Single lumbar L4 and L5 human dorsal root ganglia (DRG) were rapidly recovered from transplant donors within 90 min of cross-clamp and were immediately stored in RNAlater. Nuclei from individual ganglia were isolated and samples were enriched for neuronal nuclei by selection using an antibody to NeuN (see Figure 1—figure supplement 1). Five ganglia from one male and four female donors with ages ranging from 34 to 55 were subjected to droplet based single nucleus (sn) capture, barcoding, and reverse transcription (×10 Genomics). Combinatorial clustering methods (Stuart et al., 2019) allowed co-clustering of neuronal nuclei into a well-defined set of distinct transcriptomic groups that are well separated from their non-neuronal counterparts (Figure 1A, Figure 1—figure supplement 2). After removal of non-neuronal nuclei from the dataset, reclustering the DRG-neuron data from 1837 cells identified a range of about a dozen diverse transcriptomic classes of human somatosensory neurons (Figure 1B, Figure 1—figure supplements 2 and 3).

One of the best studied groups of somatosensory receptors in mice are nociceptive peptidergic neurons that coexpress a variety of neuropeptides including substance P, calcitonin gene-related peptide (CGRP), and pituitary adenylate-cyclase-activating polypeptide (PACAP). These neurons are typically small soma diameter, nonmyelinated, slow conducting c-fibers, but also include faster conducting lightly myelinated Aδ neurons (Sharma et al., 2020; Nguyen et al., 2019). In the human DRG dataset, TAC1 (substance P), CALCA and CALCB (CGRP), and ADCYAP1 (PACAP), are expressed in several transcriptomic classes (H1, H2, H3, H5, and H6, Figure 1C, Figure 1—figure supplements 4 and 5). For comparison the expression of the same genes in mouse DRG neurons is shown (Figure 1C, Figure 1—figure supplements 4 and 5) using data from single nuclei sequencing (Renthal et al., 2020). Just as in mice, the putative human peptidergic nociceptors express the high affinity nerve growth factor receptor NTRK1, the capsaicin and mustard oil-gated ion channels TRPV1 and TRPA1 but generally only low levels of the stretch-gated ion channel PIEZO2 (Figure 1C, Figure 1—figure supplements 4 and 5).

Although previous localization studies have suggested that in humans the neurofilament protein NEFH is expressed in all sensory neurons (Rostock et al., 2018), this gene showed graded expression in our data (Figure 1C) and marks several classes of cells just as in mice (Figure 1—figure supplement 5). Some of these (including H3 and H6) also express peptidergic markers and the pain-related voltage-gated sodium channel SCN10A (Figure 1—figure supplement 5) and thus have molecular hallmarks of Aδ nociceptors (von Buchholtz et al., 2020). However, the neuronal classes H14 and H15 expressing the highest levels of NEFH are distinct from the peptidergic neurons (Figure 1C, Figure 1—figure supplements 4 and 5), likely representing different types of large diameter, fast conducting myelinated Aβ neurons. These cell types are neurotrophin three receptor NTRK3 positive, some also contain the brain derived neurotrophic factor receptor NTRK2 but exhibit little expression of NTRK1 (Figure 1—figure supplements 4 and 5). In mice, proprioceptors are a subtype of Aβ neurons marked by the calcium binding protein parvalbumin, the transcription factor Etv1 and the voltage-gated sodium channel subunits Scn1a and Scn1b (Sharma et al., 2020; Renthal et al., 2020). In the human data, the small H15 group of NTRK3-positive cells had this expression pattern (Figure 1C, Figure 1—figure supplement 5) implying that proprioceptors have conserved transcriptomic markers in humans and mice. Similarly, small groups of both Aδ-low threshold mechanosensors (H13) and cool responsive neurons (H8) were identified by their characteristic expression profiles of functionally important transcripts (Figure 1—figure supplement 5). Thus, large groups of human and mouse DRG neurons appear to share basic transcriptomic signatures and functional potential, supporting our data as informative about neuronal diversity among human somatosensory neurons.

