neurons (29), and indicates that more robust groupings of L1 types into highly distinct transcriptomic subclasses can be delineated in human. We grouped related human t-types into L1-focused transcriptomic subclasses by quantifying the pairwise distinctness of t-types in terms of a d’ separation of likelihoods (23, 24). This formed three subclasses, with three t-types remaining ungrouped (Fig. 1B). Expression of the inhibitory subclass markers PAX6 and LAMP5 (27, 30) and the t-type marker MC4R also closely matched these subclass boundaries (Fig. 1C).

Fig. 1: Single-nucleus RNA-seq demonstrates L1 diversity and provides a reference for patch-seq transcriptomic mapping.

(A) UMAP projections of human (left) and mouse (right) gene expression for L1 t-types (single-neuron or -nucleus RNA-seq). (B) Human t-types grouped by thresholding transcriptomic distinctness d’, defining subclasses. Remaining ungrouped t-types are marked as ‘L1 VIP’ or ‘other’ based on cross-species homology results. (C) Expression of canonical and t-type-specific marker genes across L1 t-types in human (left) and mouse (right). Pink background: human subclass markers, grey: classical mouse markers. Vertical lines group t-types by subclass. Violins show normalized probability density of gene expression (shape width) and median expression (dots), with expression in log(CPM+1) normalized by gene for each species, and maximal expression in counts per million (CPM) noted at right). (D) Mouse t-types grouped with human t-types into homologous subclasses (outlined) by thresholding similarity scores (heatmap intensity, from cluster overlap in integrated transcriptomic space). Non-L1 t-types are excluded, with maximal similarity over all non-L1 types shown for reference. (E) Proportions of subclasses and unclassified t-types in L1 patch-seq data, by species. Other L1 t-types refers to t-types in human L1 with no mouse homologue in L1. Deeper t-types refers to types found in L1 in lower proportions, not meeting the criteria for core L1 t-types. All cross-species proportion differences (except L1 VIP) significant at FDR-corrected p<0.001, one vs. rest Fisher’s exact tests.

Human subclass marker genes did not clearly identify subclasses in mouse, posing a challenge for cross-species comparison. Marker genes were either not expressed in any mouse L1 type (MC4R) or were expressed broadly and overlapped with other markers (LAMP5) (Fig. 1C top). Similarly, markers previously suggested for L1 subclasses in mouse (19) showed graded or complete lack of expression in human L1 (Fig. 1C bottom). Of note, the marker Id2, suggested for distinguishing a class of interneurons including all L1 types (Pvalb−/Sst−/Vip−) (31), was consistently expressed across most L1 types in both species (with the exception of marginal expression in ungrouped human t-types).

Given the lack of conserved subclass markers across species, we instead grouped mouse t-types for cross-species analysis by using cluster distances in an integrated transcriptomic space (21) (Fig. S1C), identifying each mouse t-type with the most similar human subclass or ungrouped t-type (Fig. 1D, Fig. S1B). These matches resulted in four homology-driven subclasses (called subclasses hereafter) with proportions largely comparable across species (Fig. 1E; PAX6 is the notable exception), named by the subclass marker genes in human. Two additional human L1 t-types (SST BAGE2 and VIP PCDH20) were excluded from cross-species L1 subclasses because of their homology to t-types found in deeper layers, but not L1 of mouse neocortex. This observation suggests that there is a shift in some of the interneuron diversity across laminar boundaries between mouse and human (32). Reinforcing the validity of these subclass divisions in mouse, we noted likely matches to previous mouse L1 subclasses (19) based on marker gene expression (Fig. 1C, Table S1): neurogliaform cells (Npy+/Ndnf+) to LAMP5, canopy cells (Npy−/Ndnf+) to MC4R, and α7 cells (Ndnf−/Vip−/Chrna7+) to PAX6 (11). However, uncertainty in these matches highlights the need for further confirmation based on morpho-electric properties.

To characterize morpho-electric and transcriptomic diversity across human L1 cell types, we used a previously established pipeline for high-throughput data acquisition and analysis (26, 29) to generate and release a comprehensive L1 patch-seq dataset. Human tissue was obtained from surgical samples and processed with standardized protocols; most samples originated from the MTG, along with smaller fractions in other temporal and frontal areas (Data S1). All cells were filtered for transcriptomic (n=250) and electrophysiological quality (n=194), and a subset of neurons (n=71) with sufficient cell labeling were imaged at high resolution and their dendritic and axonal morphologies were reconstructed.

We assigned transcriptomic cell types and subclasses to patch-seq samples using a “tree mapping” classifier, a decision tree based on the transcriptomic taxonomy structure (Materials and Methods) (24). Validating these assignments, we visualized t-type labels from patch-seq and reference datasets in a joint UMAP projection using alignment methods from the Seurat package (33) and found strong correspondence (Fig. S1E). Additionally, marker genes used by the classifier showed strong correlation by t-type between patch-seq data and the snRNA-seq reference (Fig. S1D).

