Spatio-molecular gene expression reflects dorsal anterior cingulate cortex structure and function in the human brain

SUMMARY

The human brain’s dorsal anterior cingulate cortex (dACC) is critical in cognitive control, specifically reward processing and conflict monitoring, and regulates fear and pain expression. The dACC is evolutionarily older than more recently specialized neocortical areas, such as the dorsolateral prefrontal cortex (dlPFC). Its evolutionary specializations, including agranularity and presence of von Economo neurons (VENs) are linked to its unique roles in cognitive and emotional processing. We generated paired spatially resolved transcriptomics (SRT) and single-nucleus RNA-sequencing (snRNA-seq) data in ten adult neurotypical donors to define cell types and spatial domains. We used non-negative matrix factorization to project gene expression patterns from the snRNA-seq data onto the SRT data to infer spatial localization. Utilizing publicly available resources, we explored spatial topography, enrichment of disease risk, and connectivity of spatially localized cell types, including VENs. Leveraging dlPFC snRNA-seq and SRT data from the same donors, we compared laminar and molecular specialization between the regions.

Graphical Abstract

In brief

Shah et al. spatially characterize the agranular human dACC through integration of paired spatial transcriptomics and snRNA-seq data from ten neurotypical control donors, specifically focused on molecular profiles, spatial topography, enrichment of disease risk, and putative connectivity of spatially localized dACC cell types.

INTRODUCTION

The dorsal anterior cingulate cortex (dACC) plays key roles in various components of cognitive control and is particularly relevant for processes associated with response selection, conflict monitoring, and error detection.^1-4 The dACC’s role in these cognitive processes is especially pronounced in the context of reward processing. Neural firing patterns in the dACC reflect reward prediction error, which is related to its proposed function as a reward monitor,^5,6 and reward seeking is associated with changes in neural activity of dACC neurons.⁷ The dACC has a notable role in sustained attention and is linked to disorders that feature attentional deficits.^4,8-10 dACC neural activity is associated with expression of pain, anxiety, and fear and is implicated in the etiology of neuropsychiatric disorders related to these symptoms.^11-16

The dACC is a subdivision of the larger ACC, which lies in the medial wall on each cerebral hemisphere, located above and adjacent to the corpus callosum. The ACC is highly connected to both the prefrontal cortex and to subcortical regions within the limbic system, and the topography of connections to these areas differs across subdivisions of the ACC.^17,18 The ACC is typically subdivided into the more ventral aspect that contains the subgenual ACC (sgACC) and pregenual ACC (pgACC) versus the more dorsal aspect, the dACC, which is also referred to as midcingulate cortex (MCC) in several nomenclature systems.^2,17,18 Anatomically, the dACC is located dorsal to the genu of the corpus callosum and encompasses parts of Brodmann areas (BAs) 33, 24, and 32^2,17,18 (Figure S1).

Compared to more recently specialized neocortical areas, including the dorsolateral prefrontal cortex (dlPFC), the ACC is an evolutionarily older structure whose most prominent cytoarchitectural feature is the lack of a definable layer 4 (L4).^19,20 The region’s agranular structure and other notable specializations, such as the unique presence of von Economo neurons (VENs), are hypothesized to contribute to its specialized roles in cognitive and emotional processing. VENs are a type of spindle neuron present in great apes and humans, but their presence is more controversial in other animals.^21-23 VENs differ from canonical cortical pyramidal neurons in their size and shape and are thought to play an adaptive function in mediating fast communication across long distances.²⁴ Postmortem human brain studies in other cortical regions have suggested that genetic risk for various functions and neuropsychiatric disorders may manifest with cell type and laminar specificity,^25-29 but whether and how these findings are similar or differ in the dACC is not known.

Here, we generated spatially resolved transcriptomic (SRT) maps with paired single-nucleus RNA-sequencing (snRNA-seq) data to provide a valuable data resource of the human dACC. We highlight how this data resource can be used to better understand biological features unique to this region, specifically for anterior dACC (adACC/aMCC; BA24/33) of adult neurotypical donors, and to address questions about anatomical conservation with more recently evolved cortical regions that can provide insight into the biology underlying differences in function and connectivity. dACC snRNA-seq and SRT data were generated in ten adult neurotypical donors to define molecular profiles for cell types and spatial domains. Using non-negative matrix factorization (NMF), we integrated these data by defining gene expression patterns within the snRNA-seq data and inferred their expression in the SRT data. Combining these data with publicly available resources, we reveal insights about molecular profiles, spatial topography, enrichment of disease risk, and putative connectivity of spatially localized dACC cell types, including VENs. Utilizing dlPFC snRNA-seq and SRT data previously collected in the same ten neurotypical brain donors used to generate the dACC data,²⁸ we deployed cross-regional comparison of laminar and molecular specialization between dACC and dlPFC. To make this valuable molecular resource accessible to the scientific community, both raw and processed data are freely available, including through interactive web applications.

RESULTS

Spatio-molecular profiling of the human dACC

While the molecular organization of the human dlPFC has been well characterized using single-cell and spatial transcriptomics,^27,28,30,31 other cortical regions, including the evolutionarily older ACC, are less defined. The dlPFC is classified as a neocortex, containing six prominent cortical layers with a well-defined granular L4. In contrast, the ACC is an evolutionarily older type of allocortex, classified as dysgranular or agranular cortex, which is characterized histologically by a poorly defined or non-existent granular L4 (Figure 1A).^19,20 The ACC is divided into three major components—the sgACC, pgACC, and dACC. The dACC (also referred to as the caudal ACC [cACC] or midcingulate [MCC] in other nomenclature systems) can be further divided into an anterior aspect (e.g., adACC, acACC, aMCC) and a posterior aspect (e.g., pdACC, pcACC, pMCC)^2,17,18 (Figure S1). We dissected tissue blocks from fresh-frozen coronal hemislabs of postmortem human brain from ten neurotypical control donors (6 male/4 female) targeted at the anterior aspect of the dACC (Figure S1). Inclusion of dACC on the tissue blocks was confirmed by histological staining (H&E) and RNAScope using established laminar marker genes (Figures 1B; Figures S2-S5). To generate SRT data, tissue blocks were vertically scored in ~6.5 mm strips to match the width of capture areas on the Visium platform (10× Genomics). Four independent donors were processed per Visium slide, with a total of 17 total capture areas. For a subset of donors, we generated technical replicates from directly adjacent sections (2–3 additional capture areas per donor). After standard preprocessing and quality control workflows to remove low-quality Visium spots (STAR Methods, Figure S6), 73,367 spots were retained. Following tissue collection for SRT assays, 2–4 100 μm cryosections from each tissue block were collected (Figure 1B) for snRNA-seq on the Chromium platform (10× Genomics). Nuclei were isolated from these cryosections for each donor and sorted on propidium iodide (PI) to sample all cell types (STAR Methods). After standard preprocessing and quality control to remove empty droplets, doublets, and poor-quality nuclei, 35,161 high-quality nuclei across all 10 donors were retained (STAR Methods). To facilitate comparative molecular neuroanatomy studies between the evolutionarily specialized, granular dlPFC and the evolutionarily ancient, agranular dACC, we leveraged existing dlPFC data from the same donors previously used to characterize the spatial and cellular landscape of the dlPFC (Figures 1A; Table S1).²⁸

Figure 1. Experimental design to generate paired single-nucleus RNA-sequencing (snRNA-seq) and spatially resolved transcriptomics (SRT) data in the human dorsal anterior cingulate cortex (dACC).

(A) dACC (blue) and dlPFC (pink) regions outlined in midsagittal (dACC) and lateral (dlPFC) brain views, as well as in coronal hemislabs.

dACC tissue (blue) used here was sourced from the same ten neurotypical control donors previously profiled for dlPFC (pink)²⁸ to facilitate within-donor comparisons of dACC agranular cortex versus dlPFC granular cortex (bottom).

(B) Fresh-frozen coronal brain slab from donor Br3942 taken at the level of the anterior striatum overlaid with the outline of major landmarks from the Atlas of the Human Brain.³² Three Brodmann areas (BA) 33, 24, and 32 corresponding to the dACC are highlighted in yellow (top left).

H&E staining of tissue cryosections confirmed inclusion of dACC on the tissue block (top right). Following anatomical validation, cryosections from 10 neurotypical control donors were collected for Visium and Chromium assays (10× Genomics) from the same block for each donor (bottom).

Molecular signatures of spatial gene expression identify discrete spatial domains and reveal agranular laminar patterning in the human dACC

We used log2-transformed normalized gene counts as input to PRECAST, a spatial cluster detection algorithm, to find data-driven clusters in the dACC SRT data.³³ As input to the PRECAST algorithm, we identified a set of spatially variable genes (SVGs) using nnSVG,³⁴ as described in STAR Methods. These SVGs are listed in Table S2. We generated clusters at a range of k=5 to k=20 using PRECAST guided by the set of SVGs and used both average cluster purity and H+ as statistical metrics coupled with predicted laminar organization to choose a final set of clusters for downstream analyses (STAR Methods, Figures S7 and S8). Ultimately, we selected the clusters generated with k = 9 (Figure S9). Independently of the PRECAST-derived data-driven clusters, we used visualization of histological features and expression of established layer marker genes to generate a set of histologically driven domains, which we annotated as L1, L2/3, L5, L6, white matter (WM), and the corpus callosum (CC) (Figures 2A and 2B; Figure S10). Based on visual comparison to the histology-driven domains and expression of canonical marker genes, we noted that data-driven clusters 7, 3, and 8 mapped to WM or CC in the histology-driven domains. Data-driven cluster 7 corresponded with very low library size (Figure S9) and was hence removed from downstream analysis. Data-driven cluster 8 contained a small number of spots that were speckled within data-driven cluster 3; this larger cluster matched the known cytoarchitecture and localization of WM below L6. Furthermore, data-driven clusters 3 and 8 did not differentiate between WM and CC, as seen in the manual annotations of the dACC SRT data (Figure S10). Thus, we retained data-driven clusters 3 and 8 but collapsed them into a single WM spatial cluster, resulting in a final set of seven data-driven spatial domains (Figure 2C). Each of these spatial domains contained a comparable number of spots from each of the ten donors, with no individual donor dominating the composition of any spatial domain (Figure S16).

Figure 2. Signatures of gene expression in data-driven spatial domains provide molecular evidence for agranular organization of the dACC.

(A) Histological image of Visium capture area from donor Br8667 (sample ID: V12Y31-080_B1).

(B) Spot plot of Visium capture area from donor Br8667 (sample ID: V12Y31-080_B1) with spots colored by histology-driven domains, which were annotated as layer 1 (L1), L2/3, L5, L6, white matter (WM), and corpus callosum (CC).

(C) Spot plot of Visium capture area from donor Br8667 (sample ID: V12Y31-080_B1) with spots colored by the 7 spatial domains annotated from the k=9 PRECAST clusters, which were annotated as L1, L2, L3, L5, L6a, L6b, and WM.

(D) spatialLIBD spatial registration heatmap displays Pearson’s correlation values between the top 100 marker genes in each data-driven spatial domain (y axis) and in each histology-driven manually annotated spatial domain (x axis).

(E) Boxplots of ADCYAP1 log2(counts per million + 1) expression (computed manually) (y axis) for each spatial domain (x axis) in the pseudobulked dACC SRT data. Color represents the spatial domain. escheR spot plot of dACC Visium capture area from donor Br6432 (sample ID: V12N28-331_B1) with spots colored by the dACC spatial domains. Fill represents ADCYAP1 log2-normalized expression per spot.

(F) Boxplots of ADCYAP1 log2(counts per million + 1) expression (computed manually) (y axis) for each spatial domain (x axis) in the pseudobulked dlPFC SRT data.

Color represents the spatial domain. Spot plot of dlPFC Visium capture area from donor Br6432 (sample ID: Br6432_ant) with spots colored by the dlPFC spatial domains. Fill represents ADCYAP1 log2-normalized expression per spot.

(G) Same as (E), but for RORB expression.

(H) Same as (F), but for RORB expression.

(I) Same as (E), but for PCP4 expression.

(J) Same as (F), but for PCP4 expression.

(K) Multiplex RNAScope single molecule fluorescence in situ hybridization (smFISH) in dACC of donor Br8667 for ADCYAP1 (green), RORB (magenta), PCP4 (yellow), and merged all channels with DAPI (blue).

Spatial plot of phenotyped ADCYAP1 (green), RORB (purple), and PCP4 (yellow) cells across cortical laminae in dACC of donor Br8667.

Phenotyped cells are plotted using their x-y coordinates in per-cell object data. DAPI-positive cells in gray. Relative density distribution of each cell type (ADCYAP1+, RORB+, and PCP4+) across cortical laminae in dACC from n = 5 donors (Br8667, Br8492, Br6471, Br6432, and Br2720).

(L) Same as (K), but for dlPFC from the same donors.

