Distinct Cellular Phenotypes of Language and Executive Decline in Amyotrophic Lateral Sclerosis

SUMMARY

Cognitive manifestations, including impairments in language and executive functions, are seen in amyotrophic lateral sclerosis (ALS), but the underlying mechanisms remain unclear. We mapped prefrontal cortex regions from ALS patients by integrating spatial and single-nucleus transcriptomics in a cognitively stratified patient cohort. We uncover that cognitive impairment in ALS is associated with distinct patterns of neuronal dysfunction and glial-vascular dysregulation that vary by region and cognitive subtype. Executive dysfunction is linked to reduced mitochondrial and synaptic activity in deep layer dorsolateral prefrontal cortex neurons, whereas language-related deficits track with a diffuse pan-regional response involving glial and vascular abnormalities. Our analyses, validated by multiplexed imaging, further identify signatures in the prefrontal cortex that span both motor and cognitive phenotypes, including a multicellular gliosis response. The findings reveal that clinical heterogeneity in ALS is driven by phenotype-specific cellular interactions in motor and non-motor regions of the brain.

Graphical Abstract

In brief:

Beyond its motor dimension, ALS can also involve cognitive impairment, with affected domains varying across individuals. Rather than a single common disease process, this heterogeneity reflects regionally and cellularly distinct alterations in the prefrontal cortex that are only partially predicted by TDP-43 neuropathology.

INTRODUCTION

Amyotrophic lateral sclerosis (ALS) is a fatal neurodegenerative disorder characterized by loss of motor neurons leading to progressive paralysis¹. Half of all patients exhibit some degree of extra-motor involvement, and the clinical presentation exists on a spectrum from ALS cases with normal cognition and behavior to ALS with frontotemporal dementia (ALS-FTD)^2,3. A mid-spectrum range, which encompasses about 35% of ALS patients, is defined by mild cognitive and behavioral manifestations, including deficits in language, a decline in executive function, and apathy^2,3. Brain imaging and postmortem analyses have shown that these cognitive and behavioral profiles are predictive of underlying dysfunction in specific areas of the brain, highlighting potential regional susceptibilities to ALS pathophysiology^2,4–8. It remains unclear, however, whether the clinical phenotypes observed in ALS stem from a shared root mechanism or emerge from brain region-specific interactions of cellular, molecular, neuropathological, and genetic factors.

Addressing this question requires a transdisciplinary approach that bridges clinical phenotyping and neuropathology with cell-level profiling. In practice, however, studying the clinical heterogeneity of ALS has been challenging, due in part to the lack of detailed cognitive assessments in routine ALS care². Additionally, while technological advances have begun to provide a multicellular understanding of ALS^9–13, disease-related interactions and dependencies between cells remain incompletely understood. Here, we apply an integrative spatial biology analysis approach to a cognitively phenotyped ALS patient cohort to interrogate cellular and molecular changes linked to impairment. Our findings reveal cognitive subtype-specific signatures and highlight a set of complex relationships between clinical phenotype and multicellular dysfunction in ALS.

RESULTS

Study Overview and Cohort Demographics

We present a unique cohort of ALS patients who underwent neuropsychological testing using the Edinburgh Cognitive and Behavioral ALS Screen (ECAS). This clinical assessment tool quantitatively evaluates cognitive domains commonly affected by ALS, including executive function, language, and verbal fluency^2,3,14. Cognitive deficits in ALS are associated with localized cerebral and subcortical dysfunction, and the profile of cognitive change in each patient is predictive of underlying regional abnormalities^2,4–8. We leveraged this structure-function relationship to investigate molecular and cellular signatures linked to domain-specific deficits in ALS. We sampled from BA46 in the DLPFC, which supports decision-making and other executive processes^5,15–17, and BA44/45 in the inferior frontal gyrus of the left hemisphere (Broca’s area), given its strong links to language production¹⁸. We combined spatial transcriptomics (ST) and single-nucleus RNA sequencing (snRNA-seq) to obtain a comprehensive view of ALS-related changes in non-motor areas of the brain (Figure 1A).

Figure 1: Study overview and cohort demographics.

A) ALS cases were cognitively stratified using the Edinburgh Cognitive and Behavioral ALS Screen (ECAS); donors with abnormal scores were classified as cognitively impaired (ALSci). Controls are non-neurological cases of sudden death (n=17).

B) Distribution of ECAS subtest scores for each ALS donor; dotted lines indicate cutoff for impairment.

C) Relationship between cognitive status and age at death.

D) Kaplan-Meier plot showing disease duration, stratified by cognitive status.

E) Stratification by C9ORF72 genotype. The ECAS ALS-specific score is the sum of the executive, verbal fluency, and language scores.

F) Stratification by TDP-43 neuropathology.

G) UMAP of ST data. Points are Visium array spots, color-coded by cortical layer annotation. Seven arrays from a single donor, spanning both BA44/45 and BA46, separate from the rest of the dataset and were excluded from downstream analyses.

H) Representative spot plots showing transcript expression overlaid on H&E-stained tissue section. The summary layer plots are the average marker gene spatial distribution for all Visium arrays.

I) Number of cortical layer spots for each Visium array, color-coded by donor phenotype. The proportion of all spots assigned to each layer is annotated. Black bar indicates the median spot count per layer.

J) UMAP plot of paired snRNA-seq and snATAC-seq data (BA44/45 n = 24; BA46 n = 24), colored by broad cell type.

K) Number of annotated cell types color-coded by donor phenotype, with proportions indicated for each cell type.

L) Distribution of Visium spots by cortical layers (top) and snRNA-seq cell types (bottom).

Brain tissue was sourced from a single biorepository – the Edinburgh Brain Bank – ensuring consistent preservation and neuropathological assessment. Based on ECAS scores, donors were stratified into three groups: ALS patients with cognition in the normal range (hereafter ‘ALS’; n = 19), representing a pure motor presentation; ALS patients with impairment in one or more cognitive domains (‘ALSci’; n = 17); and age-matched sudden-death neurotypical controls (n = 17)¹⁹. ALSci donors do not meet the diagnostic criteria for ALS-FTD but fall within an intermediate range along the ALS-FTD spectrum^2,3.

We provide clinical phenotyping and disease metrics for this cohort (Figures 1B–F; see also Data S1 Figure 1). Language impairment was the most frequently affected domain among ALSci donors, consistent with published literature (Figure 1B)^14,20. The number of donors was evenly distributed across clinical groups, and we observed no correlation between donor age at death or disease duration with cognitive status (Figures 1C,D). Survival duration from disease onset was comparable between ALS and ALSci patients (Figure 1D). Genotyping and immunohistochemistry analyses showed that neither C9ORF72 repeat expansion nor TDP-43 pathology predicted cognitive status in our cohort (Figures 1E,F; Figures S1E,G).

A Multimodal Cellular Atlas of the ALS Prefrontal Cortex

ST mapping was performed using 10x Genomics’ Visium platform. We processed, sequenced, and analyzed 246 Visium arrays (BA44/45: n = 137, BA46: n = 109) from 45 individual donors (Data S1 Figures 2A–F). Each brain region was represented by at least two tissue sections, with most having higher replication (median of four arrays per region per donor). In total, 953,073 discrete 55 μm-wide capture spots passed quality control (BA44/45: mean 3,447 unique RNA molecules [UMIs], 1,572 genes per spot; BA46: mean 3,679 UMIs, 1,621 genes per spot) (Data S1 Figures 2A–C). Each spot was manually assigned to one of seven histological classes (layers L1-L6 and white matter [WM]) based on hematoxylin and eosin (H&E) image features. The quality of layer annotations was assessed using UMAP embedding (Figure 1G), as well as previously validated layer-specific gene markers (Figure 1H). Spot distribution across each layer was consistent between control and disease donors (Figure 1I), suggesting no gross structural changes linked to ALS. These analyses identified an outlier donor with atypical clustering, who was excluded from downstream analyses due to abnormally high interferon and antiviral gene expression (Figure 1G; see also Data S1 Figure 2G–H).

In parallel, we used 10x Genomics’ Multiome technology to obtain paired snRNA-seq and snATAC-seq profiles (BA44/45: n = 24, BA46: n = 24). We also assembled a collection of 2.6 million prefrontal cortex snRNA-seq profiles from publicly available sources^9,12,21–27, on which we trained a machine learning-based classifier to predict cell-type annotations and remove low-quality or doublet cells from our Multiome dataset (Data S1 Figure 3). After preprocessing and annotation (Data S1 Figure 4), we assigned 144,260 high-quality nuclei to seven broad cell types: excitatory neurons (EN), inhibitory neurons (IN), oligodendrocytes (OD), oligodendrocyte progenitor cells (OPC), astrocytes (AST), microglia (MIG), and endothelial cells (ENDO) (Figure 1J). Neuronal and neuroglial populations were evenly represented, with excitatory neurons (33.8%), oligodendrocytes (24.5%), and astrocytes (13.8%) comprising the most abundant populations (Figure 1K). All broad cell types were consistently detected across individuals and conditions (Figures 1K–L), and did not show substantial differences between ALS and control donors.

To assess sample integrity and exclude potential confounding effects, we performed extensive quality control analyses across both modalities. Overall, RNA Integrity Numbers (RINs) obtained from bulk RNA extractions showed no systematic association with clinical phenotype, demographic variables, or post-mortem interval (24–231 hours across the cohort; mean, 75 hours), and were not correlated with key ST or snRNA-seq quality metrics (Data S1 Figure 5), indicating minimal impact of technical or clinical covariates on downstream analyses. Finally, cell composition was broadly consistent across groups across both modalities (Figure S6). Cell2location-based deconvolution of Visium data did suggest a marginal decrease in IN subtypes in ALS (Figure S6)²⁸, but this was not corroborated by immunohistochemistry (Figure S7), likely reflecting reduced expression of canonical IN transcripts within corresponding subtypes rather than neuronal loss (Figures S6E,F).

Glial Reactivity is a Frequent Feature of the ALS Prefrontal Cortex

Identifying responses shared across ALS patients regardless of cognitive status is critical for distinguishing cognitive impairment-specific mechanisms from broader disease processes. Among all cell types, endothelial cells showed the most pronounced transcriptional changes, with differential expression analysis revealing enrichment in gene ontology (GO) terms related to immune response, apoptosis, and angiogenesis in ALS donors (Figure 2A). The rarity of endothelial cells in the snRNA-seq dataset (1.1% of captured nuclei; Figure 1K) limited our ability to resolve vasculature-associated subtypes or draw detailed inferences using this modality. This undersampling of vascular cells is a well-recognized limitation of current protocols^27,29.

Figure 2: Glial Reactivity is a Frequent Feature of the ALS Prefrontal Cortex.

A) (i) Endothelial cell snRNA-seq data as a UMAP plot colored by donor phenotype. (ii) Heatmap of gene expression differences (pseudobulk DESeq2; multiple hypothesis test corrected Wald test; p < 0.05) between control and ALS or ALSci donors, from selected ontology categories.

B) Astrocyte snRNA-seq UMAP plot colored by (i) donor phenotype, (ii) transcriptome-inferred functional state (homeostatic [GFAP^Low] or reactive [GFAP^High]), and (iii) key functional marker genes. Dot plot is colored by mean gene expression and dot size indicates the proportion of cells with non-zero counts, from 0% to 100%.

C) Oligodendrocytes classified as low-myelinating (PLP1^Low) and high-myelinating (PLP1^High), plotted as in B.

D) Microglia classified as homeostatic (P2RY12⁺) and reactive/inflammatory (TMEM163⁺), plotted as in B. Phagocytic, stress-response, and non-microglial immune cells are also plotted in (ii) without labels.

E) OPCs classified as homeostatic (NRXN1⁺), maturing (OLIG1⁺), and stressed (NCOA1⁺), plotted as in B.

F) The proportion of glial cells for each tissue sample for control, ALS, and ALSci donors was averaged for plotting and tested for significant differences (Welch’s t-test, p < 0.05), compared to controls.

G) Proportion of activated-to-total microglia (i) and stressed-to-total OPCs (ii) plotted against the proportion of reactive-to-total astrocytes for each tissue sample, colored by phenotype, and annotated with regression line and Pearson correlation.

H) Pearson correlation matrix (multiple hypothesis test corrected Wald test, p < 0.05) between the proportions of functionally classified-to-total glial cells for each sample. Black boxes highlight correlations shown in G.

I) Spatial distribution of glial functional module scores across Visium arrays (Wald test with multiple-hypothesis correction).

J) ECAS scores and donor age, grouped by presence (+) or absence (−) of reactive gliosis (derived from snRNA-seq data), defined as >80% reactive astrocytes or >20% reactive microglia (Welch’s t-test, p < 0.05).

We next annotated distinct functional classifications for each major glial cell type based on Leiden clustering, key marker gene expression, and cross-referencing with previous analyses (Figures 2B–E; see also Data S1 Figure 6). We classified astrocytes into homeostatic and reactive states based on published transcriptional profiles (Figure 2B; see also Data S1 Figures 6A–B). Oligodendrocytes were classified into high- and low-myelinating states based on expression of myelin-associated genes (Figure 2C; see also Data S1 Figure 6B). Microglia were divided into homeostatic and reactive/inflammatory states, with smaller subsets exhibiting phagocytic or heat shock-associated signatures (Figure 2D; see also Data S1 Figure 6B). A small population of non-microglial immune cells present among nuclei broadly classified as microglia was annotated separately as ‘other immune’ (Data S1 Figure 6B). Finally, we classified OPCs into homeostatic and OLIG1⁺ maturing/differentiating states, along with a distinct NCOA1⁺ subpopulation which we have named ‘stressed’ (Figure 2E; see also Data S1 Figure 6B).