Despite these broad similarities between the putative peptidergic, proprioceptive, cooling sensitive, Aβ and Aδ classes of DRG neurons in mice and humans there were differences in expression of many genes. These include molecules that modulate cellular responses to internal signals (e.g., growth factor receptors), sensory stimuli and also the mediators they may release. For example, in humans, the H8 putative cool responsive neurons expressing TRPM8 were strongly positive for the BDNF-receptor NTRK2 but hardly expressed the neuropeptide TAC1 whereas in rodents the converse was true (Figure 1—figure supplement 5). Other genes that have been shown to control sensory responses in mice exhibit a different expression pattern in human DRG neurons. For instance, Tmem100 encodes a protein that in mice has been implicated as playing an important role in functional interactions between Trpv1 and Trpa1 and contributing to persistent pain (Weng et al., 2015). By contrast it was almost undetectable in the human sequencing data (Figure 2A). Similarly, we did not detect marked expression of the sphingosine-1-phosphate receptor S1PR3 (Figure 2A) that has been suggested as a target for treating both pain and itch based on mouse work (Hill et al., 2018). More strikingly, a small group of human neurons, H5, expressing TRPA1 were resolved in our clustering (Figure 2B), whereas in mouse nuclear sequencing data no direct counterpart was detected (Figure 1—figure supplement 4). Whole cell-based sequencing (Sharma et al., 2020) of mouse DRG neurons does identify a group of peptidergic nociceptors (called CGRP-gamma) with abundant Trpa1 expression, highlighting the caution needed when interpreting differences across species and the value of alternative approaches. Nonetheless, whereas mouse CGRP-gamma neurons strongly express Calca and Ntrk1, H5 cells are essentially CALCA (CGRP) and NTRK1 negative and instead are strongly NTRK2 positive (Figure 1, Figure 1—figure supplements 4 and 5) suggesting that they may respond differently to external stimuli as well as in their signaling properties and therefore may not be a direct equivalent of mouse CGRP-gamma neurons. Thus, the availability of human transcriptomic data should help focus translational work in model organisms on promising targets with conserved expression patterns in humans.

Analysis of the gene expression patterns of the different classes of human somatosensory neurons revealed several groups for which we could not discern direct counterparts in the mouse. One small but prominent group of human DRG neurons (H9) expresses TRPM8, PIEZO2, SCN10A, and SCN11A (Figure 1—figure supplements 4 and 5, Figure 2B) and clearly segregates from the putative cool sensing cells (H8) that express TRPM8, GPR26, NTM, and FOXP2 but are devoid of both the light touch receptor and the pain-related sodium channels (Figure 2B, Figure 1—figure supplement 5). In mice, Trpm8 expression is simpler with the cool sensing, menthol responsive ion channel just expressed in cells with this latter gene expression pattern (Figure 1—figure supplement 5). Interestingly, single fiber recordings have identified human neurons that respond to both cooling and gentle touch as might be expected for cells expressing both TRPM8 and PIEZO2 (33). Moreover, recent single-cell sequencing of macaque DRG neurons also identified two populations of cells that match H8 and H9 neurons (Kupari et al., 2021). Finally, H9 neurons resemble (but also have differences from) human mechanosensory neurons that were recently engineered by transcriptional programming of stem cells (Nickolls et al., 2020).

A second larger group of human neurons H12 is marked by NTRK3 and the voltage-gated ion channel SCN1A, but is only weakly positive for NEFH, expresses moderate levels of PIEZO2 (Figure 2B, Figure 1—figure supplement 5) and appears distinct from any potential mouse counterpart. The H12 gene expression pattern is most consistent with these cells functioning as a type of mechanosensor that has no direct equivalent in mice. Similarly, we designated H4 as c-nociceptors because of their expression of nociception-related SCN10A and NTRK1 and low level of NEFH (Figure 2B, Figure 1—figure supplement 5). These neurons expressed low levels of neuropeptides, but their overall gene expression patterns did not resemble any mouse counterparts including the nonpeptidergic nociceptors (see below).

The two remaining large groups of neurons in the human dataset H10 and H11 that have no clear mouse counterpart exhibit most similarity with mouse c-type nonpeptidergic neurons (Figure 2—figure supplement 2A). At a functional level both H10 and H11 express receptors that in mice have roles in detecting pruritogens. For example, these clusters were positive for the two subunits (IL31RA and OSMR) of the interleukin 31 receptor and the histamine receptor HRH1 (Figure 2C) that mediate mast cell-related scratching in mice (Solinski et al., 2019). They also express the itch-related neuropeptide NPPB (Figure 2C), nociception-related sodium channels SCN10A and SCN11A as well as TRPV1 (Figure 2—figure supplement 2A) but not appreciable NEFH or TAC1 (Figure 1C). Therefore, it is likely that these are groups of putative unmyelinated, nonpeptidergic nociceptors with potential for triggering human itch responses. H10 and H11 characterization is revisited later in its own section of the results.

