Abstract
A subset of COVID-19 patients exhibit altered olfactory function. Here we analyze bulk and single cell RNA-Seq datasets to identify cell types in the olfactory epithelium and olfactory bulb that express cell entry molecules that mediate infection by SARS-CoV-2 (CoV-2), the causal agent in COVID-19. We find that samples from whole olfactory mucosa in species including mouse and human express two key genes involved in CoV-2 entry, ACE2 and TMPRSS2. However, neither olfactory sensory neurons nor olfactory bulb neurons express these genes, which are instead expressed in support cells, stem cells, and perivascular cells. These findings suggest that CoV-2 infection of non-neuronal cell types leads to anosmia and related disturbances in odor perception in COVID-19 patients.
One Sentence Summary Analysis of new and previously published single-cell sequencing datasets reveals that the SARS-CoV2 receptor ACE2 is expressed in olfactory support cells, stem cells and perivascular cells — but not in neurons — suggesting mechanisms through which the COVID-19 syndrome could lead to olfactory dysfunction.
Introduction
SARS-CoV-2 (CoV-2) is a pandemic coronavirus that causes the COVID-19 syndrome, which can include upper respiratory infection (URI) symptoms, severe respiratory distress, acute cardiac injury and death (1–4). CoV-2 is closely related to other beta-coronaviruses, including the causal agents in pandemic SARS and MERS (SARS-CoV and MERS-CoV, respectively) and endemic viruses typically associated with mild URI syndromes (hCoV-OC43 and hCoV-229E) (5–7). Clinical reports suggest that infection with CoV-2 is associated with high rates of disturbances in smell and taste perception, including anosmia (8–13). These observations are consistent with a case report describing a SARS patient with long term anosmia (14) and with data demonstrating that olfactory function is commonly altered after infection with endemic coronaviruses (7, 15–18). While many viruses induce transient changes in odor perception due to inflammatory responses, in at least some cases COVID-related anosmia has been reported to occur in the absence of significant nasal inflammation or coryzal symptoms (12). This observation suggests that CoV-2 might directly target odor processing mechanisms, although the specific means through which CoV-2 alters odor perception remains unknown.
The nasal epithelium is anatomically divided into a respiratory epithelium (RE) and sensory olfactory epithelium (OE), whose functions and cell types differ. The nasal RE is continuous with the epithelium that lines much of the respiratory tract and is thought to humidify air as it enters the nose; main cell types include basal cells, ciliated cells, secretory cells (including goblet cells), and brush/microvillar cells (19, 20) (Figure 1).
The OE is a laminated structure housing mature olfactory sensory neurons (OSNs), which detect odors via receptors expressed on their dendritic cilia. OSNs are supported by sustentacular cells, which act to structurally support sensory neurons, phagocytose and/or detoxify potentially damaging agents, and maintain local salt and water balance (21–23); microvillar cells and mucus-secreting Bowman’s gland cells also play important roles in maintaining OE homeostasis and function (19, 24)(Figure 1). In addition, the OE contains globose basal cells (GBCs), which are primarily responsible for regenerating OSNs during normal epithelial turnover, and horizontal basal cells (HBCs), which act as reserve stem cells activated upon tissue damage (25–27). OSNs elaborate axons that puncture the cribriform plate at the base of the skull and terminate in the olfactory bulb, whose local circuits process olfactory information before sending it to higher brain centers.
CoV-2 — like SARS-CoV — infects cells through interactions between its spike (S) protein and the ACE2 protein on target cells. This interaction requires cleavage of the S protein, likely by the cell surface protease TMPRSS2, although other proteases (such as Cathepsin B and L, CTSB/CTSL) may also be involved (4–6, 28–32). Other coronaviruses use different cell surface receptors and proteases to facilitate cellular entry, including DPP4, FURIN and HSPA5 for MERS-CoV, ANPEP for HCoV-229E, TMPRSS11D for SARS-CoV (in addition to ACE2 and TMPRSS2), and ST6GAL1 and ST3GAL4 for HCoV-OC43 and HCoV-HKU1 (6, 33–35). It has recently been demonstrated through single cell RNA sequencing analysis (referred to herein as scSeq) that cells from the human upper airway — including nasal RE goblet, basal and ciliated cells —express high levels of ACE2 and TMPRSS2, suggesting that these RE cell types may serve as a viral reservoir during CoV-2 infection (36). However, analyzed samples in that dataset did not include any OSNs or sustentacular cells, suggesting that tissue sampling in these experiments did not access the OE (37, 38). Here we query both new and previously published bulk RNA-Seq and scSeq datasets from the olfactory system for expression of ACE2, TMRPSS2 and other genes implicated in coronavirus entry. We find that non-neuronal cells in the OE and olfactory bulb, including support, stem and perivascular cells, express CoV-2 entry-associated molecules, raising the possibility that infection of these non-neuronal cell types contributes to anosmia in COVID-19 patients.
Results
To determine whether genes relevant to CoV-2 entry are expressed in OSNs or other cell types in the OE, we examined published datasets in which RNA-Seq was independently performed on mouse whole olfactory mucosa (WOM) and on purified populations of mature OSNs (39–41). The CoV-2 receptor Ace2 and the protease Tmprss2 were expressed in WOM, as were the cathepsins Ctsb and Ctsl (Figures 2A-B) (39). However, expression of these genes (with the exception of Ctsb) was much lower and Ace2 expression was nearly absent in purified OSN samples (Figure 2A, see Legend for counts, see Methods). Genes used for cell entry by other CoVs (except St3gal4) were also expressed in WOM, and de-enriched in purified OSNs. These samples reflected the presence of expected cell types, as the WOM sequencing data included gene expression profiles associated with a wide variety of RE and OE cells, while the purified OSN samples were specifically enriched for markers of mature OSNs. The de-enrichment of Ace2 and Tmprss2 in OSNs relative to WOM was also observed in two other mouse RNA-Seq datasets (40, 41) (Figure S1A).
