Mapping spatially organized molecular and genetic signatures of schizophrenia across multiple scales in human prefrontal cortex

Abstract

The dorsolateral prefrontal cortex (dlPFC) controls many cognitive and emotional processes that are disrupted in schizophrenia (SCZ). However, the spatial location of molecular changes associated with SCZ within the dlPFC remains poorly characterized. dlPFC cell types are spatially organized across six layers into microcircuits that mediate dlPFC function. While SCZ has been linked to regionally-defined cell types, spatially-resolved transcriptomics (SRT) can more directly map molecular associations of disease. We integrated protein detection for perineuronal nets, neurons and vasculature with SRT to investigate how gene expression varies across different cellular microenvironments in the human dlPFC from neurotypical control (n=31) and SCZ (n=32) brain donors. We mapped transcriptional alterations in synaptic, neuroimmune and metabolic pathways to neuropil and glia-enriched domains including the white matter. Integrative analyses linked laminar alterations predominantly to non-neuronal populations, and in situ profiling further resolved these changes at cellular resolution, indicating intracellular transcriptional changes that may underpin SCZ-associated alterations. By mapping SCZ-associated ligand-receptor pairs, we curated altered patterns of cell-cell communication that may coordinate local signaling dynamics across specialized tissue microenvironments. Finally, analyses that integrate SCZ genetic risk further highlight the significance of neuron-glia interactions, suggesting that neuronal genetic liability may signal through non-neuronal alterations across cortical domains. This multimodal atlas provides a spatial-anatomical framework for linking genetic risk to transcriptional phenotypes in the human cortex, delineating SCZ-associated gene expression landscapes across layers, single cells, and microenvironments. The data is available as a browsable PsychENCODE resource to support broad utilization that can inform mechanistic studies investigating SCZ pathology and risk.

1 |. Introduction

Schizophrenia (SCZ) is a chronic neuropsychiatric disorder affecting approximately 0.3–1% of the global population.^1–4 SCZ is characterized by a diverse spectrum of symptoms that disrupts thoughts, emotions, and behavior. Symptoms are generally classified as positive, including hallucinations and delusions; negative, including social withdrawal and blunted affect; or cognitive, including deficits to working memory, executive function, and attention.^5,6 The cognitive symptoms are particularly challenging because they have a substantial impact on daily functioning and are correlated with poorer quality of life, but remain largely unaddressed by existing treatments.^7,8 These cognitive deficits have been linked to dysfunction in the dorsolateral prefrontal cortex (dlPFC).⁵ This region has undergone considerable evolutionary expansion in non-human primates and humans,⁹ and is critical for higher-order cognitive functions, including working memory, planning, and cognitive flexibility. As a neocortical structure, the dlPFC has six distinct layers, which contain differing proportions of neuronal and glial subtypes. These cell types are spatially and functionally organized into specialized microcircuits, which regulate complex cognitive processing and provide top-down control over subcortical structures.^8,10 Neuroimaging and postmortem studies consistently suggest abnormalities in the dlPFC of individuals with SCZ,^5,11,12 but the exact mechanisms underlying these disruptions remain poorly understood. Integration of both bulk and single-nucleus transcriptomic approaches with genome-wide association studies (GWAS) has provided some molecular-genetic evidence that SCZ vulnerability is preferentially enriched within specific cortical layers and cell types in the human prefrontal cortex, including the dlPFC.^4,5,13–17 However, the cellular and/or spatial resolution of these approaches has limited the ability to finely localize risk to specific laminar compartments and resident cell types that comprise these microcircuits. Rapidly emerging multi-modal approaches that incorporate spatially-resolved transcriptomics (SRT) can overcome these limitations.^12,18

We used the Visium and Xenium SRT platforms (10x Genomics) to profile gene expression in postmortem human dlPFC tissue from individuals diagnosed with SCZ compared to neurotypical controls (NTCs). We identified SCZ-associated molecular signatures linked to synaptic biology, immune signaling, and GABAergic signaling, which we mapped to discrete cortical layers and spatially-localized cell types. We extended our analyses by integrating protein detection with the SRT data to investigate how SCZ-associated gene expression varies across cellular microenvironments. Specifically, we analyzed microenvironments that have been linked with SCZ including compartments enriched for neurons, neuropil, perineuronal nets (PNNs), or vasculature. PNNs preferentially surround parvalbumin (PVALB) inhibitory neurons, which regulate synchronization of gamma oscillations that are altered in individuals with SCZ.^19,20 PNNs maintain synapses localized to neuronal cell bodies and proximal dendrites, and regulate synaptic plasticity, which is disrupted in SCZ.^19–21 Neurovasculature is linked to SCZ pathobiology, and directly interacts with immune signaling to regulate synaptic homeostasis.^22,23

This integrative multi-omic strategy allowed us to contextualize how SCZ-associated signaling pathways are impacted across spatially-localized cortical domains, cellular organization, and local microenvironments. The findings underscore the value of spatially-resolved approaches to identify and localize novel molecular targets or refine our understanding of known targets. To provide broad access to the scientific community, we share both raw and processed data, along with the code used for our analyses. We also developed user-friendly web resources that allow researchers to explore and visualize the SRT data.

2 |. Results

2.1 |. Cross-diagnostic identification of data-driven spatial domains and cellular microenvironments in control and SCZ human dlPFC

Using the Visium platform, we generated SRT data from postmortem human dlPFC (Brodmann area 46) in 32 donors diagnosed with SCZ and 31 age- and sex-matched NTCs (Figure 1A, Supplementary Table 1). dlPFC tissue samples were dissected perpendicular to the pial surface and spanned cortical layers (L) 1–6 and white matter (WM) (Figure 1B). Instead of performing standard hematoxylin and eosin (H&E) staining, tissue sections were immunofluorescently (IF) labeled against DAPI, wisteria floribunda agglutinin (WFA), NeuN and Claudin-5, to identify cellular microenvironments associated with neuropil, PNNs, neurons, and vasculature, respectively (Figure 1B) for spatial proteogenomic (SPG) profiling across defined microenvironments. Following quality control (QC), we retained 279,806 high-quality spots (Figure S1–Figure S3) that were comparable across groups (NTC: 136,456; SCZ: 143,350) (Figure S4). Expression of SNAP25 (neurons), MOBP (oligodendrocytes), and PCP4 (L5 marker) verified intact cortical structure across samples (Figure 1B).

Figure 1. Overview of data collection and downstream processing to identify spatial domains (SpDs) corresponding to cortical layers and cellular microenvironments in the human dlPFC.

(A) Schematic of the cohort and experimental workflow integrating Visium SRT and IF images to generate SPG data. Created in BioRender. Kwon, S. H. (2026) https://BioRender.com/nryjogo

(B) Overview of SPG data for NTC and SCZ samples where Br8667 and Br5973 indicate donor labels, respectively. (Top) Tissue block images showing laminar orientation and morphology (L for cortical layers; WM for white matter). (Middle) IF images highlighting cellular components and microenvironments (DAPI marking nuclei in blue, NeuN marking neuronal nuclei in yellow, WFA marking PNNs in pink, and Claudin-5 marking blood vessels in green). Scale bar: 1 mm. (Bottom) Spot plots displaying log-transformed normalized expression (logcounts) of SNAP25, MBP, and PCP4, outlining gray matter, WM, and cortical layer 5, respectively.

(C) Unsupervised clustering using PRECAST delineated SpDs that recapitulate the characteristic laminar organization of the human dlPFC at k=7.

(D) Proportion of Visium spots assigned to each PRECAST SpD at k=7 in all donors across sex and diagnosis.

(E) Correlation of enrichment t-statistics profiles between NTC and SCZ samples across PRECAST SpDs, demonstrating similar laminar organization. Heatmap shows correlation values (Corr.).

(F) Correlation of enrichment t-statistics of PRECAST SpDs with manual histological layer annotations.^26,27 Heatmap shows correlation values (Corr.), with higher-confidence annotations marked by “X”.

(G) Top 10 most highly expressed marker genes for each PRECAST SpD at k=7. Each SpD is denoted as SpD-L where “L” represents the most confidently approximated cortical layer for a given PRECAST SpD, using the spatial registration approach shown in (F).

(H) Classification of Visium spots (cyan circles, 55-μm diameter) into cellular microenvironments based on enrichment of non-nucleated neuropil (DAPI-free, absence of blue), neuronal bodies (NeuN, yellow), PNNs (WFA, pink), and vasculature (Claudin-5, green) in NTC and SCZ samples. (Top) Visium spots overlaid on raw IF images. (Bottom) Segmentation of respective IF signal (white) was thresholded (Thre.) to classify Visium spots.

(I) Heatmaps showing distribution of SPG spots classified into four SPG-defined tissue microenvironments (neuropil, neuronal, PNN, and vasculature microenvironments) across PRECAST SpDs at k=7. Each row represents a single SpD and each column represents an SPG-defined tissue microenvironment. Values indicate the proportion of SPG spots that are localized to each SpD (column-wise summing to 1). These patterns show consistent laminar localization of the microenvironments in both NTC and SCZ samples. The higher proportions observed in L3/4 are likely due to the larger size of this SpD. A complementary within-SpD assessment is provided in Figure S12.

We spatially clustered all SRT samples with PRECAST²⁴ using the top 100 marker genes for human cortical layers as reported by Huuki-Myers et al.²⁵ as input. We evaluated various clustering resolutions (k=2–16) (Figure S5), and used the spatial domains (SpDs) generated with k=7 for downstream analysis because these seven contiguous SpDs best corresponded to the classic histological cortical layers (Figure 1C, Figure S6). Supporting our decision to cluster across all samples, the majority of variation in gene expression at k=7 was attributed to SpD compared to other factors (e.g., slide id, age, or diagnosis) (Figure S7). We annotated the SpDs by adapting a previous approach called spatial registration,^25–27 where gene expression profiles from our seven SpDs were compared against those derived from the manually annotated human cortical layers from Maynard, Collado-Torres et al.²⁷ (Figure 1F). Each SpD was assigned to a canonical dlPFC layer, denoted as SpDX-L, where X represents the domain number and L represents the layer annotation: SpD07-L1/M, SpD06-L2/3, SpD02-L3/4, SpD05-L5, SpD03-L6, SpD01-WM, and SpD04-WM. For SpD07-L1/M, the M indicates inclusion of leptomeningeal structures, particularly vasculature, within this domain (Figure 1B). The proportion of spots for each SpD was balanced across diagnosis (Figure 1D), and there was strong correlation in gene expression across all seven SpDs between diagnosis (Figure 1E). Of the two SpDs associated with WM, SpD01-WM was anatomically positioned in between the deepest gray-matter domain (SpD03-L6) and the other WM domain (SpD04-WM). Analysis of the top 10 marker genes for SpD01-WM revealed co-enrichment of representative genes associated with neuronal functions (CCN2, FIBCD1, NXPH4) and oligodendrocyte functions (TMEM125, SLC31A2, OPALIN) (Figure 1G). Differential expression (DE) analysis between SpD01-WM and SpD04-WM revealed that SpD01-WM had higher expression of neuronal marker genes, including SNAP25 and SLC17A7, and lower expression of oligodendrocyte genes, including PLP1, MBP, and MOBP (Figure S8, Figure S9, Supplementary Table 2). Among the neuronal genes enriched in SpD01-WM, we identified genes associated with deep layer neurons in L6, including CCN2, CPLX3, and NR4A2.²⁸ Given its anatomical positioning at the border of the WM as well as mixed neuronal and glial transcriptomic profiles, we refined the annotation for SpD01-WM to WM transition zone (SpD01-WMtz).

To annotate Visium spots based on their inclusion within a cellular microenvironment, we segmented the high-resolution IF images of the immunostained tissue to identify regions of interest (ROI) for each IF channel (Methods, Figure S10, Figure S11). In each IF channel, we further classified Visium spots into one of the two categories based on spot coverage by segmented ROIs, namely SPG and non-SPG spots (Figure 1H, Methods).^25,29 Specifically, spots were classified as ‘neuropil’/‘cell body’ per the DAPI channel, ‘neuronal’/‘non-neuronal’ per the NeuN channel, ‘PNN’/‘non-PNN’ per the PNN channel, and ‘vasculature’/‘non-vasculature’ per the Claudin-5 channel. These classifications were independent of each other such that a spot could be positive for more than one feature (e.g., both neuronal and PNN). Among all QC’ed spots (n=279,806 spots), 74,000 spots contained only DAPI signals and were classified as non-SPG across all channels (Figure 1H, Figure S10, Figure S11, Methods). Analyzing across all samples, we identified SPG spots enriched for neuropil (NTC: 71,804; SCZ: 76,161), neuronal (NTC: 28,957; SCZ: 26,060), PNN (NTC: 7,084; SCZ: 5,249), and vasculature (NTC: 10,708; SCZ: 11,198).

We examined the distribution of SPG spot types across SpDs in both NTC and SCZ samples (Figure 1I). For all four IF channels, the respective SPG spots were localized to SpD02-L3/4 more than any other SpDs, likely reflecting the large size of this SpD (Figure 1D). Neuropil spots were detected across all SpDs, with the highest enrichment in SpD02-L3/4 and SpD07-L1/M and lowest enrichment in the WM SpDs (SpD01-WMtz and SpD04-WM). Neuronal spots were primarily localized to gray matter (SpD06-L2/3 to SpD03-L6). We also observed a small proportion of neuronal spots localized to SpD01-WMtz (NTC: 756 spots, 2.6%; SCZ: 479 spots, 1.8%), potentially reflecting deep cortical neurons, which is consistent with the relatively elevated neuronal gene expression in SpD01-WMtz compared to SpD04-WM (Figure S9). PNN spots were enriched in SpD02-L3/4. Since PNNs preferentially surround PVALB inhibitory neurons, this data is consistent with enrichment of PVALB neurons in L4.³⁰ Vasculature spots were broadly distributed across SpDs in both NTC and SCZ, with enrichment in SpD02-L3/4 (NTC: 3,744 spots, 34.9%; SCZ: 4,087 spots, 36%) and SpD07-L1/M (NTC: 1,717 spots, 16.0%; SCZ: 1,896 spots, 16.9%) (Figure 1I). We also quantified the proportion of SPG spot types within individual SpDs across NTC and SCZ (Figure S12). Among SPG spot types in SpD07-L1/M, vasculature spots showed the highest proportion after neuropil spots, potentially due to the contribution of leptomeningeal vasculature in the sulcal regions adjacent to L1 (Figure S12, Figure 1B).³¹ Neuropil spots, including the spots in WM, consistently accounted for the largest proportion in nearly all domains (Figure S12). This pattern may reflect the high density of neuropil, whose composition includes synaptic connections, glial processes, and extracellular matrix components, which is particularly abundant in human cortical regions that are enriched for dendritic and axonal arborizations.^32–35

To validate our classification approach, we assessed the molecular composition of the annotated cellular microenvironments in the NTC samples by analyzing gene expression in pseudobulked SPG spots compared to non-SPG spots from the corresponding microenvironment (Figure S13, Supplementary Table 3, Methods). Neuropil spots, compared to non-neuropil spots, were enriched for synaptic genes including CAMK2A, DLG4, MAP2, and SHANK1–3 (Figure S13A). Neuronal spots were enriched for neuronal markers including SNAP25 and RBFOX3 (Figure S13B). PNN spots were enriched for PVALB and proteoglycan-related genes (HAPLN1 and HAPLN4), as well as ERBB4, a receptor tyrosine kinase expressed in PVALB neurons.^30,36,37 These PNN spots were also depleted of metalloproteinase genes (MMP16 and MMP28) (Figure S13C), which function as extracellular matrix (ECM) remodeling enzymes that degrade PNN components and regulate their structural maintenance.^20,38 Vasculature spots were enriched for genes associated with endothelial junctions and immune response, including CLDN5, PECAM1, and A2M (Figure S13D).

