Spatiotemporal dynamics of molecular pathology in amyotrophic lateral sclerosis

See allHide authors and affiliations

Science  05 Apr 2019:
Vol. 364, Issue 6435, pp. 89-93
DOI: 10.1126/science.aav9776

Spatiotemporal gene expression in ALS

Amyotrophic lateral sclerosis (ALS) is a progressive motor neuron disease that affects nerve cells in the brain and the spinal cord. It has proven difficult to identify the early stages of disease and where it spreads within the body. Maniatis et al. used RNA sequencing to define transcriptomic changes over the course of disease in different regions of the spinal cord of a mouse ALS model and a postmortem human ALS spinal cord. From changes in gene expression, they identified disease-associated pathways and established the key steps in motor neuron degeneration observed in ALS.

Science, this issue p. 89


Paralysis occurring in amyotrophic lateral sclerosis (ALS) results from denervation of skeletal muscle as a consequence of motor neuron degeneration. Interactions between motor neurons and glia contribute to motor neuron loss, but the spatiotemporal ordering of molecular events that drive these processes in intact spinal tissue remains poorly understood. Here, we use spatial transcriptomics to obtain gene expression measurements of mouse spinal cords over the course of disease, as well as of postmortem tissue from ALS patients, to characterize the underlying molecular mechanisms in ALS. We identify pathway dynamics, distinguish regional differences between microglia and astrocyte populations at early time points, and discern perturbations in several transcriptional pathways shared between murine models of ALS and human postmortem spinal cords.

Amyotrophic lateral sclerosis (ALS) is a progressive neurodegenerative disease in which symptoms often manifest at first in distal muscles of a single limb. Symptoms then spread throughout the body, ultimately leading to total paralysis. However, the events that initiate disease pathology, and the mechanisms through which it becomes pervasive, remain poorly understood (1). Mounting evidence indicates that dysfunction in signaling between motor neurons and glia is a key component of the disease (13). Inherent limitations of current gene expression profiling technologies, such as low throughput or lack of spatial resolution, have thus far frustrated efforts to understand how such dysfunction participates in the onset and spread of ALS pathology in the spinal cord. Given the stereotyped cellular organization of the spinal cord and importance of intercellular communication in ALS progression, we reasoned that a spatially resolved view of disease-driven gene expression changes would be needed to order these events, reveal the relevant subpopulations of cells involved in each stage in disease progression, and characterize the underlying molecular mechanisms that trigger and maintain the disease phenotype.

Spatial transcriptomics (ST) generates quantitative transcriptome-wide RNA sequencing (RNA-seq) data through capture of polyadenylated RNA on arrays of spatially barcoded DNA capture probes (4). The workflow does not require specialized equipment, does not require preselection of genes to interrogate, and has a throughput substantially greater than that of other spatially resolved methods. We applied ST to spatially profile gene expression in lumbar spinal cord tissue sections (L3 to L5) from SOD1-G93A (ALS) and SOD1-WT (control) mice at presymptomatic, onset, symptomatic, and end-stage time points (table S1). We then applied ST to profile gene expression in postmortem lumbar and cervical spinal cord tissue sections from either lumbar- or bulbar-onset sporadic ALS patients, resulting in more than 76,000 spatial gene expression measurements (SGEMs) mapping to ~1200 spinal cord tissue sections of 67 mice, and more than 60,000 SGEMs mapping to 80 postmortem spinal cord tissue sections from seven patients (table S1). All raw data have been deposited publicly, as described below. Additionally, an interactive data exploration portal is publicly available at

We annotated each SGEM with an anatomical annotation region (AAR) tag and used these tags to conduct differential expression analysis and to register the data to a common coordinate system (figs. S1 and S2 and table S1). To estimate gene expression levels accurately and detect significant regional, anatomical, and cell type changes within and between conditions, we formulated a hierarchical generative probabilistic model (supplementary methods). Our model corrects for missing data due to undersampling and bias, an issue in spatial and single-cell RNA-seq (scRNA-seq) studies. We thus reliably quantitated the spatial distribution of 11,138 genes in mouse and 9624 genes in human spinal cord sections. Furthermore, principal component analysis of the complete mouse SGEM dataset reveals that most of the variance is explained by spatial location, disease state, and genotype (fig. S3), and not by batch effects.

Our analysis reveals altered expression of several known ALS genes (e.g., Matr3, Kif5a, and Pfn1) ( (5). Our analysis also recapitulates the specific regional and temporal expression patterns for genes with described regional expression profiles (68) and roles in ALS progression. Immunofluorescence (IF) imaging of the protein products of these genes demonstrates spatial concordance with our analysis (Fig. 1 and tables S2 and S3). Furthermore, our data suggest that microglial dysfunction occurs well before symptom onset, precedes astroglial dysfunction in ALS, and is proximal to motor neurons (table S3 and fig. S4).