The presence of Ace2 and Tmprss2 transcripts in mouse WOM and their (near total) absence in purified OSNs suggest that the molecular components that enable CoV-2 entry into cells are expressed in non-neuronal cell types in the mouse nasal epithelium. To identify the specific cell types that express Ace2 and Tmprss2, we performed scSeq (using the 10x Chromium platform, see Methods) on the WOM; olfactory sensory neurons did not express Ace2 in this de novo analysis (2 of 28769 mature OSNs were positive for Ace2), while expression was observed in Bowman’s gland cells and HBCs in a small percentage of cells (Figures 2C-E and S1B-C, see methods). Because WOM scSeq de-enriches certain cell types (e.g., sustentacular cells), to complement these data we employed a previously established lineage tracing protocol to perform scSeq on HBCs and their descendants during injury-induced regeneration (27). This analysis revealed that Ace2 and Tmprss2 are expressed in subsets of sustentacular cells and HBCs, as well as activated HBCs that serve to regenerate the epithelium (Figures 2F-H and S2A). To validate these results, we re-analyzed a similar lineage tracing dataset in which identified HBCs and their progeny were subject to Smart-seq2-based deep sequencing (27). In this dataset, Ace2 was detected in more than 0.7% of GBCs, nearly 2% of HBCs and nearly 3% of sustentacular cells but was not detected in OSNs (Figures 2I and S2B). Furthermore, larger percentages of HBCs, GBCs and sustentacular cells expressed Tmprss2. Taken together, these bulk RNA-Seq and scSeq mouse data demonstrate that CoV-2 cell entry-related genes are expressed in support and stem cells in the mouse OE.
To address whether there are similar patterns of gene expression in human OE, we queried previously published bulk WOM RNA-Seq data derived from macaque, marmoset and human (43), and found expression of almost all CoV-entry-related genes in all WOM samples, consistent with our observations in mouse WOM (Figure S3A). To identify the specific cell types in human OE that express ACE2, we quantified gene expression in scSeq derived from four human nasal biopsy samples recently reported by Durante et al (44). As in the mouse scSeq datasets, neither ACE2 nor TMPRSS2 were detected in mature OSNs, whereas these genes were detected in both sustentacular cells and HBCs (Figures 3A-E and S3B-C). In contrast, genes relevant to cell entry of other CoVs were expressed in OSNs, as well as in other OE cell types. We confirmed the expression of ACE2 proteins via immunohistochemical staining of human olfactory epithelium biopsy tissue, which revealed expression in sustentacular and basal cells, and an absence of ACE2 protein in OSNs (Figure 3F). Together, these results demonstrate that sustentacular and olfactory stem cells, but not mature OSNs, are potentially direct targets of CoV-2 in the human OE.
Given that the nasopharynx is a major site of infection for CoV-2 (11), we compared the frequency of ACE2 and TMPRSS2 expression among the cell types in the human RE and OE (44). Sustentacular cells exhibited the highest frequency of ACE2 expression in the OE (2.90% of cells) although this frequency was slightly lower than that observed in respiratory ciliated and secretory cells (3.65% and 3.96%, respectively). Although all HBC subtypes expressed ACE2, the frequency of expression of ACE2 was lower in olfactory HBCs (0.84% of cells) compared to respiratory HBCs (1.78% of cells) (Figure 3B). In addition, all other RE cell subtypes showed higher frequencies of ACE2 and TMPRSS2 expression than was apparent in OE cells.
These results demonstrate the presence of key CoV-2 entry-related genes in specific cell types in the OE, but at lower levels of expression than in RE isolated from the nasal mucosa. We wondered whether these lower levels of expression might nonetheless be sufficient for infection of CoV-2. It was recently reported that nasal RE has higher expression of CoV-2 entry genes than RE of the trachea or lungs (36), and we therefore asked where OE fell within this previously established spectrum of expression. To address this question, we developed a two step alignment procedure in which we first sought to identify cell types that were common across the OE and RE, and then leveraged gene expression patterns in these common cell types to normalize gene expression levels across all cell types in the OE and RE (Figure S4). This approach revealed a correspondences between goblet cells in the RE and Bowman’s gland cells in the OE (96% mapping probability, see Methods), and between pulmonary ionocytes in the RE and a subset of microvillar cells in the OE (99% mapping probability, see Methods); after alignment, human OE sustentacular cells were found to express ACE2 and TMPRSS2 at levels similar to that observed in the remainder of the non-nasal respiratory tract (Figure 3G) (36).
Given the potential for the RE and OE in the nasal cavity to be directly infected with CoV-2, we assessed the expression of Ace2 and other CoV entry genes in the mouse olfactory bulb (OB), which is directly connected to the OSNs in OE via cranial nerve I (CN I); in principle, alterations in bulb function could cause anosmia independently of functional changes in the OE. To do so, we performed scSeq (using DropSeq, see Methods) on the mouse OB, and merged these data with a previously published OB scSeq analysis, yielding a dataset with nearly 50,000 single cells (see Methods) (45). This analysis revealed that Ace2 expression was absent from OB neurons and instead was observed only in vascular pericytes (Figures 4A-D and S5–6), which are involved in blood pressure regulation, maintenance of the blood-brain barrier, and inflammatory responses (46). Although other potential CoV proteases were expressed in the OB, Tmprss2 was not expressed in the olfactory bulb. We also performed Smart-seq2-based deep sequencing of single OB dopaminergic neurons (Figures 4E and S7, see Methods), which confirmed the absence of Ace2 and Tmprss2 expression in that cell type. In addition, re-analysis of 10 deeply sequenced datasets from different regions of the nervous system demonstrated that Ace2 expression is absent from neurons, consistent with prior immunostaining results (Figure S8)(47). Together these findings suggest that OB neurons are likely not a primary site of infection, but that vascular pericytes may be sensitive to CoV-2 infection in the OB and brain.
Discussion
Here we demonstrate that subsets of OE sustentacular cells, HBCs, microvillar cells and Bowman’s gland cells express the CoV-2 receptor ACE2 and the spike protein protease TMPRSS2. Human OE sustentacular cells express these genes at levels comparable to that observed in lung cells. However, we failed to detect ACE2 expression in mature OSNs at either the transcript or protein levels. These observations suggest that CoV-2 does not directly enter OSNs, but instead may target OE support and stem cells. Similarly, neurons in the olfactory bulb fail to express ACE2, whereas vascular pericytes do. Thus primary infection of non-neuronal cell types — rather than sensory or bulb neurons — may be responsible for anosmia and related disturbances in odor perception in COVID-19 patients.
The identification of non-neuronal cell types in the OE and bulb susceptible to CoV-2 infection suggests three possible, non-mutually-exclusive mechanisms for the acute loss of smell reported in COVID-19 patients. First, local infection of support cells and vascular pericytes in the nose and bulb could cause significant inflammatory responses whose downstream effects could alter the function of OSNs or bulb neurons. Second, damage to support cells (which are responsible for local water balance and ionic balance) could indirectly influence signaling from OSNs to the brain; our finding that a subset of microvillar cells are homologues of pulmonary ionocytes — which play a key role in maintaining extracellular chloride gradients in the lung — is consistent with this model (48). Finally, damage to sustentacular cells and Bowman’s gland cells in mouse models leads to diffuse architectural damage to the entire OE, which could abrogate smell perception (49).