To better understand the relationship between the seven data-driven spatial domains to canonical histological layers, we used a combination of approaches. We first pseudobulked spots within the data-driven and histology-driven domains, which collapse data from spot-level to spatial domain-level data within each sample. In contrast to sparse spot-level information, this information is similar to bulk RNA-sequencing for each domain, enabling the use of differential expression (DE) tools optimized for bulk data. The minimum number of spots required to be included as a pseudobulked sample was 10 spots. Prior to proceeding with DE analysis, we examined the samples in principal component (PC) space to ensure that the data-driven domains were relatively well separated and demonstrated that any variation was not driven by total unique molecular index (UMI) counts per sample (Figure S11). For these analyses, we pseudobulked samples by aggregating total UMI counts across spatial domains and samples. Comparison of the pseudobulked data-driven domains to the histology-driven annotations revealed high correlations across layers and WM (Figure 2D).

Next, we capitalized on the availability of an SRT dataset of the dlPFC²⁸ generated from the same 10 brain donors to directly investigate the molecular organization of granular versus agranular cortex. Specifically, we visualized expression of established cortical layer markers derived in the dlPFC data, including RELN (L1), LAMP5 (L2), ADCYAP1 (L3), RORB (L4), PCP4 (L5), NR4A2 (L6), and MOBP (WM) across the data-driven domains in both the dACC and dlPFC (Figures 2E-2J; Figure S12). The data-driven domains in the dlPFC include domains for WM and the 6 layers typical of this granular structure (Figures 2F-2H, and 2J; Figure S12). We also performed multiplex single molecule fluorescence in situ hybridization (smFISH) for three of these markers, ADCYAP1, RORB, and PCP4, in dACC and dlPFC tissue sections from a subset of the donors (Figures 2K and 2L). Based on the comparison to histology-driven domains (Figure 2D) and enrichment of layer marker genes (Figures 2E-2J; Figure S12), we identified single dACC data-driven domains that distinctly map to L1, L2, L3, L5, and WM. We noted that two of the dACC data-driven domains showed high correlation to the histology-driven domain annotated as L6 (Figure 2D). These two domains expressed canonical markers of L6, including NR4A2 (Figures S12I-S12L). We therefore annotated these two dACC data-driven domains as L6a and L6b. In the dlPFC SRT data, expression of the canonical L4 marker RORB is high in both L4 and L5 (Figure 2H), while RORB maps tightly to L5 in the dACC SRT data (Figure 2). Confirming these results, quantification of smFISH reveals that RORB+ cells overlap with the L5 marker PCP4+ cells in the dlPFC, but also extended superficially into L4, which contains no PCP4 expression (Figure 2L, distribution separation p=0.009). However, in the dACC, RORB+ cell distribution is limited to the L5 area defined by PCP4+ cells (Figure 2K, p=0.41). In summary, we generated molecular profiles for dACC data-driven domains using spatial gene expression data at transcriptome scale, which we annotated as L1, L2, L3, L5, L6a, L6b, and WM. While the agranular structure of the ACC has been long appreciated from a cytoarchitecture perspective, these data provide complementary molecular evidence for the absence of L4 in dACC in the human brain.

NMF reveals cell type-specific gene expression patterns shared between snRNA-seq and SRT data in the dACC

To better characterize the molecular identity of the data-driven spatial domains in the dACC, we integrated the SRT data with paired snRNA-seq data, which was generated from adjacent tissue sections in each donor. Following quality control of the snRNA-seq data (Figures S13, S14, and S15), we used the Azimuth Human Motor Cortex reference to predict and transfer cell-type annotations to our dataset.^35,36 Each cell type contained a comparable proportion of nuclei from each of the ten donors with no individual donor dominating the composition of any cell type (Figure S16). To integrate the snRNA-seq data across donors, we selected the top 2,000 features based on Poisson deviance³⁷ (Figure S17) followed by Harmony batch correction³⁸ on the generalized linear model principal components analysis (GLM-PCA) reduced dimensions. We found comparable numbers of nuclei in each cell type with good separation across cell types (Figure 3A). Similar to the SRT data, we used pseudobulked DE analysis to find marker genes for each cell type compared to all other cell types. The minimum number of nuclei required to be included as a pseudobulked sample was 10 nuclei. The cell types were well separated in principal components (PC) space and this variation was not driven by total UMI counts per sample (Figure S17). We generated volcano plots for each cell type (Figures S18, and S19) and extracted the top 30 markers (Table S3). To annotate cell types from the snRNA-seq data with the spatial domains identified using the SRT data, we used the robust cell type decomposition (RCTD) framework to deconvolve the spot-level SRT data.³⁹ Layer-specific cell types, such as L2/3 intratelencephalon (IT)-projecting neurons, L5 IT-projecting neurons, L5 extratelencephalon (ET)-projecting neurons, L6 corticothalamic (CT)-projecting neurons, and L6b, were enriched for their respective spatial domains, while astrocytes, Sncg+, Vip+, and Lamp5+ GABAergic neurons were enriched in L1 and oligodendrocytes in WM (Figure 3C). Several layer-enriched marker genes in the SRT data were also cell-type marker genes in the snRNA-seq data, including ADCYAP1, RORB, and PCP4 (Figures 2E-2J). Strong expression of these genes within their respective layer-specific cell types further confirms laminar identity assignment for the dACC data-driven spatial domains (Figure 3B). Comparing the dlPFC and dACC snRNA-seq data, for the dlPFC, expression of RORB is highest in the L4 cell type but also expressed in the L5 cell type, while in the dACC, RORB expression is localized more specifically to the L5 cell type (Figure S20). These data are concordant with the SRT data (Figures 2G and 2H) and provide additional molecular evidence at the cellular level of agranularity in the dACC.

Figure 3. snRNA-seq cell type annotation and non-negative matrix factorization (NMF) reveals distinct cell types in the human dACC.

(A) Uniform manifold approximation and projection (UMAP) representation of the single-nucleus RNA-sequencing (snRNA-seq) dataset. Each point is an individual nucleus colored by its Azimuth reference cell type. Astro, astrocyte; Endo, endothelial cell; L2_3_IT, layer 2–3 glutamatergic neuron, intratelencephalon-projecting; L5_ET, layer 5 glutamatergic neuron, extratelencephalon-projecting; L5 IT: Layer 5 glutamatergic neuron, intratelencephalon-projecting; L5_6_NP: Layer 5–6 glutamatergic neuron, near-projecting; L6_CT, layer 6 glutamatergic neuron, corticothalamic-projecting; L6_IT, layer 6 glutamatergic neuron, intratelencephalon-projecting; L6_IT_Car3, layer 6 Car3+ glutamatergic neuron, intratelencephalon-projecting; L6b, layer 6 b glutamatergic neuron; Lamp5, Lamp5+ GABAergic neuron; MicroPVM: microglia/perivascular macrophage; Oligo, oligodendrocyte; OPC, oligodendrocyte precursor cell; Pvalb: Pvalb+ GABAergic neuron; Sncg, Sncg+ GABAergic neuron; Sst, Sst+ GABAergic neuron; Vip, Vip+ GABAergic neuron; VLMC, vascular leptomeningeal cell.

(B) Heatmap showing two selected marker gene log2-normalized expression averaged by cell type and scaled per gene (columns) across the layer-specific cell types (rows): L2/3 IT, L5 ET, L5 IT, L5/6 NP, L6 CT, L6 IT, L6 IT Car3, and L6b. Grouping and color across the top indicates the cell types.

(C) RCTD spatial deconvolution heatmap displays the cell type weight for each snRNA-seq cell type averaged by each SRT data-driven spatial domain. The x axis displays the snRNA-seq cell types described in (A). The y axis displays the data-driven domains: L1, L2, L3, L5, L6a, L6b, and white matter (WM).

(D) Dot plot showing the top NMF pattern (x axis) selected for each snRNA-seq cell type (y axis). Dot size represents the proportion of nuclei with nonzero weight for each NMF pattern across each cell type. Dot color represents the aggregated nuclei-level weights for each NMF pattern across each cell type.

(E) Heatmap showing 5 selected marker gene loadings scaled by total gene loadings (columns) across the layer-associated NMF patterns (rows): L2/3 IT, L5 ET, L5 IT, L5/6 NP, L6 CT, L6 IT, L6 IT Car3, and L6b. Grouping across the top indicates the NMF patterns and the color represents the cell type each pattern is associated with.

(F) escheR spot plots of dACC Visium capture area from donor Br8667 (sample ID: V12Y31-080_B1) with spots colored by the dACC spatial domains. Gray scale fill represents weights of distinct cell type-specific NMF gene expression patterns learned from the snRNA-seq dataset and projected into the SRT data using transfer learning.

The threshold for low weight is 0. The thresholds for high weight are 0.002, 0.001, 0.023, 0.012, 0.013 for NMF3, NMF38, NMF61, NMF15, and NMF35, respectively.

Next, we used NMF as a complementary approach to identify cell type-specific gene expression programs.⁴⁰ NMF reduces the dimensions of the gene expression data to k factors that represent cell types, cell patterns, and various biological processes.^41,42 Using cross-validation, we determined that the optimal rank to factorize the snRNA-seq gene expression matrix was 75 factors (Figure S21). After factorizing the gene expression matrix, we first annotated these 75 patterns based on association with specific cell types (STAR Methods, Figure S21; Table S4). We removed NMF patterns associated with technical variables by computing the correlation with each pattern’s weight and variables related to quality control metrics, brain donor, and sex of brain donor and removed them from further downstream analyses (Figure S22). The dot plot identifies a top NMF pattern for each cell type (the vascular leptomeningeal cell [VLMC] type did not have a specific NMF pattern) and demonstrates that each pattern is highly specific to the cell type to which it was annotated (Figure 3D). For the layer-specific cell types, we extracted the top marker genes for each pattern, highlighting genes whose function has been associated with these layers (Figure 3E). Finally, we projected these NMF patterns into the SRT data using a transfer learning approach in order to predict the spatial location of distinct neuronal subtypes (Figure 3F). We found that these cell type-specific NMF patterns correspond to the laminar organization of the dACC, with patterns enriched in specific neuronal cell types aligning to their expected cortical layer (e.g., L5 NMF patterns localizing to the L5). These findings provide a foundation for exploring biologically meaningful, layer-specific gene expression programs and mapping them into SRT data.

Localization of a VEN signature to a spatially restricted area of deep L5 in dACC

A unique feature of the dACC is the presence of VENs. VENs are a specialized type of bipolar spindle neuron that is localized to L5 but only in a restricted set of human brain regions including the pgACC, dACC, and frontoinsula. VENs are large in size, have a unique morphology, and are more abundant in the human brain compared to nonhuman primates, possibly representing an evolutionary specialization.^21-23 VENs are also associated with schizophrenia,⁴³ suicidal psychosis,⁴⁴ and autism spectrum disorder.⁴⁵ We again utilized the dlPFC snRNA-seq dataset generated in the same 10 donors as the dACC snRNA-seq dataset²⁸ to explore molecular patterns that map to region-specific cell types, including VENs. Similar to previous analyses, we computed pseudobulked DE genes by comparing each cell type to all other cell types, first visualizing the pseudobulked data along the first two principal components and then calculating the correlation between the top 100 markers for each dACC cell type and each dlPFC cell type (Figure S23). Every dACC cell type was strongly correlated with at least one dlPFC cell type, except for L5 ET and L6 IT Car3, a subpopulation of L6 neurons marked by expression of CAR3⁴⁶ (Figure S23).

We further explored the L5 ET cell type, hypothesizing that it may contain VENs, which reside in L5 of dACC, but not in dlPFC (Figure 4A). We identified both NMF38 and NMF61 as patterns from the dACC snRNA-seq data associated with L5; however, we found that NMF38 was highly related to L5 IT, while NMF61 was highly related to L5 ET (Figure 3D). Further, we showed that several VEN markers: VAT1L, ADRA1A, GABRQ, and POU3F1, were associated with NMF61 (the gene program highly related to L5 ET) (Figure 3E). We extended these findings to the SRT data by using transfer learning to project the dACC snRNA-seq-derived NMF38 and NMF61 patterns into both the dACC SRT data, as well as the dlPFC SRT data, generated from the same paired ten donors²⁸ (Figures 4B and 4C). The output that we obtained is spot-level weights for these two patterns in each dataset. First, in the dACC SRT data, we determined that the spots associated with NMF38 and NMF61 were distinct for all samples (Figure S24). While both NMF patterns localized to L5, they had a distinct topography within the layer: NMF38 is localized to superficial L5 (closer to L3) while NMF61 is localized to deep L5 (closer to L6) (Figure 4D). Comparing directly across the dACC and dlPFC SRT datasets, we showed that NMF38 is expressed across L5 in both the dACC and the dlPFC, while NMF61 is highly expressed only in deep L5 of the dACC (Figure 4D). Quantifying these observations, we demonstrated that the fraction of nonzero L5 spots across donors is similar between the dACC and dlPFC for NMF38, but higher in the dACC for NMF61 (Figure 4E). Then, we calculated differentially expressed genes between the dACC L5 IT and L5 ET cell types in the snRNA-seq data, which revealed that several established VEN markers,^47,48 specifically HAPLN4, SULF2, FEZF2, and GABRQ, are more highly expressed in L5 ET compared to L5 IT (Figure 4F). We also calculated differentially expressed genes between the NMF61-positive and NMF38-positive spots in the dACC SRT data (STAR Methods), which similarly showed that established VEN markers,^47,48 specifically VAT1L, SULF2, FEZF2, and GABRQ, are more highly expressed in NMF61 compared to NMF38 (Figure 4F).