We next leveraged the gene expression data to cross-validate the image-based annotation of SPG spots. Through existing transcriptomic studies, we compiled marker gene sets for neuropil, cortical neurons, PNN-ensheathed PVALB inhibitory neurons, and vascular cells.^25,30,39,40 Annotation of the SPG spot types using the compiled marker gene lists (referred to as gene expression-based classification, Methods), we observed strong concordance between image-based and gene expression-based classification, supporting the validity of image-based SPG spot annotations (Figure S14, Methods). We next compared the proportion of image-based SPG spots with marker gene expression across spots for each SpD and found that domains with higher SPG proportions generally exhibited higher expression of the marker genes of the same type, indicating concordance between morphological classification and transcriptomic profiles (Figure S15). We observed an increased proportion of PNN spots and elevated PNN marker gene expression in SpD02-L3/4 and SpD05-L5, supporting the biological validity of our image-based classifications (Figure S14C, Figure S15C).

2.2 |. Localizing SCZ-associated differences in gene expression to cortical layers

Both bulk RNA-seq and single-nucleus (snRNA-seq) studies have identified changes in gene expression associated with SCZ in the prefrontal cortex, including the dlPFC.^15–18,41 While laminar enrichment of differentially expressed genes (DEGs) in these studies can be inferred, SRT can provide more direct evidence for spatial localization of gene expression changes. To identify changes in gene expression associated with SCZ, we applied two different DE models to quantify the overall and within-SpD transcriptional alteration, respectively. We refer to the first model as the “layer-adjusted” DE model (Figure S16). The layer-adjusted model quantified global transcriptional alterations across tissue sections while adjusting for laminar differences using the data-driven SpDs. This analysis identified 172 DEGs (referred to as ‘layer-adjusted SCZ-DEGs’) with a false-discovery rate (FDR) < 0.10 (Figure 2A, Supplementary Table 4). The majority of these layer-adjusted SCZ-DEGs were down-regulated (58%, n=99 genes), which is consistent with previous bulk and snRNA-seq findings, including Ruzicka et al..^16,41 For example, a large-scale bulk RNA-seq study (BrainSeq Phase 2; dlPFC samples only) identified 632 DEGs at FDR < 0.10, a majority of which were also down-regulated (60%, n=379 genes).⁴¹ We compared our layer-adjusted SCZ-DEGs to BrainSeq Phase 2 bulk RNA-seq DEGs and found 32 overlapping genes, including MAPK3, C3, SERPINA3, and A2M (Figure 2B). MAPK3 was among the 73 up-regulated layer-adjusted SCZ-DEGs, and its dysregulation is consistent with both genetic and transcriptomic evidence linking MAPK3 signaling with SCZ.^13,42,43

Figure 2. Identification and spatial characterization of SCZ-associated gene expression changes across cortical layers in the human dlPFC.

(A) Volcano plot showing the log₂ fold-change (logFC) of 172 layer-adjusted SCZ-associated DEGs (SCZ-DEGs) with FDR < 0.10.

(B) Scatter plot comparing t-statistics between BrainSeq Phase 2 bulk RNA-seq DEGs previously identified in the dlPFC⁴¹ and layer-adjusted SCZ-DEGs identified here.

(C) Layer-adjusted SCZ-DEGs mapped onto synaptic compartments using the SynGO database.⁵⁶ Representative genes are annotated with their associated synaptic components.

(D) Dot plot showing enrichment of layer-adjusted SCZ-DEGs across the four SPG-defined microenvironments involving neuropil, neuronal, vasculature, and PNN microenvironments. Enrichment significance was stratified by FDR < 0.05 and nominal p < 0.05. Odds ratios were indicated with the color scale.

(E) Diverging bar plot showing the numbers of up- and down-regulated layer-restricted SCZ-DEGs identified at nominal p < 0.05, highlighting their overlap with the layer-adjusted SCZ-DEGs identified at FDR < 0.10.

(F) Volcano plots showing layer-restricted SCZ-DEGs for SpD07-L1/M and SpD01-WMtz, respectively. Blue denotes down-regulation and red denotes up-regulation in SCZ. Points in darker shades indicate genes at FDR < 0.10, whereas lighter shades indicate genes at nominal p < 0.05.

(G) Diverging bar plot showing the numbers of up- and down-regulated layer-specific SCZ-DEGs at nominal p < 0.05.

(H) Dot plot showing logFC and corresponding statistical significance of top 5 most differentially expressed up- and down-regulated layer-specific SCZ-DEGs (ranked by logFC) within each SpD, compared with their logFCs estimated in the layer-adjusted DE model (Top row). Statistical significance was stratified by FDR < 0.05 and nominal p < 0.05. LogFCs were indicated with the color scale.

Layer-adjusted SCZ-DEGs were consistent with known functional alterations in SCZ. For example, neuroplasticity and glutamatergic signaling are consistently implicated in SCZ pathology,^44,45 and our SCZ-DEGs included many genes with key roles in neuroplasticity, including BDNF, FKBP5, MAPK3, NPTX2, and EGR1. Consistent with previous studies in postmortem human brain,^17,46 we found down-regulation of microglial-associated genes linked to immune function and inflammatory signaling, including TMEM119, P2RY12, GPR34, AIF1, TREM2, and C3. GABAergic signaling dysfunction in SCZ is thought to contribute to disruptions of gamma band oscillatory synchrony and excitatory-inhibitory balance in SCZ, physiological functions that are controlled by GABA neurons.^47,48 Consistent with this functional link to SCZ, layer-adjusted SCZ-DEGs included genes associated with GABA neuron functions, such as CORT and SST. We also identified a number of layer-adjusted SCZ-DEGs associated with astrocyte biology, including GFAP, ALDH1A1, AQP1, and SLC14A1, supporting snRNA-seq studies showing a dysregulated synaptic neuron-astrocyte program in SCZ.¹⁵ Finally, we identified multiple metallothionein family members (MT1X, MT1M, MT2A) as layer-adjusted SCZ-DEGs, which participate in heavy metal detoxification and may influence the ability of PNNs to buffer metal ions via metalloproteinase-based ECM remodeling.^49,50 These findings may be relevant for emerging hypotheses linking developmental neurotoxicity to oxidative stress and redox dysregulation in SCZ.^51,52 Gene set enrichment analysis (GSEA) of the layer-adjusted SCZ-DEGs further reinforced these findings, which highlighted convergence on signaling pathways involved in various immune and inflammatory responses and cellular response to metal ions. Intriguingly, the ATP metabolism and mitochondrial respiratory terms were distinctively enriched, suggesting significant metabolic alterations across cortical layers in SCZ (Figure S17).

Given established links between synaptic dysfunction and SCZ,^15,53–55 we performed enrichment analysis of layer-adjusted SCZ-DEGs against SynGO annotations.⁵⁶ SynGO identifies overrepresented synaptic terms in an experimental dataset, predicts localization of genes to pre- and post-synaptic compartments and assigns known synaptic functions (Figure 2C). We identified a subset of layer-adjusted SCZ-DEGs as synapse-related, including NPTX2, BDNF, SST, C1QA, and C1QB, spanning both the pre- and post-synaptic compartments. Corroborating the synaptic function of these genes, we found that up-regulated layer-adjusted SCZ-DEGs were significantly enriched in SPG-defined neuropil spots (Odds ratio=2.07, nominal p=0.019; Figure 2D). Finally, we observed that several complement-related genes (e.g., C1QA, C1QB) were localized to the synapse, supporting the theory that neuroimmune regulation of synapse pruning may be disrupted in SCZ.^57,58

To directly test whether SCZ-associated changes in gene expression were localized to specific layers, we developed a second DE model, referred to as the “layer-restricted” DE model. Compared to the previous layer-adjusted model that quantifies the overall transcriptional alteration, the “layer-restricted” model quantifies the alteration within individual SpDs using statistical interaction of SpDs and diagnosis status (Figure S16). This analysis identified 944 and 6,978 DE events across all SpDs at FDR < 0.10 and nominal p < 0.05, respectively (Figure 2E, Supplementary Table 5), which included 170 out of the 172 layer-adjusted SCZ-DEGs. Consistent with snRNA-seq studies showing an up-regulation of DEGs in glial cells in SCZ,¹⁶ more than half of the layer-restricted SCZ-DEGs were localized to WM-associated domains (26%, n=1,581 genes for SpD01-WMtz; 36%, n=2,521 genes for SpD04-WM) and the majority of these WM-localized SCZ-DEGs show up-regulation (2,873 up-regulated vs. 1,229 down-regulated DEGs). Particularly, SpD01-WMtz, a region enriched in both neuronal and glial cell populations (Figure 1G, Figure S9), showed up-regulation of genes involved in synaptic signaling and neurotransmission, including SCZ-implicated neurodevelopmental genes (DISC1, DCC) and genes involved in glutamatergic signaling at the synapse (GRIN2C/D, SHANK1/2). Additionally, genes encoding neurotransmitter receptors, including the serotonin receptor HTR5A were up-regulated (Figure 2F). In contrast, the majority of layer-restricted SCZ-DEGs that localized to the superficial layers, particularly SpD07-L1/M and SpD06-L2/3, were down-regulated. For example, microglia immune-complement genes (AIF1, TREM2, C3), GABA neuron signaling genes (GAD1, SST, CORT, CRH, PVALB), and genes associated with neuroplasticity (BDNF, VGF, PTN, NLGN3, EPHB1) were down-regulated in SpD07-L1/M. Together, we observed layer-restricted variations in SCZ-DEGs and their expression patterns, but common biological themes emerged, including neurotransmission imbalance, immune dysfunction, and neurodevelopment across cortical layers and WM regions.

Last, we identified layer-restricted SCZ-DEGs that were differentially expressed in only a single SpD (we refer to these genes as ‘layer-specific SCZ-DEGs’). Among the layer-restricted SCZ-DEGs, 4,073 genes were layer-specific (2,416 up-regulated and 1,657 down-regulated). The number of layer-specific SCZ-DEGs was largely proportional to layer-restricted SCZ-DEGs, with the exception of SpD02-L3/4, which showed fewer layer-specific SCZ-DEGs (Figure 2G). Outside of the WM domains, the largest number of layer-specific SCZ-DEGs were found in SpD07-L1/M. For example, we observed prominent up-regulation of NRG4 in SpD07-L1/M. Neuregulin-ERBB signaling plays key roles in neural development and synaptic plasticity, suggesting NRG4, as a member of this family, may exhibit altered signaling dynamics within the neuropil-rich molecular layer.^59–61 Analysis of the layer-specific SCZ-DEGs with largest transcriptional alteration in each SpD (Figure 2H) suggests involvement of multiple signaling pathways associated with synaptic function and neuronal activity (KCNE4, PDLIM5, EEF2K), cellular growth and differentiation (NR4A1, SMAD6, TBX3), and immune regulation (LILRB1, HAVCR2, CCL2, HLA-DMB, HLA-DQB1, CD44), selectively altered across gray-matter domains. Taken together, we demonstrate localized changes in gene expression across cortical layers and WM in SCZ, where dysregulation of synaptic and immune functions is consistently observed as a broadly shared biological theme across the cortical architecture.

2.3 |. Integration of cellular context with spatially-localized gene expression changes to resolve cell-type contributions in SCZ

While our laminar DE analyses identified spatially-localized SCZ-DEGs with both synaptic and immune functions (Figure 2C–Figure 2H), these analyses lacked cellular resolution. To provide additional biological context interpreting the spatially-localized SCZ-DEGs, we integrated our findings with an existing snRNA dlPFC atlas²⁵ and a SCZ-comparative study,¹⁶ using the gene set enrichment approaches.

We first performed cell-type enrichment analysis using layer-adjusted SCZ-DEGs and dlPFC cell-type marker genes²⁵ (Figure 3A). Up-regulated SCZ-DEGs were enriched for marker genes of non-neuronal populations, particularly astrocytes and vascular leptomeningeal cells (VLMCs). Conversely, down-regulated SCZ-DEGs were enriched for markers of microglia and immune cells. Notably, both up- and down-regulated SCZ-DEGs were enriched in endothelial cells, pericytes (PC), and smooth muscle cells (SMC), consistent with emerging evidence of vascular signaling dysregulation in SCZ.^22,23 Unexpectedly, we did not observe strong evidence that supports the enrichment of neuronal cell-type markers among layer-adjusted SCZ-DEGs. To better understand this observation, we compared layer-adjusted SCZ-DEGs against SpD marker genes and checked their laminar localization (Figure 3B). SCZ-DEGs predominantly mapped to cortical domains enriched for non-neuronal cells, including SpD07-L1/M, SpD01-WMtz, and SpD04-WM, supporting an association with non-neuronal cell types. SpD07-L1/M, a domain abundant in vascular and glial cells, showed strong enrichment for both up- and down-regulated SCZ-DEGs. Together, these analyses highlight the emerging importance of non-neuronal populations in SCZ.^15,62,63

Figure 3. Integration of cellular context into spatially-localized SCZ-associated gene expression changes.

(A) Enrichment of 172 layer-adjusted SCZ-DEGs among dlPFC cell-type marker genes from Huuki-Myers et al..²⁵ Red and blue denote up- and down-regulated SCZ-DEGs, respectively. Dot size indicates the three different significance levels for gene set enrichment: non-significant (N.S.), nominal p < 0.05, and FDR < 0.05. Dot color indicates the odds ratio. The same color and significance schemes are applied across panels in A-E.

(B) Enrichment of 172 layer-adjusted SCZ-DEGs among marker genes associated with data-driven SpDs. The color scheme for directionality of gene expression is the same as in (A).

(C) Enrichment of layer-restricted SCZ-DEGs among dlPFC cell-type marker genes from Huuki-Myers et al..²⁵

(D) Enrichment analysis comparing reference cell-type-specific DEGs from Ruzicka et al.¹⁶ with layer-restricted SCZ-DEGs.

(E) Enrichment of layer-restricted SCZ-DEGs in SPG-defined microenvironment marker genes for neuropil, neuronal, vasculature, and PNN microenvironments. SpD labels and cortical layer assignments are consistent with those in (C).

We also performed cell-type enrichment analysis with layer-restricted SCZ-DEGs (Figure 3C). SCZ-DEGs, particularly those that were up-regulated, were enriched for astrocyte markers in neuronal-rich domains (SpD06-L2/3, SpD02-L3/4, SpD05-L5, SpD03-L6) as well as WM (SpD04-WM), suggesting widespread astrocyte-associated transcriptional alterations.^15,62,64 In contrast, down-regulated SCZ-DEGs were enriched predominantly for microglia and immune cell markers across all SpDs, consistent with our cell-type enrichment of layer-adjusted SCZ-DEGs (Figure 3A) and prior studies.^17,46 Enrichment of oligodendrocyte markers was divergent, with down-regulated SCZ-DEGs enriched for oligodendrocyte markers in deeper layers (SpD05-L5, SpD01-WMtz), and up-regulated SCZ-DEGs enriched in superficial layers (SpD06-L2/3). While we identified relatively few instances of enrichment among neuronal marker genes within SCZ-DEGs across SpDs, there were several exceptions. For example, down-regulated SCZ-DEGs in SpD06-L2/3 were enriched for L6 intratelencephalic-projecting (IT) neuron markers. Together, the cell-type enrichment analysis of layer-restricted SCZ-DEGs delineated the spatial localization of implicated cell types and revealed contributions from both glial and neuronal populations to SCZ-associated gene expression changes across cortical layers and WM.