The peptidergic nociceptors, myelinated Aβ and Aδ neurons, rarer human-specific cells, and the two nonpeptidergic nociceptor clusters H10 and H11 account for all the neurons in our analysis with H10 and H11 totaling approx. 20% of the neurons. In marked contrast, mouse nonpeptidergic, small diameter neurons are far more numerous than H10 and H11 accounting for 40% of the sensory neurons in mouse DRGs (Renthal et al., 2020) and divide into four highly stereotyped transcriptional groups (Figure 1—figure supplement 4). Two of these classes of mouse neurons (NP2 and NP3) trigger itch (Mishra and Hoon, 2013; Han et al., 2013), one (NP1, expressing Mrgprd) responds to noxious mechanical stimulation (von Buchholtz et al., 2021b). NP1 neurons may have a role in mechanonociception (Gatto et al., 2019) and have recently been associated with suppression of skin inflammation (Zhang et al., 2021), which was hypothesized as relevant for human health. The fourth class corresponds with low threshold mechanosensors (cLTMRs) that are thought to mediate affective touch (Gatto et al., 2019; McGlone et al., 2014). Given this difference between the transcriptomic map of human DRG neurons and their rodent counterparts, we next used independent ISH-based analysis to test basic predictions of the sequencing. If transcriptomic characterization of human DRG neurons is accurate then one clear expectation is that TAC1, NEFH, and OSMR should be expressed by distinct and only partially overlapping populations of human DRG neurons. If it is also comprehensive, that is, not missing equivalent classes to mouse neurons NP1, NP2, and cLTMRs that constitute almost a third of mouse DRG neurons and do not significantly express Nefh, Tac1, or Osmr, then we would anticipate that the same three markers should label the vast majority of neurons. Multigene ISH demonstrates that both these predictions are true for human DRG neurons (Figure 2D, Figure 2—figure supplements 1B and 2B) with essentially every cell labeled by one of these probes but with very few exhibiting strong coexpression. Although NEFH expression could be detected in some of the cells positive for the other markers (Figure 2D, Figure 2—figure supplement 1B), many TAC1- or OSMR-positive small diameter neurons were negative for this neurofilament subunit. Moreover, TAC1 and OSMR labeled almost completely separate sets of cells (Figure 2—figure supplement 1). Notably, in keeping with our assignments based on transcriptomic data, the largest diameter neurons were strongly positive for NEFH whereas TAC1 and OSMR primarily labeled smaller cells (Figure 2D). Finally, these three markers each labeled a large group of neurons (Figure 2—figure supplement 1).

As detailed above, the expression of genes that are thought to be important for functional and morphological features of somatosensory neurons reveal similarities between groups of human and mouse neurons. They also expose differences that likely reflect distinct somatosensory adaptations in the two species. We next used co-clustering methods to test whether the wider transcriptome could reveal additional information about the relationships between classes of human and rodent DRG neurons using the same mouse dataset (Renthal et al., 2020) that we analyzed above (Figure 1, Figure 1—figure supplement 4). We used the well-established approach developed by the Satija lab (Stuart et al., 2019) as it has been shown to perform well without forcing false matches. As predicted, several classes of human neurons grouped with corresponding mouse counterparts including H15 – proprioceptors, H14 – Aβ cells, H13 – AδLTMRs, H11 – NP3 (Nppb) neurons, and H3/H6 – Aδ nociceptors (Figure 3A). This analysis suggested that H10 the other cluster that gene expression indicated are also itch related most closely resembled NP1 (Mrgprd) neurons rather than any other human or mouse class of sensory neurons. The H12 cluster, which is human specific, grouped close to larger diameter mouse neurons, whereas other clusters of human cells appeared better aligned with smaller diameter nociceptors. However, all types of peptidergic small diameter nociceptors were less organized in the co-clustering and separated from their potential mouse counterparts despite their qualitatively similar expression of some of the best-known functional markers (Figure 1—figure supplement 4).