Our data suggest that only sparse subsets of support cells and pericytes express ACE2 (although this may reflect undersampling of transcripts by scSeq). Intriguingly, exposure of the rat OE to the SDAV coronavirus causes infection of only a limited number of sustentacular cells, yet over time OSNs become denuded of cilia (presumably foreclosing the possibility of odor transduction) and the global architecture of the OE becomes disrupted (49). These observations suggest that coronavirus infection of subsets of sustentacular cells may be sufficient to cause a pathophysiological cascade that culminates in damage to OSNs and defects in olfactory function.
The natural history of CoV2-induced anosmia is only now being defined; while recovery of smell has been reported, it remains unclear whether in a subset of patients smell disturbances will be long-lasting or permanent (8–13). It is possible that infection of stem cells including both resting and activated HBCs — which express ACE2 and TMPRSS2 at lower levels than RE or sustentacular cells — could result in long-lasting olfactory dysfunction. The survival time of mature OSNs varies based upon many factors, but on average in the mouse such neurons survive weeks to short months, rather than days or years (50, 51). Although HBCs in mice appear largely quiescent, recent scSeq data suggests that ongoing neurogenesis may play a larger role than previously appreciated in the human OE (44). It is therefore possible that the dual challenge of acute loss of sustentacular cells, together with the inability to effectively renew the OE over time, may cause persistent anosmia.
Many viruses, including coronaviruses, have been shown to propagate from the nasal epithelium past the cribriform plate to infect the olfactory bulb; this form of central infection has been suggested to mediate olfactory deficits, even in the absence of lasting OE damage (16, 30, 52–55). The rodent coronavirus MHV passes from the nose to the bulb, even though rodent OSNs do not express CEACAM1, the main MHV receptor (53, 56) (Figures S1A/C and S2A), suggesting that CoVs in the nasal mucosa can reach the brain through mechanisms independent of axonal transport by sensory nerves; in this case, OB neurons express CEACAM1 (Figure 4E), which likely supports the ability of MHV to target these neurons and change odor perception. One speculative possibility is that local seeding of the OE with CoV-2-infected cells can result in OSN-independent transfer of virions from the nose to the bulb, perhaps via the vascular supply shared between the OB and the OSN axons that comprise CN I. Although CN I was not directly queried in our datasets, it is reasonable to infer that vascular pericytes in CN I also express ACE2, which suggests a possible route of entry for CoV-2 from the nose into the brain. Given the absence of ACE2 in OB neurons, we speculate that any central olfactory dysfunction in COVID-19 is the secondary consequence of pericyte-mediated vascular inflammation (46).
We note several caveats that temper our conclusions. Here, we assume that ACE2 protein levels follow the pattern revealed by analysis of mRNA; while our staining in human OE is consistent with this idea, it remains to be rigorously assessed in all cell types in the OE and bulb. Although current data suggest that ACE2 is the most likely receptor for CoV-2 in vivo, it is possible (although it has not yet been demonstrated) that other molecules such as BSG may enable CoV-2 entry independently of ACE2 (Figures 3E, 3G, S1, S2) (57, 58). We also propose that damage to the olfactory system is either due to primary infection or secondary inflammation; it is possible (although has not yet been demonstrated) that cells infected with CoV-2 can form syncytia with cells that do not express ACE2. Such a mechanism could damage neurons adjacent to infected cells. Definitive identification of the pathophysiological mechanisms underlying COVID-19-mediated anosmia will require additional research. Nonetheless, our identification of cells in the OE and OB expressing molecules known to be involved in CoV-2 entry illuminates a path forward for future studies.
Materials and Methods
Mouse bulk RNA-Seq datasets
Normalized gene expression tables were obtained from previous published datasets (39–41, 43). For the mouse data sets, the means of the replicates from WOM or OSN were used to calculate Log2 fold changes. For the mouse data from Saraiva et al. and the primate data sets (39, 43), the normalized counts of the genes of interest from individual replicates were plotted. Below is a table with detailed sample information.
Sample information for the bulk RNA-seq data analyzed in this study
WOM and HBC lineage tracing mouse 10x scSeq experiments
Mice
2 month-old and 18 month-old wild type C57BL/6J mice were obtained from the National Institute on Aging Aged Rodent Colony and used for the WOM experiments; each experimental condition consisted of one male and one female mouse to aid doublet detection. Mice containing the transgenic Krt5-CreER(T2) driver (59) and Rosa26-YFP reporter allele (60) were used for the HBC lineage tracing dataset. All mice were assumed to be of normal immune status. Animals were housed and maintained according to IACUC guidelines.
Single-Cell RNA Sequencing
The olfactory epithelium was surgically removed, and the dorsal, sensory portion was dissected and dissociated, as previously described (27). For WOM experiments, dissociated cells were subjected to fluorescence-activated cell sorting (FACS) using propidium iodide to identify and select against dead or dying cells; 100,000 cells/sample were collected in 10% FBS. For the HBC lineage tracing experiments Krt5-CreER; Rosa26YFP/YFP mice were injected once with tamoxifen (0.25 mg tamoxifen/g body weight) at P21-23 days of age and sacrificed at 24 hours, 48 hours, 96 hours, 7 days and 14 days post-injury, as previously described (27, 42). For each experimental time point, YFP+ cells were isolated by FACS based on YFP expression and negative for propidium iodide, a vital dye.
Cells isolated by FACS were subjected to single-cell RNA-seq. Three replicates (defined here as a FACS collection run) per age were analyzed for the WOM experiment; at least two biological replicates were collected for each experimental condition for the HBC lineage tracing experiment. Single cell cDNA libraries from the isolated cells were prepared using the Chromium Single Cell 3’ System according to the manufacturer’s instructions. The WOM preparation employed v3 chemistry with the following modification: the cell suspension was directly added to the reverse transcription master mix, along with the appropriate volume of water to achieve the approximate cell capture target. The HBC lineage tracing experiments were performed using v2 chemistry. The 0.04% weight/volume BSA washing step was omitted to minimize cell loss. Completed libraries were sequenced on Illumina HiSeq4000 to produce paired-end 100nt reads.