We then compared disease and control donors to determine whether the proportion of glia in functional states differed. We observed a significant increase in the proportion of reactive microglia and stressed OPCs in both ALS and ALSci donors, while that of reactive astrocytes significantly increased in ALSci donors only (Figure 2F). Interestingly, while the proportion of reactive astrocytes increases in ALSci donors, the transcriptome of reactive astrocytes in ALS and ALSci donors was similar (Data S1 Figure 7F). Also of note, the proportion of reactive glia were highly correlated within individual samples (Figure 2G; see also Data S1 Figure 6H), suggesting a coordinated response. The proportions of certain homeostatic glial subpopulations were also correlated, including homeostatic astrocytes with microglia and between maturing OPCs with high-myelinating oligodendrocytes (Figure 2H). By contrast, the proportion of homeostatic microglia and astrocytes did not correlate with that of maturing OPCs or myelinating oligodendrocytes (Figure 2H).

We next examined the spatial distribution of this shared gliotic response in the ST dataset. We identified transcripts that were both specific to individual glial cell types and differentially expressed between functional states of that glial cell type (Data S1 Figure 7G), which enabled us to quantify shifts in glial cell states by scoring each spatial spot. Reactive astrocytes and reactive microglia were significantly increased in ALS and ALSci across layers L1-L6 and WM (Figure 2I), while stressed OPCs increased from L2 through WM, indicating a pan-layer gliotic response. In contrast, myelinating oligodendrocyte scores decreased and maturing OPC scores increased in L2-L6, but both populations were largely unchanged in WM (Figure 2I). Supporting these observations, analysis of the snATAC-seq dataset revealed that reactive astrocytes and stressed OPCs from ALS donors exhibit enrichment of AP-1 binding motifs at regions with increased chromatin accessibility (Figure S3), consistent with activation of gliotic and inflammatory transcriptional programs^30,31. Together, the findings support the idea that glial responses in ALS are both coordinated and spatially organized.

While we detected gliosis signatures in both ALS and ALSci donors, they were not found in every patient (Figure 2G; Figure S4A), and, when present they were generally more widespread in ALSci (Figures 2F,G,I; Figure S4A). This patient-level heterogeneity raised the question of whether specific cognitive profiles tracked with this gliotic response. To address this, we categorized each sample in the snRNA-seq dataset as positive or negative for reactive microglia or astrocytes (AST/MIG gliosis). Across cognitive domains, ECAS scores were largely similar between donors with and without reactive gliosis, except for language function, which was significantly lower in the gliosis-positive group (Figure 2J). Stratification by brain region showed that this association was specific to BA44/45 (Figures S4B–D). Donor age was similarly not associated with gliosis status (Figure 2J).

Convergent Transcriptional and Morphological Signatures of Reactive Gliosis

We complemented our transcriptomic analyses with whole-tissue multiplexed immunofluorescence imaging using Lunaphore’s COMET platform, which sequentially applies, images, and elutes antibody pairs (Figure 3A). Our panel included markers for astrocytes, microglia, neurons, and vasculature (Figure 3A). We profiled formalin-fixed, paraffin-embedded (FFPE) tissue from both BA44/45 and BA46 and acquired 12.5 × 12.5 mm scans spanning gray and white matter (Figures 3B,C), which we subsequently annotated, cell typed, and quantified using HALO (Figure 3D). This dataset includes ~3,000,000 cells (based on DAPI-stained nuclei counts) from 17 donors spanning control, ALS, and ALSci cases, with both BA44/45 and BA46 represented for nearly all individuals (31 whole-tissue sections in total).

Figure 3: Sequential Immunofluorescence Identifies Convergent Transcriptional and Morphological Signatures of Reactive Gliosis in ALS and ALSci.

A) Sequential immunofluorescence workflow and antibody panel.

B) Full 12.5mm by 12.5mm tissue image of BA44/45 section. Donor is a female ALS patient without cognitive impairment. Scale bar, 2 mm.

C) Magnified view of the boxed region in panel B. Scale bar, 50 μm.

D) Tissue-wide cell segmentation and phenotyping.

E) GFAP-stained astrocytes across clinical groups. Scale bar, 50 μm.

F) Astrocyte reactivity quantified from the mean pan-layer astrocyte morphology for each tissue (Welch’s t-test, p < 0.05). Black bar is the mean reactivity across donors.

G) Difference in morphology-quantified astrocyte reactivity for each layer, compared to controls (Welch’s t-test, p < 0.05).

H) Donor-level Pearson correlations between imaging-based and ST-derived astrocyte reactivity scores, plotted per layer (Wald test, p < 0.05).

I) IBA1-stained cells across clinical groups. Scale bar, 25 μm.

J) Mean pan-layer proportion of IBA1⁺ cells as a fraction of DAPI⁺ nuclei (Welch’s t-test, p < 0.05). Black bar is the mean score across donors.

K) Donor-level Pearson correlations between IBA1⁺ proportion and ST-derived microglial reactivity scores, computed separately for each cortical layer (Wald test, p < 0.05).

L) Mean tissue IBA1⁺ area as a percentage of total area across gray and white matter broken by clinical group (Welch’s t-test, p < 0.05).

Astrocytic reactivity is characterized by both molecular changes and morphological alterations, including hypertrophy and altered process ramification^32,33. In ALS, reactive astrogliosis is observed around both upper and lower motor neurons and extends into cortical and subcortical regions of the frontal cortex and other non-motor areas^32,33. To quantify astrocyte morphology, we trained a classifier on GFAP-immunostained profiles using manually annotated examples from healthy and diseased donors, spanning both gray and white matter (Data S1 Figures 8A–D). Subsequently, as a metric for astrocyte reactivity, we calculated cellular roundness (also referred to as circularity) from cell area and perimeter (Figure 3E)³⁴. Compared to controls, astrocytes in disease donors were significantly less spherical (i.e., more ramified) (Figures 3E,F). This shift was observed across cortical layers and in white matter, and was largely shared across disease groups – albeit with stronger involvement of BA46 L1-L5 in ALS than ALSci – confirming a widespread gliotic response (Figure 3G; Figures S5A–C). While this morphometric descriptor may not fully resolve qualitative differences between astrocytic states, it correlated strongly with ST-derived signatures of astroglial reactivity (Figure 3H; Figures S5I–N). This convergence validates our original observations and supports transcriptional alterations and morphological remodeling as parallel features of astrocyte activation in ALS.

To investigate microglial responses, we analyzed IBA1 immunoreactivity, a general marker of microglia and CNS macrophages (Figure 3I). Overall, the number of IBA1-positive (IBA1⁺) cells, averaged across gray and white matter, did not significantly differ between disease and control donors (Figure 3J; Figures S5D–G), although the number of IBA1⁺ cells was significantly correlated with the ST-derived microglial reactivity scores in L2-L6 (Figure 3K). To complement these analyses, we examined whether microglial alterations in ALS were accompanied by morphological remodeling. Due to the difficulty in segmenting long microglial projections, we used total IBA1⁺ tissue area as a proxy for microglial tissue coverage (Data S1 Figure 8E). This revealed a significant reduction in IBA1⁺ area in ALS and ALSci samples, with the strongest effect in L1-L4 (Figure 3L; Figures S13H). We interpret the decrease in IBA1⁺ coverage area without decreases to IBA1⁺ cell abundance as a shift from a ramified, homeostatic morphology to an amoeboid, reactive morphology. This analysis cannot resolve the fraction of individual cells undergoing morphological remodeling, but indicates a broad alteration in microglial dynamics in ALS.

Spatial Expression Signatures Distinguish ALS Cognitive Subtypes

At the molecular level, the CNS is characterized by spatially correlated transcriptomic patterns that reflect cellular interactions^34,35. Disruptions in these spatial expression programs have been linked to neurological conditions^36,37, including ALS in the spinal cord¹³. To investigate ALS-related changes in non-motor areas of the brain in an unbiased manner, we employed Splotch^13,38,39, a Bayesian hierarchical model designed for ST data (Figure 4A). Splotch models gene expression at each spot as a zero-inflated Poisson distribution, with spot-level λ (transcript abundance) informed by spot- and donor-level metadata (cortical layer and ALS diagnosis), spatial autocorrelation with neighboring spots, and spot-specific effects (Figures S6A–C). After inference, we quantified both the magnitude (log₂ fold-change) and significance (Bayes factor) of gene expression changes between control and disease donors within each cortical layer by analyzing posterior distributions over model parameters (Figures S14D–E). Globally, λ parameters from the fitted Splotch model – representing denoised estimates of transcript expression at every spot – exhibited similar spatial correlation patterns in both BA44/45 and BA46, consistent with the shared laminar architecture and cellular composition of these regions (Figure S7A).

Figure 4: Spatial Expression Signatures Distinguish ALS Cognitive Subtypes.

A) The Splotch model is fit to each gene in the spatial transcriptomic data set. Transcript expression (λ) at each spatial spot on each array is modeled as a combination of layer and sample covariate effects (β), a spatial autoregressive component (φ), and a spot noise (ε) component. Distance between genes is calculated from modeled spot expression (λ), and clustered with the Leiden community clustering algorithm to identify spatially correlated gene modules. Genes are colored by spatial module in a UMAP plot.

B) Biological process (BP) GO terms with the smallest p-value for each spatial gene module. Module 13 has no BP term enrichment; the top Cellular Component (GO:CC) ontology is shown instead.

C) Heatmap showing the proportion of transcripts specific for reactive astrocytes and microglia, low-myelinating oligodendrocytes, and stressed OPCs in each spatial module. Modules with >50% of function-specific transcripts are highlighted in red.

D) Pearson correlations between donor-aggregated gene module and ECAS scores (Benjaminini-Yekutieli corrected Wald test; p < 0.01).

E) Summary of selected spatial modules. Left to right, the average score of the spatial gene module from snRNA-seq expression data is plotted for each snRNA-seq cell type, and the snRNA-seq UMAP is plotted where each dot is an individual cell colored by the module score. Average score of the spatial gene module from spatial expression data is plotted for each cortex layer, and a representative example array is shown with module scores overlaid onto each spot. Differential module expression is plotted from spatial expression data, comparing control donors to ALS and ALSci donors (Benjaminini-Yekutieli-corrected Wald test; p < 0.01).

F) Spatial module M2 genes clustered into submodules using snRNA-seq data. UMAP plot shows M2 genes colored by submodule.

G) Average expression of M2 genes across snRNA-seq-defined cell types. Expression values are min-max scaled per gene for visualization. Submodule identity is shown at right.

H) Pearson correlations between M2 submodule scores by clinical phenotype.

I) Top enriched GO:BP terms for each M2 submodule. Blank entries indicate no GO:BP enrichment.

J) M2 submodule expression relative to controls by cortical layer (Benjaminini-Yekutieli-corrected Wald test; p < 0.01).

K) Heatmap of differentially expressed genes from endothelial M2 submodules in endothelial snRNA-seq nuclei (pseudobulk DESeq2; multiple testing-corrected Wald test; p < 0.05).

Beyond identifying individual genes with differential expression, we sought groups of spatially co-expressed genes. We embedded each transcript into a k-Nearest Neighbors (k-NN) graph using Pearson correlation distance, and performed Leiden clustering to derive co-expressed spatial modules (Figure 4A; Figure S7B). This yielded 23 modules, each representing genes with overlapping spatial patterns across the cortex. Two modules, capturing chromosome-specific gene expression (Y chromosome [chrY] and mitochondrial [chrM]), were excluded from subsequent analyses (Figure 4A; Figure S7C).

We functionally characterized the remaining 21 spatial modules based on GO enrichment (Figure 4B). These included immune response-linked modules (M2 and M21), a myelination-linked module (M5), mitochondria-associated modules (M8 and M9), and several synapse- or receptor signaling-related modules (M1, M4, M6, M7, M16 and M19). Notably, M2, M12, and M20 were enriched for transcripts associated with reactive astrocytes and microglia (Figure 4C), consistent with the gliotic response we previously characterized (Figure 2 and Figure 3). This suggests that the identified co-expression modules capture both spatial and functional relationships.

For each spatial module, we calculated module scores summarizing the aggregate expression of all module transcripts for each ST spot. We then correlated spatial module scores with cognitive scores to localize their regional specificity (Figure 4D). Impairment in executive function correlated with downregulation of M8 and M9 in BA46, two modules associated with mitochondrial function, particularly oxidative metabolism (Figures 4B,D). Verbal fluency impairment, by contrast, correlated with upregulation of several modules in BA44/45 (Figure 4D). Differently, deficits in language function correlated with upregulation of gliosis-linked modules M2 and M20 across both regions (Figures 4C,D).

To explore the association between spatial modules and cellular processes, we calculated module scores within our snRNA-seq data and determined the cell type associations for each module (Figure 4E; Figure S7B). We also evaluated module scores across cortical layers and sought to identify disease-related changes by comparing control to ALS and ALSci donors (Figure 4E; Figure S7D). Because Visium spots often sample multiple cells, our spatial co-expression modules can be associated with multiple distinct cell types. We further separated them into submodules by performing k-NN graph embedding on snRNA-seq transcript correlation distance and reclustering (Figure 4F; Figure S7D). We selected spatial module M2 for further exploration as it was upregulated in ALS and ALSci donors and associated with both language and verbal fluency dysfunction (Figures 4D–E). Submodule clustering of M2 revealed transcriptional signatures related to endothelial, microglial, and neuronal nuclei (Figure 4G). Notably, we observed a progressive increase in spatial cross-correlation between M2 submodules from control to ALS to ALSci donors, suggesting that cognitive impairment is linked to the disruption and an intermixing of cellular processes that are typically spatially segregated (Figure 4H; see also Data S1 Figure 9). GO analyses further revealed that specific M2 submodules were enriched for microglial immune responses, endothelial vascular development, and neuronal protein folding functions (Figure 4I). These submodules were upregulated in ALS, with generally more pronounced effects in ALSci individuals (Figures 4J,K). These spatially coordinated vascular, neuronal, and glial changes implicate disrupted intercellular interactions as key contributors to language and fluency impairment in ALS.