To determine the extent to which our spatially-localized SCZ-DEGs recapitulate cell-type-specific transcriptional alterations identified in previously published SCZ snRNA-seq studies, we compared our layer-restricted SCZ-DEGs with cell-type-specific SCZ-DEGs identified from Ruzicka et al.¹⁶ (referred to as “reference DEGs”) (Figure 3D). Reference DEGs were stratified by directionality (Methods), and enrichment of up- and down-regulated reference DEGs was examined within our corresponding up- and down-regulated layer-restricted SCZ-DEGs. The highest concordance was observed among glial cells, including astrocytes, oligodendrocytes, and oligodendrocyte progenitor cells (OPCs). Specifically, up-regulated astrocyte-specific reference DEGs were enriched among up-regulated SCZ-DEGs, with the strongest concordance in SpD02-L3/4. This finding provides evidence for astrocyte-associated transcriptional alterations in SCZ, particularly within these specific cortical layers, where such changes may relate to circuit-level disruptions in these domains.^15,62,64 Oligodendrocyte-specific reference DEGs exhibited strong concordance among down-regulated SCZ-DEGs in deeper cortical layers including SpD05-L5 and SpD03-L6, consistent with our previous observation of oligodendrocyte marker genes enriched among layer-restricted SCZ-DEGs in these domains (Figure 3C). While our earlier enrichment analyses demonstrated limited enrichment of neuronal marker genes among SCZ-DEGs (Figure 3A, Figure 3C), registration of reference DEGs revealed transcriptional alterations across several excitatory and inhibitory neuronal subtypes, such as deep-layer excitatory neuron-specific DEGs (Ex-L6b_SEMA3D) overlapping in SpD07-L1/M. However, the magnitude of neuronal changes was generally less pronounced than those observed in glial populations. Indeed, our dataset was more sensitive to glial- and WM-associated alterations than the snRNA-seq dataset,¹⁶ highlighting the complementary strength of SRT in capturing non-neuronal transcriptional changes in situ (Figure S18). Together, these findings indicate that spatially-localized gene expression changes encompass a biologically meaningful spectrum of SCZ-associated alterations, with glial-linked changes being more salient in our SRT dataset.

To further interrogate these glial effects and infer how our layer-restricted SCZ-DEGs related to transcriptional alterations within spatially-localized cellular compartments or synaptic microenvironments, we assessed enrichment of marker genes for the four SPG spot types (neuropil, PNN, neuronal, vasculature) among our SCZ-DEGs (Figure 3E). Broadly, these microenvironment marker genes mapped onto layer-restricted SCZ-DEGs across multiple SpDs, revealing heterogenous laminar patterns. Neuropil marker genes were enriched among both up- and down-regulated SCZ-DEGs, with up-regulated SCZ-DEGs localized to mid-to-deep neuropil (SpD02-L3/4 to SpD01-WMtz) and down-regulated SCZ-DEGs to superficial-to-middle layer neuropil (SpD07-L1/M to SpD05-L5), suggesting divergent synaptic dysregulation across neuronal and glial processes. PNN marker genes were enriched among down-regulated SCZ-DEGs in middle cortical layers, particularly SpD06-L3/4, where we previously observed astrocytic alterations (Figure 3C, Figure 3D). Given that PVALB neurons are abundant in these domains and preferentially ensheathed by PNNs,^27,30,65,66 these findings may indicate inhibitory signaling disruptions downstream of PNN regulation of PVALB neurons, potentially intersecting with glial alterations detected in these domains, as suggested by a recent mouse study showing that astrocyte-PNN interactions support synaptic homeostasis.⁶⁷

Non-intuitively, marker genes for neuronal spots were enriched among SCZ-DEGs in glial-rich domains (SpD07-L1/M, SpD01-WMtz, SpD04-WM), possibly implicating aberrant neuron-glia interactions that may be linked to extranuclear transcripts within neuronal processes or sparse neuronal populations interspersed throughout these domains.^68–72 Vascular marker genes were enriched among up-regulated SCZ-DEGs in SpD03-L6 and SpD04-WM and among down-regulated SCZ-DEGs in SpD07-L1/M and SpD01-WMtz, which may signify domain-specific neurovascular alterations. Overall, layer-restricted SCZ-DEGs were largely associated with neuropil and PNN microenvironments, a finding further supported by enrichment of BrainSeq Phase 2 bulk RNA-seq DEGs⁴¹ in these domains (Figure S19). In summary, these data highlight the utility of SRT to capture subtle alterations in specialized cellular and extracellular compartments relevant to synaptic function and point towards potential dysregulation of neuron to non-neuronal interactions across microenvironment domains in SCZ.

2.4 |. Convergence of genetic risk and transcriptional regulation with SCZ-associated spatio-molecular differences

To explore regulatory elements of spatially-localized SCZ-associated gene expression changes across cortical layers, we utilized a transcription factor (TF) prediction algorithm ChEA3⁷³ to predict candidate TFs that may trans-regulate the observed layer-restricted SCZ-DEGs (Figure 4A, Supplementary Table 6, Methods). We tested enrichment of TF target gene sets within the up- and down-regulated layer-restricted SCZ-DEGs for each SpD, respectively. Candidate TFs were ranked based on their aggregated enrichment scores, where the top 10 TFs per SpD were displayed across SpDs. Many TFs recurred across multiple SpDs (e.g., NACC2, FOSB among up-regulated SCZ-DEGs; OLIG1, SOX10 among down-regulated SCZ-DEGs), suggesting their cross-laminar involvement in transcriptional dysregulation. We also observed some TFs concurrently appearing within both up- and down-regulated SCZ-DEGs. These discordant patterns likely indicate their context-dependent regulation influenced by the local cell composition and conditions. In addition, we identified TFs that are involved in cellular differentiation and maturation, including neuronal lineages (ATOH8, TSHZ1), synaptic plasticity and neurite remodeling (NR4A1, KLF9), neurovascular development (SOX18). We also identified TFs exhibiting localized associations with specific SpDs. For instance, ZNF324B showed strong relevance to up-regulated SCZ-DEGs in SpD01-WMtz, while CAMTA1, PEG3, ZNF483, PURG, and CSRNP3 demonstrated relative specificity in SpD07-L1/M. Additionally, within SpD04-WM, TFs including SGSM2, FLYWCH1, TSHZ1, and TEF showed preferential associations with down-regulated SCZ-DEGs, implicating their layer-restricted regulatory involvement in downstream targets within this domain.

Figure 4. Convergence of transcriptional regulation and genetic risk on spatio-molecular differences in SCZ.

(A) Heatmaps showing the top ranked 10 TFs predicted by ChEA3,⁷³ per SpD, associated with layer-restricted SCZ-DEGs up-regulated (top) and down-regulated (bottom) in SCZ. Due to overlapping TFs across SpDs, the initial total of 70 TFs in each group was reduced to 42 unique TFs (up-regulated) and 41 unique TFs (down-regulated).

(B) STRING TF functional networks for non-neuronal-rich (SpD07-L1/M, SpD01-WMtz, SpD04-WM) and neuronal-rich (SpD06-L2/3, SpD02-L3/4, SpD05-L5, SpD03-L6) domains. The top 10 TFs linked to up- and down-regulated layer-restricted SCZ-DEGs per SpD were selected, resulting in initial input sets of 60 TFs (non-neuronal) and 80 TFs (neuronal). After removing overlapping TFs across SpDs and unmapped entries (e.g., NKX62, NKX22), 48 TFs (non-neuronal) and 42 TFs (neuronal) remained. Networks display only connected nodes with an interaction score > 0.7. Complementary network diagrams including all TFs are shown in Figure S20 at an interaction score > 0.4 to capture broader connectivity.

(C) Heatmap of ChEA3-predicted TFs associated with layer-restricted SCZ-DEGs, intersecting with prioritized SCZ GWAS risk genes from Trubetskoy et al.¹³ across PRECAST SpDs.

(D) Box plot comparing predicted SCZ PRSs shows a higher mean PRS in the SCZ group (n=32 donors), compared to the NTC group (n=31 donors); p=1.507 × 10⁻⁴.

(E) Volcano plot of layer-adjusted PRS-associated DEGs (PRS-DEGs). The x-axis is the logFC and the y-axis is the −log₁₀ (transformed p). Points in darker shades indicate PRS-DEGs at FDR < 0.10. Points in lighter shades indicate PRS-DEGs at nominal p < 0.05.

(F) Scatter plot comparing t-statistics between PRS-DEGs and layer-adjusted SCZ-DEGs (Dx-DEGs; Dx=diagnosis)

(G) Spot plots illustrating the distribution of SCZ disease relevance normalized z-scores in representative NTC and SCZ donors, as estimated using scDRS. Each point represents a Visium spot, shaded according to SCZ risk z-score (white to black). SpD boundaries are outlined and color-coded to indicate PRECAST SpDs at k=7.

(H) Heatmap showing mean SCZ scDRS z-scores across SpDs in NTC and SCZ groups. The color scale reflects relative enrichment z-scores. Asterisks indicate SpDs with significantly increased SCZ scDRS z-scores relative to the NTC group (FDR < 0.05).

Given the substantial number of candidate TFs and their heterogenous patterns across SpDs (Figure 4A), we used STRING database,⁷⁴ which integrates functional and physical protein associations, to construct full interaction networks of TFs and characterize their overarching biological themes. Based on the domain-specific enrichment pattern of TFs in non-neuronal-rich cortical domains (Figure 4A) as well as the robust mapping of SCZ-DEGs to these regions (Figure 3A–C), we stratified ChEA3-derived TFs (top 10 ranked per SpD for each up- and down-regulated layer-restricted SCZ-DEGs, Methods) into two groups: non-neuronal-rich domains (SpD07-L1/M, SpD01-WMtz, SpD04-WM) and neuronal-rich domains (SpD06-L2/3, SpD02-L3/4, SpD05-L5, SpD03-L6) (Figure 4B, Figure S20). Within non-neuronal-rich domains, TF networks prominently featured factors associated with immune and microglial activation (SPI1, IRF8, FLI1). In neuronal-rich domains, we observed a similar network associated with stress response and inflammatory signaling (SPI1, FLI1, TFEC), along with additional networks implicated in oligodendrocyte differentiation and myelination (MYRF, OLIG1, SOX10) and immediate early genes relevant to neuronal activity and plasticity (NR4A1, FOSB, ATF3). Together, these TF networks highlight converging biological themes: glial contributions and disruptions in neuronal plasticity.

To integrate transcriptional regulation with genetic risk, we examined whether candidate TFs intersected with known genetic factors implicated in SCZ. By cross-referencing candidate TFs from ChEA3⁷³ database with previously prioritized SCZ-associated GWAS risk genes from Trubetskoy et al.,¹³ we identified and visualized 12 overlapping TFs with their enrichment ranks (Figure 4C, Methods). Among these, two inflammation-related TFs, KLF6 and IRF3, were consistently ranked among the top 50% of TFs across all SpDs, frequently within the top 20–35%, across both up- and down-regulated SCZ-DEGs. Domain-specific GWAS-prioritized TFs were also noted, although their rankings were modest, including ZNF823 (linked to down-regulated SCZ-DEGs in SpD04-WM, top 17%) and CREB3L4 (linked to down-regulated SCZ-DEGs in SpD05-L5, top 36%). ZNF804A, an established SCZ risk gene,^75–77 showed target gene modules overlapping with down-regulated SCZ-DEGs in neuronal-rich SpD06-L2/3 (top 38%) and with up-regulated SCZ-DEGs within non-neuronal-rich domains (SpD07-L1/M, SpD01-WMtz; top 38% and 13%, respectively). Many SCZ-DEGs linked to ZNF804A were involved in neuronal maturation (NR4A2, EGR1, DCC), neurotransmission (GRIN1, CACNA2D3, GABRB3, SCN2B, SCN8A), and synaptic plasticity and organization (BDNF, CNTN1, CSMD3), supporting a convergent genetic-transcriptional disruption of neuronal circuitry and plasticity in these cortical regions (Supplementary Table 6, Figure S21).

To quantitatively assess the impact of genetic risk on gene expression, we computed polygenic risk scores (PRSs) for SCZ (Methods).⁷⁸ As expected, PRS was significantly higher in the SCZ group compared to controls (p=1.507 × 10⁻⁴) (Figure 4D). Using the layer-adjusted model, we replaced diagnosis as the predictor with PRS. This analysis identified 354 PRS-associated DEGs (PRS-DEGs) controlling FDR < 0.10 (1,588 PRS-DEGs at nominal p < 0.05, Figure 4E, Supplementary Table 7, Methods). PRS-DEGs included genes involved in protein translation (e.g., EIF2A), mitochondrial respiration (COX7A2), apoptosis (MCL1), microglial function (AIF1, TREM2), vascular and immune processes (VEGFA, A2M), activity-related signaling (FOSB, JUN, SYT1), and neuronal structure and outgrowth (MAP2, GAP43). Several PRS-DEGs mapped to SCZ risk loci, including MAPK3, MSI2, R3HDM2, THOC7 and KANSL1-AS1.¹³ PRS-DEGs were moderately correlated with diagnosis-associated DEGs (Pearson’s correlation=0.555, Figure 4F) with 55 genes shared between the two sets, many of which were linked to microglial- and immune-related functions such as AIF1, TREM2, C1QA, C1QB, C3, CX3CR1, TYROBP, and CD74. GSEA on these PRS-DEGs highlighted pathways related to metabolism, gliogenesis, immune regulation, and vascular-associated processes (Figure S22), suggesting non-neuronal contributions arising from both PRS- and diagnosis-associated transcriptional alterations. Together, these results provide evidence that PRS-DEGs link genetic risk for SCZ to transcriptional dysregulation that extends beyond neuronal processes, potentially.⁶³

To further investigate how genetic risk is spatially distributed across cortical architecture, we applied the single-cell disease relevance score (scDRS) framework⁷⁹ to predict SCZ genetic risk at the level of individual Visium spots (Figure 4G, Figure S23, Methods). Comparing across SpDs, predicted genetic risk was largely localized to neuronal-rich domains (L2–6), where mean scDRS z-scores were positive, as further supported by enrichment of neuronal-rich domain marker genes among SCZ GWAS loci using multi-marker analysis of genomic annotation (MAGMA)⁸⁰ (Figure S24). In contrast, non-neuronal-rich domains (L1/M, WMtz, WM) showed overall negative mean z-scores across diagnosis. We also observed subtle case-control differences in the magnitude and distribution of neuronal genetic risk, with SCZ samples showing slightly higher mean scDRS z-scores (FDR-adjusted p < 0.05). Overall, these results indicate a laminar dichotomy between the neuronal localization of genetic susceptibility and the non-neuronal transcriptional alterations (Figure 2E), suggesting the potential of functional interplay between neuronal- and non-neuronal-rich cortical domains in SCZ.

2.5 |. Mapping SCZ-associated gene expression changes to spatially-localized cell types

To directly measure SCZ-associated changes in spatial gene expression at the cellular level, we used the Xenium platform to generate single-cell resolution SRT data in dlPFC from a subset of donors used for Visium (n=12 NTC donors, n=12 SCZ donors) (Figure 5A, Supplementary Table 1, Methods). Compared to the spot-based Visium platform, Xenium provides higher resolution using a pre-defined set of target genes to investigate cell-type composition and signaling across groups. Compared to snRNA-seq, Xenium does not require tissue dissociation or cell sorting and can capture cell types that are often underrepresented in snRNA-seq studies, including microglia.^81,82 This feature is particularly powerful in the context of SCZ, as transcriptional changes across these underrepresented cell types have been reported.^15,83,84 Previous studies showed changes in the cell-type abundance between SCZ and NTC,^18,85,86 while others reported dysregulation at the level of gene expression.^15,16,87,88 To directly investigate whether the transcriptional changes we observed reflect intrinsic dysregulation or differences in cell-type abundance across cortical layers, we designed a Xenium gene panel with 300 genes, consisting of both cell-type-specific gene markers and the SCZ-DEGs identified in the layer-adjusted model (Figure 2A, Supplementary Table 8). After QC, 1,262,965 high-quality cells were retained (NTC: 629,981; SCZ: 632,984) (Figure S25, Figure S26).

Figure 5. Cell-type-specific gene expression changes linked to SCZ in the human dlPFC.