Sequence data were processed with the 10X Genomics Cell Ranger pipeline (2.0.0 for v2 chemistry), resulting in the initial starting number before filtering of 60,408 WOM cells and 25,469 HBC lineage traced cells. The scone R/Bioconductor package was used to filter out lowly-expressed genes (fewer than 2 UMI’s in fewer than 5 cells) and low-quality libraries (using the metric_sample_filter function with arguments hard_nreads = 2000, zcut = 4).
Preliminary Filtering
Cells with co-expression of male (Ddx3y, Eif2s3y, Kdm5d, and Uty) and female marker genes (Xist) were removed as potential doublets from the WOM dataset. For both datasets, doublet cell detection was performed per sample using DoubletFinder and Scrublet (63). Genes with at least 3 UMIs in at least 5 cells were used for downstream clustering and cell type identification. For the HBC lineage tracing dataset, the Bioconductor package scone was used to pick the top normalization (“none,fq,ruv_k=1,no_bio,batch”), corresponding to full quantile normalization, batch correction and removing one factor of unwanted variation using RUV (64). A range of cluster labels were created by clustering using the partitioning around medoids (PAM) algorithm and hierarchical clustering in the clusterExperiment Bioconductor package (65), with parameters k0s=(10,13,16,19,22,25) and alpha=(NA,0.1,0.2,0.3). Clusters that did not show differential expression were merged (using the function mergeClusters with arguments mergeMethod = ‘adjP’, cutoff = 0.01, and DEMethod = ‘limma’ for the lineage-traced dataset). Initial clustering identified one Macrophage (Msr1+) cluster consisting of 252 cells; upon its removal and restarting from the normalization step a subsequent set of 15 clusters was obtained. These clusters were used to filter out 1515 cells for which no stable clustering could be found (i.e., ‘unassigned’ cells), and four clusters respectively consisting of 31, 29 and 23 and 305 cells. Doublets were identified using DoubletFinder and 271 putative doublets were removed. Inspection of the data in a three-dimensional UMAP embedding identified two groups of cells whose experimentally sampled timepoint did not match their position along the HBC differentiation trajectory, and these additional 219 cells were also removed from subsequent analyses.
Analysis of CoV-related genes in WOM and HBC lineage datasets
Analysis of WOM scSeq data were performed in python using the open-source Scanpy software starting from the raw UMI count matrix of the 40179 cells passing the initial filtering and QC criteria described above. UMIs were total-count normalized and scaled by 10,000 (TPT, tag per ten-thousands) and then log-normalized. For each gene, the residuals from linear regression models using the total number of UMIs per cell as predictors were then scaled via z-scoring. PCA was then performed on a set of highly-variable genes (excluding OR genes) calculated using the “highly_variable_genes” function with parameters: min_mean=0.01, max_mean=10, min_disp=0.5. A batch corrected neighborhood graph was constructed by the “bbknn” function with 42 PCs with the parameters: local_connectivity=1.5, and embedding two-dimensions using the UMAP function with default parameters (min_dist = 0.5). Cells were clustered using the neighborhood graph via the Leiden algorithm (resolution = 1.2). Identified clusters were manually merged and annotated based on known marker gene expression. We removed 281 cells containing mixtures of marker genes with no clear gene expression signature. The identified cell types and the number of each of the remaining 39898 cells detected were as follows. 28,769 mOSN: mature OSN; 2,607 iOSN: immature OSN; 859 INP: Immediate Neural Precursor; 623 GBC: Globose Basal Cell; HBC: Horizontal Basal Cell (1,083 Olfactory and 626 Respiratory); 480 SUS: sustentacular cell; 331 BG: Bowman’s gland; MV: Microvillar cell (563 Brush-like and 1,530 Ionocyte-like); 92 OEC: Olfactory Ensheathing Cell; 76 Resp. Secretory cells; 227 Resp. unspecified cells; 172 atypical OSN; 1,757 various immune cells, 103 RBC: Red Blood Cell. TPT gene expression levels were visualized in two-dimensional UMAP plots.
The filtered HBC lineage dataset containing 21722 cells was analyzing in python and processed for visualization using pipelines in SPRING and Scanpy (66, 67). In brief, total counts were normalized to the median total counts for each cell and highly variable genes were selected using the SPRING gene filtering function (“filter_genes”) using parameters (90, 3, 3). The dimensionality of the data was reduced to 20 using principal components analysis (PCA) and visualized in two-dimensions using the UMAP method with parameters (n_neighbors=20, min_dist=0.5). Clustering was performed using the Leiden algorithm (resolution=1.45) and clusters were merged manually using known marker genes. The identified cell types and number of each type were: 929 mOSN: mature OSN; 2073 iOSN: immature OSN; 786 INP: Immediate Neural Precursor; 755 GBC: Globose Basal Cell; HBC: Horizontal Basal Cell (7782 Olfactory, 5418 Regenerating, and 964 Respiratory); 2666 SUS: sustentacular cell; and 176 Ionocyte-like Microvillar (MV) cell.
Expression of candidate CoV-2-related genes was defined if at least one transcript (UMI) was detected in that cell, and the percent of cells expressing candidate genes was calculated for each cell type. In the WOM dataset Ace2 was only detected in 2 out of 28,769 mature OSNs (0.007 %), and in the HBC lineage dataset, Ace2 was not detected in any OSNs. Furthermore, Ace2 was not detected in immature sensory neurons (GBCs, INPs, or iOSNs) in either dataset.
Mouse HBC lineage Smart-Seq2 dataset
Single-cell RNA-seq data from HBC-derived cells from Fletcher et al. and Gadye et al (27, 42), labeled via Krt5-CreER driver mice, were downloaded from GEO at accession GSE99251 using the file “GSE95601_oeHBCdiff_Cufflinks_eSet_counts_table.txt.gz”. Processing was performed as described above, including total counts normalization and filtering for highly variable genes using the SPRING gene filtering function “filter_genes” with parameters (75, 20, 10). The resulting data were visualized in SPRING and a subset of cells were removed for quality control, including a cluster of cells with low total counts and another with predominantly reads from ERCC spike-in controls. Putative doublets were also identified using Scrublet and removed (6% of cells) (63). The resulting data were visualized in SPRING and partitioned using Louvain clustering on the SPRING k-nearest-neighbor graph using the top 40 principal components. Cell type annotation was performed manually using the same set of markers genes listed above. Three clusters were removed for quality control, including one with low total counts and one with predominantly reads from ERCC spike-in controls (likely background), and one with high mitochondrial counts (likely stressed cells). For visualization, and clustering the remaining cells were projected to 15 dimensions using PCA and visualized with UMAP with parameters (n_neighbors=15, min_dist=0.4, alpha=0.5, maxiter=500). Clustering was performed using the Leiden algorithm (resolution=0.4) and cell types were manually annotated using known marker genes.