Disease-associated Changes in the Perivascular Microenvironment

Building on the vascular-immune link from our spatial analyses, we next examined IBA1⁺ myeloid cell distribution and activation around the vasculature in ALS using our imaging dataset. We trained a classifier to identify blood vessels based on alpha smooth muscle actin (aSMA), which labels vascular smooth muscle cells, and Claudin5, a marker of vessel wall inner liner endothelial cells. Because aSMA is expressed in arterioles (and to a lesser extent venules) and largely absent from capillaries, this analysis specifically captured these larger vessel types. We demarcated the perivascular microenvironment around each aSMA-ensheathed vessel with concentric rings (40-μm radius; 20 rings at 2-μm increments) and quantified IBA1⁺ area as a function of distance from vessels. To increase statistical power, we aggregated rings into four segments (0–10 μm [most proximal], 10–20 μm, 20–30 μm, and 30–40 μm [most distal]) (Figure 5A).

Figure 5: Perivascular Microglial Reactivity Differs Between ALS and ALSci.

A) Schematic of the perivascular analysis workflow, including classifier-based vessel annotations and demarcation of the vascular microenvironment. Concentric rings (2-μm increments, 2–40 μm from the vessel wall) define vessel-proximal domains.

B) IBA1⁺ area averaged across donors and plotted against distance from vessels across cortical layers and clinical groups (Mann-Whitney U test; p < 0.05).

C) CD68 activation scores averaged across donors and plotted against distance from vessels as in Figure 5B (Mann-Whitney U test; p < 0.05).

D) IBA1- and CD68-stained perivascular space across clinical groups. Scale bar, 20 μm.

In control donors, IBA1⁺ area exhibited a clear vessel-proximal enrichment (Figures 5B,D). In ALS and ALSci, however, IBA1 distribution near the vessel was decreased relative to controls, particularly in the shallower layers of the cortex (L1-L4) and WM (Figures 5B,D; Figure S8A). The reduction persisted across all radial segments, indicating a broad loss of vessel-proximal IBA1⁺ signal.

We next asked whether vessel-adjacent myeloid cells differed across our cohort at the functional level. To this end, we focused on CD68, a lysosomal glycoprotein and marker of phagocytic activity in microglia that clustered into spatial module M2^40,41. First, donor-specific CD68 intensity thresholds (low, mid, and strong) were established to overcome inter-individual variability in marker detection related to tissue preservation (Figure S8B). We then quantified CD68 activation as the fraction of strongly CD68-positive area within the total CD68-positive area, computed per radial segment. Elevated perivascular CD68 signal was selectively observed in ALSci donors (Figures 5C,D; Figure S8C), while control and ALS individuals showed comparatively low CD68 near vessels, suggesting differential activation of vascular-associated myeloid cells in ALSci.

Together, these findings reveal changes in the perivascular microenvironment that are shared across ALS patients, as well as those that are unique to individuals with cognitive impairment.

Executive Impairment is Linked to BA46 Deep-layer Neurons

While many spatial gene modules (e.g., M2) showed links to verbal fluency and language deficits, a smaller and distinct set correlated with executive dysfunction, suggesting differential mechanisms associated with this impairment. To further dissect the cellular and molecular basis of executive deficits in ALS, we focused on M8 and M9 – the only spatial modules significantly correlated with ECAS executive function scores (Figure 4D). Both modules are enriched for mitochondria-associated genes and linked to excitatory and inhibitory neurons (Figure 4B; Figure S7D). They are also downregulated in ALS patients with executive impairment, specifically in BA46 (Figure 4D). Stratification of spatial module correlations by cortical layer further revealed that this response was localized to the deeper layers (L5 and L6) of the gray matter (Figure 6A), suggesting that executive deficits in ALS are tied to dysfunction in spatially distinct neuronal subpopulations. These findings prompted a more detailed investigation of the underpinnings of executive dysfunction in ALS.

Figure 6: Executive Impairment is Linked to BA46 Deep-layer Neurons.

A) Pearson correlation between spatial co-expression module scores and ECAS scores, stratified by region and cortical layer (Benjaminini-Yekutieli-corrected Wald test; p < 0.01).

B) Pearson correlation between ST co-expression module scores and donor age, compared with correlation to ECAS executive scores (Benjaminini-Yekutieli-corrected Wald test; p < 0.01).

C) Pearson correlation between spatial module M8 and M9 scores and donor age (left), or ECAS executive score (right), stratified by region and cortical layer (Benjaminini-Yekutieli-corrected Wald test; p < 0.01).

D) Pearson correlation between M8 and M9 scores (calculated from snRNA-seq data) and donor age (left), or ECAS executive score (right), stratified by region and broad cell type (Benjaminini-Yekutieli-corrected Wald test; no significant correlations).

E) UMAP plot of excitatory neurons, colored by subtype.

F) UMAP plot of inhibitory neurons, colored by subtype.

G) Log₂ fold change of M8 and M9 scores (calculated from snRNA-seq data), stratified by region and neuronal subtypes. (i) Comparison of older age groups to the young group. Age categories were defined to stratify donors into three equal-sized groups. (ii) ALS and ALSci versus control donors (Benjamini-Hochberg-corrected Welch’s t-test, p < 0.05).

H) snRNA-seq Log₂ fold change of CALM2 (the highest expressed module M8 transcript) and GAPDH (the highest expressed module M9 transcript), stratified by region and neuronal subtype (Benjamini-Hochberg-corrected Welch’s t-test, p < 0.05).

We first examined the influence of aging on modules M8 and M9 to distinguish disease-driven executive impairment from age-related cognitive decline^42,43. Donor age correlated with several modules, including M8 and M9, both of which showed decreased expression in older donors and in those with lower executive function scores (Figure 6B). However, the spatial patterns of age- and disease-related M8/M9 changes were distinct: age was associated with reduced expression across layers L1-L5 in both BA44/45 and BA46, whereas executive function correlated with L5-L6 expression in BA46 alone. This highlights the importance of spatial profiling in transcriptomic analyses, as the same molecular processes (e.g., modules M8/M9) can be perturbed in distinct patterns that distinguish age- from disease-driven cognitive decline.

We next asked whether specific neuronal subpopulations were linked to executive impairment in ALS. From our snRNA-seq dataset, we separately clustered ENs and INs, and assigned each cluster a subtype based on the expression of key marker genes (Data S1 Figure 10)⁴⁴. We identified 9 well-characterized EN subtypes, including L2/3, L4, L5, L6, and L6 Car3 intratelencephalic (IT) neurons, L5 extratelencephalic (ET) neurons, L5/L6 near-projecting (NP) neurons, L6 corticothalamic (CT) projection neurons, and L6b neurons (Figure 6E). For INs, we identified 9 well-characterized subtypes: VIP⁺, basket and chandelier PVALB⁺, three LAMP5⁺ subtypes, and three SST⁺ subtypes (Figure 6F), annotated as Martinotti (MC), non-Martinotti (nMC), and long-range projection (LRP) interneurons based on expression of comparable mouse marker transcripts (Figure 6F)⁴⁵. The relative proportion of excitatory and inhibitory neuronal subtypes inferred from this dataset was not significantly different between control, ALS, and ALSci donors (Data S1 Figures 10C,E).

For each neuronal subtype, we compared spatial module scores between three age groups, as well as between control and ALS donors (Figure 6G). In line with our layer-level analysis above (Figure 6C), M8 and M9 showed decreased expression in older donors in L2-L3, L4, and L5 IT ENs, and in L1 LAMP5⁺/PAX6⁺ INs across both BA44/45 and BA46 (Figure 6Gi). We additionally found that L5 IT, L6 IT, and L6 IT Car3 ENs exhibited significant downregulation of M8 and M9 genes in ALSci donors, consistent with executive impairment-associated changes previously inferred in deeper BA46 layers (Figure 6Gii). L5-L6 LAMP5⁺/LHX6⁺ and L5 PVALB⁺ basket inhibitory neurons likewise showed decreased M8 and M9 gene expression in ALSci donors (Figure 6Gii). In contrast, M8- and M9-linked transcripts were not significantly affected in other neuron subtypes located to L5 or L6, including L5/L6 NP, L6b, and L6 CT ENs, nor in other EN and IN subtypes (Figure 6G). As a specific example, the expression of CALM2 (calmodulin) and GAPDH, the highest-expressed genes in M8 and M9, respectively, were downregulated in BA46 L6 IT Car3 ENs and BA46 LAMP5⁺/LHX6⁺ INs from ALSci donors (Figure 6H), with CALM2 additionally downregulated in BA46 L5 IT and L6 IT cells. Finally, M8 and M9 scores showed a possible association with NeuN⁺ (encoded by RBFOX3) neuronal cell counts in the BA46 imaging dataset, but no relationship was observed between these modules and MAP2 distribution (Figure S9). The biological implication of this difference remains unclear, and our study is not powered to resolve it.

We conclude that while transcriptional alterations linked to increased age and executive dysfunction are similar, they are spatially distinct. Age-related changes are associated with neurons in shallower layers (L1-L5), whereas ALSci-linked changes are associated with neurons in deeper layers (L5-L6) of the cortex. Critically, executive dysfunction-linked transcriptional shifts are strongly associated with the BA46 region, unlike age-linked changes, which appear similar across both BA44/45 and BA46 (Figure 6C). Apart from age, other confounding donor variables (e.g., ALS disease duration or PMI) did not correlate with ALS or ALSci-linked spatial gene expression modules in either the ST data or in snRNA-seq-defined neuronal subtypes (Figure S10; see also Data S1 Figure 11).

Executive Dysfunction is Associated with Mitochondrial and Synaptic Alterations

These findings implicate selective deep-layer neuronal subtypes in the pathophysiology of ALS-linked executive impairment. To address the nature of this dysfunction, we examined gene expression changes in these neuronal subtypes (Figure 7A; see also Data S1 Figures 11E,F). Across all five EN and IN subtypes, transcripts differentially expressed in ALSci BA46 were enriched for three top GO terms: presynapse (GO:0098793), postsynapse (GO:0098794), and mitochondrial membrane (GO:0031966). Within both synaptic categories, pathways related to the synaptic vesicle cycle (GO:0099504) and synaptic organization (GO:0050808) were particularly prominent (Figure 7A). ALSci donors showed decreased presynaptic gene expression, particularly V-ATPase genes involved in endolysosomal and synaptic vesicle acidification, and increased expression of postsynaptic glutamate receptor subunits (e.g., GRIK1, GRIA1, and GRM1) (Figure 7A). We also detected strong downregulation of oxidative phosphorylation pathways and alterations in PINK1-Parkin-mediated mitophagy (Figure 7A). Oxidative phosphorylation and synaptic-related transcripts, including glutamatergic and GABAergic receptors, correlated within individual neuronal subtypes across donors (Figure 7B), indicating that these changes co-occur in the same tissue rather than reflecting independent shifts in different patient subgroups.

Figure 7: Executive Dysfunction is Associated with Mitochondrial and Synaptic Alterations.

A) snRNA-seq differentially expressed genes (pseudobulk DESeq2; multiple test-corrected Wald test; p < 0.05) in neuronal subtypes linked to executive impairment. GO enrichment is plotted as −log₁₀(p-value) for the top three GO functional terms. Differentially expressed genes annotated to these ontologies are plotted as log₂ fold change (ALS or ALSci vs. control). TDP-43 binding targets (eCLIP-defined) are marked by black bars at left.

B) snRNA-seq Ontology-based correlation analysis of synaptic and oxidative phosphorylation gene sets. Mean ontology scores from neuronal cells for each donor sample are plotted. Scatter plots show oxidative phosphorylation scores on the x-axis and another ontology score on the y-axis, with best-fit regression line plotted for each neuronal subtype and average Pearson correlation coefficients (r) across the five neuronal subtypes. Plots with correlations significantly different from zero (Wald test; p < 0.05) are shaded in lilac.

C) Differential module expression from ST data, stratifying ALS donors by cognitive or TDP-43 neuropathological profile (Benjaminini-Yekutieli-corrected Welch’s t-test; p < 0.01).

D) Cross-region, donor-level comparison of glial functional subtype proportions between BA44/45 and BA46 (n = 14 matched samples). Significant correlations are shaded in lilac (Wald test; p < 0.05).

E) Proportions of reactive microglia, reactive astrocytes, high-myelinating oligodendrocytes, and maturing OPCs in each donor sample are plotted against oxidative phosphorylation ontology expression scores, calculated as in Figure 7B. Significant correlations are shaded in lilac (Wald test; p < 0.05).

F) Schematic highlighting region- and cortex layer-specific cellular and molecular differences between language impairment, verbal fluency impairment, and executive impairment.