Figure 4. Identification of discrete molecular signatures for von Economo neurons (VENs) in Layer 5 of the dACC.

(A) Cytoarchitecture of dACC agranular cortex and dlPFC granular cortex at laminar resolution highlighting the relative location of VENs in Layer 5 ET. This cell type was captured by NMF61.

(B) UMAP plot from Figure 3A shows cell types from the dACC snRNA-seq data. The L5 NMF patterns derived from these data are projected into the SRT data shown in Figure 4C.

(C) First spot plot is the same as Figure 2C. The second spot plot shows Visium capture area from dlPFC donor Br8667 (sample ID: Br8667_mid) with spots colored by the 7 spatial domains annotated from the previous dlPFC study.²⁸

(D) TOP ROW: escheR spot plot of dACC Visium capture area from donor Br8667 (sample ID: V12Y31-080_B1) with spots colored by the dACC spatial domains. Spot plot of dlPFC Visium capture area from donor Br8667 (sample ID: Br8667_mid) with spots colored by the dlPFC spatial domains. Fill represents NMF38 projected into the dlPFC and dACC SRT datasets. The range for low to high weight is 0–0.001. BOTTOM ROW: Spot plot of dACC Visium capture area from donor Br8667 (sample ID: V12Y31-080_B1) with spots colored by the dACC spatial domains.

Spot plot of dlPFC Visium capture area from donor Br8667 (sample ID: Br8667_mid) with spots colored by the dlPFC spatial domains. Fill represents NMF61 projected into the dlPFC and dACC SRT datasets. The range for low to high weight is 0–0.023.

(E) Boxplots comparing NMF38 and NMF61 between dACC and dlPFC Layer 5 spatial domain. The top boxplots represent sample averages of the fraction of nonzero NMF38 spots in the dACC SRT Layer 5 and the dlPFC SRT Layer 5.

The standard deviations are 0.123 and 0.078. The bottom boxplots represent sample averages of the fraction of nonzero NMF61 spots in the dACC SRT Layer 5 and the dlPFC SRT Layer 5. The standard deviations are 0.054 and 0.005.

Statistical analysis was performed using an unpaired two-tailed t test, with significance defined as ns p > 0.05, *p ≤ 0.05, **p ≤ 0.01, ***p ≤ 0.001, and ****p ≤ 0.0001.

(F) Left: differential expression volcano plot compares pseudobulked snRNA-seq cell types L5 IT and L5 ET.

Right: Differential expression volcano plot compares pseudobulked spots with nonzero NMF38 and NMF61 projections in the dACC SRT data. Known VEN markers (VAT1L, SULF2, HAPLN4, FEZF2, and GABRQ) are highlighted in both volcano plots as upregulated in L5 ET or NMF61.

(G) Multiplex RNAScope single molecule fluorescence in situ hybridization (smFISH) in dACC (top box) and dlPFC (bottom box) for SULF2, POU3F1, GABRQ, and merged of all. Heatmaps: Top heatmap displays the average proportion of cells called as expressing one gene (row label) that are also called as expressing another gene (column label), averaged across the three dACC samples. Bottom heatmap is similar, except averaged across the two dlPFC samples. Boxplots: Top boxplots display the proportion of cells (y axis) from layer 5 that were called as expressing both POU3F1 and GABRQ (x axis). Bottom boxplots display Spearman’s correlation values (y axis) for POU3F1 and GABRQ expression (x axis).

Color indicates the region of the sample, either dACC or dlPFC. Shape indicates the brain donor of the sample, either Br6432 or Br8325. Scale bars, 30 μm.

(H) Dot plot of cortical layer-specific NMF patterns (columns) in the dACC human snRNA-seq dataset corresponding to the nuclei collection source (rows) for the mouse retroviral tracing dataset. Dot size represents the proportion of nuclei with nonzero weight for each NMF pattern across each cell type. Dot color represents the aggregated nuclei-level weights for each NMF pattern across each cell type.

(I) Dot plot of mouse retroviral tracing using projection of cortical layer-specific NMF patterns (columns). Rows represent axonal projection target regions from the retroviral tracing experiments. Cortical targets: AI, agranular insular cortex, AUDp, primary auditory cortex, ENT, entorhinal cortex, MOp, primary motor cortex, PFC, prefrontal cortex, PIR, piriform area, PTLp, posterior parietal cortex, RSP, retrosplenial cortex, SSp, primary somatosensory cortex, VISp, primary visual cortex. Subcortical targets: AMY, amygdala, P, pons, SC, superior colliculus, STR, striatum, TH, thalamus, VTA, ventral tegmental area.

Dot size represents the proportion of nuclei with nonzero weight for each NMF pattern across each cell type. Dot color represents the aggregated nuclei-level weights for each NMF pattern across each cell type.

As orthogonal validation, we used RNAScope smFISH to visualize 3 VEN markers, POU3F1, SULF2, and GABRQ, with the L5 marker gene PCP4 from 5 tissue sections across 2 brain donors. Two sections were from dlPFC (Br6432 and Br8325) and 3 sections were from dACC (one from Br6432 and two from Br8325). In each sample, we computationally focused on the L5 region and classified the number of cells expressing each gene (STAR Methods). The proportion of cells for each sample classified as expressing POU3F1, SULF2, and GABRQ is presented in Figure S25A. The dACC samples had higher coexpression of 2–3 of the VEN markers compared to the dlPFC samples, with the greatest difference observed in POU3F1 and SULF2 coexpression (Figures S25B and 4G). Additionally, dACC samples had higher Spearman’s correlation between each pair of VEN markers compared to the dlPFC samples (Figures S25). To further characterize coexpression, we calculated the average proportion of expressors for one gene that are also expressors for another gene, which shows higher co-expression for dACC samples compared to dlPFC samples, with greater differences for POU3F1 and SULF2 co-expression with the remaining two VEN markers (Figure 4G). The RNAScope images show a high degree of overlap of the 3 VEN markers in the dACC but not in the dlPFC (Figure 4G).

Although VENs are not reported as present in the rodent cortex, they do show molecular homology to an ET excitatory neuron type in the mouse cortex.⁴⁷ Aligning cell types across species based on gene expression profiles can facilitate predictions about functional properties and circuit connectivity of cell types in the human brain, where it is not possible to perform labeled cell projection studies. To extend our findings across species, we integrated our data with a mouse dataset that combined viral retrograde labeling with molecular profiling to define gene expression profiles for cells based on their axonal projection targets.⁴⁹ We first showed that NMF61 and NMF38 map to L5 ET and L5 IT, respectively, in the mouse cortex according to the Allen Brain subclassification (Figure 4H). We next showed that these two NMF patterns roughly map to different cell types based on their long-range projection targets: NMF61 maps to cell types projecting to subcortical structures, such as pons, superior colliculus, thalamus, and ventral tegmental area, whereas NMF38 maps to cell types projecting to other cortical structures, such as agranular insular cortex, primary auditory cortex, entorhinal cortex, posterior parietal cortex, primary somatosensory cortex, and primary visual cortex (Figure 4I). These projection targets are in line with the localization of NMF61 to the deep ET-projecting L5 neurons and NMF38 to the superficial IT-projecting L5 neurons (Figure 4D). Given that these two NMF patterns find highly similar molecular features between human and mouse, particularly for L5 ET and IT neurons, these data facilitate predictions about putative axonal targets of spatially localized L5 dACC neurons, including VENs.

Identification of unique spatio-molecular features in the human dACC

We first looked at molecular features across the dACC data-driven spatial domains to understand spatio-molecular features that define its laminar organization. To do this, we identified top marker genes for each of the dACC data-driven domains using the pseudobulked counts matrix generated in Section 2.1. We then used an enrichment model to identify genes that are differentially expressed between one data-driven spatial domain compared to all other spatial domains by fitting a linear mixed effects model on the pseudobulked counts matrix. We visualized volcano plots for each spatial domain (Figures S26A-S26G) and extracted the top 50 markers (Table S5). We identified several genes that were enriched in dACC spatial domains that were not previously established as laminar marker genes in the dlPFC including RSPO2 (L2), DRD5 (L5), and KCTD8, ADRA2A, and CRHBP (L6). Visualization of these genes across the dACC and dlPFC spatial domains showed that these genes are enriched in specific dACC layers but not in the corresponding dlPFC layers (Figure 5E; Figure S27). More specifically, for the L6 markers, KCTD8 localizes to L6a while ADRA2A and CRHBP localize to L6b (Figures 5E; Figure S27).

Figure 5. Within-donor differences in gene expression patterns across individual spatial domains of the dACC and dlPFC.

(A) Pseudobulked sample-level comparison: spatialLIBD spatial registration heatmap displays Pearson’s correlation values between the top 100 marker genes in each dlPFC SRT spatial domain and in each dACC SRT spatial domain. The x axis displays dlPFC spatial domains: L1, L2, L3, L4, L5, L6, and white matter (WM). The y axis displays the dACC spatial domains described in (Figure 2C). The black “X” represents high confidence.

(B) Pseudobulked gene-level comparison: the heatmap displays the number of DEGs between each spatial domain in dACC and dlPFC, scaled by the proportion of DEGs across each dACC spatial domain. The x axis displays dlPFC spatial domains: L1, L2, L3, L4, L 5, L6, and WM. The y axis displays the dACC spatial domains described in (Figure 2C).

(C) TOP ROW: Spot plot of dACC Visium capture area from donor Br8667 (sample ID: V12Y31-080_B1) with spots colored by the dACC spatial domains. Spot plot of dlPFC Visium capture area from donor Br8667 (sample ID: Br8667_mid) with spots colored by the dlPFC spatial domains. Fill represents NMF15 projected into the dlPFC and dACC SRT datasets.

The range for low to high weight is 0–0.012. BOTTOM ROW: Spot plot of dACC Visium capture area from donor Br8667 (sample ID: V12Y31-080_B1) with spots colored by the dACC spatial domains. Spot plot of dlPFC Visium capture area from donor Br8667 (sample ID: Br8667_mid) with spots colored by the dlPFC spatial domains. Fill represents NMF35 projected into the dlPFC and dACC SRT datasets. The range for low to high weight is 0–0.013.

(D) Each scatterplot shows a comparison of the log fold-change (logFC) between one spatial domain in dACC and dlPFC SRT spatial domains. Each point is a gene. The x axis of the first plot shows the logFC from enrichment model pseudobulked DE testing comparing L2 to all other spatial domains for dACC. The y axis of the first plot shows the logFC from enrichment model pseudobulked DE testing comparing L2 to all other spatial domains for dlPFC. The second plot compares L5 in the dACC to L5 in the dlPFC. The third plot compares L6a in the dACC to L6 in the dlPFC. The fourth plot compares L6b in the dACC to L6 in the dlPFC. The colors highlight genes that are classified as either dACC enriched (red), dlPFC enriched (blue), or both (purple), where the threshold is a logFC greater than 1.25. The number of genes in each category is as follows: L2 dACC vs. L2 dlPFC: 117 dACC enriched and 24 dlPFC and dACC enriched; L5 dACC vs. L5 dlPFC: 80 dACC enriched, 9 dlPFC enriched, 18 dlPFC and dACC enriched; L6a dACC vs. L6 dlPFC: 42 dACC enriched, 2 dlPFC enriched, 19 dlPFC and dACC enriched; L6b dACC vs. Lb dlPFC: 191 dACC enriched, 11 dlPFC enriched, 10 dlPFC and dACC enriched.

(E) Boxplots comparing gene expression of DRD5, KCTD8, and CPLX3 across pseudobulked spatial domains from dACC SRT data and dlPFC SRT data. The y axis displays log2(counts per million + 1) expression (computed manually) for each spatial domain (x axis), for dACC (top row) and dlPFC (bottom row). Color represents spatial domains for each dataset.

(F) Multiplex RNAScope single molecule fluorescence in situ hybridization (smFISH) in dACC (top row) and dlPFC (bottom row) from Br6471 for DRD5 (magenta), KCTD8 (yellow), CPLX3 (green), and merged of all channels with DAPI (blue). Approximate cortical layer boundaries are indicated by dashed lines. Scale bars, 200 μm.