(A) Schematic showing the cohort selection and data generation process for the Xenium data. Notably, we highlight the single-cell resolution of the Xenium data, as opposed to the spot-level resolution of the Visium data. A table showing the 300 genes selected for the custom Xenium panel is shown, providing the number of genes in each category. Created in BioRender. Kwon, S. H. (2026) https://BioRender.com/9djvx8j

(B) Cell-resolution plots of two representative tissue sections, Br8667 and Br5973 (NTC and SCZ, respectively), with cells colored by the predicted SpD label from spaTransfer after spatial smoothing. Predicted labels recapitulate the cortical layer structure of the dlPFC.

(C) Same as (B), but with each cell colored by cell types. Cell types were annotated by performing data-driven clustering with BANKSY and then annotating clusters based on expression of the 36 cell-type markers in the gene panel.

(D) Box plot of cell-type composition changes between NTC and SCZ, performed across all cortical layers. No significant changes in composition were detected between the diagnostic groups in any cell type.

(E) Scatter plot showing comparison of logFC between the Visium layer-adjusted DE analysis and the Xenium layer-adjusted DE analysis, for each of the 168 Visium layer-adjusted SCZ-DEGs.

(F) Volcano plot showing genes that are DEGs between NTC and SCZ, while adjusted for cortical layers. 19 genes are identified as differentially expressed at FDR < 0.10.

(G) Same as (F), but with cell-type adjustment rather than layer adjustment. 15 genes are identified as differentially expressed at FDR < 0.10.

(H) Volcano plots showing cell-type-specific SCZ-DEGs. At FDR < 0.10, CALB2 was up-regulated while BDNF and ADCYAP1 were down-regulated in L4/5 Ex neurons, and KLF2 and ABCG2 were down-regulated in microglia.

We used spaTransfer,⁸⁹ a label transfer method based on non-negative matrix factorization, with the Visium dataset as a reference to perform data-driven annotation of SpDs in the Xenium data. Specifically, each cell was annotated with a predicted cortical layer label identified using the Visium spot-level SpDs from PRECAST. Finally, a spatial smoothing algorithm was applied to increase spatial contiguity of predicted domains (Methods). The predicted domains recapitulated the classic histological cortical layers (Figure 5B). We found high correlation between our predicted Xenium cell-level SpDs and previous manually annotated dlPFC layers,²⁷ confirming the SpD annotation of the Xenium dataset (Figure S28). Similar to the spot-level SpDs, the predicted cell-level SpDs accounted for the majority of variation across samples in the first two principal components (PC1 and PC2) (Figure S27), which suggested that the transferred annotations were not biased by diagnosis across the top 10 PCs (cumulative variance explained 46%).

To further characterize cell-type composition across the dlPFC domains, we used unsupervised spatial clustering with BANKSY⁹⁰ and identified k=18 clusters, which were annotated based on expression of cell-type marker genes included in the Xenium gene panel (Figure S29, Supplementary Table 8). In total, we annotated 12 clusters with cell-type labels; four of which were excitatory (L2/3, L4/5, L5, and L6 ‘Ex’ neurons), two were inhibitory (one ‘In’ neuron cluster consisting of a mixture of VIP- and LAMP5-expressing cells, and the other consisting of a mixture of SST- and PVALB-expressing cells), one cluster each for astrocytes, microglia, oligodendrocytes, and endothelial cells, and two ambiguous clusters (one expressing primarily oligodendrocyte marker genes and the other expressing inhibitory neuron and endothelial cell marker genes) (Figure 5C). Due to the limited number of genes on a Xenium panel, BANKSY clustering was not able to separate the excitatory and inhibitory neuron subtypes at the k=18 resolution, and we annotated the two inhibitory neuron clusters as medial and caudal ganglionic eminence clusters, respectively (MGE for mixed SST/PVALB populations, CGE for VIP/LAMP5).^91,92 To validate the cell-type annotation, we next examined the relationship between the SpDs and cell types by computing the proportion of each cell type within the SpDs (Figure S30). For each cell type, proportions were similar between NTC and SCZ. Cell types were also enriched in the expected cell-level SpDs; for example, oligodendrocytes were predominantly enriched in the WM domain, and L2/3 Ex neurons were primarily enriched in the L2/3 domain.

Previous studies suggested alterations in cellular composition across cortical layers, particularly highlighting vulnerability in specific inhibitory neuronal populations in SCZ.^18,93,94 To investigate this, we used the cacoa R package⁹⁵ based on compositional data analysis techniques to quantify cell-type composition differences between NTC and SCZ samples following two strategies: 1) globally considering all cell-level SpDs altogether, and 2) within each individual cell-level SpDs. We found no evidence of global cell-type composition changes (Figure 5D, Figure S30). To investigate whether specific cell types differ in abundance, and which cell-level SpDs these differences occur in, we examined the proportion of each of the 12 cell types within each of the cell-level SpDs. This analysis found no evidence for layer-specific differences in cell-type composition between NTC and SCZ, except in L2/3 where there was a significant enrichment for L6 Ex neurons in SCZ compared to NTC (p=0.002). However, L6 Ex neurons represent a trivial proportion of cells in this domain (1.68% to 4.26%) (Figure S31) and hence this finding needs to be interpreted with caution. Notably, we did not observe a compositional difference in microglia between NTC and SCZ in any of the layers. Gene expression down-regulation has previously been observed in microglia,^17,46 and taken together with our results, which showed down-regulation of microglial genes (Figure 2A, Figure 2E, Figure 3A, Figure 3C), this suggests microglia are likely undergoing intrinsic expression changes in SCZ.⁸³ Overall, our compositional analysis did not provide strong evidence for substantial changes in cell-type abundance between NTC and SCZ, supporting the interpretation that previously observed SCZ-associated gene expression changes reflect intrinsic transcriptional changes.

To identify transcriptional changes associated with SCZ to validate our Visium results, we first performed layer-adjusted DE analysis on the Xenium data by quantifying expression changes associated with SCZ while controlling for laminar biology, as well as a set of technical covariates (Methods). Considering only effect sizes and not statistical significance, the majority (n=124 genes; 73.8%) of the 168 layer-adjusted SCZ-DEGs found in the Visium analysis showed consistent directionality of gene expression change in the Xenium dataset (Figure 5E, Pearson’s correlation=0.502). Specifically, 72 genes were down-regulated in both analyses and 52 genes were up-regulated in both analyses. Furthermore, in the Xenium analysis, we identified a total of 19 statistically significant layer-adjusted SCZ-DEGs with FDR < 0.10 (Figure 5F). 13 of the 19 genes were down-regulated in SCZ, again showing concordance with previous bulk and snRNA-seq studies.^16,41 We confirmed significant differential expression of 16 of the genes identified in the Visium data. These include FKBP5, NPTX2, IFITM3, SERPINA3, CHI3L1, and TNFSF10, all of which are associated with neuroplasticity and neuroinflammation. Of the 16 genes that were also found in the Visium layer-adjusted DE analysis, 15 genes showed the same direction of expression change. The only discrepancy in directionality was IFITM3, which was up-regulated in the Visium data, but down-regulated in the Xenium data. The three additional genes identified in the Xenium analysis were KLF2, SLC44A1, and MBP. These genes are primarily expressed in microglia (KLF2)⁹⁶ and oligodendrocytes (SLC44A1, MBP),²⁵ providing orthogonal evidence for cell-type-specific transcriptional alterations across non-neuronal cell types in SCZ, in concordance with our earlier findings (Figure 3A–Figure 3D). Taken together, the results from both the Visium and Xenium layer-adjusted analyses consistently identified genes associated with neuroinflammation and plasticity, providing additional evidence for disruption in these pathways in SCZ.

Given that the layer-adjusted analysis allowed us to identify SCZ-DEGs while controlling for differences in cortical layer composition between donors, we hypothesized that controlling for cell-type composition could yield additional insight into transcriptional changes. To this end, we next performed cell-type-adjusted DE analysis while also adjusting for a set of technical covariates (Methods). This analysis resulted in 15 significant DEGs at FDR < 0.10 (Figure 5G). Up-regulation of FKBP5, SERPINA3, IFITM3, GFAP, and GSTM5 suggest stress and immune-related dysregulation in SCZ. In concordance with previous results, we observed down-regulation of mitochondrial-associated genes such as UQCRFS1 and COX5A, consistent with oxidative stress reported in SCZ.⁹⁷ We also observed down-regulation of NPTX2, VGF, and ARPP19, suggesting impairment in synaptic plasticity.^98–100 Interestingly, down-regulation of the immediate early gene NPTX2 has been implicated in SCZ.¹⁰¹ Up-regulation of the myelin-related genes MBP and BCAS1 were also observed in SCZ in the cell-type-adjusted analysis (Figure 5G), supporting previous findings of myelin dysfunction in SCZ.^94,102 Taken together, the cell-type-adjusted DE model revealed intrinsic transcriptional changes in SCZ, independent of cell-type composition.

Next, we leveraged the single-cell resolution of the Xenium data to identify transcriptional changes occurring in discrete cell populations while adjusting for a set of technical covariates (Methods). Using this “cell-type-specific” DE model, we found significant gene expression changes in microglia and L4/5 Ex neurons at FDR < 0.10 (Figure 5H). In microglia, we observed significant down-regulation of KLF2 and ABCG2, genes involved in cellular stress responses, consistent with the overall reduction of microglial-associated gene expression reported in previous studies.^17,46 Interestingly, KLF2 was also down-regulated in the Xenium layer-adjusted DE analysis, highlighting the ability of the cell-type-specific model to pinpoint the specificity of gene expression changes and map them to cellular populations. In L4/5 Ex neurons, we observed up-regulation of CALB2 and down-regulation of BDNF and ADCYAP1 in SCZ. Given that CALB2 is associated with neuronal excitability, and BDNF/ADCYAP1 have roles in synaptic strengthening and dendritic remodeling, ¹⁰³ this pattern may reflect altered synaptic plasticity in deep layer neurons in SCZ. Similar to the observations for microglia, we found no evidence for compositional differences in L4/5 Ex neurons between NTC and SCZ. Taken together, these findings highlight transcriptional dysregulation in cellular stress and neuroplasticity within discrete cellular populations, rather than alterations in cell-type abundance.

Finally, we profiled expression patterns of some SCZ risk genes compiled from GWAS and other literature (Supplementary Table 8, Column G), including GRIN2A, COMT, DISC1, SV2A, and C4A.^13,104–107 Although these genes did not show significant differences across the three Xenium DE models described above (Supplementary Table 9), several genes were expressed in specific cell types. ZNF804A, a TF previously linked to SCZ,^13,77 was enriched in the MGE inhibitory neuron populations despite no significant diagnostic-group differences (Figure S32). Based on our earlier TF analysis (Figure 4D, Figure S21), we identified ZNF804A as a possible TF linked to the layer-restricted SCZ-DEGs in the Visium SpDs (SpD07-L1/M, SpD06-L2/3, SpD01-WMtz). Given the localization of these neurons, this suggests that the dysregulation of ZNF804A in MGE inhibitory neurons, which are endogenous to these layers, could drive the observed layer-restricted SCZ-DEGs, contributing to a potential cell-driven mechanism for the disruption in transcriptional changes occurring in SCZ.

2.6 |. SCZ-associated differences in gene expression across spatially-organized cellular microenvironments

Given the intrinsic transcriptional alterations observed across cortical layers and the strong enrichment of SCZ-DEGs within SPG-defined microenvironments, we next examined how these changes were represented across neuropil, PNN, and neurovascular domains, motivating microenvironment-level DE analyses. To directly investigate DE changes within each microenvironment (Figure 6A, Figure 6B, Supplementary Table 10, Methods), we pseudobulked neuropil, PNN, and neuronal spots within individual gray-matter SpDs (e.g., L1-L6). Given the broad distribution of blood vessels in both WM and gray matter, we pseudobulked vasculature spots across all SpDs (e.g., L1-L6 and WM). For each SPG spot classification, we evaluated diagnostic-group differences within each microenvironment while controlling for laminar differences and covariates, thereby defining microenvironment-restricted SCZ-DEGs. At FDR < 0.10, we identified 352, 0, 90, and 19 significant SCZ-DEGs restricted to neuropil, PNN, neuronal, and vasculature microenvironments, respectively. Neuropil had the highest number of significant DEGs, indicating a relatively greater load of detectable SCZ-associated transcriptional changes in this microenvironment. To capture a broader range of candidate transcriptional changes and enhance sensitivity, we additionally examined genes at nominal p < 0.05. Using this threshold, we identified additional SCZ-DEGs: neuropil (n=1,669 genes), PNN (n=559 genes), neuronal (n=1,789 genes), and vasculature (n=440 genes). The neuropil and neuronal microenvironments still exhibited the most transcriptional changes (Figure S36), hinting at more nuanced SCZ-associated vulnerability within these local niches that are enriched for neuronal components.^{13,16,53–55} This approach identified several SCZ-DEGs including NRXN2 in neuropil, and SST and MAPK3 in the neuronal microenvironments (Figure 6B). Additional SCZ-DEGs and their microenvironment-restricted gene expression patterns included GABRA1 for PNNs and A2M for vasculature.

Figure 6. SCZ-linked differential gene expression changes across SPG-defined microenvironments in the human dlPFC.

(A) Volcano plots showing microenvironment-restricted SCZ-DEGs identified in four SPG-defined microenvironments. SCZ-DEGs were identified at two significance thresholds: stringent (FDR < 0.10; red=up-regulated, blue=down-regulated; neuropil: 352, PNN: 0, neuronal: 90, vasculature: 19) and nominal (p < 0.05, horizontal solid line, semi-transparent red=up, semi-transparent blue=down; neuropil: 1,669, PNN: 559, neuronal: 1,789, vasculature: 440). Non-significant genes are shown in gray. Vertical dashed lines indicate reference lines at |logFC|=0.14. Representative genes are highlighted.

(B) Representative microenvironment-restricted SCZ-DEGs are highlighted with their Visium spots, paired with dot plots showing their expression changes in SCZ: (i) NRXN2 (neuropil), (ii) GABRA1 (PNN), (iii) SST (neuronal), (iv) A2M (vasculature). Each data point represents pseudobulked gene counts from an SPG-defined microenvironment within each SpD of each donor (donor-SpD-microenvironment; blue for NTC and red for SCZ group). Spot diameter: 55 μm, Scale: logcounts. Interactive plots are available for viewing and download in both the iSEE app and Samui browser (Data Availability).

(C) SynGO enrichment analyses of neuropil and PNN microenvironment-restricted SCZ-DEGs identified at nominal p < 0.05. Cellular Component (CC) terms significantly enriched at Q-value < 0.01 are shown, with colors indicating enrichment significance (−log₁₀ (Q-value)) and representative genes labeled. For PNNs, selected terms significant at relaxed threshold (Q-value < 0.05) were labeled manually to indicate significant enrichment of transcripts for postsynaptic elements. Additional supporting SynGO results for PNN using a default ‘brain-expressed’ background are provided in Figure S33.

(D) GO-ORA of neuronal and vasculature microenvironment-restricted SCZ-DEGs identified at nominal p < 0.05. BP terms were tested against matched microenvironment-specific background sets, defined by all genes expressed in the respective microenvironments. Treemaps depict the top 50 significantly enriched GO-BP terms (FDR ≤ 0.05) after semantic similarity reduction (threshold=0.7). An additional neuron-focused treemap for neuronal SCZ-DEGs, generated by selectively filtering neuron-related GO-BP terms, is provided in Figure S34.

(E) STRING-based protein-protein interaction (PPI) networks of microenvironment-restricted SCZ-DEGs were generated for each SPG-defined microenvironment (neuropil, PNN, neuronal, vasculature). Within each PPI network, functional clusters were identified using the built-in MCL algorithm (inflation=1.2). For each microenvironment, one representative cluster is shown here as an example of a microenvironment-restricted signaling subnetwork where a single representative gene within that subnetwork (BDNF, GABRA1, GRIN1, or TF) is highlighted with a colored box. Blue indicates down-regulation and red indicates up-regulation of gene expression in SCZ.