Fig. 1 Spatially and temporally resolved gene expression in the mouse spinal cord.

(A) Schematic of a hematoxylin and eosin–stained cross-section of mouse lumbar spinal cord with AARs (top). Scale bar, 500 μm. Visualization of colocalized spatial mRNA expression (the posterior means of λ) in all SOD1-WT SGEMs (bottom left). Z maximum projection of 10-μm confocal immunofluorescence image stack (N = 7 animals) (bottom right). Scale bar, 500 μm. (B) Spatial mRNA expression of Aif1 and Gfap at postnatal day 70 (P70) and P100. (C) Z maximum projections of 10-μm confocal immunofluorescence image stacks at P70 and P100 (N = 12 animals). Scale bars, 250 μm.

To further explore the spatiotemporal dynamics of microglial activation, we focused on a mechanism involving TREM2 reported in neurodegenerative disease models (3, 9). TREM2 and TYROBP form a receptor complex that can trigger phagocytosis or modulate cytokine signaling when engaged by membrane lipids, or lipoprotein complexes (9, 10). This mechanism also involves Apoe, Lpl, B2m, and Cx3cr1 and is activated by microglial phagocytosis of apoptotic neurons. Spatial gene expression analysis suggests a spatiotemporal ordering of this TREM2-mediated mechanism in this mouse model of ALS. We observe that Tyrobp expression is up-regulated presymptomatically and before Trem2 in the ventral horn and ventral white matter. Furthermore, Lpl and B2m are simultaneously up-regulated (presymptomatically) specifically in the ventral horn, whereas Apoe and Cx3cr1 are not (Fig. 2, table S3, and fig. S4). Expression of these latter genes becomes widely up-regulated in spinal cords of symptomatic mice (table S3). Apoe expression is driven by Trem2 signaling and is itself a ligand for Trem2. Therefore, Apoe and Trem2 act in an autoregulatory loop that can trigger and maintain a phagocytic microglial phenotype (3). Collectively, our analysis suggests that TREM2- and TYROBP-mediated signaling is an early step in disease-relevant changes in microglial gene expression and reveals the spatiotemporal ordering of these changes.

Fig. 2 Presymptomatic dysregulation of TREM2- and TYROBP-mediated signaling.

(A) The posterior distributions of coefficient parameters β of Tyrobp at P30, P70, P100, and P120. The coefficient parameters β capture offsets of expression (in natural logarithmic space) in distinct AARs across all tissue sections of a given condition. (B) As in (A), for Trem2. (C) Spatial mRNA expression of Tyrobp at P70. (D) Z maximum projections of 10-μm confocal immunofluorescence image stacks in ventral-lateral spinal cords at P70 (N = 6 animals). Scale bar, 250 μm.

Trem2 mutations are associated with several neurodegenerative diseases (1012) and, through mammalian target of rapamycin (mTOR) signaling in myeloid cells (10), Trem2 expression modulates autophagy. Mutations in several autophagy-related genes are associated with ALS (10). ST analysis and IF imaging show that genes involved in autophagy and the endolysosomal system are dysregulated in the spinal cord of this ALS mouse model (fig. S5 and table S3). Ablation of autophagy by conditional knockout of Atg7 in cholinergic cells, including motor neurons (ChAT-Cre+/+; Atg7fl/fl; SOD1-G93A), leads to earlier symptom onset but prolonged survival in ALS mice (2). This manipulation also partially rescues reactive gliosis in ALS mice. To investigate which pathways might link dysfunction in autophagy to gliosis and motor neuron loss in ALS, we applied our methods to these mice (Atg7 cKO). As expected, we observe that expression of Gfap and Aif1, and activity of the TREM2 microglial activation axis, are greatly reduced when autophagy is ablated in motor neurons, particularly in AARs distal to motor neuron somata (table S3).