TDP-43 aggregation, a pathological hallmark of ALS, has been implicated in disease-driving gene expression changes⁴⁶, suggesting it may contribute to the transcriptional alterations observed in these neuronal subtypes. To investigate this, we intersected differentially expressed genes from our dataset with TDP-43-associated targets, including mRNAs identified by eCLIP-seq⁴⁷, protein-protein interactors⁴⁸, and genes affected by nuclear TDP-43 depletion in postmortem FTD-ALS brains⁴⁹ (Figure 7A; Figure S11A). Of these three sets, only eCLIP-defined direct mRNA targets were significantly overrepresented. TDP-43 eCLIP targets were predominantly upregulated (Figure 7A; Figures S11A–C), and GO enrichment analysis identified the postsynapse as the most significantly affected cellular compartment (Figures S11D,E). Conversely, protein-protein interactors and nuclear TDP-43 depletion-associated genes did not show significant enrichment (Figure S11B).

TDP-43 Pathology and Cognitive Impairment Have Partially Distinct Molecular Signatures

These observations led us to ask how broadly TDP-43 pathology contributes to the transcriptional responses observed in the ALS prefrontal cortex. In our earlier analyses, several spatial modules associated with neuronal function were upregulated in both ALS and ALSci (Figure S7D), which we interpreted as gene expression changes shared across ALS patients regardless of cognitive phenotype. However, when we leveraged BA44/45 and BA46 neuropathology assessments to re-group ALS donors by TDP-43 status (positive: TDP-43^POS; negative: TDP-43^NEG) rather than by cognitive phenotype, we found that many of these shared changes were correlated with TDP-43 pathology (Figure 7C; Figure S12). Modules enriched for neuronal gene expression, particularly those linked to synaptic function (e.g., M6), showed minimal change in TDP-43^NEG donors relative to controls but were significantly upregulated in TDP-43^POS individuals (Figure 7C; Figure S12A). This contrast was not apparent when stratifying by cognitive phenotype alone (Figure 7C; Figure S12A). Together with the strong overlap between eCLIP-defined TDP-43 targets and synapse-related genes (Figure 7A; Figures S11C,D), this suggests that synapse-associated transcriptional alterations are, at least in part, mechanistically tied to TDP-43 aggregation. Spatial modules not enriched for synaptic genes, such as the gliosis-associated M2 module, were significantly upregulated in both TDP-43^NEG and TDP-43^POS donors (Figure 7C; Figure S12A). Accordingly, the proportion of reactive glia was also similar between TDP-43^NEG and TDP-43^POS cases (Figure S12B), indicating that this gliotic response is likely driven by alternative mechanisms. Supporting the validity of the neuropathological stratification, STMN2 cryptic splice form expression, a canonical molecular marker of TDP-43 dysfunction^50,51, was elevated specifically in TDP-43^POS donors (Figures S12D,E). Together, these data highlight a nuanced relationship between TDP-43 dysfunction and cognitive deficits: pathology and clinical phenotype do not fully overlap at the molecular level but instead reveal distinct features of ALS pathobiology. This aligns with previous findings that TDP-43 aggregation correlates poorly with cognitive symptoms⁵².

Neuron-glia correlates of cognitive impairment

Our analyses thus far revealed both shared and region-specific patterns of neuronal and glial dysregulation. To integrate these findings, we examined within-donor correlations across cortical regions to assess whether these patterns reflected coordinated changes. This revealed a shared pattern of gliosis across BA44/45 and BA46, primarily driven by astrocytes and microglia (Figure 7D; Figure S13A), potentially reflecting a common gliotic response to ALS across both regions irrespective of cognitive phenotype. While astrocytic and microglial responses tracked similarly between BA44/45 and BA46, myelinating oligodendrocytes (PLP1^High) and maturing OPCs showed region-specific patterns, with weak correlation in their proportions between BA44/45 and BA46 (Figure 7D). The abundance of both oligodendrocyte subtypes correlated closely with neuronal transcriptional dysregulation, particularly with synapse- and mitochondria-associated genes (Figure 7E; Figures S13B–D). Overall, this neuron-oligodendrocyte-maturing OPC coupling emerged as one of the strongest within-sample associations, occurring largely independently of neuronal subtype (Figures S13B,D). Although the basis for this regional heterogeneity remains unclear, these results uncover a complex landscape of neuron-glia interactions that correlate with cognitive impairment in ALS (Figure 7F).

DISCUSSION

Together with recent reports^9,11,12, our study provides a broader understanding of ALS beyond its motor dimension and contributes to a better understanding of the disease. By combining patient stratification with unbiased profiling of two prefrontal cortex regions, we identify spatially defined signatures associated with cognitive changes and trace molecular patterns linked to language, verbal fluency, and executive impairment. Our findings support a model of the disease in which abnormal interactions between neurons, glia, and the vasculature, rather than the simple degeneration of vulnerable neurons, are mechanistically involved in the decline of cognitive function. These interactions differ between cognitive subtypes, indicating a heterogeneous disease etiology driven by distinct mechanisms.

We identified two distinct glial signatures in the ALS prefrontal cortex. One is a classic gliosis response⁵³, shared between BA44/45 and BA46, that involves a coordinated increase in reactive microglia, reactive astrocytes, and stressed OPCs across all cortical layers and WM. The second is more regionally restricted and characterized by a shift in oligodendrocyte lineage composition, with reduced high-myelinating oligodendrocytes and a relative increase in maturing OPCs. At the transcriptional level, reactive glial states are largely shared across ALS and ALSci; what differs is the proportion of cells that adopt them. This suggests that cognitive impairment is associated not with the emergence of qualitatively distinct reactive programs, but with a broader or more sustained engagement of a shared glial response. Moreover, these glial responses appear to be independent of TDP-43 neuropathology, as they were similarly detected in TDP-43^POS and TDP-43^NEG tissues. This does not imply that TDP-43 plays no role – localized cell- or microenvironment-level interactions between TDP-43 pathology and gliosis may exist but are unresolved by our analysis. Such localized interactions may include synaptic changes that engage glial surveillance^54,55 and/or vascular alterations, which appear to be differentially affected in ALS and ALSci.

Deficits in executive function are linked to specific neuronal subtypes localized to the deeper layers of BA46. These neurons show robust molecular alterations in synaptic processes and in mitochondrial function, and their transcriptional changes correlate with shifts in oligodendrocyte and OPC abundance, suggesting that executive deficits emerge from coordinated neuron-oligodendrocyte-OPC dysfunction rather than from neuronal changes alone. In contrast, impairments in language and verbal fluency are associated with more diffuse transcriptional patterns, largely involving glial and vascular processes. Specifically, language deficits correlate with molecular signatures shared across both regions, whereas verbal fluency deficits are restricted to BA44/45 but span many of the spatial co-expression modules we identified. These diffuse patterns may reflect the complexity of the brain’s language network, which is distributed across interconnected frontal and temporal regions⁵⁶. Future studies could test this possibility by expanding sampling to additional brain regions involved in language processing.

TDP-43 proteinopathy is a pathognomonic feature of ALS⁴⁶. This RNA-binding protein is ubiquitously expressed in the brain and typically localizes to the nucleus, but in disease states it is redistributed to the cytoplasm where it forms insoluble aggregates. Nuclear TDP-43 loss is linked to perturbations in gene expression hypothesized to contribute to ALS pathogenesis⁴⁶. Our analyses revealed a nuanced relationship between TDP-43 neuropathology and clinical outcomes. Consistent with previous reports⁵², TDP-43 aggregation was not predictive of cognitive deficits in our cohort; however, some molecular responses, particularly those linked to neuronal and synaptic function, were tightly associated with TDP-43 pathology. This suggests that TDP-43 aggregation alone is insufficient to drive cognitive impairment in ALS, which likely depends on additional disease mechanisms, including the coordinated responses we observed involving neurons, glia, and vasculature. In sum, cognitive phenotypes in ALS appear to have a multicellular etiology and regional specificities, with TDP-43 dysfunction as one modulating factor.

Limitations of the Study

Although BA44/45 and BA46 were profiled using the same multimodal approach, we sampled from partially different subsets of donors due to limited availability of tissue and the stringent quality-control thresholds applied during sample preparation. This imbalance may have reduced our power to detect region-specific differences, as some signals could be masked by donor heterogeneity. Second, cognitive deficits in ALS frequently co-occur (e.g., executive and verbal fluency impairments). Consequently, some of the signatures we identified may reflect contributions from overlapping phenotypes, potentially diluting patterns that would be clearer in patients with isolated cognitive impairments. Third, although donor sex was included as a covariate in our single-nucleus analyses, it was not modeled in our spatial Splotch framework and may represent a further dimension of the disease responses described here. Finally, because the ST technology used here does not provide single-cell resolution, we did not always have the ability to pinpoint the exact cell subpopulations and interactions driving spatially restricted programs. The complementary cellular detail afforded by our multiplexed imaging pipeline helped mitigate this gap, but next-generation spatial technologies will be necessary to refine these mechanisms further.

RESOURCE AVAILABILITY

Lead Contact

Further information and requests for resources should be directed to and will be fulfilled by the lead contact, Hemali Phatnani (hphatnani@nygenome.org).

Materials availability

This study did not generate new unique reagents.

Data and code availability

• Transcriptomic data have been deposited in NCBI Gene Expression Omnibus (GEO; Superseries accession GSE315586). Paired snRNA-seq and snATAC-seq data have been deposited in NCBI GEO (GEO accession: GSE290359). Spatial transcriptomic data have been deposited in NCBI GEO (GEO accession: GSE313918) and are also available through the Target ALS data portal. Bulk RNA-seq data have been deposited in NCBI GEO (GSE314526). All are publicly available as of the date of publication.

• Whole-genome sequencing data are publicly available on the Target ALS portal (https://dataengine.targetals.org/collections/postmortem-tissue-core/).

• Multiplexed protein imaging summary data has been deposited in Synapse (syn68259959). Imaging data reported in this paper will be shared by the lead contact upon request.

• Hierarchical Bayesian modeling code was written in STAN and Python and is available on GitHub (DOI: 10.5281/zenodo.21046739; https://github.com/adaly/cSplotch).

• Spatial gene module and submodule analysis code was written in Python and has been deposited in the PyPi repository (DOI: 10.5281/zenodo.21046386; https://pypi.org/project/scself/).

• All code is publicly available as of the date of publication.

• Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request.

STAR★METHODS

EXPERIMENTAL MODEL AND STUDY PARTICIPANT DETAILS

Human samples

Postmortem tissue specimens were obtained from the Medical Research Council (MRC) Edinburgh Brain Bank (Research Ethics Committee approval: 21/ES/0087), following review by the bank’s internal ethics committee, the Academic and Clinical Central Office for Research and Development Medical Research Ethics Committee, and the Biomedical Research Alliance of New York. Tissue specimens originated from patients recruited through the Scottish Motor Neurone Disease Register (a prospective, population-based ALS study in Scotland) who had consented to the collection and use of their health-related information for research purposes. Clinical metadata associated with each donor, including ECAS scores, were extracted from the Clinical Audit Research Evaluation (CARE)-MND database. Tissue from neurotypical control cases – defined in general terms as individuals without neurological or psychiatric conditions and no substantial neuropathology at postmortem – was provided by the Edinburgh Sudden Death Brain Bank¹⁹. All applicable federal, state, and local regulations were followed when transferring materials to the New York Genome Center.

The diagnosis of patients fulfilled the revised El Escorial criteria for clinically definite or probable ALS. Comorbid neurological disorders, including frontotemporal dementia, or psychiatric history were exclusion criteria for the selection of study participants. We note also that dementia was not documented in the clinical files of any recruited patients. A total of 53 individuals were included: 36 ALS donors, of whom 19 had normal cognition (6 male, 13 female) and 17 met criteria for cognitive impairment (ALSci; 12 male, 5 female), alongside 17 neurotypical controls (8 male, 9 female). All donors were recorded as White and of Scottish ethnicity.

Cognitive performance was assessed in life using the Edinburgh Cognitive and Behavioural ALS Screen (ECAS) following standard guidelines (https://ecas.psy.ed.ac.uk). Briefly, the ECAS incorporates 16 subtests assessing a range of cognitive domains, including executive functions (reverse digit span, alternation, inhibitory sentence completion, and social cognition), verbal fluency (free and fixed), language (naming, comprehension, and spelling), memory (immediate recall, delayed percentage retention, and delayed recognition), and visuospatial functions (dot counting, cube counting, and number location), complemented by a separate semi-structured interview with a relative or carer designed to detect behavioral changes^14,57.

Performance across all tasks generates the ECAS Total Score, which adds up to an aggregated maximum of 136 points. The ECAS Total Score is composed of two fractions, the ALS-Specific and ALS Non-Specific Scores. Subtest evaluations of executive functions, language, and verbal fluency are combined to generate the ALS-Specific Score, which can reach a maximum of 100 points in cognitively healthy individuals. Memory and visuospatial functions tasks contribute to the ALS Non-Specific Score. Abnormal scores are defined as ≤77 for the ALS-Specific Score, ≤24 for the ALS Non-Specific Score (out of a maximum of 36 points), and ≤105 for the ECAS Total Score. Further subdomain-specific thresholds include ≤33 out of 48 for executive function, ≤14 out of24 for fluency, ≤26 out of 28 for language, ≤13 out of 24 for memory, and ≤10 out of 12 for visuospatial functions.

Clinical donor-level metadata is provided in Figure S1 and Table S1, and sample metadata is included in Table S2.