(F) Sankey representation of candidate SCZ-linked LR interactions from Huuki-Myers et al.²⁵ that match microenvironment-restricted SCZ-DEGs identified at nominal p < 0.05. LR pairs are shown when either the ligand or receptor intersects with one of SCZ-DEGs, with blue indicating down-regulation and red indicating up-regulation of gene expression in SCZ. The partner genes shown in gray were not significantly differentially expressed within the microenvironments in SCZ.

To infer broad functional implications of these SCZ-DEGs, we combined up- and down-regulated genes within each microenvironment and performed gene ontology (GO) enrichment analyses (Figure 6C, Methods). Given the relevance to synaptic and ECM functions, SynGO was applied to neuropil and PNN microenvironment-restricted SCZ-DEGs, while general GO-based over-representation analysis (ORA) was applied to neuronal and vasculature microenvironment-restricted SCZ-DEGs. Neuropil SCZ-DEGs were significantly enriched for presynaptic compartments and ribosomal machinery, suggesting altered local synaptic translation and plasticity.^108–110 PNN SCZ-DEGs (e.g., GABRG2, NTRK2) were enriched for synaptic components, with stronger enrichment in postsynaptic compartments, which may preferentially affect stabilization of synaptic inputs onto PVALB neurons surrounded by PNNs (Figure 6C, Figure S33).^111–114 Neuronal SCZ-DEGs featured pathways associated with mitochondrial metabolism, cellular respiration, and energy production (Figure 6C). Enrichment of neuron-specific processes included neuronal projection, calcium ion-mediated neurotransmission, and neuronal apoptotic processes (Figure S34). Vasculature SCZ-DEGs showed enrichment in metabolic pathways, metal-ion homeostasis, and immune functions, with some enrichment for amyloid precursor protein metabolism (Figure 6C).

To complement these functional annotations and identify potential signaling pathways among SCZ-DEGs, we performed STRING network analysis restricted to physical protein-protein interactions, and clustered gene modules into subnetworks using the built-in Markov Cluster (MCL) algorithm (Methods, Figure 6E, Supplementary Table 11). Across all microenvironments, more than 70% of genes within the top five MCL subnetworks converged on mitochondrial metabolism, RNA processing, and protein homeostasis, with SCZ-DEGs repeatedly shared across microenvironments and enriched for housekeeping pathways such as ATP metabolism and oxidative phosphorylation (Figure S35, Figure S36, Figure S37, Supplementary Table 11). Additionally, we identified subnetworks that recapitulate specialized biological functions of each microenvironment (Figure S38, Supplementary Table 11). Neuropil subnetworks highlighted genes associated with neuronal functions and signaling, including mTOR signaling pathways linked to BDNF-TrkB signaling (Figure 6E, MCL cluster 6), synaptic vesicle trafficking (MCL cluster 5), GABAergic signaling (MCL cluster 11), and complement-mediated neuroimmune responses (MCL cluster 8). Neuronal subnetworks also emphasized neuronal communication, particularly calcium-mediated signaling (Figure 6E, MCL cluster 5) and complement activation (MCL cluster 44). PNN subnetworks were marked by representation of GABAergic signaling (Figure 6E, MCL cluster 3) with multiple GABA receptor subunits (GABRA1/3/5, GABRG2). Vasculature subnetworks showed enrichment of SCZ-DEGs not only for vascular growth (MCL cluster 13) and NMDA receptor signaling (MCL cluster 10) but also for RNA processing linked to immune presentation and iron regulation (Figure 6E, MCL cluster 4). Collectively, these analyses organized SCZ-DEGs into physically anchored, functionally relevant subnetworks, suggesting shared metabolic perturbations and microenvironment-restricted signaling changes in SCZ.

Building on our pathway-level analyses, we next examined how SCZ-associated transcriptional changes relate to cell-cell communication (CCC). We leveraged a curated set of 90 candidate ligand-receptor (LR) pairs from Huuki-Myers et al.,²⁵ which integrated dlPFC snRNA-seq CCC analysis with SCZ-linked LR annotations. We mapped these interactions to contextualize how transcriptional dysregulation may coordinate altered intercellular communication within microenvironments. Among the 90 LR pairs, we identified 13, 17, 6, and 26 pairs in which at least one interactor matched a microenvironment-restricted SCZ-DEG in neuropil, neuronal, PNN, and vasculature microenvironments, respectively (Figure 6F, Figure S39, Figure S40, Methods). Neuropil SCZ-DEG-matched LR pairs included those associated with axon-guidance and growth factor signaling (NTN4-DCC, EFNA5-EPHB1, SEMA6B-PLXNA4, FGF-FGFR1). Neuronal SCZ-DEG-matched LR pairs included growth- and metabolism-linked signaling (PTN-PTPR, APOD-LEPR), as well as adhesion complexes (EFNA5-EPHA5/7, NXPH2-NRXN1). PNN SCZ-DEG-matched LR pairs were characterized by ECM-, glial-, and vascular-associated interactions, such as ANGPTL4-CDH5/11 and THBS2-NOTCH4. Vasculature SCZ-DEG-matched LR pairs involved trophic and immune-related interactions, including FGF-FGFR1 and PTN axes, along with adhesion pairs such as EFNA5-EPHA4 and CNTN4-APLP1. Overall, we observed recurrent trophic axes (e.g., PTN-PTPR and FGF-FGFR1) across specialized microenvironments, suggesting shared upstream signaling components that may couple neuronal, glial, and vascular responses as part of altered neuron to non-neuronal communication.^115,116 Taken together, integrating SCZ-DEGs with spatially-resolved LR interactions provides a refined view of gene expression changes in dlPFC, outlining candidate signaling mechanisms that may contribute to intercellular dysregulation within microenvironment domains in SCZ.

3 |. Discussion

Here we integrated Visium SRT and Xenium in situ profiling to assemble a spatial multi-omics atlas of the human dlPFC in SCZ. Demonstrating the utility of spatial profiling, we localized SCZ-associated transcriptional alterations across discrete cortical layers and cell types. Joint clustering across donors generated data-driven SpDs that recapitulated classic cortical cytoarchitecture and captured fine molecularly-defined features, including the superficial L1/M domain and the gray-white matter transition zone (WMtz).^25,117 By anchoring DE analyses to the SpDs, we spatially localized disease-associated DEGs (SCZ-DEGs), finding prominent enrichment in glia-vascular dominant layers (L1/M, WM, and WMtz). These findings strengthen growing evidence for non-neuronal contributions to SCZ and highlight the vulnerability of glial-rich domains,^63,118,119 which may have been underappreciated in previous SCZ snRNA-seq studies due to dissociation artifacts, loss of fragile structures (L1/M), limited cytoplasmic transcript capture, or sorting bias.^{53,82,120–126} Importantly, our Xenium data showed that abundance of neuronal and glial populations was similar across diagnosis, supporting the interpretation that the laminar- or bulk-level gene expression changes arise primarily from intrinsic cellular alterations rather than shifts in cell-type composition.^15,16,87,88 Our cell-level analysis further refined these broader gene expression changes by resolving their cellular contexts, for example, altered ABCG2 expression within microglia, illustrating how cell-level profiling complements the laminar-level observations. Together, the Visium and Xenium data present an integrated view of SCZ-associated alterations linking laminar-level pathology to single-cell perturbations.

Consistent with prior studies that implicate neuronal pathology in SCZ, we did observe discernible transcriptional alterations in neuronal-rich gray-matter domains, particularly L2/3.^16,18,87 However, direct comparison to published datasets¹⁶ indicated limited convergence across neuronal cell types. Intriguingly, we detected deep-layer (L6) excitatory neuronal signatures within superficial layers (L1–3), suggesting the presence of deep-layer-like molecular signatures outside their expected cortical layers. This pattern may indicate capture of transcripts derived from distal projections,^{27,71,127–129} or more nuanced disruptions in neuronal positioning or maturation.^130–132 Indeed, our laminar-level data captured both nuclear and cytoplasmic transcripts in the compartments beyond cell bodies, featuring neuropil, the critical interface for neuron-glia interactions, as a key microenvironment harboring pronounced gene expression changes. These findings highlight the potential importance of transcripts localized to these cellular compartments in regulating neuron-glial interactions in SCZ.^{15, 27,71,127–129} Such local transcripts may modulate neuron-glia signaling in SCZ and contribute to the propagation of intrinsic cellular changes across nearby neuronal and non-neuronal elements into broader circuit-level changes.

To interrogate how intrinsic transcriptional alterations may be relayed across neighboring cells and microenvironments,^133–137 we leveraged the IF staining that was paired with Visium SRT (e.g., Visium-SPG). We identified neuropil and neuronal microenvironments as major sites of SCZ-associated alterations, with additional changes detected in PNN- and vasculature-enriched microenvironments. To our knowledge, our study is the first to analyze transcription local to PNNs and vasculature in the human postmortem cortex, establishing a framework for molecular profiling within these specialized extracellular and neurovascular niches. Our microenvironment-level LR analysis further contextualized these findings, nominating several trophic and adhesion signaling axes as candidate signaling pathways potentially coupling neuronal, glial, and vascular processes, including FGF-FGFR1 and PTN-PTPR.^115,116 These pathways may underlie a broader alteration of neuron to non-neuronal communication, which may reconfigure local signaling to destabilize cortical homeostasis. As many of these LR interactions represent druggable signaling nodes, these findings also open opportunities for mechanistic and therapeutic exploration.

Integration with genetic datasets provided additional insights into how spatial and neuroanatomical phenotypes relate to SCZ risk. Polygenic risk predominantly localized to neuronal-rich gray-matter domains, consistent with evidence that neurons constitute the primary cellular substrate of genetic vulnerability.^12,16,18,87 However, the transcriptional alterations we detected were largely localized to glial- and vascular-rich layers. This dichotomy may suggest that while genetic liability may reside primarily in neurons, that effect may manifest via neuropil and associated glial-vascular processes as transcriptional phenotypes in non-neuronal cells. This transduction of risk from neuronal to non-neuronal domains highlights the importance of communication between these elements in maintaining cortical network stability.¹⁵ As one potential mechanism to mediate this transduction, several TFs mapping onto SCZ GWAS loci were predicted to regulate SCZ-DEGs identified in both neuronal-rich and non-neuronal-rich domains. Notably, ZNF804A^75–77 was predicted to form regulons with layer-restricted SCZ-DEGs such as BDNF and DCC, across L1/M, L2/3, and WMtz, and Xenium localized ZNF804A to MGE-derived inhibitory neurons, providing cell-type-specific context for its potential contribution to GABAergic dysfunction and circuit-level impairment in SCZ.^138–140 Complementing this TF-based approach, PRS-associated pathways enriched for non-neuronal processes also highlighted MAPK3 and KANSL1-AS1. Shared between PRS- and diagnosis-associated gene expression changes, these two genes show potentially distinct cell-type expression patterns, with MAPK3 more commonly associated with neuronal contexts and KANSL1-AS1 linked to non-neuronal cell types.^141,142 Gene expression changes for both genes were consistent across SCZ and high-PRS donors, with MAPK3 up-regulated and KANSL1-AS1 down-regulated. Both genes have been noted as candidate genes mapping to SCZ risk loci in previous genomic studies, as well as studies of other brain disorders.^13,141–145 Taken together, these observations highlight the potential relevance of TF- and PRS-associated genes, including ZNF804A, MAPK3, and KANSL1-AS1, and warrant further exploration of their contributions to the molecular basis linking neuronal genetic risk and non-neuronal transcriptional phenotypes.

Beyond these genetic and regulatory links, our transcriptional signatures converged on pathways involved in circuit-level signaling and neuroimmune-stress responses. Down-regulation of activity-dependent synaptic genes (BDNF, VGF, NPTX2) and inhibitory neuron markers (SST, CORT, PVALB) supports the notion that impaired trophic and inhibitory signaling underlies cortical circuit dysfunction in SCZ.^99,146,147 BDNF-TrkB signaling critically supports the function of multiple inhibitory neuron subtypes,^148–154 and our finding of dysregulated NTRK2, which encodes the BDNF receptor TrkB, in PNN microenvironments highlights inhibitory-circuit vulnerability associated with PVALB neurons in SCZ.^113,114 We also found spatially divergent glial-vascular changes, suggesting complex neuroimmune disturbances in SCZ, including up-regulation of astrocytic programs (GFAP, SERPINA3), attenuation of microglial signatures (AIF1, TREM2, C3), and heterogeneous oligodendrocytic and vascular-associated signals. FKBP5, a key mediator of glucocorticoid-mediated stress signaling,^155–157 was also up-regulated in SCZ. Pathways related to mitochondrial metabolism and RNA processing were repeatedly dysregulated.^{18,87,158–160} Altogether, these findings point to an interconnected network of trophic-inhibitory signaling and stress-immune dysregulation that may contribute to cortical vulnerability in SCZ. Future studies should address how these spatially- and cellularly-resolved gene programs are mechanistically coordinated during disease progression.

In summary, we provide a resource-scale atlas that maps the spatio-molecular neuroanatomy of the dlPFC in SCZ. By analyzing across laminar, cellular, and microenvironment compartments, we revealed how disease-associated transcriptional programs were organized across the dlPFC in a spatial-anatomical context. Collectively, the data reframes SCZ pathology as spanning cell-intrinsic changes, local cellular milieu, and laminar circuitry, with genetic risk converging on these hierarchical axes. We provide publicly available web browsers (Samui, iSEE) to facilitate utility of the data in interrogating pathways and mechanisms implicated in SCZ, providing a resource for hypothesis generation and data-driven discovery.

3.1 |. Limitations of the study

Several limitations of this study should be acknowledged. Biological heterogeneity of SCZ and modest statistical power, particularly within neuronal-rich cortical domains and PNN microenvironments, may constrain the generalization of certain findings. Some of the CCC results and data-driven LR pairs including NTN4-DCC and THBS2-NOTCH4 have limited experimental evidence for direct binding and therefore require cautious interpretation. As with all postmortem research, potential confounds including medication exposure, comorbidities, demographic variation, and polygenic risk may complicate interpretation, and our cross-sectional design captures only a snapshot of illness state, making it difficult to separate cause from consequence. Technical aspects of the platforms also come with inherent limitations,^161,162 including relatively low gene coverage and gene-level resolution, which currently hinder comprehensive spatial eQTL or transcriptome-wide association analyses. Despite these caveats, the current study provides a foundation for dissecting the molecular landscape of SCZ pathology in the human dlPFC and points toward translational opportunities. Future work leveraging higher-resolution spatial technologies, long-read sequencing, and multi-modal integration will be critical to advance mechanistic understanding of transcriptional changes in SCZ and facilitate translation to effective therapeutic strategies.^163–166

4 |. Methods

4.1 |. Postmortem human tissue samples

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 and at the University of North Dakota School of Medicine and Health Sciences under WCG IRB protocol #20111080. Additional samples were consented through the National Institute of Mental Health Intramural Research Program under NIH protocol #90-M-0142, and were acquired by LIBD via material transfer agreement. Details of tissue acquisition, handling, processing, dissection, clinical characterization, diagnoses, neuropathological examinations, and QC measures have been described previously.¹⁶⁷ Fresh frozen coronal brain slabs from 64 donors were selected for dlPFC sub-dissections. The cohort was divided into two groups of 32 donors, one composed of individuals diagnosed with SCZ and the other composed of NTC. One NTC sample failed anatomical validation and was removed from the study, resulting in a total of 63 dlPFC tissue blocks for the study (one per donor). Cohorts were balanced for age (NTC: mean 46.80 years; SCZ: mean 47.55 years) and sex (NTC: 16 males, 15 females; SCZ: 18 males, 14 females). Fresh frozen coronal slabs at the level of the anterior striatum were rapidly sub-dissected with a hand-held dental drill, perpendicular to the pial surface, targeting Brodmann Area (BA) 46. Cortical brain blocks were approximately 8 X 8 X 5 mm in dimensions, were collected across a sulcus to preserve the integrity of L1, and encompassed the complete laminar structure, spanning cortical L1-L6 and WM. Following dissection, tissue blocks were stored at −80°C in sealed cryogenic bags until cryosectioning. Demographics for the 63 donors are listed in Supplementary Table 1. Experimenters were blind to clinical diagnosis of the donors.