To better understand disease-relevant changes in gene regulation and interactions between cell types, we carried out an unbiased coexpression analysis of our mouse ST data. We identified 31 major coexpression modules (13) (Fig. 3A and table S4) of diverse spatiotemporal and pathway activities, a subset of which is affected in Atg7 cKO (Fig. 3B, figs. S6 and S7A, and table S5). In the context of published scRNA-seq data (14), many of the modules are composed of genes preferentially expressed in specific cell types (fig. S7B). We grouped the genes of each module on the basis of their cell type–specific expression pattern, resulting in submodules (13) (Fig. 3C and tables S6 and S7). Submodules that are enriched for a given cell type can display distinct spatiotemporal expression patterns depending on the parent module from which they are derived. Such differences represent functionally distinct subpopulations within that cell type. For example, submodules containing Prdx6 and Gfap (submodule 8.9) or Slc7a10 and Bcan (submodule 29.41) represent regional astrocyte subpopulations that behave differently across the course of disease (fig. S8A). In ALS animals, the Prdx6 submodule displays increasing activity as disease progresses, whereas the Slc7a10 submodule is attenuated with progression. Intriguingly, the disease-related dynamics of the Prdx6 astrocyte submodule in ALS animals are rescued in Atg7 cKO, but those of the Slc7a10 astrocyte submodule are not. Our spatial analysis thus identifies gene expression programs characteristic of regional astrocyte populations (15) that display distinct, disease-relevant spatiotemporal dynamics and are differentially dependent on cholinergic autophagy.

Fig. 3 Spatiotemporal dynamics of gene expression during disease progression in ALS.

(A) Biclustering of the mouse SGEMs reveals spatially and temporally coexpressed genes. The dashed vertical purple line in the dendrogram denotes the break. Identifiers of coexpression modules are listed. (B) Average spatiotemporal expression dynamics of modules 8 and 11. (C) Hierarchical clustering of genes in module 8 using independent gene expression data of mouse central nervous system cell types. The dashed purple line in the dendrogram denotes the break. Identifiers for submodules having at least 10 genes are listed. Selected genes are noted. (D) Analysis of enriched KEGG pathways among the genes for submodules in (C) [one-tailed Fisher’s exact test with Benjamini-Hochberg correction; FDR (false discovery rate) < 0.1].

This approach also allows us to identify coordinated activities across cell types that reflect major aspects of disease and to highlight expression programs within distinct cell types associated with these processes (fig. S7B). In one such example, a microglial expression program containing Trem2, Tyrobp, Aif1, and other reactive microglial genes displays spatiotemporal dynamics reflecting patterns of progressive gliosis in ALS (submodule 8.17) (3, 16, 17). A second microglial expression program containing Lrp1 and Gba (submodule 8.4) exhibits spatiotemporal dynamics correlated with those of the Trem2 microglial submodule 8.17. Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway enrichment analysis of these submodules shows that they are both enriched for lysosomal factors (Fig. 3D). However, whereas the Trem2 submodule includes many factors involved in the complement cascade, Fc receptor–mediated signaling, and phagocytosis, the Lrp1 submodule includes many sphingolipid signaling factors. Both the Trem2 and Lrp1 submodules exhibit correlated spatiotemporal dynamics with astrocyte submodule 8.9. Lastly, the activity of these microglial and astroglial submodules is correlated with activities of oligodendrocyte precursor (OPC) submodule 8.24 and mature oligodendrocyte submodules 8.18 and 8.19 (fig. S8B). Notably, the spatiotemporal pattern of module 8 expression is rescued by ablation of autophagy in cholinergic neurons in Atg7 cKO mice, demonstrating the key role that neural interactions with module 8 glia play in these activities. Collectively, the pathway activities encompassed by module 8 reveal signaling within and between cell types during early glial activation in intact tissue and the mechanisms that maintain and spread the reactive phenotype in the ALS mouse model.

A microglial expression program that includes Sall1 (submodule 1.12) is increased in the white matter of control and presymptomatic ALS animals (fig. S8C). Sall1 is expressed by homeostatic microglia, and loss of Sall1 expression in microglia results in a phagocytic, inflammatory phenotype (3). The expression pattern of the Sall1 submodule illustrates glial interactions that are spatiotemporally and mechanistically distinct from those of module 8. By the end stage in ALS animals, this submodule has infiltrated the gray matter and is attenuated in the white matter. This expression program is coordinated with astrocyte submodule 1.15 and oligodendrocyte submodule 1.8 (fig. S8D). Expression of Sall1 in this submodule suggests that the late-stage expansion of the microglial population responsible for this expression program differs from the reactive microglial populations present in modules 8 and 6 (fig. S8C). Further, the collective spatiotemporal expression pattern of module 1 in ALS animals is consistent with late-stage defects in myelination (6). Notably, Atg7 cKO does not rescue the dynamics of this expression module, which encompasses the activities of multiple cell types. Thus, spatiotemporal correlation of the microglial, astroglial, and the oligodendrocyte expression programs in module 1 suggests the coordination of signaling mechanisms in glial behavior related to axonal pathology in late-stage ALS animals. Taken together, these examples demonstrate how disease-related processes are exemplified with the combination of scRNAseq and ST.