METHOD DETAILS

Neuropathology Assessments

Brain regions of interest, including BA44/45 and BA46, were fixed in 10% neutral-buffered formalin for a minimum of 24 hours to ensure tissue integrity and morphology. Post-fixation, tissue processing involved a dehydration step through an ascending ethanol series (70%–100%), followed by clearing with three rounds of xylene treatment. Specimens were then embedded in molten paraffin wax over three consecutive 5-hour stages and, after solidifying at room temperature, sectioned onto Superfrost Plus slides on a Leica microtome at 4 μm thickness with overnight drying at 40°C. All staining procedures were performed using a Leica Bond RX autostainer instrument. Aggregated TDP-43 was detected using an antibody against-phosphorylated TDP-43 (2B Scientific, #CAC-TIP-PTD-MO1, 1:2000; Leica Bond RX: protocol F ER1 – 30 mins), combined with 3,3′-diamino-benzidine (DAB) visualization. TDP-43 pathology was staged using a semiquantitative four-tiered scoring system, where a score of 0 indicates the absence of pathology, and scores 1 through 3 represent progressively increasing levels of TDP-43 pathology burden (mild, moderate, and severe, respectively). Scores were defined as follows: 1 (mild) for up to 5 affected cells, 2 (moderate) for 6–15 affected cells, and 3 (severe) for more than 15 affected cells, all within at least a single 20× field of view per tissue section. This scoring scheme was applied independently to both neurons and glia using haematoxylin counterstaining for morphology-based cell classification⁵. Tau: Thermo Fisher, #MN1020, 1:1000; Leica Bond RX: protocol F, no pretreatment. Beta-amyloid: BioLegend, #800712, 1:16000; Leica Bond RX: protocol F wet start, formic acid pretreatment (5 mins). Alpha-synuclein: Leica, #NCL-L-ASYN, 1:20; Leica Bond RX: protocol F ER2 (30 mins).

Brain regions were classified based on the neuronal pathology burden into TDP-43 aggregation-positive (TDP-43^POS, score ≥1) and TDP-43 aggregation-negative (TDP-43^NEG, score of 0) groups for analysis. Glial pathology burden scores were reported but were not used for this classification. For donor-level analyses (e.g. Figure S1F), donors classified as TDP-43^POS in either BA44/45 or BA46 were identified as TDP-43^POS. For brain region-level analyses, brain regions were classified separately, such that individual donors may have been assigned with both TDP-43^POS and TDP-43^NEG classifications.

Tissue sectioning and bulk RNA-seq library preparation

Frozen post-mortem brain tissue was sectioned in a cryostat maintained at −18 °C and 20–50 mg of trimmed tissue was collected per preparation. Total RNA was extracted from sectioned tissue in TRIzol reagent and further purified using RNeasy Mini columns (QIAGEN). RNA concentration was quantified by Qubit fluorometry (Thermo Fisher Scientific), and RINs were determined using a BioAnalyzer (Agilent Technologies). For library preparation, 500 ng of total RNA was processed using a KAPA Stranded RNA-Seq Kit with RiboErase (KAPA Biosystems). Libraries were sequenced on an Illumina NovaSeq X system to a target depth of 40 million paired-end reads (2×100 bp).

Bulk RNA-seq Data Processing and Analysis

Sequenced libraries were demultiplexed, aligned with STAR to a standard reference genome (hg38; GENCODE v32/Ensembl98), and counted with htseq-count. Raw counts were converted to transcripts per million (TPMs) by dividing counts by transcript lengths and then dividing by an appropriate library size factor. TPMs were log1p transformed, and the resulting count matrix was rotated for Principal Component Analysis (PCA) by truncated singular value decomposition.

Nuclei isolation from frozen post-mortem brain tissue and library preparation

Nuclei isolation protocol was adapted from previously described methods^58,59. All procedures were carried out on ice unless otherwise stated. Frozen post-mortem brain tissue was retrieved from storage at −80 °C and equilibrated on dry ice for 20–30 minutes before being transferred to a cryostat at −18 °C. Tissue samples were left for 15–20 minutes inside the cryostat before sectioning. We collected 20–50 mg of tissue per preparation, taking care to avoid excessive sampling of white matter. Tissue cuts were placed in a pre-chilled Dounce homogenizer containing 700 μL of ice-cold homogenization buffer (320 mM sucrose, 5 mM CaCl₂ [Millipore Sigma; #21115;], 3 mM magnesium acetate [Millipore Sigma; #63052], 10 mM Tris HCl [pH 7.8; prepared from Thermo Fisher’s #AM9850G and #AM9855G], 0.1 mM EDTA [pH 8.0; Thermo Fisher, #AM9260G], 0.1% IGEPAL CA-630 [Millipore Sigma; #I8896], 1 mM DTT [Millipore Sigma; #646563], and 0.4 U/μl Protector RNase inhibitor [Millipore Sigma; #03335402001]). Homogenization was performed with 10 strokes of the loose pestle followed by 10 strokes of the tight pestle. After a 5-minute incubation on ice, the homogenized tissue was passed through a 70-μm Flowmi Cell Strainer (Bel-Art; #H13680–0070) and combined with an equal volume of a solution made of 50% OptiPrep (#07820, STEMCELL Technologies), 5 mM CaCl₂, 3 mM magnesium acetate, 10 mM Tris HCl (pH 7.8), 0.1 mM EDTA (pH 8.0), and 1 mM DTT. Using a 15-mL conical tube, this mixture was then layered onto a density gradient made up of 750 μL of a 30% OptiPrep solution supplemented with 134 mM sucrose, 5 mM CaCl₂, 3 mM magnesium acetate, 10 mM Tris HCl (pH 7.8), 0.1 mM EDTA (pH 8.0), 1 mM DTT, 0.04% IGEPAL CA-630, and 0.17 U μl⁻¹ RNase inhibitor) over 300 μL of 40% OptiPrep solution supplemented with 96 mM sucrose, 5 mM CaCl₂, 3 mM magnesium acetate, 10 mM Tris HCl (pH 7.8), 0.1 mM EDTA (pH 8.0), 1 mM DTT, 0.03% IGEPAL CA-630, and 0.12 U μl−1 RNase inhibitor. Nuclei were separated by centrifugation at 3,000 × g for 20 minutes at 4 °C using a pre-chilled swinging-bucket rotor with the brake off. Approximately 150 μL of nuclei were collected from the 30%/40% gradient interphase and subsequently washed in PBS containing 1% BSA and 1 U μl⁻¹ RNase inhibitor. Following a 5-minute incubation on ice, the nuclei were centrifuged at 450 × g for 3 minutes at 4 °C, and the pellet resuspended in 15 μL of diluted 10x Genomics Nuclei Buffer. Nuclei counting was performed using a Countess II FL instrument (Thermo Fisher, #AMQAF1000) by staining with SYTOX Green (Thermo Fisher; #S7020). Additional quality control measures included Trypan Blue staining (#EBT-001, NanoEntek) to evaluate adequate permeabilization (cutoff: >95% of ‘dead cells’) and the assessment of nuclear morphology by light microscopy before proceeding with transposition.

Matched snRNA-seq and snATAC-seq libraries were prepared using the Chromium Next GEM Single Cell Multiome ATAC + Gene Expression Kit (10x Genomics, #PN-1000283) according to the protocol provided by 10x Genomics. Nuclei (16,100 per reaction; targeted recovery: 10,000 nuclei) from a single donor were used for each library preparation. Partitioning of nuclei was performed on a Chromium X instrument (10x Genomics, #1000331). Quality control of cDNAs and libraries was conducted using Agilent High Sensitivity DNA Kits and a BioAnalyzer instrument. Final library quantification prior to pooling utilized KAPA Biosystems’ Library Quantification Kit for Illumina Platforms (KK4824). Libraries were sequenced on a NovaSeq X system using a 2 × 100 bp configuration run on a 10B flow cell. BA44/45 and BA46 libraries (n = 48) were prepared and sequenced in independent batches.

Tissue processing, RNA Capture, and Library Preparation for Spatial Transcriptomics

Frozen tissue blocks, kept at −80°C prior to processing, were equilibrated to −16°C inside a Leica CM3050S cryostat, embedded in Tissue Plus OCT Compound (Fisher Healthcare, #4585), and sectioned in a perpendicular orientation with regard to the pial surface. Tissue sections (10 μm-thick) from quality controlled blocks (H&E staining and imaging) were collected onto prechilled Visium Spatial Gene Expression Slides (10x Genomics, #1000185) while briefly warming the back of the slide with a gloved finger to improve adherence and minimize artifacts.

Visium slides capture polyadenylated RNAs from intact tissue sections and localize each RNA molecule to a specific 55-μm barcoded spot across a 6.5 × 6.5 mm capture area – sufficient to span all six cortical layers as well as adjacent white matter. We sought to capture cortical layers 1 through 6 and white matter within each spatially barcoded array. For each subject, four technical replicates, sampled from tissue cuts approximately 20 μm apart, were collected across two Visium slides. A deliberate effort was made to section control and ALS subjects onto the same Visium slide when possible, with sample processing proceeding in parallel from that point onward. Where necessary due to tissue sectioning artifacts or poor yield, additional replicates were collected and processed on a later date.

Generation of spatially resolved transcriptomes was performed in accordance with 10x Genomics guidelines (CG000239, Revisions F and G). Briefly, tissue sections were fixed in chilled methanol, stained using an abbreviated H&E protocol, and imaged on a Zeiss Axio Observer Z1 fitted with an EC Plan-Neofluar 10x/0.3 M27 objective and a Zeiss Axiocam 506 Mono camera. Tissue sections were permeabilized for 12 minutes following optimization (10x Genomics, CG000238, Revision E). The number of cDNA amplification cycles for each Visium array was calculated using qPCR, as instructed by 10x Genomics.

Visium libraries were prepared using unique Illumina-compatible PCR primers with Dual Index Kit TT Set A (10x Genomics, #1000215). Size distribution profiles of the final libraries were assessed by running Agilent High Sensitivity DNA chips (#5067–4626) on a 2100 Bioanalyzer system (Agilent, #G2939BA). Libraries outside the expected range (450 ± 50 bp) or that contained adapter dimer contaminants were discarded and reprocessed. Total yields were initially measured on a Qubit 3.0 fluorometer before final qPCR quantification (KAPA Library Quantification Kit for Illumina Platforms, KAPA Biosystems, #KK4824). Per batch, libraries from 48–52 Visium arrays were pooled at a concentration of 10 nM prior to sequencing. We processed 246 arrays from 46 unique donors.

SST⁺ Cell Quantification

SST labelling (Santa Cruz Biotechnology, #SC-74556, 1:200) was carried out on a Leica Bond RX autostainer (protocol F ER1 – epitope retrieval duration: 20 minutes) following standard IHC procedures. Raw H&E images were first analyzed to differentiate between white matter and gray matter regions. Each field of view (FOV) was classified based on its three dominant colors, excluding the lightest and darkest 20% of pixels to avoid artifacts from holes in tissue, as well as the cells and SST stainings themselves. The dominant colors were determined using k-means clustering (k = 3) and the resulting cluster centers were used as representative colors. A FOV was classified as gray matter if, in RGB space, the blue channel intensity exceeded the red channel by at least 25 units for any of the three dominant colors.

For FOVs identified as gray matter, color deconvolution was performed to separate the cellular component from the SST signal. Cell segmentation was conducted using a Cellpose machine learning model trained from scratch using human-in-the-loop training on several manually-annotated FOVs⁶⁰. The resultant cell masks were dilated by 50% of the cell’s radius to include surrounding cytoplasmic regions. The SST image was converted to grayscale, and pixel values below 90% of the median FOV pixel intensity were discarded to remove background noise. The remaining SST signal was thresholded based on size, retaining only regions larger than 20 pixels in area. Finally, colocalization of the thresholded SST signal and the dilated cell masks was performed, and any mask containing SST signal was classified as SST-positive.

Sequential immunofluorescence using COMET platform

Formalin-fixed, paraffin-embedded (FFPE) tissue sections were processed through two consecutive xylene baths (10 minutes each), followed by two 100% ethanol baths (3 minutes each), two 96% ethanol baths (3 minutes each), and a single 70% ethanol bath (3 minutes). Slides were subsequently rinsed in distilled water for 30 seconds prior to antigen retrieval using the 10x Genomics Decrosslinking Solution (catalog no. 2000566, 10x Genomics) at 95 °C for 1 hour. Following retrieval, sections were washed in Multi-Staining Buffer (BU06, Lunaphore) and maintained in this buffer until loading onto the COMET platform (Lunaphore).

Automated, sequential immunofluorescence was performed on the COMET system, which executes iterative cycles of staining, imaging, and antibody elution. For each cycle, two primary antibodies were applied for 4 minutes using a pressurized microfluidic chip (MK03, Lunaphore), followed by incubation with two secondary antibodies for 2 minutes at 37 °C prior to imaging. Each cycle stained and imaged DAPI (Lunaphore, DR100) to allow alignment and generation of the assembled image. The elution step ensured complete removal of both primary and secondary antibodies, enabling iterative rounds of staining and imaging without residual signal carryover.

The antibodies used for this multiplex immunostaining were validated manually using a standard immunofluorescence protocol and subsequently organized into a panel for application to the tissue samples. The full antibody panel is as follow: IBA1 (Abcam, ab178847, EPR16589, used at 1:2000), GFAP (Novus, NBP1-05198, used at 1:5000), Olig2 (Abcam, AB109186, used at 1:100), MAP2 (Origene, TA309162, G9824pm used at 1:2000), NeuN (Millipore Sigma, MAB377, A60, used at 1:200), Neurofilament Light (Millipore Sigma, N0142, used at 1:1500), AQP4 (Proteintech, 16473-1-AP, used at 1:2000), ALDH1L1 (Altas, HPA050139, used at 1:200), CD68 (DAKO, M0814, KP1, used at 1:100), Claudin5 (Invitrogen, 35–2500, 4C3C2, used at 1:100), aSMA (Lunaphore, MR10100, 1A4, used at 1:500), secondary antibodies Alexa fluorophore 488 (Invitrogen, A78948, used at 1:500), 555 (Lunaphore, DR555RB used at 1:200), and 647 (Lunaphore, DR647MS, used at 1:200).