4.2 |. Tissue processing and anatomical validation

Tissue blocks were mounted onto round chucks with OCT (Tissue-Tek Sakura), acclimated to the cryostat (Leica CM3050s) to −14°C, ~50 μm of tissue was trimmed from the block to achieve a flat surface, and several 10-μm sections were collected for anatomical validation (AV). Three AV measures were implemented: 1) visual inspection of the blocks to assess inclusion of the sulcus as well as WM and gray matter; 2) H&E staining to assess cellular integrity of all cortical laminae and the WM; and 3) multiplex RNAscope single molecule fluorescence in situ hybridization (smFISH) to ensure presence of molecular markers for all 6 layers and the WM. H&E staining was performed as previously described and images were acquired using an Aperio CS2 slide scanner (Leica) equipped with a 20x/0.75NA objective and a 2x doubler. RNAscope smFISH was performed as previously described,¹⁶⁸ and images were acquired using the Nikon AX-R confocal microscope equipped with a 2x/0.1 NA objective or a Vectra Polaris slide scanner (Akoya Biosciences) at 20x magnification (0.75 NA). Probes for established marker genes (Advanced Cell Diagnostics, ACD) were used to identify cortical layers in the human cortex,²⁷ including SLC17A7 (Cat No. 415611) indicative of L1–6; RELN (Cat No. 413051) and AQP4 (Cat No. 482441) indicative of L1; HPCAL1 (Cat No. 846051) indicative of L2; FREM3 (Cat No. 829021) indicative of L3; RORB (Cat No. 446061) indicative of L4; TRABD2A (Cat No. 532881) indicative of L5; NR4A2 (Cat No. 582621) indicative of L6; and MBP (Cat No. 411051) indicative of WM. After AV, blocks were again acclimated to the cryostat (Leica CM3050s), mounted onto chucks, scored with a razor into 6.5 X 6.5 mm squares, and 10-μm sections were mounted onto the Visium Spatial Gene Expression Slide (Cat No. 2000233, 10x Genomics). Each Visium slide included four sections from four individual donors, which were also balanced across slides for diagnosis and sex. Adjacent 10-μm tissue sections were mounted onto glass slides for further validation studies (e.g., RNAscope).

4.3 |. Visium-SPG data generation and image preprocessing

Visium-SPG staining

IF staining was performed according to the manufacturer’s instruction (10x Genomics, CG000312 Rev D) with slight modifications to avoid staining artifacts introduced by ribonucleoside vanadyl complex (RVC) after methanol fixation. In brief, following cryosectioning, tissue was fixed in pre-chilled methanol, treated with BSA-containing blocking buffer without RVC, and incubated for 30 minutes at room temperature with staining reagents using biotinylated WFA (Sigma Aldrich, Cat No. L1526, 1:200 and 1:600) for perineuronal nets and anti-NeuN antibody conjugated to Alexa Fluor 555 (Sigma Aldrich, Cat No. MAB377A5, 1:50) for neurons. Two different lots of WFA were optimized prior to Visium-SPG staining to account for their lot-to-lot variations. Following a total of 5 washes with RVC-containing buffer, streptavidin conjugated to Alexa Fluor 647 (Thermofisher, Cat No. S32357, 1:400) and anti-Claudin-5 antibody conjugated to Alexa Fluor 488 (Thermofisher, Cat No. 352588, 1:50) was applied for 30 minutes at room temperature to fluorescently label WFA-reactive perineuronal nets and endothelial cells for vascular structures. DAPI (Thermofisher, Cat No. D1306, 1:3000, final 1.67 μg/ml) was used for nuclear counterstaining. The slide was coverslipped with 85% glycerol supplemented with RiboLock RNase inhibitor (Thermofisher, Cat No. EO0384, final 2 U/μl) and scanned on a Vectra Polaris slide scanner (Akoya Biosciences) at 20x magnification (0.75 NA). The exposure time per channel was determined manually, slide by slide, for six different band filters: DAPI (mean 3.2, range 2.3–3.9 msec), Opal 520 (mean 237.8, range 200–290 msec), Opal 570 (mean 240.3, range 200–290 msec), Opal 620 (mean 6.4, range 2–7 msec), Opal 690 (mean 397.2, range 120–550 msec), and autofluorescence (100 msec, fixed) (Supplementary Table 12).

Visium-SPG cDNA synthesis and library preparation

Immediately after slide imaging, each tissue section was permeabilized for 12 minutes and processed for cDNA synthesis, amplification, and library construction according to the Visium Spatial Gene Expression User Guide (10x Genomics, CG000239 Rev F). The resulting whole transcriptome Visium libraries were sequenced on Illumina NovaSeq 6000 with S4 200 cycle kit using a loading concentration of 300 pM (read 1: 28 cycles, i7 index: 10 cycles, i5 index: 10, read 2: 90 cycles).

Spectral unmixing

For image data preprocessing, we applied spectral unmixing methods following the slide scanner manufacturer’s instructions to reduce noise from multi-plexed spectral profiles of Alexa fluorophores and mitigate autofluorescence from postmortem human tissue, such as lipofuscin.¹⁶⁹ Image preprocessing and spectral unmixing was performed using Phenochart Whole Slide Viewer and inForm Automated Image Analysis Software (both from Akoya Biosciences). After acquiring image data using a Vectra Polaris slide scanner, the files were subsequently processed in Phenochart. Annotations were drawn to outline the entire tissue section for each QPTIFF file, ensuring no overlap between stamped tiles when processed in batch mode using inForm. The annotated QPTIFF images were then processed for spectral unmixing in inForm. Spectral fingerprints were previously generated from single-positive control samples to establish reference spectral profiles for DAPI and individual Alexa fluorophores and were stored in the inForm spectral library.^25,29 We selected spectral fingerprints for DAPI, Alexa 488, Alexa 555, and Alexa 647 to isolate fluorescence signals corresponding to nuclei, Claudin-5, NeuN, and WFA, respectively. The same unmixing algorithm was applied uniformly to all samples to ensure consistency in image preprocessing and outputs. This process decomposes multi-spectral profiles into spectrally unmixed, multi-channel TIFF tiles using inForm. Each multi-component TIFF tile represented a stamped tissue area, with fluorescence signals assigned to separate channels for DAPI, Alexa 488, Alexa 555, and Alexa 647. Autofluorescence, including lipofuscin, was segregated into a dedicated “AF” channel.

Image stitching

The spectrally unmixed multichannel tiles of the entire slide (~600 TIFF files) were stitched using the VistoSeg::InformStitch function to recreate the multichannel whole slide TIFF image. This function utilizes the X and Y coordinates embedded in each tile's filename to accurately position the tiles. Additionally, an offset parameter was used to exclude any extraneous dead space beyond the fiducial frames of all capture areas.

Image splitting

Next, this stitched image was split along the y-axis into four individual capture area images in multichannel TIFF format using the VistoSeg::splitSlide_IF function.

4.4 |. Visium gene expression data processing and analysis

Visium raw gene expression data processing and QC

The FASTQ files of all 63 Visium SPG samples were preprocessed and aligned with their corresponding immunofluorescence images following 10x Genomics SpaceRanger pipeline (v2.1.0),¹⁷⁰ and later contained in a SpatialExperiment 1.12.0 object.¹⁷¹ To identify spots that were compromised and hence deemed low quality, we used SpotSweeper (v0.99.2)¹⁷² to remove spots with extremely low library complexity (sum_umi < 100, sum_gene < 200) as well as local spot-level outliers and compromised regions (Figure S2, Figure S3). Out of 63 Visium SPG samples, there were two regional artifacts found and annotated using SpotSweeper (Figure S3). Low-quality spots and regional artifacts were later discarded for downstream analysis. Overall, we discarded 3001 (1.06 %) spots, with 4 to 1,077 spots per sample with a median of 14 (Figure S4). Empirical cumulative distributions functions (CDFs) of three QC metrics, including total number of unique molecular identifiers (UMIs), total number of unique genes detected, and proportion of mitochondria genes, were compared between two diagnostic groups (Figure S4). No visual difference between the CDFs of the two diagnostic groups suggests that there was no evidence of bias with respect to the diagnosis. QC metrics, spot-level descriptive statistics are visualized using Scater (v1.34.0)¹⁷³; spot plots are made with escheR (v1.3.2).¹⁷⁴

Data-driven SpD identification

To identify SpDs in a data-driven and scalable fashion, we applied the spatially-aware clustering algorithm PRECAST²⁴ to all 63 samples. PRECAST provides native data integration to account for complex batch effects, and hence is effective to predict SpDs across experimental batches and diagnostic groups. We chose 723 genes for feature selection as input to PRECAST, which consisted of the top 100 laminar marker genes of SpD09 in Huuki-Myers et al.,²⁵ and examined multiple possible numbers of clusters (k=2–16) (Figure S5). After examining the predicted SpDs of k=2–16, we found that PRECAST SpDs of k=7 matched most closely with laminar organization of the classic histological layers in a spatially congruent fashion, and was used in the following downstream analysis.

Marker gene identification for each PRECAST SpD

To characterize the transcriptomic profiles of PRECAST SpDs, we created pseudobulked profiles based on combinations of donor and individual domains from the data-driven SpDs at k=7 using spatialLIBD::registration_pseudobulk. We first calculated the proportion of variance in gene expression explained by key variables, including PRECAST SpDs, Visium slide id, donor diagnosis status, sex, implemented using scater::plotExplanatoryVariables¹⁷³, assuming that data points in the pseudobulked dataset were independent (Figure S7). Then, the pseudobulk profiles at the donor and SpD levels was then used to identify DEGs across the PRECAST SpDs using two models, an enrichment test and a pairwise test, as previously described in Maynard, Collado-Torres et al.,²⁷ while adjusting for donor age, sex and Visium slide ids. The enrichment test examines differences in expression between one SpD and all other SpDs, implemented in spatialLIBD::registration_stats_enrichment. The pairwise test examines the differences in expression between every pair of SpDs, implemented in spatialLIBD::registration_stats_pairwise (Figure S8, Figure S9).

PRECAST SpD annotation via spatial registration

To annotate the data-driven SpDs, we used spatialLIBD::layer_stat_cor to spatially register seven PRECAST SpDs to the previously established manual annotation of cortical structures.²⁷ Specifically, the function computes the Pearson correlation of enrichment t-statistics from each data-driven domain of seven PRECAST SpDs and each manual annotated layer, where the top 100 genes of each manually annotated layer are considered to calculate the Pearson correlation. We then used annotate_registered_clusters to further assess the confidence of the annotation for each pair of PRECAST SpDs and manually annotated cortical layers. For PRECAST SpDs that were registered to multiple manually annotated cortical layers, we found that they formed hybrid-layer identities. The calculation of the confidence score is detailed in Huuki-Myers et al.²⁵ and an “X” is placed on the heatmap for the SpDs we were most confident in. The spatial registration procedure was also performed to compare PRECAST SpDs between the two diagnostic groups, further validating that the PRECAST SpDs captured the same cortical structures across diagnostic groups.

4.5 |. Visium-SPG image processing

Segmentation for each IF channel

We utilized Visium-SPG and stained tissue sections for DAPI (nuclei), NeuN (neurons), WFA (PNN), and Claudin-5 (blood vessels) (Methods 4.3, Figure S10, Figure 1A). Following staining, we performed imaging and preprocessing to obtain raw multichannel images per tissue section. We employed a custom image processing pipeline to segment and identify contours in the DAPI, NeuN and Claudin-5 channels (Figure S10). Using Python libraries such as SciPy,¹⁷⁵ NumPy,¹⁷⁶ OpenCV,¹⁷⁷ and scikit-image,¹⁷⁸ we processed images by converting them to grayscale and applying binary thresholding with Otsu's method (implemented via scikit-image::threshold_otsu) to isolate the signal. Contours were detected using OpenCV::findContours function, filtered based on area to remove noise. Contours with an area < 50 pixels for DAPI, area < 100 pixels for NeuN and Claudin-5 regions smaller than 5,000 pixels were removed to exclude artifacts. Binarized images were saved for further analysis. We extracted intensity histograms (x-axis - pixel intensity, y-axis - density) (Figure S11B) for all samples and noticed that the WFA signal was predominantly located in the tail of the intensity histogram. Consequently, we applied the T-point thresholding method for isolating PNN microenvironments (Figure S11C).^179,180 We developed a custom function in MATLAB to implement this method, specifically designed to exclude the initial peak at 0 in the intensity histogram corresponding to the dark background. The WFA signal, situated in the right tail of the histogram, was isolated by drawing a straight line from the histogram's maximum to its minimum on the right side (red line, Figure S11B). The threshold for extracting PNN microenvironments was determined at the x-axis value where the distance between this line (connecting the maximum and minimum of the histogram) and the histogram was maximal (green line in Figure S11B). Subsequently, to separate the WFA signal from the background tissue in the extracted PNN microenvironments (Figure S11C), adaptive thresholding was applied. Further refinement involved masking the signal with autofluorescence segmentations (obtained using simple Ostu’s thresholding) to eliminate regions corresponding to meninges and blood vessels (Figure S11D). Finally, a size threshold of 30,000 pixels or greater was imposed on the segmentations to remove spots associated with fluorophore aggregates, thus refining the identification of genuine WFA signals (Figure S11E).

Quantification and spot classification

Segmentations from all channels were combined into a single .mat file per sample and, along with the raw image (Figure S10A), were provided to VistoSeg::CountNuclei¹⁸¹ function to extract spot-level fluorescent metrics across all channels. Metrics for each channel included number of segmented regions per SPG spot (if any pixel of the segmented region lies inside the SPG spot), number of segmented regions per SPG spot (if the centroid of the segmented region lies inside the SPG spot), proportion of spot pixels covered by segmented signal, and mean intensity of all segmented pixels per spot. We used the proportion spot covered with segmented pixels in each channel to classify spots into SPG vs. non-SPG, using the thresholds PDAPI < 5%, 30% > PNeuN > 5%, PWFA > 5%, 20% > PClaudin > 5%, (Figure S10B). Given the multimodal distribution, all spots with intensities above the background level of 0 were included for all IF channels. Spots covered with less than 5% DAPI signal were classified as neuropil, spots covered with 5–30% NeuN signal were classified as neuronal, spots covered with more than 5% WFA signal were classified as PNN and spots covered with 5–20% Claudin-5 signal were classified as vasculature (Figure S10). Notably, when visualizing the proportion of spots from each SpD across microenvironments, we observed similar distributions in both SCZ and NTC samples, as expected, supporting the robustness of our classification scheme (Figure S12). We validated spot classification in the NTC samples to avoid potential confounding effects of diagnosis using a two-pronged approach.

Validation of spot classification using marker genes respective to SPG-defined microenvironments

We performed pseudobulk DE analysis comparing SPG and non-SPG spots for each IF channel within NTC samples (n=31 donors) (Supplementary Table 3). Pseudobulking was conducted by aggregating counts per sample and spot classification, resulting in 62 data points per channel. As expected, DEGs included CAMK2A in neuropil spots (Figure S13A), SNAP25 in neuronal spots (Figure S13B), PVALB enriched in PNN spots (Figure S13C), and CLDN5 in vasculature spots (Figure S13D). These findings corroborate the accuracy of our classification methodology.

Validation of spot classification by comparison with gene expression of reference marker gene list

To further validate our image-based spot classification, we compared our results with gene expression patterns of marker genes corresponding to neuropil, neuronal, PNNs and vasculature (Figure S14), as defined in publicly available reference datasets.^25,30,39,40 For each channel, spots were classified as SPG or non-SPG based on the summed expression of marker genes, using empirically defined thresholds: > 1,000 counts to determine neuropil spots (Figure S14A), > 250 counts to determine neuronal spots (Figure S14B), > 5,000 counts to determine PNN spots (Figure S14C), and > 2,500 counts to determine vasculature spots (Figure S14D). These gene expression-based classifications were visually compared to those derived from image-based metrics. We observed strong concordance between the image- and gene-based classifications across all IF channels, with image-defined SPG spots exhibiting higher expression of their corresponding marker gene sets.