Next, we deployed two separate approaches to directly compare spatial and cell type-specific gene expression across dACC and dlPFC. We refer to the first method of comparison as the pseudobulked sample-level comparison. For the pseudobulked sample-level comparison, we computed the correlation between the top 100 marker genes using an enrichment model across all pairs of spatial domains between dACC and dlPFC (Figure 5A). This analysis confirmed strong correlation between L1, L2, L3, L5, and WM domains between dACC and dlPFC. Assessing correlation to the dlPFC L4 spatial domain, we noted that it was moderately associated with dACC L3 and L5. dlPFC L6 was more highly correlated with dACC L6a compared to L6b. We refer to the second method of comparison as the pseudobulked gene-level comparison, which was designed to directly leverage the fact that we have data from both regions from individual donors. Specifically, we used a mixed-effect model (STAR Methods) that uses a random effect to take into account samples that originated from the same donor. For the pseudobulked gene-level comparison, we calculated the total number of differentially expressed genes between each pair of spatial domains in the dACC and dlPFC, scaled by the total number of differentially expressed genes (DEGs) in each dACC spatial domain (Figure 5B). The pseudobulked gene-level analysis showed a small proportion of DEGs between L1, L2, L3, L5, and WM domains between dACC and dlPFC, indicating a strong similarity between these domains. We found complementary results between the two levels of comparison, as dlPFC L4 had the highest minimum number of DEGs compared to all other dlPFC spatial domains. This result indicates that there was no singular dACC spatial domain that was unequivocally mapped to dlPFC L4. Also similar to the pseudobulked sample-level of comparison, dlPFC L6 had a smaller proportion of DEGs with dACC L6a compared to dACC L6b.

To further explore the spatial domain differences between dACC and dlPFC, we compared the log fold-change (logFC) values from the enrichment model DE results. L5 in dACC and L5 in dlPFC have a similar relationship to L6a in dACC and L6 in dlPFC, while L6b in dACC and L6 in dlPFC have a very low correlation (Figure 5D). logFC plots comparing the regions highlight the more substantial differences in the deeper layers (L5 and L6) compared to higher similarity in the more superficial layers (L1, L2, and L3) (Figures 5D; Figure S28). While correlation across the regions for L2 was strong, we noted several notable differences, including BDNF and RSPO2. Consistent with our data in Figures 2G and 2H, this analysis provides additional molecular evidence for the lack of an analogous L4 layer in the dACC. As noted previously, dlPFC L4 had the highest minimum number of DEGs compared to all other dlPFC spatial domains, showing that it did not have a direct pairing in the set of dACC spatial domains (Figure 5B). Moreover, in the scatterplots, we note that at the gene level, RORB and PVALB, which are established as L4 markers in granular cortical areas, are enriched in dACC L5. We also noted that another class of dACC enriched genes in L5 are expressed in VENs (POU3F1 and HAPLN4), providing additional evidence for localization of molecularly defined VENs in the dACC versus dlPFC. In L6, we noted enrichment of SSTR2 and TBR1 when comparing L6a dACC to L6 dlPFC. We also noted enrichment of CPLX3, ADRA2A, and NOS1 when comparing L6b dACC to L6 dlPFC, highlighting significant differences in deeper layers across dACC and dlPFC. Furthermore, we showed moderate correlations when comparing dlPFC L4 with dACC L3 and dACC L5 (Figure S28).

Since L6a and L6b cluster as distinct spatial domains in the dACC, we further probed differences between these sublayers. We note that the spatial deconvolution results show that the snRNA-seq cell-type L6b clearly and specifically maps to the L6b SRT spatial domain, highlighting the distinction (Figure 3C). We first computed pairwise model DE genes between L6a and L6b in the dACC pseudobulked SRT data (Table S6). Several established neuronal markers were more highly expressed in L6a, including CCK, NPTXR, and NCAM2, while several glial markers were more highly expressed in L6b, including MBP, MOG, and GFAP (Figure S26H). These data likely reflect the closer proximity of L6b to the boundary of the WM and the inclusion of a WM transition zone in the dACC. From the dACC snRNA-seq NMF factorization, we identified two NMF patterns associated with L6 cell types (NMF15 and NMF35). NMF15 was associated with L6 CT neurons, whereas NMF35 was associated with L6b neurons (Figure 3D). Directly comparing these dlPFC and dACC cell types revealed that both the L6 CT and L6b cell types were associated with the dlPFC Excitatory L6 cell type (Figure S23). Similar to our approach in Section 2.5, we projected these two patterns into the dACC and dlPFC SRT datasets (Figure 5C). Spot plots localizing NMF35 in the dACC revealed that this pattern was localized to the boundary of L6a and L6b, bleeding into both regions, while NMF15 was restricted to L6a. In contrast, the dlPFC spot plots showed that NMF15 is localized to L6, while NMF35 is only very sparsely expressed in L6. The data reveal that the gene expression patterns present in L6b of dACC are not strongly expressed in dlPFC.

To extend our results and provide orthogonal validation for selected findings from these analyses, we performed multiplex smFISH for DRD5, KCTD8, and CPLX3 in dACC and dlPFC tissue sections derived from the same brain donor. DRD5, KCTD8, and CPLX3 were identified as relatively specific markers for dACC L5, L6a, and L6b domains, respectively. In dACC, we confirmed robust and specific expression in the expected layers for all three markers. However, in dlPFC, while KCTD8 and CPLX3 were expressed in the deep layers, expression was not as restricted compared to the dACC. For DRD5, expression in dlPFC was not robust, nor localized, to L5 (Figure 5F).

Spatial domain and cell-type mapping and inference of clinical and functional datasets

dACC dysfunction has been implicated in many neuropsychiatric disorders. To evaluate heritability across disorders and their relevant behavioral traits,^50-66 we used stratified linkage disequilibrium score regression (s-LDSC). We used the Z score of per-SNP heritability to calculate FDR-adjusted p values and quantify the statistical significance of the enrichment for each trait within spatial domain and cell type marker genes (Figures 6A-6D; Table S7; Table S8). Five spatial domains and ten cell types were significantly associated (FDR <0.05) with at least one trait. L2 and L3 spatial domains were associated with bipolar disorder, while the L6a spatial domain and L5 IT, L6 CT, and L6b excitatory neurons were associated with intelligence. As expected, microglia were associated with Alzheimer’s disease.⁶⁹ Inhibitory VIP+ neurons were associated with depression, further supporting a role for GABAergic dysregulation in major depressive disorder (MDD).⁷⁰

Figure 6. Stratified linkage disequilibrium score regression (s-LDSC) identifies differential disease risk heritability across cell types and spatial domains, and functional enrichment identifies differential enrichment across cell types and spatial domains.

(A) Plot showing stratified linkage disequilibrium score regression (s-LDSC) coefficient z-scores for heritability of various polygenic traits (y axis) across dACC SRT spatial domains (x axis). Coefficients found to be significantly different from the baseline model following FDR correction (FDR <0.1) are shown. Results are colored by s-LDSC regression coefficient Z score. The black “X” indicates stronger statistical significance (FDR <0.05). The full table of results is shown in Table S7.

(B) Heatmap showing statistical significance from Fisher’s exact test of psychiatric-related DEGs from a previous human dACC bulk RNA-seq study⁶⁷ (x axis): MDD Down: downregulated in major depressive disorder; PTSD Down: downregulated in post-traumatic stress disorder; MDD Up: upregulated in major depressive disorder; PTSD Up: upregulated in post-traumatic stress disorder; to dACC spatial domains (y axis). Significant results with confidence intervals for the odds ratio are reported here: L1/MDD Down (3.556, ∞), WM/MDD Down (1.165, ∞).

(C) Heatmap showing statistical significance from Fisher’s exact test of pain-related DEGs from a previous mouse ACC bulk RNA-seq study⁶⁸ (x axis): Pain down: downregulated in pain; Pain up: upregulated in pain; to dACC spatial domains (y axis). Significant results with confidence intervals for the odds ratio are reported here: L6b/Down (4.053, ∞), L6a/Down (1.627, ∞), WM/Down (1.204, ∞).

(D) Same as (A) for dACC snRNA-seq cell types. The full table of results is shown in Table S8.

(E) Same as (B) for dACC snRNA-seq cell types. Significant results with confidence intervals for the odds ratio are reported here: MicroPVM/MDD Down (3.657, ∞), Lamp5/MDD Down (1.656, ∞), Pvalb/MDD Down (1.546, ∞), L2/3 IT/MDD Down (1.279, ∞), Sst/MDD Down (1.131, ∞), L6b/MDD Down (1.288, ∞), L6 IT/MDD Down (1.071, ∞), Astro/MDD Up (1.041, ∞), OPC/MDD Up (1.278, ∞), Lamp5/PTSD Down (2.198,∞), Sst/PTSD Down (1.904, ∞), Astro/PTSD Down (1.220, ∞)

(F) Same as (C) for dACC snRNA-seq cell types. Significant results with confidence intervals for the odds ratio are reported here: Oligo/Down (1.821, ∞), OPC/Down (1.063, ∞), L5/6 NP/Down (1.169, ∞), Astro/Down (1.158, ∞).

As a complementary approach to heritability analyses, we leveraged transcriptional signatures from case-control bulk RNA-sequencing of the human dACC⁶⁷ to test whether genes differentially expressed in post-traumatic stress disorder (PTSD) or MDD were enriched within specific spatial domains and cell types using a one-sided Fisher’s exact test (Figures 6B-6E). Downregulated MDD DEGs showed significant enrichment (p < 0.05) in both L1 and WM. Among dACC cell types, downregulated MDD DEGs were enriched across multiple neuronal and non-neuron populations, including inhibitory neurons (LAMP5+, PVALB+, and SST+), excitatory subclasses, and glial/immune-related cell types such as microglia. Downregulated PTSD DEGs similarly showed enrichment in inhibitory neuronal populations and astrocytes. In contrast, upregulated MDD DEGs were enriched in astrocytes and oligodendrocyte precursor cells (OPCs). Together, these findings point to coordinated alterations across both neuronal and non-neuronal populations in the pathophysiology of MDD and PTSD in the dACC. Disease-associated transcriptional changes in the dACC reflect distributed dysfunction across inhibitory circuits together with prominent involvement of glial and immune-related processes, rather than being restricted to a single cell class.

Given the prominent role of the dACC in pain perception and processing,¹⁴ we next asked whether transcriptional signatures associated with chronic pain states map onto specific spatial domains and cell types in the dACC. We used a bulk RNA-sequencing dataset that analyzed gene expression changes in the ACC in a mouse model of repeated activation of ACC innervation from the basolateral amygdala that is associated with behaviors linked to neuropathic pain and depression.⁶⁸ Using the same enrichment framework as above, we assessed whether genes differentially expressed in the ACC pain model mapped onto human dACC spatial domains and cell types. We found that downregulated genes were concentrated in the deep-layer domains L6a and L6b, and in WM. At the cellular level, we found that downregulated genes were enriched across multiple cell classes, with significant associations in oligodendrocytes, astrocytes, OPCs, and L5/6 NP neurons (Figures 6C-6F). These findings suggest that transcriptional signatures associated with chronic pain map preferentially onto glial and myelination processes in the human dACC, with additional contributions from deep-layer cortical circuits. While deep layer neurons are well positioned to influence subcortical systems involved in sensory processing, affect and motor control,^71,72 the strong enrichment in WM and oligodendrocyte lineage cells and astrocytes highlights a central role for neuroglial interactions and WM-associated processes in pain-related circuit dysfunction. More broadly, this analysis underscores the utility of our spatial and snRNA-seq atlases as a valuable tool for linking circuit-level dysfunction to molecular and cellular pathology in neuropsychiatric and neurological disorders.

DISCUSSION

To investigate spatio-molecular features of the human dACC that could reflect the region’s unique structure and function, we generated paired snRNA-seq and SRT data. To provide evidence for the molecular rationale underlying evolutionary specializations in the dACC, we deployed comparative molecular neuroanatomy analyses using data generated in the same ten donors in the dlPFC, which used the same molecular profiling approaches.²⁸

The agranular organization of the dACC has long been appreciated from a cytoarchitectural perspective, but the molecular basis for this organization is not well understood. While molecular differences that support ACC agranularity were previously investigated using snRNA-seq data generated in sgACC and dlPFC of paired donors,⁷³ SRT is advantageous because it provides gene expression at exact x-y coordinates, providing anatomically precise, rather than inferred, spatial positioning. Moreover, snRNA-seq fails to capture extranuclear transcripts, and given the expanded amount of neuropil in the human brain,⁷⁴ cytoplasmic transcripts are likely to significantly contribute to overall molecular laminar identity. We used the dlPFC SRT and snRNA-seq data to determine that canonical L4 markers, including RORB and PVALB, are substantially expressed in both L4 and L5 spatial domains and corresponding cell types. However, in the dACC, expression of these genes is limited to the L5 spatial domain and cell types. Expression of these markers is not enriched in the L5-adjacent L3 spatial domain nor in L3 cell types, confirming that the spatial clustering for dACC with no L4 spatial domain accurately reflects the agranular structure of the region. Ability to query the integrated snRNA-seq and SRT allowed us to provide compelling molecular evidence for the absence of L4 in dACC in the human brain.