QUANTIFICATION AND STATISTICAL ANALYSIS

Prefrontal Cortex Atlas and Classifier

Published 10x Genomics snRNA-seq sequencing data was acquired from NCBI GEO or Synapse repositories: GSE157827²², GSE167494²³, GSE174332⁹, GSE174367²⁴, GSE207334²⁵, GSE219280¹², PRJNA434002²⁶, syn45351388²¹, and syn51123517²⁷ (Table S4).

Sequencing data was aligned to a standard reference genome (hg38; GENCODE v32/Ensembl98) and counted (cellranger v7.1.0). Individual cells were filtered so that the number of transcripts (UMIs) was between 500 and 50000, and the maximum fraction of mitochondrially-encoded transcripts was 0.2. Prefrontal cortex atlas was restricted to donor samples from patients between 50 and 90 years of age at death taken from the prefrontal cortex. The resulting dataset consists of 2,603,756 cells from 407 donors.

Each cell was scaled to a total of 10,000 counts, the count data was transformed by log(x + 1), and each gene was scaled by dividing by the interquartile range (Equation 1; RobustScaler). Genes which would have a variance greater than 1 after scaling were further scaled so that their variance was equal to 1. Cells were embedded into a k-nearest neighbors (k-NN) graph from 50 principal components and 25 neighbors, then clustered into 114 separate clusters using the leiden algorithm (resolution set to 2.5). Labels for cell type and cell subtype were hand annotated for each cluster based on known marker genes (Table S5).

$$\text{Scaled}\left({\mathbf{X}}_{\mathbf{g}}\right)=\left\{\begin{array}{ll}{\mathbf{x}}_{\mathbf{g}}/\sqrt{\text{Var}\left({\mathbf{X}}_{\mathbf{g}}\right)}& \text{Var}\left({\mathbf{X}}_{\mathbf{g}}\right)\ge \text{IQR}{\left({\mathbf{X}}_{\mathbf{g}}\right)}^2\\ {\mathbf{x}}_{\mathbf{g}}/\text{IQR}\left({\mathbf{X}}_{\mathbf{g}}\right)& \text{Var}\left({\mathbf{X}}_{\mathbf{g}}\right)<\text{IQR}{\left({\mathbf{X}}_{\mathbf{g}}\right)}^2\\ {\mathbf{X}}_{\mathbf{g}}& \text{IQR}\left({\mathbf{X}}_{\mathbf{g}}\right)=0\end{array}\right.$$ Equation 1

This standardized, transformed, scaled data was used as input to a neural network (NN) autoencoder/classifier that predicts cell type, subtype, and denoised transcript expression data. The NN encoder module takes 36601 transcript counts as input, and is fully connected to a 1000 node hidden layer, a 250 node hidden layer, then a 100 node hidden layer, each of which has a hyperbolic tangent (tanh) activation function. The 100 node hidden layer is connected to a 50 node (h_width) hidden layer with a tanh activation function that represents a shared latent space embedding. This hidden layer is shared between three decoder modules. The transcript expression predictor connects the 50 node shared latent space embedding to a 50 and then 250 node hidden layers with a tanh activation function, then a 36,601 node output layer with a rectified linear (ReLU) activation function. The cell type classifier connects the 50 node shared latent space embedding to a 50 node hidden layer with a tanh activation function, and connects that hidden layer to a 9 node output softmax layer that predicts cell type label. The cell subtype classifier connects the 50 node shared latent space embedding to a 50 node hidden layer with a tanh activation function, and connects that hidden layer to a 14 node output softmax layer that predicts cell subtype label.

The NN model was constructed in pyTorch and trained with the Adam optimizer. Model hyperparameters h_width, weight decay, and learning rate were tuned by training on randomly sampled 465,000 cell training sets for 200 epochs and scoring on 155,000 cell validation sets. Transcript expression was trained to minimize mean squared error, and cell type and subtype classifiers were trained to minimize cross-entropy. Full training was performed using 1.95m cells for 1500 epochs, and scored on 650k cells held out. Held out test cells were not used during cross-validation. Models were evaluated based on coefficient of determination (R²) for transcript expression, and on F1 score for cell type and subtype classifiers.

Multiome Data Processing

Sequenced 10x Genomics gene expression and ATAC libraries were aligned to a standard reference genome (hg38; GENCODE v32/Ensembl98) and counted (Cellranger-ARC v2.0.2; Table S6). ATAC peaks were aligned between samples for quantification (Cellranger-ARC aggr v2.0.2). Individual cells were filtered so that the number of gene expression transcripts (UMIs) was between 500 and 50000, the maximum fraction of mitochondrially-encoded transcripts was 0.2, and the number of quantified ATAC fragments in peaks was between 250 and 100,000. A total of 165,884 cells passed this initial filtering. Cell types and subtypes were determined by the prefrontal cortex cell type classifier, and validated by comparing to published marker genes (Table S5). Doublets and low-quality cells were removed, leaving 144,260 usable cells from 48 libraries for analysis.

To maintain the difference in count depth between cell types, individual cells were standardized for count depth by dividing all transcript counts for the cell by a size factor, to reach the median count depth for the assigned cell type. Mitochondrially encoded genes (MT-) and MALAT1 were excluded from size factor calculations but were otherwise retained for analysis. Standardized counts were log(x+1) transformed, and each gene g was then scaled by dividing by the interquartile range (IQR). Genes which would have a variance (Var) greater than 1 after scaling were further scaled so that their variance was equal to 1, giving log standardized counts (Equation 1).

Principal Component Analysis (PCA) was performed using a truncated singular value decomposition with 25 components on the log standardized counts. Cells were embedded into a 25-neighbor k-Nearest Neighbors (k-NN) graph using these 25 components. Cells were projected into 2 dimensions with Uniform Manifold Approximation and Projection (UMAP) using the k-NN embedding (Minimum distance of 0.5; initialized from spectral graph embedding).

For ATAC analysis, the 144,260 usable cells were further filtered for cells with at least 2500 ATAC peak counts, yielding 119,153 cells with both RNA and ATAC modalities. Individual cells were standardized for ATAC count depth by dividing all peak counts for the cell by a size factor, to reach the median ATAC count depth for the assigned cell type. Standardized ATAC counts were then log(x+1) transformed to give log standardized ATAC counts. PCA, k-NN embedding, and UMAP projection were performed identically to gene expression data, using log standardized ATAC counts.

Glial Cell Functional Annotation

snRNA-seq log standardized counts for each glial cell type were separated for PCA (25 components), k-NN graph embedding (25 neighbors), and UMAP as for the full dataset. Cells from each cell type were separately clustered into a large number of groups with the leiden algorithm (resolution of 2), and then each cluster was assigned to one of a few functional groups based on relative expression of key functional marker genes (Table S7).

The proportion of glial cells within each functional category was calculated for each snRNA-seq donor sample by taking the number of functionally annotated glial cells and dividing it by the total number of glial cells of that type (e.g. the number of homeostatic astrocytes divided by the total number of astrocytes).

Gell cell-type-specific functional markers were identified by differential expression performed on raw integer counts from snRNA-seq data with pyDeseq2^61,62, standardizing using only non-zero counts (‘poscounts’). Statistical testing was against a null hypothesis Log 2 Fold Change of 0, rejecting the null for adjusted p-values ≤ 0.05. Significantly changing genes between functional groups of cells were filtered for genes which were specific for that glial cell type, defined as having the highest mean expression in that cell type, having at at least 25% of the total transcript count within the snRNA-seq dataset in that cell type, and not significantly changing between other functional annotations in glial cell types. Genes which met this criteria were annotated as cell type specific functional markers (Table S8) and used to calculate gliosis scores for spatial spots.

Spatial Gliosis Scores

Annotated integer cellranger count data was standardized by dividing all spots by a size factor, so that the total depth at each spot of a cortex layer (e.g. L1) was standardized to the median depth of all spots of that layer. Mitochondrially encoded genes and MALAT1 were excluded from size factor calculations. Transcript counts were then log(x+1) transformed, and standardized to a 0–1 range for each gene, where 0 is fixed to the 1st percentile, 1 is fixed to the 99th percentile, and values are linearly scaled between 0 and 1. Values below the 1st percentile are clipped to 0, and values above the 99th percentile are clipped to 1 (Equation 2). For each gliosis functional group, a gliosis score is calculated for each array spot by taking the mean of the 0–1 scaled gene scores for every cell type specific functional marker.

$$\text{MinMaxScaled}(x)=\left\{\begin{array}{ll}\frac{x-{\mathbf{X}}_{q=0.01}}{{\mathbf{x}}_{q=0.99}-{\mathbf{X}}_{q=0.01}}& {\mathbf{X}}_{q=0.01}\le x\le {\mathbf{X}}_{q=0.99}\\ 0& x<{\mathbf{X}}_{q=0.01}\\ 1& x>{\mathbf{X}}_{q=0.99}\end{array}\right.$$ Equation 2

Spatial spot-level gliosis scores are aggregated by taking the median gliosis score for all spots in a layer (L1-L6 and WM) for each array, and rescaling the median layer-array gliosis scores to a 0–1 range. These scaled layer-array gliosis scores are used for statistical testing, and are further aggregated into mean gliosis scores per ALS phenotype for plotting.

Published Reactive Astrocyte Markers

Reactive astrocyte markers were retrieved from previously published studies. In some cases, markers were obtained directly from differential expression tables corresponding to reactive astrocyte cell clusters (e.g., astR1 and astR2 clusters⁶³; Cluster 1 and Cluster 2⁶⁴; Cluster 6⁶⁵; Cluster ASC2²⁴). In other cases, markers were extracted from published compilations of astrocyte reactivity genes (e.g., Table S9⁶⁶). In addition, we derived marker sets by reanalyzing two independent datasets and identifying genes differentially expressed between homeostatic and reactive astrocyte populations^12,44. All markers used in this study are listed in Table S9.

Neuronal Cell Subtype Annotation

snRNA-seq log standardized counts for excitatory (EN) and inhibitory neurons (IN) were further filtered by transcript count to neurons with a minimum of 2500 UMI counts, yielding 41,933 ENs and 15,553 INs. ENs and INs were separately processed for PCA (25 components), k-NN graph embedding (25 neighbors), and UMAP as for the full dataset. ENs and INs were then clustered into transcriptionally distinct clusters with the leiden algorithm (resolution of 3), and clusters were assigned to a neuronal subtype based on marker expression (Table S7).

Cell Proportions Per Sample

For each single nucleus sample (n=48), the number of cells in each functional group for a glial cell type are divided by the total number of cells of that type (e.g., 60 homeostatic astrocytes and 40 reactive astrocytes have proportions of 0.6 and 0.4). The correlation between these ratios is calculated as the within-sample correlation of glial cell function proportions. Welch’s t-test was used as the statistical test between proportions in different groups. Wald test was used as the statistical test for correlations. Multiple comparisons were corrected with Benjamini-Hochberg as appropriate.

Single-cell Differential Expression Analysis

Transcript differential expression analysis was performed on pseudobulked raw integer counts with pyDeseq2^61,62, standardizing using only non-zero counts (‘poscounts’) and excluding mitochondrially encoded genes and MALAT1 from size factor calculations. Statistical testing was against a null hypothesis Log 2 Fold Change of 0, rejecting the null for adjusted p-values ≤ 0.05.

Gene Ontology Enrichment Analysis

Gene ontology enrichment was performed with Metascape using the default complete proteome as the background of each analysis⁶⁷, and g:Profiler⁶⁸, which tests gene lists for enrichment against background a multiple hypothesis test corrected hypergeometric test. Here, gene lists are differentially expressed genes against a background of genes which are not all-zero in counts, or genes in a module against a background of genes which were included in the Splotch spatial model.

Single-cell Differential Peak Analysis

Differential accessibility analysis was performed on pseudobulked raw integer counts with pyDeseq2, standardizing using only non-zero counts (‘poscounts’). Pseudobulked samples were required to have a minimum of 10 cells; those with fewer were excluded from this analysis. Statistical testing was against a null hypothesis Log 2 Fold Change of 0, rejecting the null for adjusted p-values ≤ 0.05.

ATAC Gene Ontology Enrichment Analysis

ATAC peaks were linked to nearby gene promoters (within 1000bp of annotated transcription start site). Peaks that were not within 1000bp of annotated transcription start sites were excluded. Ontology enrichment was performed as in Gene Ontology Enrichment Analysis using linked genes as the query list and all genes with peaks annotated near promoters as the background list.

ATAC Motif Search

HG38 genomic sequences of differentially accessible ATAC peaks were extracted with bedtools⁶⁹ and searched for motifs from the JASPAR Homo sapiens motif collection⁷⁰ with Simple Enrichment Analysis⁷¹, filtering based on a q-value for false discovery rate less than 0.05.