Since patch-seq sampling was not uniform across cortical layers, we also measured the laminar distribution of L1 t-types using a spatially resolved, robust and reliable single-cell profiling technique (multiplexed error-robust fluorescence in situ hybridization or MERFISH (32); Materials and Methods). MERFISH laminar distributions were compatible with those from layer dissections of snRNA-seq, confirming that human L1 t-types are predominantly found in L1 or on the L1/L2 border, and demonstrating t-type-specific distributions across deeper layers and within L1 for certain types (Fig. S2B–D). Proportions of t-types within L1 were also generally matched between patch-seq, snRNA-seq, and MERFISH (Fig. S2A), with one exception: patch-seq had fewer SST BAGE2 cells and more PAX6 CDH12 and VIP TSPAN12 compared to snRNA-seq (p<0.05, False discovery rate or FDR-corrected Fisher’s exact test). MERFISH had intermediate proportions of SST BAGE2 with no significant differences compared to patch-seq (p>0.18), perhaps indicating differences caused by technical factors in snRNA-seq only such as imprecise layer dissections.

Morpho-electric diversity in human L1

Organizing the patch-seq dataset by transcriptomic subclass revealed the exceptionally diverse morphology and physiology of human L1 interneurons. Morphologically, subclasses were distinguished by vertical orientation of axons and dendrites, axon extent and shape, and dendrite branching (Figs 2A, D, S4; Data S2). Electrophysiologically, subclasses were distinguished by subthreshold properties such as sag (steady-state hyperpolarization reduced from transient peak) as well as several suprathreshold properties including firing rate, single action potential kinetics and adaptation of spike kinetics during trains of action potentials (Figs 2B, C; Data S2). Spike adaptation properties showed a strong inverse relationship with sag across the dataset (Fig. S3A). Sag is often mediated by HCN channels (34) and spike broadening by specific K+ channels (35, 36), so this finding may indicate a functional relationship between these channels in all subclasses of human L1 neurons.

Fig. 2: Human L1 transcriptomic subclasses are morpho-electrically diverse.

(A) Example human morphologies for L1 t-types are displayed by subclass. Neurons are shown aligned to an average cortical template, with histograms to the right of the morphologies displaying average dendrite (darker color) and axon (lighter color) branch length by cortical depth for all reconstructed cells in L1 and L2 (shading shows +/− 1 SD about mean, soma locations represented by black circles). (B) Electrophysiology summary view by t-type and subclass. Top shows example spiking response (scalebar 20 mV, 0.5 s). Cell-by-cell summary traces shown below, with black t-type average, dashed dataset average, individual cells in color. Top to bottom: phase plane (dV/dt vs V) plot of first action potential; instantaneous firing rate (IFR) normalized to peak; hyperpolarizing response normalized to peak. Spiking plots (example, phase-plane, IFR) at 40pA above rheobase, hyperpolarizing response at membrane potential closest to −100 mV (scalebar 0.5 s). Counts in Table S2. (C-D) Electrophysiological and morphological features distinguishing L1 subclasses (KW test, FDR-corrected p<10⁻⁷ for electrophysiology, <10⁻³ for morphology; Data S2). Boxplots show subclass statistics (box marks quartiles, whiskers extend 1.5xIQR past box), with individual cells arranged horizontally by t-type. Significant pairwise comparisons marked by lines above (FDR-corrected p<0.05, Dunn’s post-hoc test). Illustrative electrophysiology traces (scalebar 20 mV, 0.5 s) or layer-aligned morphologies shown for high and low values of each feature. Image inset shows that sparse dendrites in human PAX6 cells are not due to inability to resolve dendrites.

LAMP5 cells, the largest subclass, corresponded to the classical neurogliaform cell type (37), with highly branched, descending dendrites and horizontally elongated axons, either with a rectangular or triangular shape (Fig. 2A). Their electrophysiological phenotype was relatively undistinguished, with firing-rate adaptation and sag present but small (Fig. 2B). PAX6 cells had similar axons to LAMP5 cells, occasionally with descending branches, and sparse downward dendrites, along with minimal sag and high initial firing rate at the onset of response to depolarizing current injection (Fig. 2A–B). MC4R cells had extremely compact ball-like axonal arbors, along with strong sag; on this basis, they were tentatively identified as a match to the recently discovered ‘rosehip’ type (further characterized below) (20). L1 VIP (TSPAN12 t-type) cells had descending axon collaterals (13) with a consistent stellate-like dendrite morphology and high sag. The two cell types with no matching t-types within mouse L1, SST BAGE2 and VIP PCDH20, showed extremely diverse dendritic and axonal structure, often with substantial horizontal or descending axon branches – even avoiding L1 entirely in the case of some BAGE2 cells. These t-types appeared more uniform electrophysiologically, with relatively small spikes, high adaptation and sag, but were sparsely sampled and thus difficult to fully characterize. In a few instances, we also observed differentiation between t-types within the same subclass. Within the LAMP5 subclass, sag and initial firing rate decreased from the NMBR t-type to DBP to LCP2 t-types (Fig. 2C: strip plots within boxplot; Spearman correlation FDR-corrected p<10⁻³). Whereas other LAMP5 t-types were mostly restricted to L1 and superficial L2, the LCP2 t-type was found distributed across all cortical layers, with axonal arbors becoming less elongated and overlapping less with dendritic arbors for deeper cells (Fig. S2B,D; S3B). PAX6 cells were distinguished by whether the initial high-frequency firing formed a discrete burst (TNFAIP8L3) or continuously adapted (CDH12), and the two MC4R t-types were distinguished by the magnitude of sag and irregularity of firing (Fig. 2B).