VENs are an understudied cell type that are present in a selected set of large-brained animals and are unique to only a handful of frontal cortical regions.⁴⁷ We identified an NMF pattern from the snRNA-seq data that corresponded to VENs. Projection of this pattern into the SRT data enabled us to better characterize the spatial localization of VENs, which we mapped to deep L5 in the dACC. Both data modalities were necessary for this discovery, since we leveraged the finer resolution of the snRNA-seq data to find gene programs and then leveraged the spatial topology from the SRT data to map the single-nucleus resolution gene patterns. In addition to mapping sub-localization within L5, we were able to infer connectivity patterns of VENs using functional data from rodent models, specifically that the NMF61 pattern (L5 ET/VENs) projects subcortically to the pons, superior colliculus, thalamus, and ventral tegmental area, while the NMF38 pattern (L5 IT) projects to other cortical regions. This information is valuable because the ACC is highly connected to both the prefrontal cortex and to subcortical regions within the limbic system, and the topography of connections to these areas differs across subdivisions of the ACC.^17,18

Substantial donor-to-donor variability in gene expression is a significant caveat of postmortem human tissue studies that complicates comparative analyses across brain regions. To leverage the fact that the dACC data were generated in the same donors as our previous dlPFC study,²⁸ we deployed comparative, within-donor comparisons of the two brain regions, which revealed greater divergence between regions in the deeper layers compared to the superficial layers. Specifically, we noted a substantial difference in L5 molecular composition; this difference was partially driven by gene expression patterns in deep L5, corresponding to VENs, which were not present in the paired dlPFC tissue. While this study focused on profiling the neurotypical dACC, the developed framework can be extended to case-control studies to identify direct evidence for the role of VENs in neuropsychiatric and neurological disorders. We noted substantial differences in the molecular composition of L6—both in spatial localization of known L6 marker genes, and also identifying both L6a and L6b markers in the dACC that are entirely divergent from the dlPFC. The dACC, dlPFC, and hippocampus are highly connected in the human brain, and similar cross-region comparison computational strategies can be adapted to assess molecular coherence across these connected regions to investigate cellular communication patterns. We note that the data generated in our human hippocampus snRNA-seq and SRT data resource also used the same ten neurotypical donors used to generate the dACC and dlPFC data.^28,40 To make this data resource easily accessible for future analyses and data analysis methods development or benchmarking, we made all raw and processed data freely available through GEO and several interactive web applications.

Limitations of the study

Several limitations should be considered when interpreting these results. First, the use of postmortem human brain tissue introduces inherent caveats that warrant cautious interpretation of RNA expression patterns. In addition, our analyses are restricted to the anterior dorsal anterior cingulate cortex (adACC/aMCC; BA24/33) from adult neurotypical donors, and extrapolation to other ACC subregions, developmental stages, or disease states should be made with caution. Second, SRT data were generated using the 10× Genomics Visium platform, which does not achieve single-cell resolution. As a result, individual spots can contain mixtures of cell types, particularly at laminar boundaries, complicating spatial domain definitions. While integration with snRNA-seq data helps mitigate this limitation, residual ambiguity in cell-type localization may remain. Third, the sample size is modest (n=10 donors), and inter-individual variability may influence the observed patterns despite efforts to control for technical and biological covariates. More broadly, transcriptomic measurements may be affected by batch effects and differences across sequencing platforms, which represent inherent limitations of large-scale molecular profiling studies. Fourth, the identification of spatial domains using NMF is sensitive to model parameters such as rank selection and initialization. Although we employed cross-validation approaches to guide the selection of k = 75 and assessed stability across runs, alternative parameterizations may yield partially different domain structures. Fifth, validation using smFISH is limited by the number of genes and regions that can be assayed, and thus provides only partial coverage of the full transcriptomic landscape. Finally, the associations between spatial domains, cell types, and neuropsychiatric traits or external transcriptional signatures (including cross-species comparisons) are based on statistical enrichment and correlation, and therefore do not establish causality. In particular, mappings from model organism datasets to human dACC cell types and spatial domains should be interpreted as inferential rather than direct evidence of conserved biological mechanisms. Despite these limitations, the convergence of spatial, snRNA-seq, and external dataset analyses supports the robustness of the major conclusions.

STAR★METHODS

EXPERIMENTAL MODEL AND STUDY PARTICIPANT DETAILS

Postmortem human brain tissue from the Lieber Institute for Brain Development (LIBD) Human Brain and Tissue Repository was collected through the following sites and protocols at the time of autopsy with informed consent from the legal next of kin: the Office of the Chief Medical Examiner of the State of Maryland, under the Maryland Department of Health IRB protocol #12–24, the Departments of Pathology at Western Michigan University Homer Stryker MD School of Medicine, at the University of North Dakota School of Medicine and Health Sciences, and the County of Santa Clara Medical Examiner-Coroner Office in San Jose, CA, all under the WCG protocol #20111080. Demographics for the 10 neurotypical control donors of European ancestry are listed in Table S1. Details of tissue acquisition, handling, processing, dissection, clinical characterization, diagnoses, neuropathological examinations, and quality control measures have been described previously.⁷⁸ Fresh frozen coronal slabs at the level of the anterior striatum (caudate nucleus and putamen) were rapidly sub-dissected with a hand-held dental drill, perpendicular to the pial surface, targeting the dACC Brodmann Areas (BA) 33 and 24. Dissected blocks contained the following landmarks: corpus callosum ventrally, the anterior portion of the cingulate sulcus dorsally, white matter of the forebrain laterally, and the interhemispheric fissure medially. Brain blocks were approximately 8 × 8 × 5 mm in dimensions, were collected across the callosal sulcus to preserve the integrity of layer 1, and encompassed complete laminar structure, spanning all cortical layers of BA 33 and 24 and white matter. Tissue blocks were stored at −80°C in sealed cryogenic bags until cryosectioning.

METHOD DETAILS

Tissue processing and anatomical validation

Following block dissection, smaller tissue blocks were embedded in cold OCT and flash frozen in isopentane on dry ice and stored at −80°C in sealed cryogenic bags. Larger tissue blocks were flash frozen on dry ice and stored at −80°C in sealed cryogenic bags. At the time of cryosectioning, tissue blocks were mounted onto round chucks with OCT, acclimated to the cryostat (Leica CM3050s) at −14°C, 50 μm of tissue was trimmed from the block to achieve a flat surface, and several 10 μm sections were collected for quality control (RNAScope, H&E). Three anatomical validation measures were implemented: 1) visual inspection of the blocks to assess inclusion of the callosal sulcus, white and gray matter; 2) H&E staining to assess cellular integrity of cortical lamina and the white matter; and 3) multiplex RNAScope single molecule fluorescence in situ hybridization (smFISH) to ensure presence of molecular markers for cortical layers and the white matter (Figures S2-S5). H&E staining was performed as previously described and images were acquired using an Aperio CS2 slide scanner (Leica) equipped with a 20×/0.75NA objective and a 2× doubler. RNAScope smFISH was performed as previously described,⁷⁹ and images were acquired using the Nikon AX-R confocal microscope, equipped with a 2×/20× NA objective or a Vectra Polaris slide scanner (Akoya Biosciences). Probes for established marker genes (ACD Bio) were used to identify cortical laminae in the human cortex,²⁷ including SLC17A7 (Cat No. 415611) indicative of all present layers; RELN (Cat No. 413051) indicative of layer 1; and MBP (Cat No. 411051) indicative of white matter. Subsequent to the anatomical validation experiments to verify the presence of all cortical layers, blocks were again acclimated to the cryostat, mounted onto chucks, and scored with a razor to isolate Brodmann Area 33 and 24 and white matter in 6.5 × 6.5 mm squares. Adjacent 10 μm tissue sections were mounted onto the Visium Spatial Gene Expression Slide (catalog no. 2000233, 10× Genomics) and onto glass slides for further anatomical validation (RNAScope). Four independent donors were processed on any individual Visium slide, with a total of n=17 capture areas processed. For some donors, biological replicates (2-3 additional capture areas per donor) were generated.

Spatially resolved transcriptomics (SRT) data generation

Scored 10 μm sections from each of the 10 dACC tissue blocks (Figure S2-S5), were collected onto a 10× Visium Gene Expression slide, and processed according to manufacturer’s instructions (protocol number CG000239, Rev G, 10× Genomics) as previously described.^27,28 Samples were processed with 4 different donors mounted on each Visium slide. Three of the donors were processed 3 times (3 spatially adjacent replicates each), one of the donors was processed twice (2 replicates) and the remaining 6 donors were processed once (1 replicate each) (Table S1). Prior to each Visium assay, H&E staining was performed (protocol CG000160, revision B, 10× Genomics) and high-resolution, brightfield images were acquired on a Leica CS2 slide scanner with a 20×/0.75NA objective and a 2× doubler. Following imaging, tissue sections were permeabilized for 18 min, reverse transcription was performed directly on the Visium slide and cDNA was collected from the slide, followed by library construction. The resulting libraries were quality controlled and sequenced on Illumina Nova-Seq 6000 following manufacturer’s instructions using a loading concentration of 300 pM at a minimum depth of 60,000 reads per Visium spot. Samples were sequenced to a median depth of 275,278,056 reads, corresponding to a median 67,211 of mean reads per spot, a median 2,257 of median unique molecular indices (UMIs) per spot, and median of 1,380 median genes per spot.

Single-nucleus RNA-sequencing (snRNA-seq) data generation

Single-nucleus RNA-sequencing (snRNA-seq) was performed using 2–3 100μm cryosections collected from each donor brain block with 10× Genomics Chromium Single Cell Gene Expression V3 technology. Approximately 70–100mg of tissue was collected from each donor, placed in a pre-chilled 2mL microcentrifuge tube (Eppendorf Protein LoBind Tube, Cat #22431102), and stored at −80°C until the time of experiment. To isolate nuclei, cryosections were combined with chilled Nuclei EZ Lysis Buffer (MilliporeSigma #NUC101) into a glass dounce. Sections were homogenized using 10–20 strokes with both loose and tight-fitting pre-chilled pestles. Homogenates were filtered through 70 μm mesh strainers and centrifuged at 500 g for 5 min at 4C using a benchtop centrifuge. Nuclei pellets were resuspended in fresh EZ lysis buffer, centrifuged again, and resuspended in wash/resuspension buffer (1× PBS, 1% BSA, 0.2U/μL RNase Inhibitor). Final nuclei were washed in wash/resuspension buffer and centrifuged a total of 3 times. Alexa Fluor 488-conjugated anti-NeuN antibody (MilliporeSigma cat. #MAB377×) diluted 1:1000 in nuclei stain buffer (1× PBS, 3% BSA, 0.2U/μL RNase Inhibitor) was used to label nuclei by incubating at 4°C with continuous rotation for 1 h. Proceeding NeuN labeling, nuclei were washed once in stain buffer, centrifuged, and resuspended in wash/resuspension buffer. Nuclei were labeled with propidium iodide (PI) at 1:500 in wash/resuspension buffer and subsequently filtered through a 35μm cell strainer. Fluorescent activated nuclear sorting (FANS) was performed using a Bio-Rad S3e Cell Sorter at the Lieber Institute for Brain Development. Gating criteria were selected for whole, singlet nuclei (by forward/side scatter), G0/G1 nuclei (by PI fluorescence), and neuronal nuclei (by Alexa Fluor 488 fluorescence). First, 9,000 nuclei were sorted based on PI + fluorescence to include both neuronal and non-neuronal nuclei from each donor. Second, 9,000 additional nuclei were sorted into a separate tube based on both PI+ and NeuN+ fluorescence to facilitate enrichment of neurons. This resulted in a final N=20 for snRNA-seq (1 PI+ and 1 PI+NeuN+ sample for all 10 donors) with a total of 18,000 sorted nuclei per donor. Samples were collected over multiple rounds, each containing 3–4 donors for 6–8 samples per round. All samples were sorted into reverse transcription reagents from the 10× Genomics Single Cell 3Reagents kit (without enzyme). Reverse transcription enzyme and water were added to bring the reaction to full volume. cDNA synthesis and subsequent library generation were performed according to the manufacturer’s instructions for the Chromium Next GEM Single Cell 3^′ v3.1 (dual-index) kit (CG000315, revision E, 10× Genomics). Samples were sequenced on a NovaSeq 6000 (Illumina) at the Johns Hopkins University Single Cell and Transcriptomics Sequencing Core.

QUANTIFICATION AND STATISTICAL ANALYSIS

SRT data processing and analysis

Visium H&E image processing

Sample slide images were first processed using VistoSeg.⁸⁰ VistoSeg was used to divide the Visium sample slides into individual images using VistoSeg’s splitSlide function. This takes one large image and separates them into distinct capture areas, one for each area on the slide, labeled at A1, B1, C1, and D1. These individual capture area images were used as one of the inputs for SpaceRanger (10× Genomics).