Furthermore, when analyzing mean summed marker gene expression per SpD, we found that domains with a higher proportion of image-based SPG spots also exhibited higher marker gene expression (Figure S15). We observed enrichment of neuropil spots in SpD06-L2/3 through SpD01-WMtz (Figure S15A), neuronal spots in SpD06-L2/3 through SpD01-WMtz (Figure S15B), PNN spots in SpD02-L3/4 and SpD05-L5 (Figure S15C), and vasculature, showed substantial proportions in SpD02-L3/4 through SpD04-WM (Figure S15D). However, some domains, such as SpD07-L1/M for neuropil and vasculature, showed meaningful proportions of SPG spots but lower marker gene expression (Figure S15A, Figure S15D). Collectively, these data support the validity of our approach for annotating SPG spot types based on their cellular characteristics while demonstrating minimal differences in density across SpDs between NTC and SCZ samples. Together, these results establish our ability to perform differential gene expression testing within these cellular environments across diagnosis.

4.6 |. DE analysis between SCZ and NTC donors

Identification of layer-adjusted, layer-restricted, and layer-specific SCZ-DEGs across diagnosis

With the donor-SpD pseudobulked data described above, we investigated SCZ-associated changes in gene expression using limma (v3.62.2).¹⁸² To identify SCZ-DEGs, we employed two linear regression models, the layer-adjusted model and the layer-restricted model, to quantify the overall and laminar-specific dysregulation, respectively (Figure S16). In the layer-adjusted model, we included diagnosis status as the variable of interest and PRECAST SpDs as one of the covariates, accounting for the excess variation attributed to differences across layers. Other confounders such as age, sex, and slide id were adjusted in the model. The layer-adjusted analysis was implemented using spatialLIBD::registration_stats_enrichment.²⁶ In comparison, the layer-restricted model included an additional statistical interaction term between diagnosis status and PRECAST SpDs, capturing the laminar-specific effect. For each PRECAST SpD, a contrast matrix was constructed and tested to quantify the DE, implemented using limma::topTable. As a result, we detected 927 unique SCZ-DEGs when controlling FDR < 0.10, or 5,183 unique SCZ-DEGs at nominal p < 0.05 (Supplementary Table 5). By excluding layer-restricted SCZ-DEGs detected in more than one SpD, we identified a total of 4,073 genes that are uniquely significant for each of the PRECAST SpDs, namely layer-specific SCZ-DEGs (Figure S16).

Identifying microenvironment-restricted SCZ-DEGs across SPG-defined microenvironments

To identify SCZ-associated transcriptional alterations within our SPG-defined microenvironments for neuropil, PNNs, neuronal, and vasculature, we conducted DE tests using the four SPG spot types (Section 2.1, Methods 4.5) while adapting the layer-adjusted DE model (Methods 4.6). For each SPG spot type, we constructed a new donor-SpD pseudobulked dataset by subsetting to include SPG spots only. Non-SPG spots were excluded prior to the donor-SpD pseudobulking procedure. To assure statistical validity, we excluded donor-SpD pseudobulked data points, i.e., SpDs that were specific to each donor or that contained less than 10 SPG spots. For neuropil, PNN, and neuronal microenvironments, we restricted the analysis to gray-matter-associated SpDs, including SpD07-L1/M, SpD06-L2/3, SpD02-L3/4, SpD05-L5 and SpD03-L6 in the layer-adjusted DE model. For the vasculature microenvironments, all seven SpDs were included in the pseudobulked data. In the layer-adjusted DE model, we used donor diagnosis as the variable of interest while adjusting sex, age, and SpD as covariates. As SPG spot classifications were channel-independent, each comparison was made separately.

4.7 |. Enrichment analysis of SCZ-DEGs

Gene ontology analyses of layer-adjusted and microenvironment-restricted SCZ-DEGs

GSEA was performed using clusterProfiler::gseGO (v4.14.6)¹⁸³ to evaluate enrichment of GO terms for biological processes (BPs) among layer-adjusted SCZ-DEGs. Genes were ranked based on their t-statistics in descending order, and enrichment scores were computed against the GO-BP gene sets obtained from org.Hs.eg.db (v3.20.0).¹⁸⁴ Significantly enriched GO-BP terms were selected using an adjusted p-value threshold of 0.05. Treemaps clustered similar terms based on semantic similarity and were used to visualize the GO terms using rrvgo::treemapPlot (v2.4–4) (Figure S17).¹⁸⁵

GO over-representation analysis (ORA) was performed for microenvironment-restricted SCZ-DEGs identified in the neuronal and vasculature microenvironments. ORA was conducted to identify enrichment of GO-BP terms, using a background gene set specific to each microenvironment. Microenvironment-restricted SCZ-DEGs were defined at nominal p < 0.05, and enriched GO-BP terms were selected at FDR ≤ 0.05. Treemaps displayed the top 50 terms after semantic-similarity reduction (threshold=0.7), and an additional neuron-focused treemap was generated by keyword filtering neuron-related terms (Figure 6C, Figure S34, Figure S37).

SynGO (version 1.2; 20231201; https://www.syngoportal.org) was used to annotate synaptic location (Cellular Component, CC) and function (BP) for both the layer-adjusted SCZ-DEGs (FDR < 0.10; Figure 2C) and the microenvironment-restricted SCZ-DEGs for neuropil and PNNs (nominal p < 0.05; Figure 6C). Unless noted otherwise, analysis used matched, analysis-specific background gene sets. Annotated genes and enriched terms (FDR < 0.01; ≥ 3 matching input genes) were reported with representative gene counts or enrichment statistics, as appropriate. Evidence filters were set to the inclusive default (“no filters; include all annotations”).

Enrichment of SCZ-DEGs among cell type, SpD, and SPG spot type marker genes

To localize the cell-type and SpD contributions among the layer-adjusted and layer-restricted SCZ-DEGs, we conducted a one-sided Fisher’s exact test, implemented with spatialLIBD::gene_set_enrichment (v1.18.0),²⁶ to examine enrichment of gene sets. The marker genes were compiled respectively from (i) cell-type markers previously reported in Huuki-Myers et al.,²⁵ (ii) PRECAST SpD markers detected using the layer DE model for marker gene identification (Methods 4.4; Marker gene identification for each PRECAST SpD), and (iii) SPG markers identified via pseudobulk DE analysis comparing SPG and non-SPG spots (Methods 4.5; Validation of spot classification using marker genes respective to SPG-defined microenvironments). Only those genes with log₂ fold-change (logFC) > 0, and FDR-adjusted p < 0.05 were included in the marker gene sets. The selected marker gene sets were compared against the up- and down-regulated gene sets of the layer-adjusted and layer-restricted SCZ-DEGs respectively.

Enrichment of layer-restricted SCZ-DEGs among reference cell-type-specific SCZ-DEGs

To cross-reference and spatially localize the cell-type-specific SCZ-DEGs, we used spatialLIBD::gene_set_enrichment to examine the enrichment of layer-restricted SCZ-DEGs among the reference cell-type-specific SCZ-DEGs reported in Ruzicka et al..¹⁶ In line with the previous reporting, we included only SCZ-DEGs whose absolute log fold-change (|logFC|) was > 0.1 and adjusted p < 0.05, and separated the SCZ-DEGs by up- and down-regulation. Enrichment tests were performed only between gene sets of the same directionality (i.e., up-regulated with up-regulated, down-regulated with down-regulated)

TF enrichment analysis

To identify candidate TFs potentially regulating the observed spatially-localized gene expression changes, particularly our layer-restricted SCZ-DEGs in SCZ, we performed TF enrichment analysis using ChIP-X Enrichment Analysis version 3 (ChEA3).⁷³ For each SpD, we ran ChEA3 on the sets of up- and down-regulated layer-restricted DEGs, separately. For each TF, enrichment was quantified and reported using the mean-rank approach suggested by ChEA3. To facilitate cross-domain comparisons, enrichment ranks were normalized by the total number of TFs tested (n=1,632 TFs) to compute a rank percentile (rank/n*100), where lower percentile values indicate stronger enrichment. We then focused on TFs overlapping with SCZ GWAS-prioritized genes curated in Trubetskoy et al. (Supplementary Table 12, “Prioritised” sheet, n=120 genes).¹³ Normalized values were visualized as heatmaps.

STRING network analyses

We performed network-based analyses using the STRING database (version 12.0, https://cn.string-db.org/)⁷⁴ to characterize functional and/or physical interactions separately for (i) ChEA3-derived candidate TFs and (ii) microenvironment-restricted SCZ-DEGs (Figure 4B, Figure 6E). For both analyses, we selected the organism as Homo Sapiens for human proteins and used active interaction sources including text mining, experimental evidence, curated databases, co-expression, genomic neighborhood, gene fusion, and co-occurrence.

For the TF analysis, we selected and combined the top 10 highly ranked TFs associated with up- and down-regulated layer-restricted SCZ-DEGs, yielding a total of 20 TFs per SpD. These were then pooled into two composite groups according to domain characteristics: non-neuronal-rich domains (SpD07-L1/M, SpD01-WMtz, SpD04-WM; total 60 TFs, with overlap across SpDs) and neuronal-rich domains (SpD06-L2/3, SpD02-L3/4, SpD05-L5, SpD03-L6; total 80 TFs, with overlap across SpDs). TF networks were analyzed using the ‘full STRING network’ option integrating both functional and physical interactions, with interaction confidence thresholds set at high (> 0.7) and additionally at medium (> 0.4) to encompass broader transcriptional dysregulation patterns. The default STRING background gene set was used for these TF analyses.

For microenvironment-restricted SCZ-DEGs, the SCZ-DEGs identified within neuropil, neuronal, PNN, and vasculature microenvironments at nominal p < 0.05 were analyzed as ‘physical subnetworks’ to specifically capture direct protein-protein interactions and associated signaling pathways. Matched background gene sets were applied, and interactions were filtered at a high-confidence threshold (> 0.7). Functional modules were defined using the Markov Cluster (MCL) algorithm with an inflation parameter set at 1.2 and further characterized using built-in gene ontology tools available within the STRING web portal (Figure S38).

4.8 |. SCZ polygenic risk score calculation

QC of genotype data

The 63 donors included in this study were originally genotyped across multiple heterogeneous batches using various Illumina SNP BeadChip microarrays. To ensure consistency, genotype data processing was performed in two stages: batch-level imputation and study-level merging. First, quality control (QC) was performed separately on each original genotyping batch using PLINK (v1.90).¹⁸⁶ Pre-imputation QC excluded variants with minor allele frequency (MAF) < 0.5% (--maf 0.005), variant missingness > 5% (--geno 0.05), and Hardy-Weinberg equilibrium (HWE) P < 1×10⁻⁵ (--hwe 1e-5). Each batch was then phased and imputed separately using the NHLBI TOPMed imputation service and the hg38 reference panel.¹⁸⁷ Post-imputation, we only retained variants within each batch if they met an imputation quality metric R² > 0.9 and a batch-level MAF > 0.05. Subsequently, the 63 specific donors were extracted from their respective post-QC imputed batches and merged into a single dataset. To generate the final analytical set, we performed another round of QC on the merged dataset using PLINK to remove variants with missing call rates > 5% (--geno 0.05), HWE P < 1×10⁻⁶ (--hwe 1e-6) and MAF < 5% (--maf 0.05). To examine the genetic ancestry of the donors, we used the global ancestry pipeline (https://github.com/nievergeltlab/global_ancestry) to calculate the European ancestry proportion. Among all 63 donors, 60 donors have a predicted European ancestry proportion greater than 0.9; three donors, including Br2378, Br2039, and Br5622, had a European ancestry proportion between 0.6 and 0.9. When calculating European ancestry proportion, the variant exclusion criteria included missing call rates exceeding 1% (--geno 0.01), HWE below 1×10⁻⁶ (--hwe 1e-6), SNP missingness exceeding 1% (--mind 0.01), heterozygosity, and MAF above 1% (--maf 0.01).

PRS calculation

We computed polygenic risk scores (PRS) using the standard LD-pruning and p-value thresholding approach implemented in PLINK2.¹⁸⁸ SNPs were pruned using p-value-informed clumping with an LD threshold of r²=0.1 within a 1,000-kb window. Eighteen PRS were generated from the pruned SNP sets using the following p-value thresholds: 5e-08, 1e-06, 1e-07, 1e-04, 0.001, 0.01, 0.05, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9 and 1. Weights for PRS calculation were defined as the natural logarithm of the odds ratios estimated from the meta-analysis of the PGC3 SCZ GWAS of European ancestry samples.¹³ The PRS used for DE analyses were selected at the p-value threshold (0.001) that yielded the smallest association p-value between PRS and SCZ case-control status in our dataset.

PRS-associated DE test

To determine PRS-DEGs, we adapted the layer-adjusted DE model described in Methods 4.6. In the layer-adjusted DE model, normalized PRSs replaced donor diagnosis as the variable of interest, while the first three genetic PCs were included as covariates in addition to SpD, age, sex and slide id. FDR-adjusted p were calculated to control the FDR. GSEA was performed on the PRS-DEGs using their t-statistics and visualized as a treemap (Figure S22).

4.9 |. Spot-level polygenic risk enrichment via scDRS

Single-cell disease relevance score (scDRS, v1.0.4) framework⁷⁹ was used to quantify SCZ polygenic risk enrichment across SpDs. First, MAGMA⁸⁰ gene-based analysis was performed on the SCZ GWAS summary statistics¹³ using 1KGP3 as the LD reference panel, generating per-gene association statistics. This resulted in a gene list with MAGMA p-values and was further used as input to the scdrs munge-gs command to construct the SCZ-associated gene set filtered at FDR < 0.05. For the expression data, the SpatialExperiment object was converted to an AnnData .h5ad file containing log-normalized counts. A covariate file was generated including sequencing depth, detected genes, mitochondrial content, sex, age, and RIN, to control for technical and biological confounders. The scdrs compute-score command was run with the argument --n-ctrl 1000, generating 1,000 sets of matched control genes via Monte Carlo sampling (with mean/variance matching) to compute normalized scDRS scores for each spot. Finally, downstream analysis with the scdrs perform-downstream command was performed, specifying SpD labels (spd_label) as the grouping variable and enabling gene-level analyses.

4.10 |. Xenium data generation

Xenium custom panel design

A custom panel targeting 300 genes was designed to characterize spatial gene expression patterns at enhanced cellular resolution using the 10x Genomics Xenium platform (Supplementary Table 8). The 300-gene panel consisted of four gene categories (Figure 5 Aiii): (1) 168 layer-adjusted DEGs, (2) 85 cortical layer marker genes based on findings from Huuki-Myers et al.,²⁵ (3) 18 SCZ-associated risk genes reported from prior literature and GWAS (references included in Supplementary Table 8), and (4) 35 broad brain cell-type gene markers, including excitatory neurons, inhibitory neurons, astrocytes, oligodendrocytes, microglia, endothelial cells, and immune cells. Of the 172 SCZ-DEGs, only 168 were included as 4 genes were recommended for exclusion by the manufacturer, including 3 long non-coding RNAs (NIPBL-DT, LINC01736, LINC01535) and SNHG25. With the exclusion of 6 overlapping genes across multiple categories, the final panel consisted of 300 unique genes, which was assessed and reviewed via the Xenium Panel Designer (10x Genomics).