The filtered dataset of mouse HBC-derived cells contained 1450 cells. The percent of cells expressing each marker gene was calculated as described above. Of the 51 OSNs identified, none of them expressed Ace2, and only 1 out of 194 INPs and iOSNs expressed Ace2. In contrast, Ace2 and Tmprss2 were both detected in HBCs and SUS cells.
Human nasal scSeq dataset
Human scSeq data from Durante et al. (44) was downloaded from the GEO at accession GSE139522. 10X Genomics mtx files were filtered to remove any cells with fewer than 500 total counts. Additional preprocessing was performed as described above, including total counts normalization and filtering for highly variable genes using the SPRING gene filtering function “filter_genes” with parameters (90, 3, 10). The resulting data were visualized in SPRING and partitioned using Louvain clustering on the SPRING k-nearest-neighbor graph. Four clusters were removed for quality control, including two with low total counts (likely background) and two with high mitochondrial counts (likely stressed or dying cells). Putative doublets were also identified using Scrublet and removed (7% of cells). The remaining cells were projected to 40 dimensions using PCA. PCA-batch-correction was performed using Patient 4 as a reference, as previously described (68). The filtered data were then re-partitioned using Louvain clustering on the SPRING graph and each cluster was annotated using known marker genes, as described in (44). For example, immature and mature OSNs were identified via their expression of GNG8 and GNG13, respectively. HBCs were identified via the expression of KRT5 and TP63 and olfactory HBCs were distinguished from respiratory HBCs via the expression of CXCL14 and MEG3. Identification of SUS cells (CYP2A13, CYP2J2), Bowman’s gland (SOX9, GPX3), and MV ionocytes-like cells (ASCL3, CFTR, FOXI1) was also performed using known marker genes. For visualization, the top 40 principal components were reduced to two dimensions using UMAP with parameters (n_neighbors=15, min_dist=0.4).
The filtered human scSeq dataset contained 33358 cells. Each of the samples contained cells from both the olfactory and respiratory epithelium, although the frequency of OSNs and respiratory cells varied across patients, as previously described (44). 295 cells expressed ACE2 and 4953 cells expressed TMPRSS2. Of the 865 identified OSNs, including both immature and mature cells, none of the cells express ACE2 and only 2 (0.23%) expressed TMPRSS2. In contrast, ACE2 was reliably detected in at least 2% and TMPRSS2 was expressed in close to 50% of multiple respiratory epithelial subtypes. The expression of both known cell type markers and known CoV-related genes was also examined across respiratory and olfactory epithelial cell types. For these gene sets, the mean expression in each cell type was calculated and normalized by the maximum across cell types.
Immunohistochemistry
Human olfactory mucosa biopsies were obtained via IRB-approved protocol at Duke University School of Medicine, from nasal septum or superior turbinate during endoscopic sinus surgery. Tissue was fixed with 4% paraformaldehyde and cryosectioned at 10 m and sections were processed for immunohistochemistry, as previously described (44). Primary antibodies were applied overnight at 4 degrees. Primary antibodies included rabbit anti-ACE2 (Abcam, ab15348; RRID:AB_301861) and mouse Tuj1 against neuron-specific tubulin (BioLegend, 801201; RRID:AB_2313773). Biotinylated secondary antibodies (Vector Labs), avidin-biotinylated horseradish peroxidase kit (Vector) followed by fluorescein tyramide signal amplification (Perkin Elmer) were applied per manufacturer’s instructions. For dual staining, Tuj1 was visualized using AlexaFluor 594 Goat anti-Mouse (Jackson ImmunoResearch, Ca# 115-585-146; RRID:AB_2338881). Sections were counterstained with 4’,6-daimidino-2-phenylindole (DAPI) and coverslips were mounted using Prolong Gold (Invitrogen) for imaging, using a Leica DMi8 microscope system. Images were processed using Fiji ImageJ software (NIH). Scale bars were applied directly from the Leica acquisition software metadata in ImageJ Tools. Unsharp Mask was applied in ImageJ, and brightness/contrast was adjusted globally. The Tuj1 antibody was validated, as previously described (44). Anti-ACE2 was raised against a C-terminal synthetic peptide for human ACE2, and was validated by the manufacturer to not cross-react with ACE1 for immunohistochemical labeling of ACE2 in fruit bat nasal tissue as well as in human lower airway. Recombinant human ACE2 abolished labeling with this antibody in a previous study in human tissue, further demonstrating its specificity (69).
Mapping scSeq datasets to each other
Data from Deprez et al. (37) were downloaded from the Human Cell Atlas website (https://www.genomique.eu/cellbrowser/HCA/; “Single-cell atlas of the airway epithelium (Grch38 human genome)”). A subset of these data was combined with a subset of the Durante data for mapping between cell types. For the Deprez data, the subset consisted of samples from the nasal RE that belonged to a cell type with >20 cells, including Basal, Cycling Basal, Suprabasal, Secretory, Mucous Multiciliated cells, Multiciliated, SMS Goblet and Ionocyte. We observed two distinct subpopulations of Basal cells, with one of the two populations distinguished by expression of Cxcl14. The cells in this population were manually identified using SPRING and defined for downstream analysis as a separate cell type annotation called “Basal (Cxcl14+)”. For the Durante data, the subset consisted of cells from cell types that had some putative similarity to cells in the Deprez dataset, including Olfactory HBC, Cycling respiratory HBC, Respiratory HBC, Early respiratory secretory cells, Respiratory secretory cells, Sustentacular cells, Bowman’s gland, Olfactory microvillar cells.
To establish a cell type mapping:
Durante (44) and Deprez (37) data were combined and gene expression values were linearly scaled so that all cells across datasets had the same total counts. PCA was then performed using highly variable genes (n=1477 genes) and PCA-batch-correction (68) with the Durante data as a reference set.
Mapping was then performed bidirectionally between the two datasets. Each cell from “Dataset 1” ‘voted’ for the 5 most similar cells in the “Dataset 2”, using distance in PCA space as the measure of similarity. A table T counting votes across cell types was then computed, where for cell type i in the Dataset 1 and cell type j in the Dataset 2,
Thus, if Dataset 1 has N cells, then T would count 5*N votes (∑Tij = 5N)
The table of votes T was Z-scored against a null distribution, generated by repeating the procedure above 1000 times with shuffled cell type labels.