Visium raw data processing

The individual images from VistoSeg were then aligned with the slide capture areas in the Loupe Browser (10× Genomics). This allows for the alignment of the image with the spots captured on the slide. Sample slides were then processed using the SpaceRanger (10× Genomics) version 1.3.1 which takes the json output from the Loupe Browser, the sample image, and associated FASTQ files in order to generate spatial feature counts for a sample. The SpaceRanger counts were input into the spatialLIBD Bioconductor package⁷⁶ to create a SpatialExperiment object.⁸¹

Visium data quality control

We removed genes with zero UMI counts across all spots, spots with all zero counts, and spots outside of the tissue defined by the Loupe Browser. We used the addPerCellQC() function from the scuttle Bioconductor package to generate spot-level quality control metrics, including library size defined as the total number of UMIs per spot, number of detected genes, and mitochondrial expression rate.⁸² We used the isOutlier() function from the scuttle Bioconductor package to explore spots with low library size and/or spots with low numbers of detected genes based on a 3× median-absolute-deviation (MAD) threshold.⁸² These low quality spots were enriched in white matter (Figure S6). To prevent not removing spots just from the white matter, we used an extreme filter to remove spots with less than 20 detected genes and/or spots with less than 20 total UMI counts. The threshold of 20 was chosen to conservatively remove only spots that we were confident had no expression. The final spots that were discarded are in Figure S6. Across many tissue sections, the gradient of mitochondrial expression rate followed gradients in histology, so we did not use this metric to remove spots (Figure S6). We removed one spot with 100% mitochondrial expression rate. We calculated the log2-transformed normalized expression matrix after removing spots using logNormCounts() from the scater Bioconductor package.⁸²

Feature selection, spatially variable genes

We used nnSVG, which implements nearest neighbor Gaussian processes, to detect spatially variable genes (SVGs).³⁴ We ran nnSVG within each sample and then combined the gene sets. Within each sample, we removed lowly expressed genes using filter_genes() to retain genes with at least 3 counts in at least 0.5% of spatial locations. Then, we re-calculated the log2-transformed normalized expression matrix for each sample after removing spots using logNormCounts() from the scater Bioconductor package.⁸² We used the nnSVG() function from the nnSVG Bioconductor package on each tissue section to rank the genes in terms of spatial variance. Then, we combined the gene sets following the recommendations in the nnSVG tutorial by calculating the mean of the ranks of each gene across all the samples to generate an overall ranking and then calculating the number of samples where each gene is within the top 1,000 SVGs for that sample. Using the second metric, we created a table of 1,487 “replicated” genes that were highly ranked in at least two samples. Finally, the genes were ranked within the set of “replicated” genes using the average overall ranking. The gene list is available in Table S2.

Unsupervised clustering of spatial transcriptomics data

Guided by the set of SVGs, we used PRECAST to integrate the samples and identify spatial domains.³³ We converted from the SpatialExperiment object to a Seurat object to a PRECAST object using CreateSeuratObject() from the Seurat CRAN package³⁶ and CreatePRECASTObject() from the PRECAST CRAN package. To prepare for PRECAST model fitting, we added an adjacency matrix using AddAdjList() and parameter settings using AddParSetting() with the default arguments. We used the PRECAST() function to identify clusters with the number of clusters ranging from 5 to 20. An example of PRECAST annotations with 5–20 spatial domains is shown in one sample, V12N28-334_C1 (Figure S7).

Evaluation of unsupervised clustering of spatial transcriptomics data

We used two metrics to evaluate the nnSVG-guided PRECAST spatial domains to help inform which k value to choose for downstream analyses (Figure S8). First, we calculated the purity of each domain for each gene, defined as the proportion of observations in its neighborhood from a different domain, with the function neighborPurity() from the bluster Bioconductor package.⁸³ We group the data by cluster and use the average purity of genes per cluster. The cluster purity metric showed similar purity from k=5 to 8 clusters, with a drop in purity after 9 clusters. Second, we used the metric H+, which measures the discordance of unsupervised clusters as an internal validity metric.⁸⁴ The H+ metric showed a decrease in discordance from k=5 to 7 clusters, with similar discordance values from 6 to 10.

Spatial domain annotation

Histology-driven spatial domains of the data were generated and annotated by manual classification of spots into domains using the Samui Browser, a web-based interactive tool for gene expression visualization overlaid with histology.⁸⁵ For spot classification, one sample from each donor was annotated based on anatomical features on the H&E images and expression of established gene markers for cortical layers (e.g., AQP4 for L1, HPCAL1 for L2/3, PCP4 for L5, KRT17 for L6) and MBP for white matter. The difference between white matter of the forebrain adjacent to L6 vs. white matter of the corpus callosum was annotated entirely based on histology and the H&E images because the same white matter marker gene, MBP, was used to annotate both domains due to its abundant expression. The final histology-driven annotations are shown in Figure S10. We also annotated the unsupervised spatial domains that were generated by PRECAST (k=9). We chose k=9 because this resolution generated contiguous spatial domains that resembled L1, L2, L3 and L5, based on established marker genes in the literature and in manually annotated layers of the human cortex (RELN for L1, LAMP5 for L2, ADCYAP for L3, PCP4 for L5 and NR4A2 for L6). This resolution (k=9) also produced three domains with established markers of WM (MOBP). However, one of these three domains was defined by low library size and they did not correspond to any anatomically distinct WM regions and were not identified during RNAScope anatomical validation. (Figure S9). We removed spots from this spatial domain from further downstream analysis. The remaining two WM domains were collapsed based on similarities in marker genes for WM. We assessed clustering stability with leave-one-out cross-validation. Because we had removed cluster 7 (WM-CC) from the final clustering, we removed the same spots before re-clustering with leave-one-out with k=8 and then computed the NMI against the final set of clusters used in the manuscript. This analysis shows that the clustering results are stable since the NMI is around 0.75 with a small amount of variation. None of the individual samples being left out greatly decreases the NMI from the mean of 0.75.

To increase confidence in annotation of the final domains, we performed spatial registration to show the similarity between the domains that were generated on manual classification of spots and the data-driven PRECAST domains. We utilized the DE enrichment model t-statistics (see next sections “Pseudobulking” and “Pseudobulking differential expression analysis”) from all data types for the spatial registration. To calculate a correlation matrix between the statistics, we used layer_stat_cor() from the spatialLIBD Bioconductor package,⁷⁶ specifying that the function keeps the top 100 marker genes (top_n). We used layer_stat_cor_plot() from the same package to visualize the correlation matrix as a heatmap.

Pseudobulking

In order to detect differentially expressed genes (DEGs) between the spatial domains, we pseudobulked the data, similar to a previous paper,²⁷ where we aggregated total UMI counts across all spots within a spatial domain and within a tissue section. We used registration_pseudobulk() from the spatialLIBD Bioconductor package⁷⁶ to pseudobulk the spatial domains and calculate a matrix approximate to log2(CPM +1) via cpm() from edgeR of the pseudobulked counts matrix. This resulted in a counts matrix with 13,576 genes and 119 samples. We visualized the percent variance explained and the first two principal components computed from the log2-transformed normalized pseudobulked matrix using getVarianceExplained(), plotExplanatoryVariables(), and plotPCA() functions from the scater Bioconductor package⁸² (Figure S11). There was 1 white matter pseudobulked sample with an extremely large PC2 value of 339 (sample V12N28-332_B1), so we removed this prior to downstream analyses.

Pseudobulked differential expression analysis

We conducted differential expression (DE) analysis using the pseudobulked samples via enrichment statistics to test if one cluster has greater expression than the rest, similar to a previous paper.⁴⁰ We used the functions wrapped into registration_wrapper() from the spatialLIBD Bioconductor package to compute gene enrichment t-statistics. We utilized the false discovery rates and log2 fold-change values from the gene enrichment tests to create volcano plots using EnhancedVolcano() from the EnhancedVolcano Bioconductor package⁶⁹ with the absolute log2 fold-change cut-off as 1.5 and the statistical significance cutoff as −log10(0.05) (Figure S26). We used sig_genes_extract() from the spatialLIBD Bioconductor package to extract the top 50 significant genes from each spatial domain (Table S5).

To compute pairwise DE genes between L6a and L6b, we used the same functions to extract the pairwise DE genes and visualize the results in a volcano plot (Figure S26H).

snRNA-seq data processing and analysis

snRNA-seq data processing and quality control

We removed empty droplets as the first step of quality control for the snRNA-seq data. We used barcodeRanks() from the DropletUtils Bioconductor package^86,87 to incorporate unique sample-informed thresholds for each sample. We used the barcode rank statistics to find cliff and knee points for each sample (Figure S13). To identify empty droplets, we used emptyDrops() from the DropletUtils Bioconductor package with 25,000 iterations (niters) and the knee points as lower bound on the total UMI count (lower) for each sample.

We removed genes with zero UMI counts across all nuclei and nuclei with all zero counts. Then, we wanted to remove low quality nuclei using doublet detection, mitochondrial expression rate, and library size before moving on to the next step in the analysis pipeline. We used the addPerCellQC() function from the scuttle Bioconductor package to generate nuclei-level quality control metrics, including library size, number of detected genes, and mitochondrial expression rate.⁸² We used the isOutlier() function from the scuttle Bioconductor package to explore low library size nuclei, nuclei with low numbers of detected genes, and nuclei with high mitochondrial expression rate based on a 3× median-absolute-deviation (MAD) threshold.⁸² We discarded these low quality nuclei (Figure S14, all low sum nuclei were also low detected nuclei). We used scDblFinder() from the scDblFinder Bioconductor package⁸⁸ to remove doublets from the dataset (Figure S15).

snRNA-seq feature selection and dimensionality reduction

We used the devianceFeatureSelection() function from the glmpca CRAN package³⁷ to calculate the Poisson deviance on the counts matrix, taking into account the donor batch variable. We used the highest Poisson deviance to select the top 2,000 features (Figure S17). To approximate GLM-PCA for dimensionality reduction, we used the nullResiduals() function from the glmpca CRAN package with Poisson likelihoods on Pearson residuals and then ran PCA using the runPCA() function from the scater Bioconductor package.⁸²

snRNA-seq clustering and cell type annotation

On the GLM-PCA reduced dimensions, we used RunHarmony() from the harmony CRAN package³⁸ for batch correction. The harmony batch corrected data were plotted using UMAP to ensure the samples were not separated out and clustered together (Figure S17). For cell type annotation, we used the Azimuth Human Motor Cortex reference data to predict cell types.³⁵ We used the function RunAzimuth() from the AzimuthR package, which uses the Seurat data organization tool.^36,89 This annotation resulted in 20 cell types, of which we removed the Sst+ Chodl+ GABAergic neuron (“Sst Chodl”) cell type due to its small proportion (n=34) (Figure S16). The Sst Chodl cell type is included in the snRNA-seq iSEE app⁷⁵ and publicly available RSingleCellExperiment object.⁹⁰ We used a UMAP plot to visualize the separation of the final 19 cell types (Figure 3A) using ggplot2.⁹¹

Pseudobulked DE analysis

Similar to the SRT DE analysis, we conducted DE analysis using pseudobulked samples via enrichment statistics to test if one cell type has greater expression than the rest. We examined the pseudobulked samples with PCA to ensure well-mixed data before proceeding with the DE analysis.

(Figure S17). We used registration_wrapper() from the spatialLIBD Bioconductor package to compute gene enrichment t-statistics58. We utilized the false discovery rates and log2 fold-change values from the gene enrichment tests to create volcano plots using EnhancedVolcano() from the EnhancedVolcano Bioconductor package⁹² with the absolute log2 fold-change cutoff as 1.5 and the statistical significance cut-off as −log10(0.05) (Figures S18 and S19). We used sig_genes_extract() from the spatialLIBD Bioconductor package to extract the top 30 significant genes from each cell type (Table S3). We visualized the log2-normalized expression of a couple of the top marker genes for each layer-specific cell type using the ComplexHeatmap Bioconductor package⁹³ (Figure 3).

Spatial deconvolution

We utilized the robust cell type decomposition (RCTD) framework to deconvolve the spot-based SRT data based on the paired snRNA-seq data. To calculate the cell type weights for each spot, we used runRctd() from the spacexr Bioconductor package,³⁹ specifying that the function uses the maximum number of cell types per spot as 5 (max_multi_types). We used Heatmap() from the ComplexHeatmap package⁹³ to visualize the relationship between cell types and spatial domains as a heatmap (Figure 3C).

Dorsolateral prefrontal cortex (dlPFC) data

Dorsolateral prefrontal cortex (dlPFC) SRT dataset

We leveraged dlPFC SRT data from the same ten donors to facilitate cross-region comparisons between the dACC and dlPFC. The dlPFC SRT data were downloaded using fetch_data() from the spatialLIBD Bioconductor package with the option “spatialDLPFC_Visium”.⁷⁶ This dataset contains 30 samples of dlPFC, where each donor has three samples. We relabeled the BayesSpace_harmony_09 annotation to reflect spatial domain annotations. Specifically, we associated 2 to Layer 1, 3-Layer 2, 5-Layer 3, 8-Layer 4, 4-Layer 5, 7-Layer 6, 6-WM, 9-WM, and removed 1 which corresponded to meninges.

Dorsolateral prefrontal cortex (dlPFC) snRNA-seq dataset

Additionally, we leveraged dlPFC snRNA-seq data from the same ten donors to facilitate cross-region comparisons between the dACC and dlPFC. The dlPFC snRNA-seq data were downloaded using fetch_data() from the spatialLIBD Bioconductor package with the option “spatialDLPFC_snRNAseq”.⁷⁶ This dataset contains 19 tissue blocks from 10 donors from dlPFC. We used the cellType_layer annotations for our comparisons. We used the same functions to compute pseudobulked DE marker genes for each cell type compared to all other cell types and used the same spatial registration process to compute the correlation between dlPFC cell type markers and dACC snRNA-seq cell type markers (Figure S23).