We applied our ST workflow to cervical and lumbar tissue from four human ALS patients who presented clinically with bulbar symptom onset and three patients who presented with lower limb symptom onset. These individuals exhibit expected regional expression patterns, resembling those from the mouse dataset (fig. S9). Consistent with previous studies (18, 19), our observations appear to support the notion that the severity of ALS pathology is related to proximity to site of symptom onset—human spatial data show variability in gene expression in the anterior horn related to such proximity (table S10). For instance, the gene encoding acetylcholinesterase (ACHE), the activity of which has been linked to neuromuscular defects in ALS (20), shows reduced expression at locations proximal to spinal segments innervating the site of symptom onset (Fig. 4, A and B). As with the mouse spatial data, we conducted an unbiased coexpression analysis on the human spatial data, resulting in 28 human expression modules of which some are preserved between mouse and human (figs. S10, A and B, and S11 and table S8). The spatial mapping patterns for these human expression modules demonstrate AAR characteristic patterns, some of which vary along the rostrocaudal axis (Fig. 4C), or differ between white matter and gray matter (fig. S10B), or with proximity to site of symptom onset (fig. S10B). For example, consistent with previous studies (18), human expression module 1, which includes genes involved in VEGF and glutamatergic signaling, is attenuated in lumbar sections from patients that have a lower limb site of onset (fig. S10B). In addition, human expression module 3 is attenuated across spinal cord sections at sites proximal to symptom onset (Fig. 4C). Furthermore, this attenuation is most pronounced in the posterior white matter and anterior horns. KEGG analysis shows that human module 3 is enriched for several pathways, including sphingolipid, retrograde endocannabinoid, and WNT signaling (fig. S10C and table S9). This finding, along with human and mouse submodules that display disease-relevant dynamics and are enriched for sphingolipid signaling pathways, underscores the importance of this mechanism in ALS pathology. Indeed, altered glycosphingolipid levels and their metabolism have been reported in spinal cords of ALS patients and in murine models of SOD1 ALS (21). Modulators of sphingolipid signaling have been proposed as potential therapeutics for ALS and improve the disease in murine models of ALS (21, 22). Our data detail the dynamics of sphingolipid signaling in multiple cell types, spinal cord regions, and disease stages and suggest target opportunities for designing therapeutics. Further, the spatiotemporal nature of our data provides insight for potential treatment strategies modulating the activity of this pathway.

Fig. 4 Spatiotemporal transcriptome of human postmortem spinal cord tissue from ALS patients.

(A) The posterior difference distributions of the ventral horn coefficients for ACHE per patient (D1 to D4). Differences are calculated between distal and proximal regions with respect to the onset location. (B) Spatial mRNA expression of ACHE in human postmortem lumbar and cervical spinal cord. (C) Average spatiotemporal expression dynamics for human coexpression modules 3, 25, and 27 are visualized.

Taken together, we provide a comprehensive spatiotemporal, transcriptome-wide gene expression dataset combining resolution, replication, and biological perturbation. Our procedure allows us to draw inferences from murine models and test them in clinical samples. As such, we expect the work presented here to be a resource to spur further mapping of the central nervous system and its modes of dysfunction.

Supplementary Materials

Materials and Methods

Figs. S1 to S16

Tables S1 to S10

References (2443)

References and Notes

  1. Materials and methods are available as supplementary materials.
Acknowledgments: We thank Target ALS Multicenter Postmortem Core for providing human postmortem tissue; G. Akbalik, J. Gregory, I. Hubbard, D. Kim, and K. Wei for manual anatomical annotation; T. Maniatis for Atg7 cKO mice; the Flatiron Institute for computational resources; and NGI Stockholm and SciLifeLab for infrastructure support. Funding: The study was supported by Target ALS, The ALS Association (grant no. 15-LGCA-234), The Tow Foundation, the Knut and Alice Wallenberg Foundation, and the Simons Foundation. Author contributions: H.P., S.M., and S.V. designed the experiments. S.M. and S.V. performed the experiments, with help from C.B., K.K., M.C., Ž.A., S.S, G.S.-C., and A.M. T.Ä. and R.B. developed and implemented the Bayesian generative model and the interactive data exploration portal. S.M., T.Ä., S.V., and D.F. analyzed the data. A.W. implemented the SGEM annotation tool. All authors discussed the results and wrote the manuscript. Competing interests: J.L. is an author on a patent applied for by Spatial Transcriptomics AB/10x Genomics Inc. covering the described technology. Data and materials availability: Raw and processed mouse data and images have been deposited at NCBI’s Gene Expression Omnibus (GEO) Repository under project ID GSE120374. Raw human data have been deposited at New York Genome Center and are available upon request submitted to A code implementing the used statistical model is available at (23). All processed data and images used in the analyses have been deposited to
View Abstract

Stay Connected to Science

Navigate This Article