Given the potential for the observed neuronal diversity to be determined in part by diversity in tissue donor characteristics, we tested all morpho-electric features for effects of donor medical condition, sex, and age (Fig. S6A,B; Data S3). Most effects were small and in features not linked to L1 diversity, with the exception of higher dendritic branching in cells from tumor patients compared to epilepsy patients (Fig. S6A; this result was not explained by brain area or subclass).

Cross-species differences in L1

Evolutionary expansion of L2/3 in primates was previously linked to changes in cytoarchitecture, including thinning out of cell density and increased soma size, accompanying the specialization of pyramidal cell types (29). Therefore, we first quantified cytoarchitecture differences in L1 of human tissue samples (NeuN-stained slices from patch-seq tissue blocks) compared to mouse samples (Fig. 3A). In contrast to cross species differences in L2/3, human L1 was thicker but not less dense, with smaller cell bodies compared with several mouse neocortical areas (Fig. S6D). To study cross species differences in the morpho-electric properties of L1 neurons, we compiled a comparison patch-seq dataset from mouse L1 neurons (n=272, 255 with electrophysiology and 43 with morphology) consisting of previously published data from a cross-layer analysis of interneurons in VISp (24) and additional recordings in L1 and L2/3 of visual cortex and the temporal association area (TEa; held out as a validation set since this region is often compared to human MTG (38, 39)). Despite the differences in L1 cytoarchitecture, morphologies of L1 neurons generally showed remarkable similarity across species when comparing across matched homology-driven subclasses, with two exceptions (Fig. 3C, S4, S5; Data S4). No mouse L1 neurons had morphologies resembling the unmatched human L1 t-types (VIP PCDH20 and SST BAGE12, homologous to deeper mouse t-types), and the MC4R subclass was less morphologically distinct, providing evidence for type-specific divergence across species (Fig. S5). Although human neurons across subclasses were slightly larger in horizontal extent, no differences were observed in vertical dimensions or dendritic diameter (Fig. S3C). Mouse VIP cells had descending axon branches as in human VIP cells, but with greater variability of structure. Mouse LAMP5 cells had dense neurogliaform-like axonal arbors, confirming their match via Ndnf/Npy expression. Unlike human axons, mouse axons rarely extended above dendrites (Fig. 3D left), perhaps reflecting sublaminar structure found only in the thicker human L1. Human neurites were also structured differently, with smaller contraction ratios (higher tortuosity) compared to the straighter but more heavily branched mouse dendrites (Fig. 3D; Data S4).

Fig. 3: Comparison of human and mouse L1.

(A) Examples of NeuN labelling of neurons in human MTG and mouse VISp. (B) Comparisons of mouse versus human L1 thickness, neuron density, soma area and neuron count in 1mm wide ROIs of L1. Metrics plotted per ROI for L1 thickness, density and neuron count, and per cell for soma area. Boxplots show quartiles, stars indicate post-hoc Dunn’s test results at p<[0.05, 0.01, 0.001] (calculated for MTG vs TEa only). Counts in Materials and Methods. (C) Example layer-aligned morphologies from mouse and human L1 subclasses. One example shown from each t-type, scalebar for both species. (D) Morphological (left) and electrophysiological (right) features with differences between human and mouse L1 cells (species effect from 2-way ANOVA on ranks, FDR-corrected p<10⁻⁷ for morphology, p<10⁻³⁸ for electrophysiology). For features with a species-subclass interaction (p<0.05 FDR-corrected for electrophysiology, not significant for morphology), stars indicate post-hoc Dunn’s test results at p<[0.05, 0.01, 0.001]. Counts in Table S2. Representative examples from LAMP5 subclass shown below each plot (L to R: layer-aligned reconstructions; AP frequency as a function of current injection; response to hyperpolarizing current nearest −100 mV, scalebar 0.5s/10mV; first action potential at rheobase, scalebar 1ms/20mV). (E) Electrophysiological feature differences between human L1 and mouse VISp L1 (left) were validated by testing against mouse TEa (right). Features selected by largest effect size against TEa (MW r, rank-biserial correlation). Stars indicate significance (FDR-corrected MW test, p<[0.05, 0.01, 0.001]). (F) Nucleated patch recordings quantified A-type K+ conductance of L1 neurons. Example traces show voltage commands (black) and recorded currents (orange) from measurement protocol (top), along with example soma size measurement. Boxplots show fast conductance density in both species, with example traces shown for each group (scalebars 400pA/200ms). Only LAMP5 neurons were sampled in mouse. (G) Features distinguishing L1 subclasses in human and mouse organized by relevance to each species. Bars show size of subclass effect (ε² from KW test), with features ranked by the difference between human and mouse effects. Unfilled bars indicate p>0.05 (FDR-corrected).

As in previous studies, electrophysiological properties showed stronger differences across species (40). Mouse cells had less sag, lower amplitude spikes, and higher rheobase across types and within at least 2 of 4 matched subclasses (Fig. 3D; Data S4). We replicated these findings in a comparison between MTG and a smaller TEa dataset, verifying that cross species differences were not due to regional differences between MTG and VISp (Fig. 3E). Proportions of L1 t-types also varied little across brain regions in mouse and human single neuron/cell RNAseq reference datasets (Fig. S2B) (41). We also explored dependence of L1 interneuron morpho-electric properties on brain region within our human data and found moderate effects on a set of features including dendrite extent and input resistance (Fig. S6C). For these features, more sparsely sampled regions either failed to differ from MTG cells or had smaller dendritic extent and larger input resistance, diverging more strongly from mouse features, suggesting again that the cross-species differences were not inflated by regional sampling.