The resulting Z-scores were similar between the two possible mapping directions (Durante -> Deprez vs. Deprez -> Durante; R=0.87 Pearson correlation of mapping Z-scores). The mapping Z-scores were also highly robust upon varying the number of votes-cast per cell (R>0.98 correlation of mapping Z-scores upon changing the vote numbers to 1 or 50 as opposed to 5). Only cell-type correspondences with a high Z-score in both mapping directions (Z-score > 25) were used for downstream analysis.
To establish a common scale of gene expression between datasets, we restricted to cell type correspondences that were supported both by bioinformatic mapping and shared a nominal cell type designation based on marker genes. These included: Basal/suprabasal cells = “respiratory HBCs” from Durante et al., and “basal” and “suprabasal” cells from Deprez et al. Secretory cells = “early respiratory secretory cells” and “respiratory secretory cells” from Durante et al., and “secretory” cells from Deprez et al. Mutliciliated cells = “respiratory ciliated cells” from Durante et al., and “multiciliated” cells from Deprez et al.
We next sought a transformation of the Durante data so that it would agree with the Deprez data within the corresponding cell types identified above To account for differing normalization strategies applied to each dataset prior to download (log normalization and rescaling with cell-specific factors for Deprez et al. but not for Durante et al.), we used the following ansatz for the transformation, where the pseudocount p is a global latent parameter and the rescaling factors fi are fit to each gene separately. In the equation below, T denotes the transformation and eij represents a gene expression value for cell i and gene j in the Durante data:
The parameter p was fit by maximizing the correlation of average gene expression across all genes between each of the cell type correspondences listed above. The rescaling factors fi were then fitted separately for each gene by taking the quotient of average gene expression between the Deprez data and the log-transformed Durante data, again across the cell type correspondences above.
Juvenile and Adult mouse whole olfactory bulb scRNAseq dataset
Juvenile mouse data
Single-cell RNAseq data from whole mouse olfactory bulb (45) were downloaded from mousebrain.org/loomfiles_level_L1.html in loom format (l1 olfactory.loom) and converted to a Seurat object. Samples were obtained from juvenile mice (age postnatal day 26-29). This dataset comprises 20514 cells passing cell quality filters, excluding 122 cells identified as potential doublets.
Tissue dissection and single-cell dissociation
A new dataset of whole olfactory bulb scSeq was generated from adult male mice (8–12 weeks-old). Briefly, dissected olfactory bulbs (including the accessory olfactory bulb and fractions of the anterior olfactory nucleus) were dissociated in 750 µl of dissociation media (DM: HBSS containing 10mM HEPES, 1 mM MgCl2, 33 mM D-glucose) with 28 U/mL Papain and 386 U/mL DNase I (Worthington). Minced tissue pieces were transferred to a 5 mL round-bottom tube (BD). DM was added to a final volume of 3.3 mL and the tissue was mechanically triturated 5 times with a P1000 pipette tip. After 1-hour incubation with rocking at 37°C, the suspension was triturated with a 10 mL pipette 10 times and 2.3 mL was passed through 40 µm cell strainer (BD). The suspension was then mechanically triturated with a P1000 pipette tip 10 times and 800 µL were filtered on the same strainer. The cell suspension was further triturated with a P200 pipette tip 10 times and filtered. 1 mL of Quench buffer (22 mL of DM, 2.5 mL of protease inhibitor prepared by resuspending 1 vial of protease inhibitor with 32 mL of DM, and 2000U of DNase I) was added to the suspension and centrifuged at 300g for 5 min. Cells were resuspended with 3 mL of Quench buffer and overlaid gently on top of 5 mL of protease inhibitor, then spun down at 70g for 10min. The pellet was resuspended using DM supplemented with 0.04 % BSA, and spun down at 300g for 5 min. Cells were suspended in 400 µL of DM with 0.04 % BSA. All mouse husbandry and experiments were performed following institutional and federal guidelines and approved by Harvard Medical School’s Institutional Animal Care and Use Committee (IACUC).
Drop-seq experiments
Drop-seq experiments were performed as previously described (70). Microfluidics devices were obtained from FlowJEM and barcode beads were obtained from chemgenes. Two 15 min Drop-seq runs were collected from a single dissociation preparation obtained from 2 mice. Two such dissociations were performed, giving 4 total replicates.
Sequencing of Drop-seq samples
4 replicates of Drop-seq samples were pooled and sequenced across 3 runs on an Illumina NextSeq 500 platform. Paired end reads from the fastq files were trimmed, aligned, and tagged via the Drop-seq tools (1-2.0) pipeline, using STAR (2.4.2a) with genomic indices from Ensembl Release 82. The digital gene expression matrix was generated for 8,000 cells per replicate.
Preprocessing of Drop-seq samples
Cells with low numbers of genes (500), low numbers of UMIs (700) or high numbers of UMIs (>10000) were removed (6 % of cells). Potential doublets were identified via Scrublet and removed (3.5 % of cells). Overall, this new dataset comprised 27004 cells.
Integration of whole olfactory bulb scRNAseq datasets
Raw UMI counts from juvenile and adult whole olfactory bulb samples were integrated in Seurat (71). Briefly, raw counts were log-normalised separately and the 10000 most variable genes identified by variance stabilizing transformation for each dataset. The 4529 variable genes present in both datasets and the first 30 principal components (PCs) were used as features for identifying the integration anchors. The integrated expression matrix was scaled and dimensionality reduced using PCA. Based on their percentage of explained variance, the first 28 PCs were chosen for UMAP visualisation and clustering.
Graph-based clustering was performed using the Louvain algorithm following the standard Seurat workflow. Cluster stability was analysed with Clustree on a range of resolution values (0.4 to 1.4), with 0.6 yielding the most stable set of clusters (72). Overall, 26 clusters were identified, the smallest of which contained only 43 cells with gene expression patterns consistent with blood cell types, and was excluded from further visualisation plots. Clustering the two datasets separately yielded similar results. Moreover, the distribution of cells from each dataset across clusters was homogenous (Figure S5) and the clusters corresponded previous cell class and subtype annotations (45). As previously reported, a small cluster of excitatory neurons (cluster 13) contained neurons from the anterior olfactory nucleus. UMAP visualisations of expression level for cell class and cell type markers, and for genes coding for coronavirus entry proteins, depict log-normalized UMI counts. The heatmap in Figure 4B shows the mean expression level for each cell class, normalised to the maximum mean value. The percentage of cells per cell class expressing Ace2 was defined as the percentage of cells with at least one UMI.