Non-negative matrix factorization (NMF)

NMF factorization

We used the log2-transformed normalized expression of the snRNA-seq dataset to run NMF. To decide on the number of ranks, we used the cross_validate_nmf() function from the singlet R package.⁹⁴ We determined the optimal rank to be k=75 (Figure S21) and then used the nmf() function from the RcppML R package to factorize the expression matrix.⁹⁵

NMF annotation

To remove NMF patterns associated with technical variables, we created a heatmap to visualize the association of each pattern with various technical variables, including brain donor, quality control metrics, and sex of brain donor. For continuous variables, such as mitochondrial percentage, sum UMI counts, and detected UMI counts, we computed the Pearson correlation with each NMF pattern’s weights. For categorical variables, we created dummy variables of either 0 or 1 to represent exclusion/inclusion in the category and then computed the correlation with each NMF pattern’s weights. The NMF patterns found to be associated with these variables were not used for further analyses: NMF29 and NMF31 for sex; NMF62 for mitochondrial percentage; NMF4, NMF8, and NMF12 for sum/detected UMI counts; and NMF6, NMF22, and NMF64 for brain donor (Figure S22).

The NMF patterns were aggregated across cell types, and the mean of each NMF pattern within each cell type is displayed in a heatmap (Figure S21). For each NMF pattern, the top 10 genes contributing to the loadings matrix (w matrix in the RcppNL::nmf() output) were chosen as the marker gene set (Table S4). We evaluated these marker genes to assure that they matched with the associated cell types for the NMF patterns as shown in the heatmap. The top NMF pattern for each cell type was selected for downstream analyses. For the few cell types with multiple, equally specific NMF factors, such as oligodendrocytes, one of the top NMF factors was chosen at random.

NMF pattern transfer to dACC and dlPFC SRT datasets

The annotated patterns learned from the dACC snRNA-seq were projected into various datasets using a similar procedure, first used in the human hippocampus study.⁴⁰ First, we describe transfer into the dACC SRT gene expression data. The loadings matrix and the gene expression matrix from the dACC SRT data were subset to the genes present in both matrices and ordered to match each other.

Then, we used the project() function from the RccpML R package to project the loadings into the log2-transformed normalized dACC SRT expression data.⁹⁵ The resulting patterns correspond to the same patterns annotated from the dACC snRNA-seq data. This process was used to transfer the patterns into other datasets, namely dlPFC SRT data and dlPFC snRNA-seq data. We quantified the average number of nonzero spots per sample for NMF patterns of interest, and the significance was defined using results of a two-tailed t test when p-values were <0.05 (represented as *p < 0.05, **p < 0.01, ***p < 0.001, or not significant (NS)).

NMF pattern transfer to mouse single-nucleus methylation sequencing (snmC-seq) dataset

The annotated patterns learned from the dACC snRNA-seq were projected into a mouse model linking axon projections to snmC-seq in neurons. The snmC-seq data were downloaded from GEO (accession code GSE230782) and reproduced a workflow used in a previous study to process the data.⁴⁰ We subset the dataset to include only ACC neurons and then log-transformed and negated the extracted CH gene body counts, approximating gene expression with the inverse of non-CpG cytosine methylation as recommended by the original authors. We then exponentiated the counts to create all non-negative values, which satisfy the assumptions of NMF. To convert into human morphology, we matched mouse genes with human orthologs and removed those without matches. This matrix is analogous to the dACC SRT gene expression data in the previous section. The loadings matrix and this matrix were subset to the genes present in both matrices and ordered to match each other. Then, we used the project() function from the RccpML R package to project the loadings into the final matrix.⁹⁵

NMF pattern differential expression

We identified differentially expressed genes between two NMF patterns using the following procedure. In the dACC SRT dataset, we used NMF38 and NMF61 transferred from the dACC snRNA-seq NMF patterns. Spots that had a nonzero entry in the patterns matrix for NMF38 were classified as expressing NMF38 (n=4,191), and spots that had a nonzero entry in the patterns matrix for NMF61 were classified as expressing NMF61 (n=1,022) (Figure S24). We used a similar DE framework as above to find pairwise DE genes and corresponding t-statistics between the two classifications with registration_stats_pairwise() from the spatialLIBD Bioconductor package.⁷⁶

Cross-region comparisons

von economo neurons (VENs)

We examined the distribution of von Economo neurons (VENs) in the dACC and compared this to the dlPFC. We used make_escheR() from the escheR Bioconductor package⁹⁶ to visualize spot plots of NMF38 and NMF61 expression in the dACC and dlPFC SRT data. We also created sample-average boxplots of NMF38 and NMF61 expression to compare the SRT and snRNA-seq data from both regions. Briefly, we calculated the average NMF38 and NMF61 expression for each donor within the dACC SRT Layer 5 and the dlPFC SRT Layer 5 and visualized these data points with boxplots. We also calculated the average NMF38 expression for each donor within the dACC snRNA-seq L5 IT cell type and the dlPFC snRNA-seq Excitatory L5 and visualized these data points with boxplots. Finally, we calculated the average NMF61 expression for each donor within the dACC snRNA-seq L5 ET cell type and the dlPFC snRNA-seq Excitatory L5 and visualized these data points with boxplots.

RNAScope validation

Three smFISH experiments were performed on tissue sections from postmortem human dACC and dlPFC from a subset of donors (Br2720, Br6471, Br8492, Br8667, Br6432 and Br8325) using RNAScope Multiplex Fluorescent Reagent Kit v2 (Advanced Cell Diagnostics) and 4-Plex Ancillary Kit as previously described.⁷⁹ Tissue was cryosectioned at 10 μm on a Leica cryostat from the same dACC tissue blocks used for Visium and independent dlPFC tissue blocks from the same donor. The following probe panels were used: 1) Hs-PCP4 (assigned Opal dye 690), Hs-RORB (assigned Opal dye 620), Hs-ADCYAP1 (assigned Opal dye 570 using catalog no. 446111, 446061, 582501, respectively (Figures 2K and 2L); 2) Hs-POU3F1 (assigned Opal dye 570), Hs-SULF2 (assigned Opal dye 620), Hs-GABRQ (assigned Opal dye 690), and Hs-PCP4 (assigned Opal dye 520) using catalog no. 483181, 502241, 483171, 446111, respectively (Figures 4G; 4Figure S25); 3) Hs-KCTD8 (assigned Opal dye 570), Hs-DRD5-O1 (assigned Opal dye 690), Hs-CPLX3 (assigned Opal dye 520) using catalog no. 563921, 818241, 487681, respectively (Figure 5F). Within each experiment, images were acquired with the same acquisition settings across all tissue sections capturing cortical layer L1 to WM using a Nikon AX-R confocal microscope, equipped with a Nikon APO lambda D 20×/0.80 NA objective. Acquired tiled z series images were used to generate stitched maximum intensity projection images. Images were then spectrally unmixed using Nikon NIS-Elements software to isolate and subtract the lipofuscin autofluorescence signal prior to analysis.

RNAScope quantification - experiment 1

Images from RNAScope experiment 1 (PCP4/RORB/ADCYAP1/AQP4) were quantified with HALO (Indica labs) as previously described.⁹⁷ HALO settings files can be found on GitHub under processed-data, then under 24-HALO-Analyses (https://github.com/LieberInstitute/spatialdACC/tree/main/processed-data/24-HALO-Analyses). To normalize cell density calculations, cell centroids were extracted for each DAPI+ region of interest (ROI) across all HALO-quantified RNAscope images. To standardize cortical depth across tissue sections, raw x-coordinates were scaled to a relative depth (xrel), where 0 and 1 represent the image boundaries. A manual orientation audit was performed to ensure anatomical consistency; samples where the initial coordinate system placed the white matter (WM) at the origin were mathematically inverted (1 - xrel). This transformation ensured that xrel = 0 consistently represents L1 and xrel = 1 represents the WM boundary across all donors and regions. ROIs were classified into mutually exclusive categories based on marker positivity (single-positive, multi-positive, or negative). Spatial distributions for ADCYAP1+, RORB+, and PCP4+ single-positive populations were visualized with Gaussian kernel density estimation (KDE) via ggplot2::geom_density.

Profiles were plotted on a reversed vertical axis (aspect ratio = 2) to maintain anatomical orientation. For these visualizations, each marker-positive population was self-normalized so that the total area under each curve equals 1. This highlights the spatial peaks and laminar specificity of each marker independently of their absolute abundance. All analyses were performed in R using data.table and ggplot2. To statistically evaluate the spatial separation between cell types, we calculated the median relative depth (xrel) for RORB+ and PCP4+ single-positive populations within each donor. These donor-level medians served as biological replicates (n=5 donors per region). We utilized a paired-sample t test to compare the median depths of these two markers within each brain region. All statistical tests were two-sided, with significance defined as p < 0.05.

RNAScope quantification - experiment 2

Images from RNAScope experiment 2 (POU3F1/SULF2/GABRQ/PCP4) were quantified with HALO (Indica labs) as previously described.⁹⁷ HALO settings files can be found on GitHub under processed-data, then under 24-HALO-Analyses (https://github.com/LieberInstitute/spatialdACC/tree/main/processed-data/24-HALO-Analyses). Given the enrichment of VENs in L5, we restricted analysis to this layer using PCP4 expression to set the boundaries of L5. Briefly, we performed k-means clustering at k=3 on both copy count and signal intensity for PCP4 and defined cells positive for PCP4 (PCP4+) as the top 2 clusters (3 minimum copy counts and minimum signal intensity of 9). Next, we spatially plotted PCP4+ cells and drew a polygon around them for each image using custom R scripts.⁹⁷ For each cell (defined by DAPI nuclear IF stain with an additional 3 μm extrapolation for cytoplasm) within the PCP4+ L5 polygon, we quantified the copy number of each gene and cell area. For each sample separately, we scaled the copy count by cell area divided by median cell area for that sample as recommended in a previous paper.⁹⁸ As copy counts may become unreliable estimates for highly expressed genes with saturated fluorescent signals, we used a Gaussian mixture model to call each cell as expressing or not expressing each gene within each sample, using Mclust() with k=2 from the mclust R package.⁹⁹ The proportion of cells for each sample classified as expressing POU3F1, SULF2, and GABRQ is presented in Figure S25A. We also calculated the proportion of total cells in each sample that were called as expressed two of the three genes or all three genes (Figure S25B). Additionally, we calculated the Spearman’s correlation for each sample for each pair of genes. For example, the Spearman’s correlation between POU3F1 and SULF2 copy counts in the left region of Br8325 for dACC was 0.37 (Figure S25C). We summarized these correlations into one plot, as shown in Figure S25D. To further describe the coexpression, we visualized the average proportion of cells called as expressing one gene that are also called as expressing another gene (Figure 4G). For example, to get the value of 0.45, we counted the total number of cells expressing SULF2 in one dACC sample. Out of those cells, we found the number of cells also expressing POU3F1 and calculated the proportion of SULF2 cells that also expressed POU3F1. We took the average across the 3 dACC samples to get 0.45.

Pseudobulked sample-level dACC and dlPFC comparison

One way we compared the spatial domains between the dACC and the dlPFC was on the pseudobulked sample-level. We used registration_wrapper() from the spatialLIBD Bioconductor package to compute gene enrichment t-statistics for the dlPFC SRT dataset.⁷⁶ We utilized the DE enrichment model t-statistics from the dlPFC SRT and dACC SRT datasets for spatial registration. To calculate a correlation matrix between the statistics, we used layer_stat_cor() from the spatialLIBD Bioconductor package,⁷⁶ specifying that the function keeps the top 100 marker genes (top_n). We used layer_stat_cor_plot() from the same package to visualize the correlation matrix as a heatmap.

Pseudobulked gene-level dACC and dlPFC comparison

A second way we compared the spatial domains between the dACC and the dlPFC was on the pseudobulked gene-level. The usual workflow, using limma::duplicateCorrelation(), is an approximation to a mixed effects model, which requires shared correlation across genes and computes very quickly. Instead, we used the dream workflow, which estimates random effects separately for each gene. We combined the pseudobulked expression matrices from both regions into one matrix, indicating the region in the metadata with the variable “Tissue”. We only kept the genes that were in both matrices in the combined gene expression matrix. In addition to the “Tissue” variable, we also had “Donor” and “Layer” variables in the metadata for each pseudobulked sample. For each “Layer” and “Tissue” combination, we calculated the number of differentially expressed genes using the dream workflow.^100,101 We created a DGEList object from the combined pseudobulked expression matrix and normalized it using calcNormFactors() from the edgeR Bioconductor package.¹⁰² We created a linear contrasts matrix with the format “~Layer_Tissue + (1 ∣ Donor)”, where “Layer_Tissue” is a variable combining “Layer” & “Tissue” and “Donor” is a random effect. The fixed effect variable to be tested is thus “Layer_Tissue”. We used voomWithDreamWeights(), followed by dream() and variancePartition::eBayes() to estimate precision weights and regression coefficients and apply empirical Bayes shrinkage on the linear mixed model.^100,103 We counted a gene to be significantly DE if the absolute value of the Z score, the p-value transformed into a signed Z score, was larger than 1.645 and the absolute value of the log2-fold change was greater than 1.5. We visualized these results with a heatmap, where the number of DEGs for each comparison was scaled by the total number of DEGs for the dACC spatial domain in that comparison. To clarify, the number of DEGs between Layer 1 in dACC and Layer 3 in dlPFC was scaled by the total number of DEGs across all Layer 1 dACC comparisons to dlPFC spatial domains.