We next investigated causal factors underlying cross-species electrophysiology differences. Simulations of passive biophysical models based on reconstructed morphologies showed that input resistance variability in human but not mouse L1 cells can be explained by morphology, suggesting the lower input resistance in mouse may be partly due to active ionic conductances (Fig. S3G). The small cross-species effect of morphology on input resistance in the models could be explained by differences in dendritic branching, especially near the soma, which would affect the effective membrane area for leak conductance. Indeed, we found a higher peak of total dendrite cross-sectional area at ~50 μm from the soma as well as slightly higher total volume in mouse cells (Fig. S3F) supporting this explanation. We also looked for correlated differences in ion channel gene expression and morphological features compared against membrane properties in the patch-seq dataset. Differences in spike shape and threshold could be explained by potassium channel differences, along with related features like rheobase and delayed spiking. Indeed, the expression of genes (KCND2, KCND3, and KCNH7) associated with fast inactivating, A-type K+ channels (Kv4.2, Kv4.3 and the ERG3 channel Kv11.3 (42, 43)) was higher in mouse neurons and was correlated with several action potential features (Fig. S3E). To test for corresponding differences in K+ channel conductance, we measured macroscopic currents in nucleated patches following whole-cell recording in a subset of cells. Compared with human neurons, mouse neurons showed much higher A-type K+ conductance but comparable slow inactivating (D-type) conductance (Fig. 3F). Considering blocking Kv4 channels in mouse neurogliaform cells decreases AP threshold and latency of first AP onset (44), these differences in A-type K+ channel conductance, along with lack of Kv1.1 expression, may contribute to the lack of late spiking observed in human L1 neurogliaform cells as well (45).

Finally, we asked whether the strong morpho-electric variability observed between human L1 subclasses is also present in L1 of mouse neocortex. Ranking electrophysiological and morphological features by the amount of variability between subclasses they explain, we found that the two species had a similar amount of variability (number of significantly different features and their effect size) but varied along different sets of features (Fig. 3G). The most distinct features in human, like sag and spike shape adaptation, showed little variability in mouse, and unlike in human, mouse subclasses varied physiologically in ISI adaptation and spike after-hyperpolarization (AHP) properties, and morphologically in relative vertical positioning of the axonal arbor (most features largely driven by L1 VIP subclass: Fig. S3, S5).

Distinctive neuronal phenotypes in human L1

Despite the quantitative similarity in L1 heterogeneity across species, we noted two particularly distinctive phenotypes found in human L1 only. The MC4R rosehip cells and the bursting PAX6 TNFAIP8L3 t-type were both qualitatively distinct from other human L1 types and did not appear similar to any mouse L1 types. To further highlight this contrast, we investigated each of these highly distinctive types in turn by quantifying the distinctive morpho-electric features and marker genes, then searching for comparable cells in the mouse L1 dataset.

Rosehip cells

The MC4R subclass, putative rosehip cells, comprises two transcriptomically similar t-types, SST CHRNA4 and ADARB2 MC4R, both highly distinct from other L1 types including the LAMP5 LCP2 t-type originally identified with the rosehip phenotype (20). MC4R morphologies were all confirmed to qualitatively match the distinctive rosehip axonal structure and boutons (Fig. 4A, S4), and were quantitatively distinct from other L1 types in terms of maximum axonal path distance and branch frequency (Fig. 4B). We also noted two examples of MC4R cells (both within the ADARB2 MC4R t-type) with elaborate descending axons reaching the lower half of L3, and confirmed that the characteristic large, dense axonal boutons were visible on both the central axonal arbor and descending axons when present (Fig. 4A right). Electrophysiologically, both t-types that comprise the MC4R subclass showed strong sag, but only the ADARB2 MC4R t-type showed the distinctive irregular firing (45) and stronger and faster sag (Fig. 2B, 3A–B). Cells in the ADARB2 MC4R t-type also had somas and axons localized near the L1/L2 border (Fig. S2C–D). We explored the expression of genes related to neuron physiology (ion channels and G-protein-coupled receptors or GPCRs) and found markers distinguishing the entire rosehip subclass from the rest of L1, including HTR1F (5HT receptor 1F), along with markers distinguishing the rosehip subtypes: GRM5 (metabotropic glutamate receptor 5) and RELN (Reelin) showed lower expression in ADARB2 MC4R neurons (Fig. 4C). Together, these differences in gene expression and physiology indicate that there are distinct rosehip neuron subtypes within human L1.

Fig. 4: MC4R rosehip cells.

(A) Characterization of MC4R subtypes as rosehip cells. Left: UMAP projection of transcriptomic data from MC4R and nearby subclasses. Right: example cells from each subtype. Morphologies show characteristic axonal arbors and boutons (insets: 63x MIP images, scalebars 10 μm; compare to panel D). Electrophysiology traces show sag response (hyperpolarization near −100 mV, rheobase, and rheobase +40 pA if present; scalebar 0.5s/10mV). (B) Electrophysiological and morphological features distinguishing MC4R t-types (FDR-corrected MW test vs. rest of L1, p<10⁻⁴). Boxplots show statistics of MC4R subtypes and other subclasses (box marks quartiles, whiskers extend 1.5xIQR past box). Significant pairwise comparisons (to MC4R t-types only) marked by lines above (FDR-corrected p<0.05, Dunn’s test post-hoc to KW test). (C) Gene expression of MC4R subclass (highlighted) and other L1 t-types, for between- and within-subclass marker genes (snRNA-seq). Violins show expression in log(CPM+1), normalized by gene (maximal expression noted at right). (D) Characterization of mouse L1 cells with moderate sag or irregular firing, the human rosehip type’s distinct properties (boxplots as in (B), all pairwise comparisons tested). Example morphology and electrophysiology shown for mouse Lamp5 Ntn1 Npy2r cell with highly irregular firing, but lack of rosehip-like morphology (Electrophysiology traces as in (A); image inset shows axonal boutons, scalebar 10 μm).

