Abstract
Autism Spectrum Disorders (ASD) are clinically and genetically heterogeneous. Nevertheless, a characteristic gene expression signature in postmortem cortex is shared across most ASD cases. Knowing whether this signature reflects changes in cells of particular types would help to determine the molecular, cellular, and anatomical etiology of ASD. To investigate we took advantage of cell type marker genes defined by recent single cell transcriptome sequencing from mouse or human cortex. We find that microglial markers showed significantly and substantially higher median expression in postmortem ASD than in control cortex for both mouse and human markers (21% higher for orthologous markers defined in mouse; 93% higher for human markers). In contrast, neuronal markers showed reduced median expression: 12% and 16% lower for mouse interneurons and pyramidal neurons, 40% lower for human neurons, and 36% lower for a class of human excitatory projection neurons. Cell type density alterations in ASD brains are indicated because distributions of these cell type markers are shifted concordantly. Importantly, orthologous mouse neuronal marker genes encoding proteins localized to diverse neuronal compartments all showed reduced median expression in ASD. Our results provide a framework for revealing the basis of transcriptomic differences in ASD. We propose comparing cell type density alterations in post-mortem tissue with and without these distinctive gene expression changes to reconcile results from stereological and transcriptomic studies.
Introduction
Autism Spectrum Disorders (ASD) are heterogeneous and highly heritable neurodevelopmental disorders characterized by verbal and nonverbal communication deficits, impaired social interactions, and repetitive behaviors from early childhood1–3. Despite its strong heritability, genetic susceptibility to ASD in the population is distributed over a thousand genes4, which complicates efforts to generalize from deep functional studies of rare but highly penetrant variants. Given this panoply of genetic causes, and a wide variety of proposed causal mechanisms for ASD, it was unexpected that most post-mortem brains from people with ASD exhibit similar gene expression profiles5. By studying this characteristic gene expression signature we seek to reveal neurological features that are common to much of the ASD spectrum.
We are only beginning to understand this molecular signature. Alternative splicing and the attenuation of regional gene expression differences in cortex appear to be key components5, 6. Nevertheless, there are also early signs that larger differences exist with regard to the density of certain cell types or subcellular components. In particular, a set of co-expressed genes underexpressed in ASD is enriched in synaptic genes5, ASD risk genes5, 7 and markers of glutamatergic neurons5, 7, and contains some interneuron markers5. A set of co-expressed genes overexpressed in ASD is enriched for immune-related genes, markers for activated microglia, and markers for astrocytes5. Another study found an overrepresentation of genes overexpressed in ASD in immune cell types and astrocytes and an overrepresentation of genes underexpressed in ASD in neuronal cell types, particularly in interneurons8.
A parsimonious neurological interpretation of cell type marker enrichment, nevertheless, remains unclear. Such enrichments might be driven by a subset of markers of that type and could therefore reflect changes in cell signaling, physiological state, functional pathways, structure, and more. For example, defects in synapse formation in ASD9, 10 may underlie the underexpression of a subset of neuronal marker genes expressed at the synapse. On the other hand, if most neuronal markers are underexpressed in ASD, this could also suggest a reduction in neuronal density. Here we use density to refer to cell number density (the number of cells in a volume of tissue)11 or other ratios such as volume density (the total volume of cells in a volume of tissue)12, 13 or cross-sectional area stained with a marker (e.g. Nissl-stained area to unstained area in a section)14. Indeed, there are scattered and sometimes conflicting reports of changes in neuronal density, which may reflect small sample size and the heterogeneity of ASD14–19. Identifying the subset of post-mortem brains in which these changes are expected to occur and the cell types involved in these alterations could guide future confirmatory stereological studies in a more targeted manner.
Previous studies have used a biased subset of cell type marker genes from a biased subset of cell types isolated with pre-selected markers followed by bulk gene expression profiling8, 20. However, clustering cells by type without prior information about their markers is now feasible with single cell transcriptomics21. Objective, comprehensive, and selective cell type markers can be defined from these clusters22. In this study, we examine how orthologs of mouse cell type markers23 are shifted in their expression in postmortem ASD brains. Furthermore, we obtain similar findings with human cell type markers from a second single cell transcriptome study6.