Smart-Seq2 sequencing of manually sorted olfactory bulb dopaminergic neurons
Tissue dissociation and manual cell sorting
Acute olfactory bulb 300 µm slices were obtained from Dat-Cre/Flox-tdTomato (B6.SJL-Slc6a3tm1.1(cre) Bkmn/J, Jax stock 006660 / B6.Cg– Gt(ROSA)26Sortm9(CAG-tdTomato)Hze, Jax stock 007909) P28 mice as previously described (73). As part of a wider study, at P27 these mice had undergone brief 24 h unilateral naris occlusion via a plastic plug insert (N = 5 mice) or were subjected to a sham control manipulation (N = 5 mice); all observed effects here were independent of these treatment groups. Single cell suspensions were generated using the Neural Tissue Dissociation Kit – Postnatal Neurons (Miltenyi Biotec. Cat no. 130-094-802), following manufacturer’s instructions for manual dissociation, using 3 fired-polished Pasteur pipettes of progressively smaller diameter. After enzymatic and mechanical dissociations, cells were filtered through a 30 µm cell strainer, centrifuged for 10 minutes at 4° C, resuspended in 500 µl of ACSF (in mM: 140 NaCl, 1.25 KCl, 1.25 NaH2PO4, 10 HEPES, 25 Glucose, 3 MgCl2, 1 CaCl2) with channel blockers (0.1 µM TTX, 20 µM CNQX, 50 µM D-APV) and kept on ice to minimise excitotoxicity and cell death.
For manual sorting of fluorescently labelled dopaminergic neurons we adapted a previously described protocol (74). 50 µl of single cell suspension was dispersed on 3.5mm petri dishes (with a Sylgard-covered base) containing 2 ml of ACSF + channel blockers. Dishes were left undisturbed for 15 minutes to allow the cells to sink and settle. Throughout, dishes were kept on a metal plate on top of ice. tdTomato-positive cells were identified by their red fluorescence under a stereoscope. Using a pulled glass capillary pipette attached to a mouthpiece, individual cells were aspirated and transferred to a clean, empty dish containing 2 ml ACSF + channel blockers. The same cell was then transferred to a third clean plate, changing pipettes for every plate change. Finally, each individual cell was transferred to a 0.2 ml PCR tube containing 2 µl of lysis buffer (RLT Plus - Qiagen). The tube was immediately placed on a metal plate sitting on top of dry ice for flash-freezing. Collected cells were stored at −80C until further processing. Positive (more than 10 cells) and negative (sample collection procedure without picking a cell) controls were collected for each sorting session. In total, we collected samples from 10 mice, averaging 50 tdTomato-positive cells collected per session. Overall, less than 2.5 hours elapsed between mouse sacrifice and collection of the last cell in any session.
Preparation and amplification of full-length cDNA and sequencing libraries
Samples were processing using a modified version of the Smart-Seq2 protocol(75). Briefly, 1 µl of a 1:2,000,000 dilution of ERCC spike-ins (Invitrogen. Cat. no. 4456740) was added to each sample and mRNA was captured using modified oligo-dT biotinylated beads (Dynabeads, Invitrogen). PCR amplification was performed for 22 cycles. Amplified cDNA was cleaned with a 0.8:1 ratio of Ampure-XP beads (Beckman Coulter). cDNAs were quantified on Qubit using HS DNA reagents (Invitrogen) and selected samples were run on a Bioanalyzer HS DNA chip (Agilent) to evaluate size distribution.
For generating the sequencing libraries, individual cDNA samples were normalised to 0.2ng/µl and 1µl was used for one-quarter standard-sized Nextera XT (Illumina) tagmentation reactions, with 12 amplification cycles. Sample indexing was performed using index sets A and D (Illumina). At this point, individual samples were pooled according to their index set. Pooled libraries were cleaned using a 0.6:1 ratio of Ampure beads and quantified on Qubit using HS DNA reagents and with the KAPA Library Quantification Kits for Illumina (Roche). Samples were sequenced on two separate rapid-runs on HiSeq2500 (Illumina), generating 100bp paired-end reads. An additional 5 samples were sequenced on MiSeq (Illumina).
Full-length cDNA sequencing data processing and analysis
Paired-end read fastq files were demultiplexed, quality controlled using FastQC (https://www.bioinformatics.babraham.ac.uk/projects/fastqc/) and trimmed using Trim Galore (https://www.bioinformatics.babraham.ac.uk/projects/trim_galore/). Reads were pseudoaligned and quantified using kallisto (76) against a reference transcriptome from Ensembl Release 89 (Gencode Release M17 GRCm38.p6) with sequences corresponding to the ERCC spike-ins and the Cre recombinase and tdT genes added to the index. Transcripts were collapsed into genes using the sumAcrossFeatures function in scater.
Cell level quality control and cell filtering was performed in scater (77). Cells with <1000 genes, <100,000 reads, >75% reads mapping to ERCC spike-ins, >10% reads mapping to mitochondrial genes or low library complexity were discarded (14% samples). The population of olfactory bulb cells labelled in DAT-tdTomato mice is known to include a minor non-dopaminergic calretinin-positive subgroup (78), so calretinin-expressing cells were excluded from all analyses. The scTransform function in Seurat was used to remove technical batch effects.