In mouse L1, we identified cell types corresponding to the MC4R subclass based on similarity in transcriptomes, but this match was weak compared with other homologous types (Fig. 1D). Similarly, there were no cell types observed with the morphological signatures consistent with human rosehip cells (Fig. S5) and only partial matches to the electrophysiological signatures: the homologous mouse MC4R subclass had moderate sag but no irregular spiking (Fig. 4D). Irregular spiking resembling the human ADARB2 MC4R rosehip t-type was present only in a subset of LAMP5 cells (primarily Lamp5 Ntn1 Npy2r) that did not have other rosehip-like features (Fig. 4D). Although not directly matching the rosehip phenotype, the mouse homology-driven MC4R subclass carried similarities with the canopy cell (19). Matched characteristics included the moderate sag, along with gene expression (Fig. 1C, Ndnf+/Npy−) and wide dendritic extent, but mouse MC4R cells did not have the canopy’s namesake L1a-dominant axon (Fig. S7B–C). The mouse MC4R subclass thus may be a unique neurogliaform-like cell population, but certainly lacks distinct boundaries that can be clearly resolved either by cross-species comparison or reference to previous mouse L1 classifications.

Bursting PAX6 TNFAIP8L3 cells

The other highly distinctive firing pattern we noted in human L1 was in the PAX6 TNFAIP8L3 t-type, which fired in high-frequency bursts at the onset of stimulation, followed by quiescence or regular firing at higher stimulus amplitudes. Spiking and dendritic structure were qualitatively distinct between this t-type and the neighboring PAX6 CDH12 t-type, despite some similarity of axonal structure and subthreshold electrophysiology (Fig. 5A). Both the initial firing rate at rheobase and the after-depolarization potential (ADP) following the final spike quantitatively distinguished PAX6 TNFAIP8L3 cells from all other L1 cells (Fig. 5B), as did the number of dendritic branches and large horizontal dendritic extent, over 550 microns wide.

Fig. 5: Burst spiking PAX6 TNFAIP8L3 cells.

(A) Reconstructed morphologies and example electrophysiology for bursting and non-bursting PAX6 t-types (TNFAIP8L3 and CDH12) (hyperpolarization near −100 mV, depolarization below rheobase, spiking at rheobase and rheobase +40 pA; scalebar 0.5s/10mV). Inset shows UMAP projection of transcriptomic data from PAX6 subclass. (B) Electrophysiological and morphological features distinguishing PAX6 TNFAIP8L3 t-type (FDR-corrected MW test vs. rest of L1, p<0.05 for morphology, <0.01 for electrophysiology). Boxplots show statistics of PAX6 subtypes and other subclasses (box marks quartiles, whiskers extend 1.5xIQR past box). Significant pairwise comparisons (to PAX6 TNFAIP8L3 only) marked by lines above (FDR-corrected p<0.05, Dunn’s test post-hoc to KW test). (C) Gene expression of human PAX6 subclass (highlighted) and other L1 t-types, for α7 type and bursting-related marker genes (snRNA-seq). Violins show expression in log(CPM+1), normalized by gene (maximal expression noted at right). (D) Characterization of mouse L1 cells with initial doublet firing in terms of the human PAX6 TNFAIP8L3 type’s distinct properties. Example morphology and electrophysiology shown from PAX6 subclass (Lamp5 Krt73 t-type). Depolarizing sag ratio is the normalized size of the hump at stimulus onset just below rheobase. (Electrophysiology traces as in (A); boxplots as in (B), all pairwise comparisons tested)

In mouse L1, the homologous PAX6 subclass was extremely rare, comprising only a few cells in the Lamp5 Krt73 t-type. These cells tended to fire in doublets at stimulus onset rather than a full burst, sometimes followed by a delayed ADP (Fig. 5C). Some cells in the Lamp5 Fam19a1 Pax6 t-type also showed this firing pattern (Fig. S7D), likely the same subset that align transcriptomically to the human PAX6 subclass (Fig. 1C). Mouse doublet-firing cells also showed a depolarizing ‘hump’ for current injection just below rheobase (Fig. 5D), which together with the marker gene signature (Ndnf−/Vip−/Chrna7+) identifies them as mouse α7 cells, a type previously defined in mouse by these physiological/gene features (19). This hump was suggested to indicate activation of T-type calcium channels, likely the same mechanism underlying the bursting in human cells (45). Bursting was also previously noted in a subset of mouse “Single Bouquet Cells” (SBC) (25), a group defined by loose morphological criteria. This class likely overlaps with the doublet-firing t-types (46), suggesting they may burst under different physiological conditions.