Materials and Methods
Datasets and Gene Lists
Expression Data
We used normalized, regressed, log2 transformed post-mortem gene expression values from the superior temporal gyrus of 12 ASD cases and 12 matched controls, ranging from 15 to 60 years old and of both sexes6. The cases and controls in that study were 57% of the individuals chosen from a larger unpublished RNA sequencing cohort on the basis that they display the strongest of this characteristic gene expression signature.
Cell Type Marker Genes
Mouse cell type marker genes were taken from Table S1 of Zeisel et al.23. They were based on RNA sequencing of single cells isolated on a Fluidigm C1 AutoPrep System from the somatosensory cortex and CA1 of hippocampus from mice of both sexes between postnatal days 21 and 3123. Human cell type marker genes, as well as marker genes of neuronal communities, were taken from Table S3 of Darmanis et al.24 and were based on RNA sequencing of single cells isolated on a Fluidigm C1 from freshly resected temporal cortex of several humans of both sexes aged 21 through 63 years24. The numbers of these marker genes are provided in Supplementary Table 1.
Localization of Neuronal Marker Genes
We used S1 pyramidal marker genes from a single cell transcriptomic study from mouse brain tissue23 having 1:1 mouse to human orthologs as defined in Ensembl. The human Ensembl identifiers for these transcripts gave linked subcellular protein localization information for their encoded proteins from the UniProt database25. Markers were separated into groups based on whether they contained the following Gene Ontology Cellular Component identifiers: "GO:0005634" for nuclear protein localization, "GO:0030425" for protein localization to dendrites, and "GO:0030424" for protein localization to axons26. Most transcripts do not have only one localization identifier. Since there was a considerable overlap between transcripts with the axonal localization identifier and the dendritic localization identifier, these were analyzed together as a single group. To improve specificity, transcripts with both nuclear localization identifiers and axonal or dendritic identifiers were removed from the analysis. We analyzed the median log2(expression fold change) of these different subsets of mouse orthologous S1 pyramidal neuron and interneuron marker genes as described below. Caveats of this method include that it assumes that a change in density of neuronal processes (and only processes) would be reflected in the expression levels of transcripts whose proteins are localized to neuronal processes but would not be reflected to a comparable extent in the expression levels of transcripts whose proteins are localized to the cell soma27, 28, that there is conservation of protein localization for these markers from mouse to human, and that the ontology’s cellular localization evidence may require further validation26.
Examining Cell Type Gene Expression Changes in ASD
For the mouse cell type markers, we cross-referenced the mouse gene symbols to Ensembl Mouse Gene identifiers, and then took the 1:1 mouse-human orthologs that were also included in the ASD sequencing study6 using Ensembl BioMart29. For the human cell type markers, we kept those markers that were in the ASD sequencing study. We plotted the distributions of their log2 fold changes with R. We tested the two-tailed significance of the median log2 fold change for each cell type by constructing a null distribution by permuting case-control labels 1000 times and then subtracting the mean log2 (FPKM) in controls from the mean log2 (FPKM) in cases. These were corrected for multiple testing using the “p.adjust” function in R30, applying the Benjamini and Hochberg False Discovery Rate (FDR) method31 with an FDR cut-off of 5%. Confidence intervals were calculated from 1000 bootstraps.
Results
We first tested a null hypothesis that the median expression fold change for each of eight high-level cortical cell types defined in mouse23 was not significantly different from the change observed when case-control labels are randomly shuffled Figure 1). To do so we considered first a set of marker genes defined by single cell sequencing in mouse cortex, before then obtaining similar results for marker genes of six high-level cell types defined in human cortex (Figure 2; Supplementary Tables 2–3). Compared to controls, the median expression of mouse-defined gene markers for S1 pyramidal neurons was significantly lower (by 10–21% at an FDR<5%31) in ASD post-mortem human cortex tissue; similarly, median expression of mouse-defined interneuron markers was reduced in ASD brains by 8–16%. By contrast, median expression of mouse-defined microglia markers was increased in ASD brains by 15–28%. Ependymal, endothelial, and mural marker genes also exhibited slightly increased expression ( at a slightly relaxed False Discovery Rate31 of 5.1%).