Stratified linkage disequilibrium score regression (s-LDSC)

Genome annotation

To define the gene set for the dACC SRT data, we used a specificity score approach.²⁹ We created a pseudobulked gene expression matrix using aggregateAcrossCells() from the scuttle Bioconductor package⁸² and removed samples that had less than 50 spots. We used filterByExpr(group = spatial domains) from the edgeR Bioconductor package¹⁰⁴ to find an adequate expression cutoff and used that cutoff to remove genes that did not have sufficiently high expression across spatial domains. We also removed duplicated genes and any non-protein coding genes. We then computed normalized CPM values for each gene in each cluster using cpm() from the edgeR Bioconductor package.¹⁰⁴ The gene set for each spatial domain was made up of the genes with the top 10% of expression. We added a 100Kb window upstream and downstream of each gene in the gene set for each spatial domain. This created the genome annotation for each spatial domain.

The same approach was used for the cell types in the dACC snRNA-seq data.

GWAS enrichment

We used the stratified linkage disequilibrium score regression (s-LDSC) approach from a previous study⁴⁰ to evaluate the enrichment of heritability for a variety of traits.^50-66 Following the workflow from the LDSC resource website (https://alkesgroup.broadinstitute.org/LDSCORE), we ran s-LDSC for each genome annotation with the baseline LD model v2.2. This model has 97 annotations to control for the LD between variants with other functional annotations in the genome. We used HapMap Project Phase 3 SNPs as regression SNPs, and 1000 Genomes SNPs of European ancestry samples as reference SNPs.^105,106 To quantify enrichment, we used the Z score of per-SNP heritability metric. From the z-scores, we calculate adjusted p-values using the Benjamini & Hochberg procedure. To visualize the results, we included traits and domains with FDR <0.1 and added an additional “X” if the FDR <0.05.

Bulk data integration

Clinical enrichment of MDD and PTSD bulk RNA sequencing dataset

We utilized an existing bulk RNA sequencing dataset to test for clinical enrichment of MDD and PTSD associated DEGs within spatial domains and cell types. This dataset from Jaffe et al. (available in the online supplementary files) consists of DEGs associated with MDD and DEGs associated with PTSD compared to neurotypical donors from postmortem human brains within the dACC, dlPFC, and the amygdala.⁶⁷ We first consider the tests of enrichment within dACC spatial domains using the dACC MDD DEGs. For the MDD DEGs, we chose the subset from the dACC (the columns were dACC_logFC_MDD and dACC_adjPVal_MDD). We kept only the genes overlapping between the results table from the bulk dataset and the enrichment model results table for the spatial domains. We further separated the MDD analysis into upregulated and downregulated DEGs. The gene set for MDD upregulated was based on an adjusted p-value threshold of 0.1 with positive log fold-change, while the gene set for MDD downregulated was based on an adjusted p-value threshold of 0.1 with negative log fold-change. The background gene set for each was the remaining genes in the results table. For each spatial domain, the gene set was based on an adjusted p-value of 0.05 and a log fold-change greater than 1. The background gene set for each spatial domain was the remaining genes in the enrichment model results table. We used Fisher’s exact test with the function fisher.test(alternative = “greater”) from the stats base R package to test for enrichment and present the p-values in a heatmap. The same process was used to test for clinical enrichment of the spatial domains in the PTSD DEGs gene sets (the columns were dACC_logFC_PTSD and dACC_adjPVal_PTSD). Finally, the same process was used to test for clinical enrichment of the snRNA-seq cell types with both the MDD and PTSD DEGs gene sets.

Pain functional enrichment from mouse ACC bulk RNA sequencing dataset

We utilized an existing bulk RNA sequencing dataset to test for functional enrichment of pain associated DEGs from mouse ACC within spatial domains and cell types. This dataset consists of DEGs associated with pain from mouse ACC.⁶⁸ We downloaded the data from GEO (accession code GSE227159). We first considered the tests of enrichment for the spatial domains with the pain DEGs. To convert into human morphology, we matched mouse genes with human orthologs and removed those without matches. We kept only the genes overlapping between the results table from the mouse dataset and the enrichment model results table for the spatial domains. We further separated the pain functional enrichment analysis into upregulated and downregulated DEGs. The gene set for upregulated was based on an adjusted p-value threshold of 0.1 with positive log fold-change, while the gene set for downregulated was based on an adjusted p-value threshold of 0.1 with negative log fold-change. The background gene set for each was the remaining genes in the results table. For each spatial domain, the gene set was based on an adjusted p-value of 0.05 and a log fold-change greater than 1. The background gene set for each spatial domain was the remaining genes in the enrichment model results table. We used Fisher’s exact test with the function fisher.test(alternative = “greater”) from the stats base R package to test for enrichment and present the p-values in a heatmap. The same process was used to test for functional pain enrichment of the snRNA-seq cell types with both the upregulated and downregulated DEGs gene sets.

Supplementary Material

Supplementary data related to this article can be found online at https://doi.org/10.1016/j.celrep.2026.117500.

KEY RESOURCES TABLE

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Biological Samples
Human postmortem brain hippocampal tissue	Lieber Institute for Brain Development	N/A
Chemicals, peptides, and recombinant proteins
O.C.T. Compound	Tissue-Tek Sakura	product code 4583
10% neutral buffered formalin	Sigma-Aldrich	catalog no. HT501128
Opal Fluorophore 520	Akoya Biosciences	catalog no. FP1487001KT
Opal Fluorophore 570	Akoya Biosciences	catalog no. FP1487001KT
Opal Fluorophore 620	Akoya Biosciences	catalog no. FP1487001KT
Opal Fluorophore 690	Akoya Biosciences	catalog no. FP1487001KT
Fluoromount-G	ThermoFisher	catalog no. 00–4958-02
Critical commercial assays
Visium Spatial Gene Expression Slide & Reagent kit	10× Genomics	PN-1000184
Chromium Single Cell 3′ Reagent, dual-index v3.1	10× Genomics	PN-1000268
RNAscope Multiplex Fluorescent Kit, v2	Advanced Cell Diagnostics	catalog no. 323100
RNAscope 4-Plex Ancillary Kit	Advanced Cell Diagnostics	catalog no. 323120
RNAscope Probe- Hs-PCP4	Advanced Cell Diagnostics	catalog no. 446111
RNAscope Probe- Hs-RORB	Advanced Cell Diagnostics	catalog no. 446061
RNAscope Probe- Hs-ADCYAP1	Advanced Cell Diagnostics	catalog no. 582501
RNAscope Probe- Hs-POU3F1	Advanced Cell Diagnostics	catalog no. 483181
RNAscope Probe- Hs-SULF2	Advanced Cell Diagnostics	catalog no. 502241
RNAscope Probe- Hs-GABRQ	Advanced Cell Diagnostics	catalog no. 483171
RNAscope Probe- Hs-KCTD8	Advanced Cell Diagnostics	catalog no. 563921
RNAscope Probe- Hs-DRD5-O1	Advanced Cell Diagnostics	catalog no. 818241
RNAscope Probe- Hs-CPLX3	Advanced Cell Diagnostics	catalog no. 487681
Deposited data
Raw sequencing read data	This paper	GEO accessions GSE296731 and GSE296789
SingleCellExperiment and SpatialExperiment R objects	This paper	public GLOBUS endpoint (https://research.libd.org/globus/)
Software and algorithms
HALO	Indica Labs	https://indicalab.com/halo/
VistoSeg	Tippani et al.	https://doi.org/10.1017/S2633903X23000235
Loupe Browser	10× Genomics	https://www.10xgenomics.com/support/software/loupe-browser/downloads
SpaceRanger (1.3.1)	10× Genomics	https://www.10xgenomics.com/support/software/space-ranger/downloads/previous-versions
SpatialExperiment (1.20.0)	R package	https://www.bioconductor.org/packages/release/bioc/html/SpatialExperiment.html
SingleCellExperiment (1.32.0)	R package	https://bioconductor.org/packages/release/bioc/html/SingleCellExperiment.html
ComplexHeatmap (2.26.1)	R package	https://bioconductor.org/packages/release/bioc/html/ComplexHeatmap.html
patchwork (1.3.2)	R package	https://cloud.r-project.org/web/packages/patchwork/index.html
spatialLIBD (1.22.0)	R package	https://www.bioconductor.org/packages/release/bioc/html/spatialLIBD.html
ggplot2 (4.0.2)	R package	https://cloud.r-project.org/web/packages/ggplot2/index.html
ggrepel (0.9.8)	R package	https://cloud.r-project.org/web/packages/ggrepel/index.html
SpaceRanger (1.3.1)	10× Genomics	https://www.10xgenomics.com/support/software/space-ranger/downloads/previous-versions
SpatialExperiment (1.20.0)	R package	https://www.bioconductor.org/packages/release/bioc/html/SpatialExperiment.html
SingleCellExperiment (1.32.0)	R package	https://bioconductor.org/packages/release/bioc/html/SingleCellExperiment.html
ComplexHeatmap (2.26.1)	R package	https://bioconductor.org/packages/release/bioc/html/ComplexHeatmap.html
patchwork (1.3.2)	R package	https://cloud.r-project.org/web/packages/patchwork/index.html
spatialLIBD (1.22.0)	R package	https://www.bioconductor.org/packages/release/bioc/html/spatialLIBD.html
ggplot2 (4.0.2)	R package	https://cloud.r-project.org/web/packages/ggplot2/index.html
ggrepel (0.9.8)	R package	https://cloud.r-project.org/web/packages/ggrepel/index.html
ggspavis (1.16.0)	R package	https://bioconductor.org/packages/release/bioc/html/ggspavis.html
schex (1.16.1)	R package	https://www.bioconductor.org/packages/release/bioc/html/schex.html
scran (1.38.1)	R package	https://www.bioconductor.org/packages/release/bioc/html/scran.html
scry (1.22.0)	R package	https://www.bioconductor.org/packages/release/bioc/html/scry.html
scuttle (1.20.0)	R package	https://www.bioconductor.org/packages/release/bioc/html/scuttle.html
EnhancedVolcano (1.28.2)	R package	https://www.bioconductor.org/packages/release/bioc/html/EnhancedVolcano.html
edgeR (4.8.2)	R package	https://www.bioconductor.org/packages/release/bioc/html/edgeR.html
escheR (1.10.0)	R package	https://www.bioconductor.org/packages/release/bioc/html/escheR.html
spacexr (1.2.0)	R package	https://bioconductor.org/packages/release//bioc/html/spacexr.html
RColorBrewer (1.1.3)	R package	https://cran.r-project.org/web/packages/RColorBrewer/index.html
gridExtra (2.3)	R package	https://cran.r-project.org/web/packages/gridExtra/index.html
RcppML (0.5.6)	R package	https://cran.r-project.org/web/packages/RcppML/index.html
singlet (0.99.5)	R package	https://zdebruine.github.io/singlet/
Azimuth (0.5.0)	R package	https://azimuth.hubmapconsortium.org/
Seurat (5.0.1)	R package	https://cran.r-project.org/web/packages/Seurat/index.html
DropletUtils (1.22.0)	R package	https://bioconductor.org/packages/release/bioc/html/DropletUtils.html
nnSVG (1.8.0)	R package	https://www.bioconductor.org/packages/release/bioc/html/nnSVG.html
PRECAST (1.8)	R package	https://cran.r-project.org/web/packages/PRECAST/index.html
scDblFinder (1.16.0)	R package	https://bioconductor.org/packages/release/bioc/html/scDblFinder.html
variancePartition (1.40.2)	R package	https://bioconductor.org/packages/release/bioc/html/variancePartition.html
R programming language	R Core Team	https://www.r-project.org/
NIS-Elements HC (High Content Analysis) v6.10.02	Nikon	https://www.microscope.healthcare.nikon.com/en_AOM/products/software/nis-elements/nis-elements-hchttps://www.microscope.healthcare.nikon.com/en_AOM/products/software/nis-elements/nis-elements-hc
Other
Code for analyses in the present study	This paper	https://github.com/LieberInstitute/spatialdACC and 10.5281/zenodo.15830481
Interactive web-based apps	This paper	https://research.libd.org/spatialdACC/#interactive-websites

Highlights.

• Explore laminar and cell type organization of human dACC using snRNA-seq and SRT data

• von Economo neuron molecular program highlights dACC’s cytoarchitectural uniqueness

• Integrate matched dACC and dlPFC datasets for direct cross-regional comparison

• Identify cell types implicated in neuropsychiatric conditions and pain processing