Using these insights from the cross-species alignment, we explored the expression of related genes in the human PAX6 t-types (Fig. 5C). Both t-types matched the α7 marker gene signature (along with the SST BAGE2 t-type; Ndnf−/Vip−/Chrna7+), and strongly expressed the T-type calcium channel alpha subunit gene CACNA1G, highlighting T-type calcium channels as a potential factor in the burst and doublet firing across species (Fig. S7A). Given the lack of bursting in PAX6 CDH12 cells, other ion channel genes differentially expressed between the two human PAX6 t-types likely also play a role, including TRPC3, a non-specific cation channel that can regulate resting membrane potential (47).

Cross-modality relationships of L1 subclasses and t-types

Given the multiple observations of distinctness between human types in contrast with continuous variation between mouse types, we explored this difference more comprehensively by defining a common quantitative framework for distinctness across modalities. We generalized the d’ metric used for transcriptomic distinctness (23, 24), quantifying the performance of classifiers trained to distinguish pairs of t-types based on electrophysiological and morphological features. The resulting t-type similarity matrices (Fig. 6A) showed comparable subclass structure in both electrophysiology and transcriptomics, with smaller d’ values within subclass blocks and higher values outside. Of note, d’ metrics were highly correlated between modalities, demonstrating that cell types with distinctive transcriptomes have similarly distinctive electrophysiological properties (Pearson r=0.59, p=0.00016; Fig. 6B). The single within-subclass pair with a high d’ was LAMP5 NMBR and LAMP5 LCP2, which sit at opposite ends of the LAMP5 continuum. We also calculated d’ similarity matrices at the subclass level to allow comparison between species in all three modalities (Fig. 6C). These results confirmed the generally higher transcriptomic distinctness of subclasses in human and show that in mouse the VIP subclass was highly distinct in all modalities, with other subclasses generally less distinct.

Fig. 6: Quantifying distinctness of L1 t-types and cross-modality structure.

(A) Pairwise distinctness of human L1 t-types, from classifiers using electrophysiology (left) and gene expression (right). d’ (d-prime) is a metric of separation of distribution means, scaled relative to the standard deviation. Groups with N<4 excluded (hatched area). (B) Correlation of pairwise d’ values between transcriptomic and electrophysiological feature spaces. Pearson r=0.59, p=0.00016, shading shows bootstrapped 95% CI of regression. Within-subclass pairs shown in orange to confirm subclass structure. (C) Pairwise distinctness d’ of L1 subclasses across species and data modality. Groups with N<10 excluded. (D) Clustering of human L1 cells in electrophysiology subspaces, with correspondence to L1 subclasses. Points show all L1 neurons, with the subclass of interest in color. Background color shows cluster membership likelihoods from 2-cluster Gaussian mixture model trained on unlabeled data. F1 scores: LAMP5 0.81, MC4R 0.69, PAX6 0.89, all others 0.5 (L1 VIP and ungrouped t-types). All features normalized and Yeo-Johnson transformed to approximate Gaussian distribution.

To visualize the subclass-level distinctness in terms of specific electrophysiological features, we found the pair of features that most distinguished each subclass and showed that clusters defined by these features closely match the transcriptomic subclass boundaries (Fig. 6D). We also tested the effectiveness of assigning subclass labels to neurons based on the full electrophysiological feature set. A multi-class classifier evaluated by cross-validation on the primary dataset had 82% accuracy balanced across subclasses (Fig. S8A). To mimic the out-of-sample issues that could be encountered for future L1 datasets collected under different conditions, we also tested classifier performance on data held out of our primary analysis due to equipment and protocol differences. After excluding features for which the distributions strongly differed from the primary dataset, we found comparable classification performance (81%, Fig. S8B), reinforcing the utility of the human L1 subclasses for understanding L1 variability even in the absence of transcriptomic information to assign subclass identity.

Discussion

Using patch-seq, we identified a coherent view of human L1 interneurons in which neuronal subclasses defined by transcriptomic distinctness have similarly distinct morpho-electric phenotypes. Although mouse L1 neurons had a similar range of diversity in most features, the features that distinguished cell types were different between species and human L1 neurons spanned a wider range of sizes. In addition to cell types not found in mouse L1 (VIP PCDH20 and SST BAGE2), two human cell types emerged with especially distinct phenotypes that were not matched in their putative homologues in mouse: the compact, high-sag MC4R ‘rosehip’ subclass and the large, burst-spiking PAX6 TNFAIPL83 t-type. Human and mouse neurons also showed consistent differences in certain morphological and physiological properties across all subclasses, despite a general similarity in cell size. These results indicate a general conservation of L1 inhibitory neuron diversity, but with distinct specializations in cell properties and subclass/cell-type proportions, likely leading to differences in the regulation of higher order input to the human cortical circuit.

Categorizing L1 neuron types

We provide support for previous classification schemes in mouse consisting of four primary types in L1 but also clarify a need for precise data-driven criteria for those types. Homology-driven subclasses were nearly aligned with cell type classifications based on single marker genes (19), but this alignment was often ambiguous and lacked coherence across modalities. Similarly, other coarse single-modality cell-type distinctions in L1 (late spiking vs non late spiking, NGFC vs SBC) likely grouped multiple distinct subclasses and shifted the exact boundaries (48). Our results suggest that a larger role for continuous variability should be considered when studying cell type diversity in mouse L1. A continuous transition between Ndnf+/Npy+ neurogliaform cells and Ndnf+/Npy− canopy cells was previously noted (24), corresponding to similarity between the Lamp5 Ntn1 Npy2r t-type (LAMP subclass) and Lamp5 Fam19a1 Tmem182 (MC4R subclass). We also observed continuity between mouse PAX6 (partial α7 cell match) and MC4R (partial canopy cell match) types, with α7-like doublet spiking in some MC4R cells. Additionally complicating the view of canopy cells, the best-match cells (MC4R subclass) had some properties at odds with the original definition, expressing Chrna7 and missing the canopy-like L1a axons which were observed primarily in Npy+ LAMP5 cells (Fig. S7A–C). These ambiguous subclass boundaries are a strong contrast to the clear cross-modality subclass distinctions in human L1, perhaps related to the smaller number of well-resolved transcriptomic types in mouse.