Similar statistically significant results were observed, with even stronger effects, using the smaller set of cell type gene markers obtained from single cell transcriptomics in human cortex24 (Figure 2; Supplementary Table 3). Human cell type marker median gene expression for neurons was decreased in post-mortem human ASD cortex tissue by 30–49%. Conversely, microglia marker median expression was increased in ASD by 48–109°%, along with oligodendrocyte precursor cell (OPC) marker genes, whose median expression was increased in ASD by 8–22%. We then repeated this analysis for four previously defined excitatory or inhibitory neuronal subpopulations in human24. Gene markers of subpopulation 4, representing a type of excitatory neuron, were decreased in median expression in ASD brains by 22–40% (Figure 2C, Supplementary Table 4).
Having found that a majority of gene markers of several cortical cell types change in expression in ASD, we next asked whether these observations were more consistent with alterations in a specific pathway or cell state, or else with larger changes in cell type density. We would expect that alterations in specific pathways exclusively would produce expression changes amongst a minority subset of genes where expression of some markers (those contributing to the pathway or cell state) change while others do not. Instead, we found that expression of these marker genes shifted concordantly (Figure 3), suggesting cell type density changes in ASD.
Expression changes of neuronal marker genes could reflect alterations not in cell number density, but in the density of subcellular compartments such as synapses. To consider this, we next examined whether protein products of neuronal marker genes that are localized in different neuronal compartments alter their mRNAs’ expression in ASD coordinately. S1 pyramidal and interneuron mouse markers were classified based on whether their protein products are localized in the nucleus of neurons or in dendrites and axons (Figure 4A, Supplementary Table 5). Mouse S1 pyramidal gene markers whose protein products are localized in the nucleus, and axons and dendrites, are similarly and significantly lower in transcript expression in ASD brains by 9–32°% and 6–32 %, respectively. The corresponding changes for mouse interneuron markers were 2–18 % and 6–24°%, respectively. Approximately three-quarters of genes for each of these localization categories show a decreased expression fold change (Figure 4B). These findings suggest that reduced expression of neuronal marker genes reflects decreased cell density.
Discussion
In this study, the majority of neuronal and microglial cell marker genes exhibit coordinated expression level change in ASD brains. Previous enrichments of microglial markers in overexpressed-in-ASD ‘modules’ of co-expressed genes and enrichments of neuronal markers in underexpressed-in-ASD ‘modules’ of co-expressed genes5 were consistent with altered cellular processes besides changes in cell density. For example, increased microglial activation, observed in living young men with ASD32, might result in overexpression of M2 microglial activation genes even in the absence of a change in microglia density33. Furthermore, the enrichment of synaptic marker genes in under-expressed modules in ASD5 may be due to dysfunctional synaptic formation9, 10. Indeed, gene markers related to neurotransmission are underexpressed in adult ASD post-mortem cortex34. Nevertheless, that study failed to find differences in expression of neuronal and glial marker genes in ASD, probably due to the few markers used. Although there is some overlap between the cell type marker genes used here and microglia activation genes33 or neurotransmission genes34, we propose that the parsimonious interpretation of these results is that ASD brains contain alterations in cell density. We do so for the following reasons: (1) marker genes shift their expression together in the same direction in the ASD versus control comparison, rather than being driven by a minority subset as expected when only a specific pathway is altered, and (2) the median fold changes for transcripts encoding proteins localized in different neuronal compartments are all reduced, which is unexpected if only neuronal processes changed in their density.
For microglia, the issues of cell state and density may be correlated, because microglial activation can trigger proliferation35, 36. Microglial activation, increased density, and increased somal volume in ASD cortex have all been reported37, 38. The role of microglial activation and proliferation in ASD is unclear but increased microglia-neuronal spatial clustering in the dorsolateral prefrontal cortex in ASD suggests that active microglia are recruited to neurons39.
Decreases in neuronal density may not necessarily be caused by a reduction in relative cell number (i.e. the proportion of cell type nuclei versus all nuclei). We cannot exclude the possibility that these cell type marker gene expression changes in ASD may be due to differences in cell size, because gene expression of individual transcripts increases in proportion to cell volume40, 41. Hence, decreased neuronal density may be explained by less cytoplasmic space being taken up by neurons as a proportion of total cytoplasmic space.