10x Genomics Visium Data Processing and Annotation

Each sequenced 10x Visium library was aligned to a standard reference genome (hg38; GENCODE v32/Ensembl98) and counted with spaceranger (v3.0.0; Table S3). We initially filtered any spot with fewer than 100 UMI transcript counts. Visium spots were manually assigned to one of the six cortical layers or the white matter using 10x Genomics Loupe Browser (version 8.0). Layer annotations were based on cortical architecture and cellular morphology. The white matter was identified by the absence of neuronal cell bodies and the linear arrangement of glial nuclei along neuronal processes. Cortical layer 6 was defined as a heterogeneous neuronal layer spanning from the white matter to the beginning of cortical layer 5, characterized by its large pyramidal neurons oriented perpendicularly to the pial surface. Cortical layer 4 was identified by the presence of small, granular neurons adjacent to the large pyramidal neurons of cortical layer 5. Cortical layer 3 was labeled as the pyramidal cell layer located between granular cortical layers 2 and 4. Cortical layer 1 was distinguished by its paucity of cells and lack of large neuronal somas. Spots containing meninges contiguous to cortical layer 1 were excluded from downstream analysis. Additionally, spots overlying tissue artifacts, such as freeze fractures, folds, or tears, were excluded. After annotation, we have 953,073 individual spots across 246 arrays.

To maintain the difference in count depth between cell types, individual spots were standardized for count depth by dividing all transcript counts for the cell by a size factor, to reach the median count depth for the annotated cortex layer. Size factors were clipped to a minimum of 0.25 to mitigate the risk of overinflating low-count spots. Standardized spatial counts were then scaled by dividing by the interquartile range (RobustScaler). Genes which would have a variance greater than 1 after scaling were further scaled so that their variance was equal to 1, giving standardized spatial counts.

PCA was performed using a truncated singular value decomposition with 25 components on the standardized counts. Cells were embedded into a 25-neighbor k-NN graph using these 25 components. Cells were projected into 2 dimensions with UMAP using the k-NN embedding (Minimum distance of 0.5; initialized from spectral graph embedding).

Splotch Modeling of Spatial Gene Expression Across Clinical Groups

Raw sequencing data was filtered to remove mitochondrially encoded genes, lncRNAs, pseudogenes, and genes expressed in less than 1% of spots (this filtering was applied separately to BA44/45 and BA46 datasets). Spots with fewer than 100 UMIs across remaining genes were discarded prior to Splotch analysis. Finally, spots without a valid annotation were discarded, and any spots without at least one immediate neighbor in the Visium grid were discarded to prevent singularities in the spatial autocorrelation component of the Splotch model. This yielded expression data for 13,188 genes measured across 401,976 spots (286,581 ALS, 115,386 control) with an average depth of 3,679 UMIs in BA46, and 13,116 genes measured across 551,106 spots (347,548 ALS, 203,558 control) with an average depth of 3,447 UMIs in BA44/45.

Splotch analysis was conducted as described previously^13,38,39. Briefly, Splotch employs a zero-inflated Poisson likelihood function (where the zero-inflation accounts for technical dropouts) to model the expression (λ) of each gene in each spot. This expression level is in turn modeled using a generalized linear model (GLM) with the following three components: the characteristic expression of the spot’s donor ID and ALS phenotype (β), autocorrelation with the spot’s spatial neighbors (ψ), and spot-specific variation (ε). Furthermore, the characteristic expression rate (β) is hierarchically formulated to account for the experimental design, enabling the investigation of the effect of sample covariates on differential gene expression. By performing posterior inference on the model given the observed spatial expression data, we identified clinical group level trends in spatial gene expression (β) that best explained our observations.

Compared to other computational methods for analysis of spatial transcriptomics data, Splotch i) enables quantification of expression differences between conditions and anatomical regions, ii) is designed to rigorously account for the uncertainty of low counts, and iii) analyzes multiple tissue sections simultaneously, stratifying patients based on clinical phenotype in order to quantify biological and technical variation.

Gene List Overlap and Enrichment Analysis

eCLIP-seq-defined TDP-43 target mRNAs were obtained from Tam et al.⁴⁷. To generate a stringent target list, we restricted the analysis to genes supported by five or more eCLIP-seq hits, irrespective of intronic or exonic annotation. This resulted in a total of 1,505 genes included in the analysis.

Enrichment was quantified using a representation factor, defined as the ratio of the observed overlap to the expected overlap under random sampling. Statistical significance was assessed using a permutation-based approach. Specifically, 10,000 random permutations were performed in which two gene sets of equal size to the original sets were sampled without replacement from the background gene universe, defined as all genes detected in a given neuronal subpopulation. An empirical p-value was calculated as the proportion of permutations in which the random overlap was greater than or equal to the observed overlap. This allowed us to test whether the observed overlap was significantly greater than expected by chance.

Parameter inference

Splotch has been implemented in the probabilistic programming language Stan, and is available at https://github.com/adaly/cSplotch. For all analyses, Bayesian inference was performed over the parameters using Stan’s adaptive Hamiltonian Monte-Carlo (HMC) sampler with default parameters. Four independent chains were sampled, each with 175 warm-up iterations and 175 sampling iterations (700 total), and convergence was monitored using the R-hat statistic.

Splotch differential expression analysis

To quantify difference in expression between two conditions using Splotch, the Bayes factor between posterior distributions over characteristic expression coefficients β estimated by the model was calculated. Without loss of generality, difference in expression was quantified between conditions represented by β⁽¹⁾ and β⁽²⁾, which may differ across any combination of genes, donor covariates, or cortical layers. A random variable Δ_β = β⁽¹⁾ – β⁽²⁾ was defined, which captures the difference between β⁽¹⁾ and β⁽²⁾. If Δ_β is tightly centered around zero, then the two distributions are very similar to each other, and the null hypothesis of identical expression cannot be rejected. To quantify this similarity, the posterior distribution Δ_β|𝒟 (where 𝒟 represents the data used to train the model) was compared to the prior distribution Δ_β using the Savage-Dickey density ratio⁷² that estimates the Bayes factor between the conditions (Equation 3):

$$\text{BF}\approx \frac{p\left({\Delta }_{\beta }=0\right)}{p\left({\Delta }_{\beta }=0\mid 𝒟\right)},$$ Equation 3

where the probability density functions were evaluated at zero. If expression is different between the two conditions, then the posterior Δ_β|𝒟 will have very little mass at 0, and the estimated Bayes factor will be large (by convention, BF > 5 indicates substantial support). Conversely, for similar expression regimes, the posterior will place a mass equal to or greater than that of the prior at zero, and the Bayes factor will be ≤ 1. p(Δ_β|𝒟) is calculated using the posterior samples obtained in the previous section.

Assigning Genes to Spatial Modules

Spatial modules are inferred from the mean posterior estimates (λ) of the depth-normalized expression of each gene in each spatial spot (a matrix of dimension 13,188 genes × 401,976 spots in BA46; 13,116 genes × 551,106 spots in BA44/45). Mean posterior gene expression estimates λ for each gene were clipped to the 99th percentile to remove the effect of outlier spots. Gene expression values were then scaled by dividing by the gene interquartile range, and genes which would have a variance greater than 1 after scaling were further scaled so that their variance was equal to 1 (Equation 1). Pearson correlation coefficients were then calculated across all spots for all pairs of genes, resulting in a 13,188 × 13,188 correlation matrix in BA46 and 13,116 × 13,116 correlation matrix in BA44/45.

Correlation distance between 13,253 genes was calculated by taking the mean Pearson correlation distance (1 - Pearson correlation coefficient) for a gene-gene pair from both brain regions. This correlation distance matrix was used as the distance matrix for a gene-gene k-NN graph embedding (k=10). Genes were projected into 2 dimensions with the UMAP algorithm, and clustered into 23 clusters using the leiden algorithm (resolution of 2). Each cluster is considered a spatially-correlated gene module. Clusters 21 and 22 contain all chrM and chrY-encoded transcripts, respectively.

Spatial modules and submodules are listed in Table S11. Gene Ontology enrichment for each spatial module is listed in Table S12. Python functions for module assignment are available in the scself (single-cell self-supervised; v0.4.8) package, installable through python’s pip package manager.

Module Scores

For each gene, standardized transcript count (snRNA-seq) or modeled expression lambda (spatial) was standardized to a 0–1 range for each gene, where 0 is fixed to the 1st percentile, 1 is fixed to the 99th percentile, and values are linearly scaled between 0 and 1 (Equation 2). Values below the 1st percentile are clipped to 0, and values above the 99th percentile are clipped to 1. For each module, a module score is calculated for each cell (snRNA-seq) or array spot (spatial) by taking the mean of the 0–1 scaled gene scores for every gene in the module.

Spatial spot-level module scores are aggregated by taking the median module score for all spots in a layer (L1-L6 and WM) for each array, and rescaling the median layer-array module scores to a 0–1 range. These scaled layer-array module scores are used for statistical testing, and are further aggregated into mean module scores per ALS phenotype for plotting.

snRNA-seq module scores are calculated for each cell as above, and aggregated to the donor tissue sample by taking the mean module score for all cells of a specified type. These aggregated mean scores are rescaled to a 0–1 range and used for statistical testing, and are further aggregated into mean module scores per ALS phenotype for plotting.

Welch’s t-test was used for statistical tests between module scores in different groups, controlling for false discovery with Benjamini-Hochberg for snRNA-seq or Benjaminini-Yekutieli for ST data.

Gene Ontology Expression Scores

Gene ontology expression scores were calculated as in Module Scores, using genes that are annotated with specified GO terms. GO term annotations used are listed in Table S13.

Assigning Genes to Cell Type Submodules

A gene-gene correlation matrix for all genes in a spatial module was constructed from standardized snRNA-seq transcript expression. This was subtracted from 1 to get a correlation distance matrix, which was used as the distance matrix for a gene-gene k-NN graph embedding (k=10). Genes were projected into 2 dimensions with the UMAP algorithm, and clustered into 2 to 5 clusters using the leiden algorithm (resolution of 0.75). Each cluster was interpreted as a gene submodule, and this process was repeated for all 23 spatial modules. Modules 21 and 22 (chrM and chrY) did not yield more than one submodule. Submodule scores were calculated for individual cells and spatial spots as above. Spatial modules and submodules are listed in Table S11.

Spatial Co-occurrence Analysis

Pearson correlation was calculated for spatial module and spatial submodule scores across all spatial spots. Modules and submodules are assigned to cortex layers by aggregating into layer-array module scores as above, and then taking the average module score across all layers. The module is assigned to the layer with the highest module score, and any other layer within 95% of that value. If all scores are within 10%, the module is assigned to the pan-layer label ‘ALL’.

Spatial Cell Type Deconvolution

Per-spot estimates of cell type abundances were generated using Cell2location²⁸. The prefrontal cortex atlas assembled in Prefrontal Cortex Atlas and Classifier was used to train a negative binomial (NB) regression model of gene expression for each of the 5 major glial cell types, for the 4 inhibitory neuron subtypes, and for the 3 excitatory neuron layer subtypes (Table S4), which was used for all subsequent deconvolution experiments. Deconvolution with Cell2location was performed separately for each donor, treating individual ST arrays as batches and assuming an average of five cells per Visium spot with all other hyperparameters set to default values. Mean estimates of the abundance of each cell type in each spot were extracted and used for differential abundance testing.

STMN2 Cryptic Exon Quantitation

For each 10x Multiome sample, aligned transcripts from BAM files were filtered for reads mapped to the STMN2 gene (chr8, 79610814-79666175; hg38). Unique transcript barcode-UMI combinations were counted to quantify the total number of STMN2 transcripts in the library. The filtered transcripts were further filtered for any aligned overlap to the ALS-linked STMN2 cryptic exon (226bp; chr8, 79616822-79617048; hg38), and unique transcript barcode-UMI combinations were counted to obtain the total number of cryptic exon-containing (or intron-retained) STMN2 transcripts in the library.

For each 10x Visium array, aligned transcripts were processed using the same approach: reads mapping to the STMN2 locus were first identified and counted, followed by filtering for overlap with the STMN2 cryptic exon. Transcript counts were assigned to array locations by lookup against the 10x Visium v1 barcode table, and those array locations were annotated by cortical layer for downstream analysis.

Cell Segmentation, Phenotyping, and Morphology Quantification

Images were processed and analyzed using HALO and HALO AI (v4.1.5944.221; Indica Labs). Cortical layers were annotated following the same strategy used for spatial transcriptomics sections, relying on the combined distribution of DAPI, GFAP, NeuN, and NfL to delineate white matter through layer 1. Because layers 3 and 4 are difficult to distinguish reliably and exhibit highly similar marker profiles, these layers were aggregated into a single composite layer (L3/L4) for all analyses (Summarized imaging data is provided in Table S10).

Nuclear and cell detection were performed using DAPI as the nuclear signal, with cell boundaries estimated by applying a standardized nuclear inflation procedure. Cell-type identification was then assigned based on marker expression within these inflated cellular volumes: GFAP for astrocytes, IBA1 for microglia, and NeuN overlapping with DAPI for neurons. This approach enabled consistent and unbiased cell-counting analyses across samples.

All 11 fluorescence channels were manually intensity-calibrated for each image to define three expression categories – low, medium, and high – used for downstream quantification. For astrocyte morphological analysis, we trained a HALO AI membrane segmentation module. 276 astrocytes were manually outlined across 10 independent samples, corresponding to a total annotated membrane area of 90,269.84 μm², together with 42 background fields covering 1,578,967.42 μm². The model was trained for 5,000 iterations and converged to a final cross-entropy loss of 0.295, enabling robust and unbiased astrocyte detection and membrane segmentation.

Cellular roundness (circularity) was computed from the HALO-derived cell area and perimeter measurements using the standard formula:roundness = 4π × (area / perimeter²), providing a quantitative descriptor of astrocyte morphology.

Quantification of Perivascular Niches

Vascular structures were identified using classifier-based annotations generated in HALO AI (Indica Labs), producing tissue-wide masks of the vascular compartment. Briefly, we trained a supervised classifier in HALO AI using manually curated vascular annotations. Classifier training was performed on a total of 1,153 annotations derived from 10 independent donors spanning all three clinical groups, to ensure robustness across disease states and inter-donor variability. Training annotations included αSMA⁺ vessels, Claudin5-positive vessels lacking αSMA signal, and background examples. Annotations were distributed across all cortical layers and imaging fields to capture morphological and staining variability.