We also demonstrate the benefits of detailed transcriptomic data over small sets of genetic markers for both accurately characterizing cell type divisions and establishing cross-species homologies that facilitate comparative analysis. Within species, reliance on marker genes can overstate the distinctness of cell types, as with the NGFC/canopy distinction, or even lead to misidentification, as with the original description of human L1 rosehip cells which were assigned an incorrect t-type based on observed marker gene patterns (20). Across species, the lack of conserved L1 markers was striking and perhaps unique to layer 1. Even using the full transcriptome, previous work on this homology found variable results for L1 types with different brain areas and methods (21). In this work, we chose to quantify similarity of t-types across species in a way that better captured ambiguity, finding strong matches for some types and weaker for others. In general, all transcriptomic homology matches were supported or contradicted by morpho-electric comparisons. Ambiguous matches may indicate areas of evolutionary change which could be illuminated by comparative or developmental analyses of additional species phylogenetically related to mouse and human. For instance, the weak cross-species transcriptomic similarity of the MC4R subclass (Fig. 1C) and lack of phenotypic similarity together suggest that human and mouse MC4R cells could represent distinct innovations in each taxon, rather than a true homology.

In human L1, the strong alignment of subclass distinctions across modalities suggests that cells can be classified using only morphology or electrophysiology with reasonable accuracy. Condition-dependent variation of certain electrophysiological features can present a challenge to this approach though, especially for classifications relying on small numbers of features with especially strong qualitative variability. We failed to observe two electrophysiological phenotypes that had been noted in past work: late spiking in human NGFCs (40), and full bursting in mouse SBCs (25), qualitative features for which conflicting observations have also been reported (19, 49). Although potential contributing factors are numerous (age differences, donor characteristics, recording conditions including internal solutions, temperature and equipment differences), we showed that for classification with a large feature set, identifying and excluding affected features can rescue reliable performance.

Cell types, evolution, and function

We propose that the divergence in the L1 interneuron repertoire between mouse and human partly reflects the increasingly complex role of L2/3 in the primate neocortical circuit (29, 50, 51); new types of pyramidal cells might have necessitated new types of dendritic inhibition. Compared with mouse, human L2/3 excitatory neurons are more transcriptomically distinct and show larger sublaminar distinctions in gene expression, dendritic morphology and physiology. Like L1, there are also neuronal types in L2/3 of human MTG with no clear homologue in L2/3 of mouse neocortex (29). Although these observations suggest similar degrees of evolutionary divergence in L2/3 excitatory neurons and L1 inhibitory interneurons, the L1 circuit might also have adapted to changes in deep layer pyramidal neuron populations, including the decreased proportions in primates of L5 extratelencephalic pyramidal neurons, the prominent apical dendrites of which are targeted by L1 inhibition in rodents (21, 22, 52).

Two types of interneurons stood out as especially distinct within human L1, with the potential to contribute to functional differences between human and mouse – the MC4R rosehip cells and the PAX6 TNFAIP8L3 t-type. Rosehip cells were previously shown to inhibit pyramidal cell apical dendrite shafts in L2/3 (20); the rosehip subtypes, with distinct electrophysiology could plausibly perform similar but distinct inhibitory functions or selectively modulate different pyramidal neuron subtypes in L2/3 (29). For example, the irregular firing of the MC4R rosehip t-type suggests that this cell type is modulated by input in the beta frequency band (20) whereas the regular firing of the CHRNA4 rosehip t-type suggest a lack of beta-band modulation. The strong bursting dynamics and distinctive morphology of the PAX6 TNFAIP8L3 t-type also clearly point to a unique functional role compared to neurons in neighboring subclasses or even the more closely related PAX6 CDH12 t-type. Their extended dendrites are well positioned to integrate local pyramidal cell inputs across a broad spatial footprint and long-range axonal inputs across topographic boundaries, and the bursting would provide a strong immediate activation in response to strong or coincidental input. The clear identification of a cross-species homology for the PAX6 subclass can help in deciphering its function, combining functional insights from manipulating mouse cells with indirect insights from the more distinctive morphologies of human cells. Conflicting connectivity patterns have been observed for coarser cell types that likely include some PAX6 cells: mouse α7 cells synapse onto nearby L2 pyramidal neurons (8), whereas rat SBCs synapse onto L2/3 interneurons (53). More focused investigation of PAX6 cell connectivity is thus needed to illuminate the function of this subclass, and in turn the functional implications of specialization within this subclass in human L1.