Studies analyzing neuronal composition and density in ASD using immunohistochemistry and stereology have been conflicting, perhaps due to small sample sizes and the disorders’ heterogeneity. Patches of cortical disorganization and a small increase in neuronal density have been reported in children with ASD16. In the fusiform gyrus, reduced neuronal density, neuronal numbers, and cell volumes were reported in ASD, but these changes were not detected in whole cortical grey matter or primary visual cortex42. In the dorsolateral and mesial prefrontal cortex of children, an increase in neuronal counts but not in neuronal volume was observed43. However, in Brodmann areas 44 and 45, smaller pyramidal neurons, albeit similar pyramidal neuronal numbers, were reported in ASD44. An interesting possibility is that the underlying structure of neurons in the cortex is different in ASD, resulting in decreased density in the cortical section sequenced. Increased center-to-center minicolumn width has been observed in different cortical areas in ASD, including a trend for increased width in the primary auditory cortex (Brodmann 41, overlapping the area used in this study)19. Wider minicolumn spacing in ASD may explain decreased expression of neuronal markers, as it would imply a lower density of neurons.
These results are consistent with previous transcriptomic studies, and expand on them. These have shown an enrichment of interneuron and glutamatergic neuron markers in an ASD underexpressed gene module in the cortex and enrichment of activated microglia and astrocyte markers in an ASD overexpressed module5, 7. Later work reported an over-representation of overexpressed genes in immune cells and astrocytes and an over-representation of underexpressed genes in interneurons and putative cortical projection neurons in the Voineagu et al. dataset8. Another study also observed overexpression of microglia and astrocyte markers and underexpression of neuronal markers in the prefrontal cortex, but used few markers45. Here, we used objectively defined gene markers, allowing us to pinpoint excitatory neuronal markers as being specifically downregulated. We also included non-neuronal and glial cell types, which previous studies did not include. Instead of analyzing the enrichment of cell type markers in underexpressed and overexpressed genes in ASD, we tested whether these markers change expression as a group. For the first time, to our knowledge, this method allows us make inferences about the composition of cell types in the ASD brain.
The transcriptomic data used in this study was obtained from the superior temporal gyrus. Since this brain region is important for auditory processing and speech46–48 and has been implicated in ASD49, 50, it is important to determine how and if cell type density changes here lead to behavioral deficits. Although these gene expression patterns may be regionally specific, there is strong overlap between differentially expressed genes in the temporal and frontal cortex, suggesting that gene expression changes in the cortex might be global and not representative of a particular region5.
In this study, we used a cell type transcriptomic approach to infer cellular composition differences in ASD cortex. The median gene expression of inhibitory or excitatory neuronal marker genes was significantly reduced in ASD. This was also the case when subdividing mouse orthologous neuronal markers based on the subcellular localization of the proteins they encode. The median gene expression of microglia markers was significantly increased in ASD. These gene expression patterns suggest changes in cell type number and/or volume density. Given the conflicting results in published immunohistochemistry and stereology studies, understanding the underlying causes of these global gene expression signatures in ASD should be prioritized. Groups with expertise in stereology might now independently validate this inference from markers to cell type density. Since the patients used in this analysis were a subset displaying strong characteristic differential gene expression changes in ASD, future integrated transcriptomics and stereological studies could compare density changes in post-mortem tissue with and without these distinctive gene expression changes. Moreover, this study provides a proof of concept for using single cell transcriptomic analyses derived marker genes to study cell type composition changes in disease, and could be easily extended to the study of other diseases such as schizophrenia, where cell type marker enrichment has been reported in gene co-expression modules51.
Acknowledgements
RBM was supported by the Clarendon Fund; RBM, TGB, and CPP were supported by the Medical Research Council; and TGB was supported by the Human Brain Project. The research leading to these results has received funding from the European Union Seventh Framework Programme (FP7/2007–2013) under grant agreement no. 604102 (Human Brain Project).Conflict of Interest
The authors declare no financial conflicts of interest.