Expression of CoV-relevent genes in scSeq datasets from various brain regions and sensory systems
An analysis of single-cell gene expression data from 10 studies was performed to investigate the expression of genes coding for coronavirus entry proteins in neurons from a range of brain regions and sensory systems. Processed gene expression data tables were obtained from scSeq studies that evaluated gene expression in retina (GSE81905) (79) inner ear sensory epithelium (GSE115934) (80, 81) and spiral ganglion (GSE114997) (82), ventral midbrain (GSE76381) (83), hippocampus (GSE100449) (84), cortex (GSE107632) (85), hypothalamus (GSE74672) (86), visceral motor neurons (GSE78845) (87), dorsal root ganglia (GSE59739) (88) and spinal cord dorsal horn (GSE103840) (89). Smart-Seq2 sequencing data from Vsx2-GFP positive cells was used from the retina dataset. A subset of the expression matrix that corresponds to day 0 (i.e. control, undisturbed neurons) was used from the layer VI somatosensory cortex dataset. A subset of the data containing neurons from untreated (control) mice was used from the hypothalamic neuron dataset. From the ventral midbrain dopaminergic neuron dataset, a subset comprising DAT-Cre/tdTomato positive neurons from P28 mice was used. A subset comprising Type I neurons from wild type mice was used from the spiral ganglion dataset. The “unclassified” neurons were excluded from the visceral motor neuron dataset. A subset containing neurons that were collected at room temperature was used from the dorsal root ganglia dataset. Expression data from dorsal horn neurons obtained from C57/bl6 wild type mice, vGat-cre-tdTomato and vGlut2-eGFP mouse lines was used from the spinal cord dataset. Inspection of all datasets for batch effects was performed using the scater package (version 1.10.1) (77). Publicly available raw count expression matrices were used for the retina, hippocampus, hypothalamus, midbrain, visceral motor neurons and spinal cord datasets, whereas the normalized expression data was used from the inner ear hair cell datasets. For datasets containing raw counts, normalization was performed for each dataset separately by computing pool-based size factors that are subsequently deconvolved to obtain cell-based size factors using the scran package (version 1.10.2) (90). Violin plots were generated in scater.
Supplementary Materials
Materials and Methods
Fig. S1. Related to Fig. 2. Analysis of mouse bulk RNA-Seq and WOM scSeq dataset.
Fig. S2. Related to Fig. 2. Analysis of mouse HBC lineage scSeq datasets.
Fig. S3. Related to Fig. 3. Bulk RNA-seq datasets for primates and additional figures of human scSeq data.
Fig. S4. Related to Fig. 3. Comparing cell types and expression levels across respiratory and olfactory epithelial datasets.
Fig. S5. Related to Fig. 4. Whole OB scSeq data integration and clustering.
Fig. S6. Related to Fig. 4. Expression of cell type markers in whole OB scSeq data.
Fig. S7. Related to Fig. 4. Sample quality control of manually sorted OB dopaminergic neurons.
Fig. S8. Related to Fig. 4. Expression of genes coding for CoV-2 entry-related proteins in neurons forming part of deeply sequenced scSeq datasets from various brain regions and sensory systems.
Funding
SRD is supported by grants RO11DC016222 and U19 NS112953 from the National Institutes of Health and by the Simons Collaboration on the Global Brain. JN was supported by NIH grant RO1DC007235. DR was supported by Programma per Giovani Ricercatori Rita Levi Montalcini granted by the Italian Ministry of Education, University, and Research. KVdB is a postdoctoral fellow of the Belgian American Educational Foundation (BAEF), and is supported by the Research Foundation Flanders (FWO), grant 1246220N. ML and MG were supported by a Leverhulme Trust Research Grant (RPG-2016-095) and a Consolidator Grant from the European Research Council (725729; FUNCOPLAN). JM is supported by a Medical Research Council grant (K013807) and a Medical Research Council Clinical Infrastructure Award (M008924). ICM is supported by a BBSRC New Investigator Grant (BB/P022073/1) and the BBSRC National Capability in Genomics and Single Cell Analysis at Earlham Institute (BB/CCG1720/1). BJG is supported by grant R01DC016859.
Author contributions
DHB, TT, RC, HC, RF, YC, ML, ICM, RAH and BJG designed and performed experiments. DHB, TT, CW, ML, KvdB, BG, RC, HC, DD, KS, and HRdB performed analysis. DR, SD, EP, JSM, BJG, MSG JN, SRD, designed and supervised experiments. DHB, TT, CW, ML, HM, DWL, BJG, MSG, JN and SRD wrote the manuscript
Competing interests
DL is an employee of Mars, Inc. None of the other authors have financial interests related to this manuscript.
Data and materials availability
Reanalyzed datasets are obtained from the URLs listed in supplementary materials. All data is currently being deposited and will be made publicly accessible from the NCBI GEO at accession GSExxxx.
Supplementary Materials
URLs for re-analyzed datasets
https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4680959/bin/srep18178-s2.xls
https://doi.org/10.1371/journal.pone.0113170.s014
https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE52464
https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE139522
https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE99251
https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE120199
Acknowledgments
We thank members of the Datta lab, James Schwob, Bernardo Sabatini, Andreas Schaefer, Kevin Franks, Michael Greenberg and Vanessa Ruta for helpful comments on the manuscript. We thank James Lipscombe and Andres Crespo for technical support.
Footnotes
Includes previously unpublished datasets (whole olfactory mucosa scSeq, HBC scSeq, olfactory bulb scSeq, olfactory bulb dopaminergic neuron deep sequencing), ACE2 immunohistochemistry in human tissue, and analysis of SARS-CoV-2 cell entry genes in the olfactory bulb and brain; research groups of John Ngai, Matthew Grubb, Bradley Goldstein, Hiro Matsunami (and colleagues) added as collaborators.
https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4680959/bin/srep18178-s2.xls
https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE139522
https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE120199
References
- 1.↵
- 2.
- 3.↵
- 4.↵
- 5.↵
- 6.↵
- 7.↵
- 8.↵
- 9.
- 10.↵
- 11.↵
- 12.↵
- 13.↵
- 14.↵
- 15.↵
- 16.↵
- 17.
- 18.↵
- 19.↵
- 20.↵
- 21.↵
- 22.
- 23.↵
- 24.↵
- 25.↵
- 26.
- 27.↵
- 28.↵
- 29.
- 30.↵
- 31.
- 32.↵
- 33.↵
- 34.
- 35.↵
- 36.↵
- 37.↵
- 38.↵
- 39.↵
- 40.↵
- 41.↵
- 42.↵
- 43.↵
- 44.↵
- 45.↵
- 46.↵
- 47.↵
- 48.↵
- 49.↵
- 50.↵
- 51.↵
- 52.↵
- 53.↵
- 54.
- 55.↵
- 56.↵
- 57.↵
- 58.↵
- 59.↵
- 60.↵
- 61.
- 62.
- 63.↵
- 64.↵
- 65.↵
- 66.↵
- 67.↵
- 68.↵
- 69.↵
- 70.↵
- 71.↵
- 72.↵
- 73.↵
- 74.↵
- 75.↵
- 76.↵
- 77.↵
- 78.↵
- 79.↵
- 80.↵
- 81.↵
- 82.↵
- 83.↵
- 84.↵
- 85.↵
- 86.↵
- 87.↵
- 88.↵
- 89.↵
- 90.↵