In addition to these unique types in human, we observed several subclass-independent cross-species differences in morpho-electric properties between species which likely have functional consequences. Forms of morpho-electric variability found in human and not mouse L1 include sag and spike adaptation properties. These vary both between and within subclasses in human and may contribute to differences in the spectral selectivity of the L1 circuit. For example, the higher voltage sag response in human L1 (and variation between LAMP5 subtypes) may contribute to differences in the temporal summation of synaptic input, as in L2/3 pyramidal neurons (54, 55). The separation of dendritic and axonal arbors in human L1 cells, not found in mouse, indicates the possibility of more sublaminar structure in L1 microcircuits in human. Similarly, given the sublaminar selectivity of thalamic projections to Vip+ versus Ndnf+ cells in mouse L1 (56), the human L1 VIP types which are located in deeper layers in mouse may be especially relevant to thalamocortical microcircuit structure. The higher input resistance and lower rheobase in human imply increased sensitivity to low-intensity synaptic input, which could have direct computational functions or help compensate for circuit differences, such as changes in excitatory to inhibitory cell ratios between mouse and human (21, 22, 32). In particular, these cross-species morpho-electric differences may indicate varying demands on the neurogliaform subclass, which was otherwise largely conserved across species. Neurogliaform GABAergic (both synaptic and extra-synaptic) and gap junction-mediated transmission depend critically on the spatial properties of the axonal arbor (37, 57, 58), but the increase in arbor size in human LAMP5 cells (~1.2x) was much less than the 1.6x increase in pyramidal cell apical dendrite extent in L1. Changes in input sensitivity could enable the conservation of their ‘blanket’ inhibitory function while also permitting some increased spatial/topographic selectivity – neurogliaform circuit connectivity has been shown to be both tightly controlled (59, 60) and to exert strong effects on pyramidal cell sensory processing (1, 6).

Much experimental evidence has documented the importance of neuromodulatory control on L1 function (40, 61–63). Linking distinct subclasses in L1 to detailed gene expression data (including emerging spatial transcriptomics results) can lead to refined hypotheses for cell-type specific neuromodulation. In particular, human MC4R cells uniquely expressed several modulatory receptor genes (Fig. S9A), including the melanocortin receptor MC4R, which plays a role in energy homeostasis in hypothalamus (64), serotonin receptor HTR1F and metabotropic glutamate receptor GRM1 (along with differential expression of GRM5 between MC4R t-types). Cholinergic activation of L1 cells (61), suggested to control attention, may also differentially modulate MC4R cells relative to other L1 types based on their stronger CHRNA6 and CHRNA4 expression.

Our findings have a few limitations related to studying human cell types using the patch-seq approach. Certain cell types were under-sampled, including those with no mouse homologue in L1 (VIP PCDH20 and SST BAGE2), limiting our ability to fully characterize their properties. Future work utilizing viral genetic labeling approaches to label difficult to target types could help in this regard (65). Our cross-species comparisons were limited to differences between mouse and human and thus could not elucidate evolutionary progression of the L1 interneuron repertoire or truly reveal human specializations. Additional investigations of L1 cell types in other species would help, whether based on single-cell transcriptomics alone or full characterization via patch-seq. Even in mouse, our analysis focused on comparisons to human and thus didn’t fully resolve different views of mouse L1 subclasses. A definitive answer may require a more comprehensive patch-seq dataset, analyzed from a cross-species perspective and perhaps in concert with multi-region snRNA-seq and spatial transcriptomics (66). Finally, directly relating any differences in intrinsic physiology, morphology, or gene expression to functional consequences requires further investigation, including characterizing synaptic connectivity and neuromodulation.

Nonetheless, the comprehensive human L1 patch-seq dataset reported here has generated a wealth of hypotheses to inspire future work, along with tools to support that work. Our analysis provides tools for the classification of L1 diversity in both human and mouse, insights into functional relationships underlying physiological differences, and the clear identification of subclasses and subtypes that are likely to be of particular interest in functional studies, all important steps for deciphering the function of this enigmatic layer of neocortex. These data may also provide insight into various disease states, considering the importance of inhibition in circuit dysfunction (67, 68). This approach represents a roadmap for annotating functionally related properties onto transcriptomically-defined cell type taxonomies that could be applied across the primate brain – a crucial step towards linking cell type diversity to functional diversity within a neural circuit.

Data and materials availability:

The primary datasets used in this study are hosted at the archives DANDI (electrophysiology: dandiarchive.org), BIL (imaging and reconstructions: brainimagelibrary.org) and NeMO (transcriptomics: nemoarchive.org, RRID:SCR_016152). These datasets are cataloged for easy access at the BICCN Data Inventory (RRID:SCR_022815; brain-map.org): knowledge.brain-map.org/data/97XR43ZTYJ2CQED8YA6/collections for human data, and knowledge.brain-map.org/data/1HEYEW7GMUKWIQW37BO/collections for previously published mouse visual cortex data. A full list of the datasets used in this study is provided below. Note that human sequence-level data at NeMO is access controlled (see https://nemoarchive.org/resources/accessing-controlled-access-data). Processed cell-by-gene-level data, along with other intermediate datasets (extracted features) and all custom analysis and visualization code, are archived for reproducibility at Zenode (https://doi.org/10.5281/zenodo.8310226), and also available at github.com/AllenInstitute/patchseq_human_L1. Other software libraries essential to the analysis are publicly available at github.com/AllenInstitute/ipfx (electrophysiology), github.com/AllenInstitute/neuron_morphology and github.com/AllenInstitute/skeleton_keys (morphology), and github.com/AllenInstitute/scrattch/ (transcriptomics).