For the purposes of this analysis, perivascular domains were defined exclusively around αSMA-ensheathed blood vessels. From these vascular masks, concentric radial rings were generated at 2 μm increments extending from the vessel boundary to a maximum distance of 40 μm, yielding 20 discrete radial bins per vessel. Microglial/macrophage IBA1 signal was quantified using the Area Quantification FL analysis module in HALO. For each radial ring, the module computed the percentage of IBA1-positive area relative to the total tissue area contained within that ring. This analysis was performed independently for each cortical layer and donor. CD68 signal was quantified using the same Area Quantification FL analysis module in HALO. CD68 intensity thresholds (low, mid, and strong) were defined prior to analysis based on visual assessment of signal intensity. Thresholds were established by one investigator and independently reviewed by a second investigator to ensure agreement and consistency. This manual thresholding strategy was adopted to account for inter-individual variability in signal intensity attributable to differences in tissue preservation and background noise across samples, as donors exhibited variable signal-to-noise characteristics. This signal intensity variability was observed across all markers, not solely in the CD68 channel.

Although the underlying tissue annotations comprised hundreds of individual vascular and microglial/macrophage objects per layer, the Area Quantification module outputs summary values per donor per radial ring. Testing each ring independently would therefore result in a large number of comparisons and a substantial multiple-testing burden. To address this while preserving biologically meaningful spatial resolution, individual rings were aggregated into four 10-μm radial segments corresponding to distinct perivascular zones: 0–10 μm (rings 1–5; vessel-adjacent zone), 10–20 μm (rings 6–10), 20–30 μm (rings 11–15), and 30–40 μm (rings 16–20). Within each segment, all ring values from donors belonging to the same clinical group were pooled.

This aggregation reduced the original 20-ring representation to four biologically interpretable spatial domains while retaining variability across donors and radial positions. For each cortical layer and radial segment, a global three-group comparison across control, ALS, and ALSci donors was first performed using one-way analysis of variance (ANOVA). Pairwise comparisons relative to the control group were subsequently assessed using the Mann–Whitney U test. Effect sizes were quantified using Cohen’s d, calculated on pooled values within each radial segment.

General procedures

Unless otherwise noted, data analysis was performed in Python, using the scanpy⁷³ ecosystem, plots were generated with matplotlib⁷⁴, and figures were assembled using Adobe Illustrator.

Supplementary Material

1

Data S1. Comprehensive donor- and sample-level metadata and quality control for the multimodal dataset, related to Figures 1, 2, 3, 4, and 6.

3

Table S1. Donor metadata, including ECAS scores and neuropathology assessments, related to Figure 1.

4

Table S2. Sample metadata (QC metrics, cell-type composition, and assay coverage), related to Figure 1.

5

Table S3. 10x Visium library metadata (per-library QC and layer annotations), related to Figure 1.

6

Table S4. Published snRNA-seq data cohorts used for model training, related to Data S1 Figure 3.

7

Table S5. List of key marker genes used for broad cell-type identity annotation, related to Figure 1.

8

Table S6. 10x Multiome library metadata (per-library QC metrics and cell type counts), related to Figure 1.

9

Table S7. List of key cell subtype marker genes, related to Figures 2 and 6.

10

Table S8. List of DESeq2-derived gliosis marker genes used to score gliosis, related to Figures 2.

11

Table S9. List of reactive astrocyte markers compiled from published studies, related to Figures 2.

12

Table S10. Count of segmented cells in each multiplexed imaging slide by annotated anatomical region, related to Figure 3.

13

Table S11. Gene assignments to spatial co-expression modules and submodules, related to Figure 4.

14

Table S12. GO enrichment for spatial co-expression modules, related to Figure 4.

15

Table S13. Genes associated with specific GO terms used for ontology module calculations, related to Figure 7.

16

Table S14. List of TDP-43 binding targets used from published eCLIP dataset, related to Figures 7 and S11.

Document S1. Figures S1–S13.

Key resources table

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Antibodies
Rabbit polyclonal anti-ALDH1L1	Proteintech	Cat#: HPA050139; RRID:AB_2681031
Rabbit polyclonal anti-AQP4	Proteintech	Cat#: 16473-1-AP; RRID:AB_2827426
Mouse polyclonal anti-aSMA	Lunaphore	Cat#: MR10100; RRID:AB_3751426
Mouse monoclonal anti-CD68	Agilent	Cat#: M0814; RRID:AB_2314148
Mouse monoclonal anti-Claudin5	Thermo Fisher Scientific	Cat#: 35-2500; RRID:AB_2533200
Chicken polyclonal anti-GFAP	Novus	Cat#: NBP1-05198; RRID:AB_1556315
Rabbit monoclonal anti-IBA1	Abcam	Cat#: ab178847; RRID:AB_2832244
Chicken polyclonal anti-MAP2	Origene	Cat#: TA309162; RRID: RRID:AB_3751976
Mouse monoclonal anti-NeuN	Millipore Sigma	Cat#: MAB377; RRID:AB_177621
Mouse monoclonal anti-Neurofilament Light	Millipore Sigma	Cat#: N0142; RRID:AB_477257
Rabbit monoclonal anti-Olig2	Abcam	Cat#: ab109186; RRID:AB_10861310
Mouse monoclonal anti-pTau	Thermo Fisher Scientific	Cat#: MN1020; RRID:AB_223647
Mouse monoclonal anti-pTDP-43	2B Scientific	Cat#: CAC-TIP-PTD-MO1; RRID:AB_3251193
Mouse monoclonal anti-Somatostatin	Santa Cruz Biotechnology	Cat#: SC-74556; RRID: AB_2271061
Mouse monoclocal anti-alpha-Synuclein	Leica Biosystems	Cat#: NCL-L-ASYN; RRID: AB_442103
Alexa Fluor 488 Donkey Anti-Chicken	Invitrogen	Cat#: A78948; RRID:AB_2921070
Alexa Fluor 555 Goat Anti-Rabbit	Lunaphore	Cat#: DR555RB; RRID: AB_3751977
Alexa Fluor 647 Goat Anti-Mouse	Lunaphore	Cat#: DR647MS; RRID: RRID: AB_3751978
Biological samples
Brain Tissue Samples from ALS and Control cases	Edinburgh Brain Bank	See Table S1
Chemicals, peptides, and recombinant proteins
2-propanol	Sigma-Aldrich	Cat#: I9516; CAS 67-63-0
20X SSC Buffer	Thermo Fisher Scientific	Cat#: AM9770
Bluing Buffer	Agilent	Cat#: CS70230-2
Bovine Serum Albumin	Miltenyi Biotech	Cat#: 130-091-376; CAS 9048-46-8
Calcium Chloride	Sigma-Aldrich	Cat#: 21115-100ML; CAS 10043-52-4
DAPI	Lunaphore	Cat#: DR100; CAS 28718-90-3
Decrosslinking solution	10x Genomics	Cat#: 2000566
Dithiothreitol (DTT)	Sigma-Aldrich	Cat#: 646563-10X.5ML; CAS 3483-12-3
EDTA pH 8	Thermo Fisher Scientific	Cat#: AM9260G
Elution Buffer Kit	Advanced Cell Diagnostics Inc.	Cat#: BU07-L
Eosin Y solution	Sigma-Aldrich	Cat#: HT110216-500ML
Ethanol	Sigma-Aldrich	Cat#: E7023; CAS 64-17-5
IGEPAL CA-630	Sigma-Aldrich	Cat#: I8896-50ML; CAS 9002-93-1
Imaging Buffer	Advanced Cell Diagnostics Inc.	Cat#: BU09
Magnesium Acetate	Sigma-Aldrich	Cat#: 63052-100ML; CAS 142-72-3
Mayer’s Hematoxylin	Agilent	Cat#: S330930-2
Multistaining Buffer	Advanced Cell Diagnostics Inc.	Cat#: BU06
OCT	Fisher Healthcare	Cat#: 4585
Optiprep	STEMCELL Technologies	Cat#: 07820
Potassium Hydroxide	Sigma-Aldrich	Cat#: P4494-50ML; CAS 1310-58-3
Protector RNAse Inhibitor	Roche	Cat#: 03335402001
Quenching Buffer Kit	Advanced Cell Diagnostics Inc.	Cat#: BU08-L
Sucrose	Sigma-Aldrich	Cat#: S0389-500G; CAS 57-50-1
SYTOX Green Nucleic Acid Stain	Thermo Fisher Scientific	Cat#: S7020
Tris (1M) pH 7.0	Thermo Fisher Scientific	Cat#: AM9850G
Tris (1M) pH 8.0	Thermo Fisher Scientific	Cat#: AM9855G
Trypan Blue	NanoEntek	Cat#: EBT-001; CAS 72-57-1
Xylene	Sigma-Aldrich	Cat#: 534056; CAS 1330-20-7
Critical commercial assays
70-μm Flowmi Cell Strainer	Bel-Art	Cat#: H13680-0070
Agilent High Sensitivity DNA Kit	Agilent	Cat#: 5067-4626
C-Chip Hemocytometers	INCYTO	Cat#: DHC-N015
Chromium Next GEM Chip J Single Cell Kit	10x Genomics	Cat#: PN-1000234
Chromium Next GEM Single Cell Multiome ATAC + Gene Expression Reagent Bundle	10x Genomics	Cat#: PN-1000283
Dual Index Kit TT Set A	10x Genomics	Cat#: PN-100215
KAPA Library Quantification Kit for Illumina Platforms	KAPA Biosystems	Cat#: KK4824
Library Construction Kit	10x Genomics	Cat#: PN-10000190
NovaSeq 6000 S4 Reagent v1.5 Kit (200 Cycles)	Illumina	Cat#: 20028313
NovaSeq X 25B Reagent Kit (300 Cycles)	Illumina	Cat#: 20104706
Single Index Kit N Set A	10x Genomics	Cat#: PN-1000212
SPRIselect, Liquid	Beckman Coulter	Cat#: B23318
Visium Spatial Gene Expression Slide and Reagent Kit	10x Genomics	Cat#: PN-1000184
Visium Spatial Tissue Optimization Slide and Reagent Kit	10x Genomics	Cat#: PN-1000193
Deposited data
Raw and analyzed Multiome (snRNA-seq/snATAC-seq) data	This paper	NCBI GEO: GSE290359
Raw and analyzed Visium ST data	This paper	NCBI GEO: GSE313918
Raw and analyzed RNA-seq data	This paper	NCBI GEO: GSE314526
Raw prefrontal cortex snRNA-seq data	Lau et al., 2020 [REF22]	NCBI GEO: GSE157827
Raw prefrontal cortex snRNA-seq data	Sadick et al., 2022 [REF23]	NCBI GEO: GSE167494
Raw prefrontal cortex and motor cortex snRNA-seq data	Pineda et al., 2024 [REF9]	NCBI GEO: GSE174332
Raw prefrontal cortex snRNA-seq data	Morabito et al., 2022 [REF24]	NCBI GEO: GSE174367
Raw prefrontal cortex snRNA-seq data	Ma et al., 2022 [REF25]	NCBI GEO: GSE207334
Raw prefrontal cortex and motor cortex snRNA-seq data	Li et al., 2023 [REF12]	NCBI GEO: GSE219280
Raw prefrontal cortex snRNA-seq data	Velmeshev et al., 2019 [REF26]	NCBI: PRJNA434002
Raw prefrontal cortex snRNA-seq data	Gittings et al., 2023 [REF21]	Synapse: syn45351388
Raw prefrontal cortex snRNA-seq data	Mathys et al., 2023 [REF27]	Synapse: syn51123517
Software and algorithms
HALO | Quantitative Image Analysis for Pathology, v4.1.5944.221	Indica Labs	https://indicalab.com/halo/
HORIZON Viewer	Lunaphore Technologies	https://lunaphore.com/products/horizon/
ZEN Microscopy Software	Zeiss	https://www.zeiss.com/microscopy/en/products/software/zeiss-zen.html#LanguageSwitchOverlayCloseButton
Spaceranger (v3.0.0) | Visium ST data processing	10x Genomics	https://www.10xgenomics.com/support/software/space-ranger
Cellranger ARC (v2.0.2)	10x Genomics	https://www.10xgenomics.com/support/software/cell-ranger-arc
Cellranger (v7.1.0)	10x Genomics	https://www.10xgenomics.com/support/software/cell-ranger
Splotch	Daly et al., 2025 [REF39]	https://github.com/adaly/cSplotch
scself | Spatial module discovery and scoring	This work	https://github.com/GreshamLab/sc_self_supervised
PyDeseq2	Muzellec et al., 2023 [REF61]	https://github.com/scverse/PyDESeq2
SEA | simple enrichment analysis of motifs	Bailey and Grant, 2021 [REF71]
Cellpose	Pachitariu and Stringer, 2022 [REF60]	https://github.com/mouseland/cellpose
Scanpy	Wolf et al., 2018 [REF73]	https://github.com/scverse/scanpy
Matplotlib	Hunter et al., 2007 [REF74]	https://github.com/matplotlib/matplotlib

Highlights:

• Multimodal spatial profiling with clinicopathological stratification of ALS brain

• Language and fluency deficits track with diffuse glio-vascular responses

• Executive impairment maps to synaptic and mitochondrial changes in deep-layer neurons

• ALS cognitive heterogeneity has multicellular origins beyond TDP-43 proteinopathy