Xenium slide preparation

Xenium tissue slides were prepared and run on the Xenium analyzer (10x Genomics, Cat No. PN-1000481) in groups of two using the 10x Genomics (Pleasanton CA) v1 workflow. Slides were processed in pairs (6 sections per experimental run), with each pair consisting of one slide containing 2 NTC and 1 SCZ tissue sections, and the other slide containing 2 SCZ and 1 NTC tissue sections, balancing the number of NTC and SCZ samples (3 each per run) and minimizing batch effects across the 4 experimental runs. Procedures related to cryosectioning and slide storage were followed according to the demonstrated protocol (10x Genomics, CG000579 rev C). Briefly, tissue blocks were acclimated to −16°C on a CM3050S Cryostat (Leica Biosystems) and Xenium slides were placed in the cryostat. Sections were mounted within the sample positioning area (10.45 mm x 22.45 mm) and slides were stored in a slide mailer at −80°C for 3–4 days before downstream fixation and permeabilization steps (10x Genomics, CG000581 Rev C). Briefly, slides were allowed to incubate at 37°C for 1 minute and then were fixed for 30 minutes in 40 mL of 3.7% formaldehyde (ThermoFisher Scientific), washed in 40 mL 1xPBS (ThermoFisher Scientific) and then permeabilized in 1% SDS solution (Millipore Sigma) diluted in 0.2-mm filtered nuclease-free water (Integrated DNA Technologies) for 2 minutes. Following removal of the SDS solution, slides were washed 3x in 1xPBS and then immersed in 40 mL chilled 70% methanol (Millipore Sigma, Cat No. 34860) diluted in 0.2-mm filtered nuclease-free water. After 60 minutes, slides were removed and rinsed 2x in 1xPBS (40 mL/wash) and placed in the Xenium Cassettes. We completed probe hybridization, ligation, and amplification steps according to manufacturer’s instructions outlined in CG000582 Rev E. In brief, our Xenium Standalone Custom 300-member gene panel (10x Genomics, Cat No. PN-1000563) was resuspended in 700 μL TE Buffer (ThermoFisher Scientific), vortexed, and maintained at room temperature. Probes were diluted in Xenium Probe Hybridization Buffer and applied to Xenium slides for overnight (16–24 hours) at 50°C. The next day, hybridization reactions were washed 2x in 0.05% 1xPBS-Tween-20 (ThermoFisher Scientific, Cat No. 28320) and incubated in 500 mL Xenium Post Hybridization Wash Buffer at 37°C for 30 minutes. After 3 washes of 0.05% 1xPBS-Tween-20, probes were ligated at 37°C for 2 hours in a reaction consisting of 87.5% Xenium Ligation Buffer, 2.5% Xenium Ligation Enzyme A and 10% Xenium Ligation Enzyme B. After 3 washes of 0.05% 1xPBS-Tween-20, rolling circle products were amplified in a reaction consisting of 90% Xenium Amplification Mix and 10% Xenium Amplification Enzyme for 2 hours at 30°C. After amplification was completed, slides were washed 2x in TE buffer, 3x in 1xPBS (1 mL/wash) and quenched for autofluorescence with 1% Reducing Agent B (10x Genomics) diluted in 1xPBS for 10 minutes at room temperature. Following 3 rinses of 100% Ethanol (Millipore Sigma), slides were dehydrated for 5 minutes at 37°C and then rehydrated in 1xPBS followed by a 2-minute incubation of 0.05% 1xPBS-Tween-20. Nuclei were then stained in Nuclei Staining Buffer (10x Genomics)and washed 3x in 0.05% 1xPBS-Tween-20. Slides were then stored overnight in the dark at 4°C and imaged the following day.

Xenium analyzer

Imaging of all slides took place on the Xenium Analyzer (10x Genomics, Cat No. PN-1000569) located at the Johns Hopkins Sidney Kimmel Comprehensive Center (SKCCC). Reagents and procedures needed to run the Xenium v1 workflow were followed according to the manufacturer’s instructions (10x Genomics, CG000584 Rev F). The most up-to-date software version (v3.1.0.0) and onboard analysis versions (v.3.1.0.4) were selected for all instrument runs. All runs passed a ‘readiness test’ confirming that the system was working optimally and that the instrument was ready for use. Reagents bottles necessary for imaging were prepared accordingly: Bottle 1 – Milli-Q ultrapure water (ThermoFisher Scientific, Cat No. 50131948), Bottle 2 – 0.05% 1xPBS Tween-20, Bottle 3 – Milli-Q ultrapure water and Bottle 4–0.1% Tween-20, 50mM KCL (Teknova, Cat No. P0330), 50% DMSO (Millipore Sigma, Cat No. 41639) diluted in nuclease-free water (ThermoFisher Scientific, Cat No. AM9932). Xenium reagent bottles, Xenium cassettes, an Object Wetting Consumable, pipette tip racks (10x Genomics, Cat No. PN-1000461) and decoding reagent plates (10x Genomics, Cat No. PN-1000461) were loaded according to details provided by the manufacturer. Following successful loading of the instrument, panel ROI selection was performed and the run was initiated. Runs took on average 48 hours to complete.

4.11 |. Xenium data analysis

Data QC and normalization

Data QC was performed on the Xenium data to remove low quality cells from downstream analyses. Within each sample, low quality cells were identified based on percentage negative control features, number of detected genes, and total counts observed in each cell. We used the negative control probes, negative control codewords, and unassigned codewords from the Xenium panel. We removed all cells with expression of any of the above three features in the 99th percentile. Cells that were outliers in terms of number of genes detected (5 median absolute deviations, MADs, below the median) or in terms of total counts (5 MADs above or below the median) were identified using scater::isOutlier and removed. The raw counts from the retained cells were then library size-normalized and log₂-transformed (log-normalized) using scuttle::logNormCounts.

SpD clustering and annotation

To identify SpDs in the Xenium data, label transfer was performed using the spaTransfer::transfer_labels function.⁸⁹ The Visium dataset described above was used as the reference dataset and the annotated PRECAST SpD7 labels (at k=7) were transferred into the Xenium dataset. Briefly, spaTransfer computes non-negative matrix factorization (NMF) on the log-normalized counts matrix of the Visium reference dataset. This results in a loadings matrix and a factors matrix, similar to PCA. spaTransfer then uses the loadings matrix to project the log-normalized counts matrix of the target Xenium dataset and learn the Xenium factors matrix through non-negative least squares. The reference factors are used to fit a multinomial model with the formula PRECAST SpD7 labels ~ NMF factors. The trained multinomial model is then given the Xenium factors and asked to predict SpD labels for each cell in the Xenim dataset. Finally, spatially-aware neighbourhood smoothing similar to Banksy::smoothLabels is used to increase spatial contiguity of predicted domains. For our purposes, we performed NMF with 50 factors and performed spatial smoothing with each cell’s 10 nearest neighbours based on physical distance. We then validated the transferred SpDs using spatial registration against manually annotated human cortical layers from Maynard, Collado-Torres et al.,²⁷ similar to the approach for the Visium dataset described above.

Cell-type clustering and annotation

Clustering was performed with the Banksy package⁹⁰ in order to identify spatially-aware cell types in the Xenium data. All 24 samples were clustered jointly following the BANKSY multi-sample tutorial. To create the BANKSY neighbourhood gene expression matrix, we set the k_geom parameter to 10 and compute_agf to be FALSE. The lambda parameter which controls the weighting on the spatial information versus the gene expression information was set to 0.1, and the clustering resolution was set to 0.7. This yielded 18 total clusters which we then annotated with cell-type labels.

The annotations were based on expression of marker genes which were included in the Xenium panel. The marker genes included in the Xenium panel spanned 7 major cell types: astrocytes, endothelial cells, inhibitory neurons, excitatory neurons, microglia, microglia/immune cells, and oligodendrocytes. The inhibitory and excitatory marker genes also allowed us to separate these cells into subclasses: SST and PVALB inhibitory neurons, VIP and LAMP5 inhibitory neurons, L2/3 excitatory neurons, L4/5 excitatory neurons, L5 excitatory neurons, and L6 excitatory neurons. Annotation of the 18 clusters resulted in a total of 12 cell types, with two of the clusters being ambiguous, likely due to the gene panel not having the resolution to confirm the identity of these subpopulations. Investigation of QC metrics within these groups gave no indication that these clusters are driven by low quality cells, so they were retained in downstream analyses.

Identifying case-control DEGs & functional annotation

DE across cortical layers

Similar to the Visium DE analysis, the counts matrix was first pseudobulked to the donor-domain level, using the predicted SpD annotations from spaTransfer. Any pseudobulk samples with fewer than 10 cells were dropped. We first performed layer-adjusted DE analysis using limma. Since the pseudobulking procedure introduced “replicates” for each donor, we used limma::duplicateCorrelation to account for possible correlation between these replicates. We adjusted for SpDs, age, sex, and the xenium slide for each sample.

DE across cell types

The cell-type-adjusted DE analysis was performed similarly to the layer-adjusted DE analysis. The counts matrix was pseudobulked to the cell type-donor level, using the cell-type annotations described above. We used limma through the spatialLIBD::registration_stats_enrichment function, again accounting for correlation between replicates with limma::duplicateCorrelation. We adjusted for cell type, age, sex, and the Xenium slide for each sample. In order to identify expression changes occurring within specific populations, we leveraged the cell-type information to perform cell-type-stratified DE analysis. For each of the 12 cell types we annotated, we first subsetted the counts matrix to only cells from one type. We then pseudobulked these cells to the donor level, dropping samples with fewer than 10 cells. Similar to as described above, spatialLIBD::registration_stats_enrichment was run, this time adjusting for age, sex, and the xenium slide for each donor.

Differential cell-type abundance and density analysis

To assess whether the abundance of cell types differed between NTC and SCZ, we employed compositional data analysis techniques using the cacoa R package.⁹⁵ Briefly, the cacoa::estimateCellLoadings function first transforms the abundances of the 12 cell types into a 11-dimensional simplex space using the isometric log-ratio transformation. A surface that best separates the NTC and SCZ samples in this 11-dimensional simplex space is then estimated, and the contribution of each cell type to the surface’s orientation is computed. To test for the statistical significance of each cell type, an empirical background distribution is constructed by resampling cells and samples. We first ran this analysis without stratifying the data to SpDs to identify global changes in cell-type abundance. We then ran a layer-stratified version of this analysis by subsetting to cells within a SpD and testing for differences in cell-type abundance within that domain.

4.12 |. Cell-cell communication within SPG-defined microenvironments

To evaluate how SCZ-associated transcriptional changes intersect with LR signaling within SPG-defined microenvironments, we compared our microenvironment-restricted SCZ-DEGs to a curated set of candidate LR pairs in Supplementary Table 5 from Huuki-Myers et al..²⁵ Specifically, we considered LR pairs that were annotated for SCZ genetic risk and further subsetted to those that were identified as mediating cell-cell communication in a dlPFC snRNA-seq dataset. We then subsetted our microenvironment-restricted SCZ-DEGs to those with nominal p < 0.05. For each of the four microenvironments, we assessed whether any of these genes were present as either a ligand or a receptor within the curated list of LR pairs. When a gene matched one of the LR interactors, we plotted the corresponding LR pair in a Sankey plot (Figure 6E), along with its interaction partner, independent of whether the partner gene was significantly differentially expressed in SCZ.

4.13 |. Interactive data visualization and exploration

To assist interactive exploration of the Visium SRT data, along with immunostaining data, we provided the web-based data visualization tools, including Samui Browser¹⁸⁹ for visualizing fluorescent images, nuclei segmentation, as well as gene expression data, and iSEE apps.¹⁹⁰ These tools can be used to explore donor-SpD pseudobulked gene expression data generated from all spots or subsets of spots classified into four SPG-defined microenvironments (neuropil, neuronal, PNN, vasculature). Further details on accessing the Samui browser and iSEE apps can be found via https://research.libd.org/spatialDLPFC_SCZ/#interactive-websites.

Supplementary Material

Supplement 2

Supplementary Table 1. Demographic information and descriptive statistics of donors included in the Visium and Xenium SRT experiments (sex, age, RNA integrity number [RIN], and postmortem interval[PMI]).

Supplement 3

Supplementary Table 2. Pairwise DE test results comparing each of the PRECAST SpDs at clustering resolution k=7.

Supplement 4

Supplementary Table 3. DE test results between SPG spots and their matched non-SPG spots to identify highly expressed genes in neuropil, neuronal, vasculature, and PNN microenvironments within NTC samples.

Supplement 5

Supplementary Table 4. Layer-adjusted DE statistics for all genes, adjusted for PRECAST SpDs at k=7.

Supplement 6

Supplementary Table 5. Layer-restricted DE statistics for all genes, derived from comparisons of matched PRECAST SpDs between NTC and SCZ groups at k=7. Layer-restricted DEGs not shared across SpDs are considered layer-specific DEGs.

Supplement 7

Supplementary Table 6. Predicted TFs across SpDs and directions (up- and down-regulation) of layer-restricted SCZ-DEGs, based on ChEA3 integrated mean-rank results.

Supplement 8

Supplementary Table 7. PRS-based DE statistics adjusted for PRECAST SpDs at k=7.

Supplement 9

Supplementary Table 8. List of 300 genes included in the Xenium in situ panel, with annotations describing their panel role and representative references where applicable.

Supplement 10

Supplementary Table 9. DE statistics for the 300-gene Xenium panel, evaluated under three modeling approaches in case-control comparisons.

Supplement 11

Supplementary Table 10. Microenvironment-restricted DE statistics across four SPG-defined microenvironments (neuropil, neuronal, PNN, vasculature), with an intersection summary of shared DEGs.

Supplement 12

Supplementary Table 11. STRING-MCL cluster descriptions for microenvironment-restricted SCZ-DEGs, with additional GO-BP annotations for the top five MCL clusters.

Supplement 13

Supplementary Table 12. Exposure times for each fluorescence channel across six different band filters used during Visium imaging with the Vectra Polaris scanner.

Glossary

AV

Anatomical validation

BA

Brodmann area

BANKSY

Building aggregates with a neighborhood kernel and spatial yardstick

CCC

Cell-cell communication

CDF

Cumulative distributions functions

CGE

Caudal ganglionic eminence

ChEA3

ChIP-X enrichment analysis version 3

DE

Differential expression

DEG

Differentially expressed gene

dlPFC

Dorsolateral prefrontal cortex

ECM

Extracellular matrix

eQTL

Expression quantitative trait locus

FC

Fold change

FDR

False discovery rate

GO

Gene ontology

GO-BP

Gene ontology - biological process

GO-CC

Gene ontology - cellular component

GO-ORA

Gene ontology - overrepresentation analysis

GSEA

Gene set enrichment analysis

GWAS

Genome-wide association study

H&E

Hematoxylin & Eosin

HWE

Hardy-Weinberg equilibrium

IF

Immunofluorescence

L

Cortical layer

LR

Ligand-receptor

M

Meninge

MAD

Median absolute deviation

MAF

Minor allele frequency

MAGMA

Multi-marker analysis of genomic annotation

MCL

Markov clustering

MGE

Median ganglionic eminence

NTC

Neurotypical control

OPC

Oligodendrocyte progenitor cell

OR

Odds ratio

PNN

Perineuronal net

PC

Pericyte

PC

Principal component

PCA

Principal component analysis

PPI

Protein-protein interaction

PRECAST

Probabilistic embedding, clustering, and alignment for integrating spatial transcriptomics

PRS

Polygenic risk score

PRS-DEG

PRS-associated differentially expressed gene

PVALB

Parvalbumin

QC

Quality control

ROI

Region of interest

scDRS

Single cell disease relevance score

SCZ

Schizophrenia

SCZ-DEG

Schizophrenia-associated differentially expressed gene

SMC

Smooth muscle cell

SMR

Summary data based mendelian randomization

snRNA-seq

SIngle-nucleus RNA-sequencing

SpD

Spatial domain

SPG

Spatial proteogenomics

SRT

Spatially resolved transcriptomics

STRING

Search Tool for the Retrieval of Interacting Genes/Proteins

TF

Transcription factor

UMI

Unique molecular identifier

VLMC

Vascular leptomeningeal cell

WFA

Wisteria floribunda agglutinin

WM

White matter

WMtz

White matter transition zone