UMAP plots (Figure 3A) provide a visual representation of similarity between cells with related transcriptomic properties. However, co-clustering methods do not provide a quantitative measure of the similarity between individual cells or clusters of cells. A given cell cluster can be viewed as a multivariate probability distribution in gene expression space. While not commonly employed in gene expression analysis, in probability and information theory, the similarity/dissimilarity between two probability distributions is most commonly measured by calculating their Kullback–Leibler (KL) divergence. Recent advances have allowed KL divergence to be estimated between two samples in continuous multivariate space (Perez-Cruz, 2008) as is observed in dimensionality reduced gene expression data of cell populations. Therefore, we made use of KL divergence estimation to quantitate the similarity between human DRG-neuron clusters and all their potential mouse counterparts (Figure 3B). As expected, clusters that co-segregate in the UMAP analysis showed greatest similarity but additional relationships not apparent from the visual representation of the co-clustering were also seen. For example, the small cluster of human ‘cool’ responsive neurons H8 showed greatest similarity to mouse Trpm8 cells and several groups of human cells (H1, H2, and H5) that gene expression predicted should be c-type peptidergic nociceptors, indeed best matched these cells (Figure 3B). Interestingly, no class of human neurons showed appreciable similarity to mouse cLTMRs. Among the groups of cells that had human-specific gene expression patterns, H5 best matched c-peptidergic nociceptors, H9 (the putative cool and mechanical responsive cells) showed only weak similarity to any mouse neuron class. H12, which we considered likely to be mechanosensors best matched mouse proprioceptors and H4 neurons appeared distantly related to several classes of nociceptor but without a clear match in mice. We also extended this analysis to mouse clustering where similar cell populations had not been combined (Figure 3—figure supplement 1). Although 19 clusters of mouse neurons were now analyzed, neither mouse cLTMRs nor the potentially human-specific classes H4, H9, or H12 better matched a cell types of the other species. One important caveat to this type of analysis remains that any functional conclusions based on shared transcriptomic features still need to be verified experimentally.

From sequence analysis, we identified a range of potential markers to better explore the diversity of human DRG neurons using ISH. To maximize information, we chose a multiplexed approach that allows localization of up to 12 probes (Figure 4) revealing the different classes of sensory neurons identified in the transcriptomic data. For example, TRPM8 expressing neurons clearly segregate into two distinct types (Figure 4A, Figure 4—figure supplement 1). One set of cells (H8) share other transcriptomic properties with mouse cooling responsive cells: TRPM8, the cool and menthol receptor is not coexpressed with the ion channels SCN10A or PIEZO2 (Figure 4A), but unlike in mice these cells are NTRK2 positive (Figure 4—figure supplement 1). By contrast, other cells (H9) coexpress the pain and light touch related ion channels (SCN10A and PIEZO2) with TRPM8 (Figure 4A, Figure 4—figure supplement 1). Similarly, putative proprioceptive neurons (H15) were distinguished by their expression of NEFH, PIEZO2, and PVALB and lack of NTRK2 (Figure 4B, Figure 4—figure supplement 1). One surprise (Figure 4A, B) was that in small fields of view, several examples of all three of these rare neuron types could be identified in human DRGs. However, much of the rest of the ganglion was devoid of these cell types and instead the neurons there had distinct sets of markers. Therefore, it appears that transcriptomic classes of human DRG sensory neurons may not be stochastically distributed in the ganglion. Indeed, when we examined the distribution of nociceptors and myelinated neurons at lower magnification (using strong selective probes), broad clustering of similar types of neurons was apparent, quantifiable, and statistically significant (Figure 4C, D). We carried out a similar analysis in mouse DRG neurons using Nefh and Scn10a probes and found that there too, cells expressing these markers are not uniformly distributed (Figure 4—figure supplement 2) suggesting spatial clustering of related DRG neurons may be a common feature across species.

Perhaps the most intriguing classes of human somatosensory neurons revealed by our transcriptomic approach are the H10 and H11 classes that primarily share features with the mouse nonpeptidergic nociceptors NP1–3 (Figures 2 and 3, Figure 2—figure supplement 2, Figure 5—figure supplement 1). ISH showed that the H10 and H11 classes of neurons, identified by their expression of OSMR were small diameter neurons comparable in size to the TAC1-expressing peptidergic nociceptors (Figure 5A). Two qualitatively different ratios of SCN10A and OSMR were apparent in these cells (Figure 5A) hinting at their distinct identities. Our data (Figure 2, Figure 2—figure supplement 2, Figure 5—figure supplement 1A) show that H10 and H11 neurons express a number of genes that are known markers of mouse NP3 cells and functionally important for triggering pruritic responses (Gatto et al., 2019; Solinski et al., 2019). They are also distinguished from each other by expression of genes that likely play roles in itch and other aspects of somatosensation (Figure 5B, Figure 2—figure supplement 2, Figure 5—figure supplement 2). For example, although not prominently expressed, the human chloroquine responsive receptor MRGPRX1 (27) localized selectively to H10 neurons (Figure 5B) perhaps suggesting a relationship to mouse NP2 cells. By contrast, Janus kinase 1 (JAK1), a mediator of itch through various types of cytokine signaling, including through OSMR (Oetjen et al., 2017), and the neuropeptide SST are particularly strongly expressed in H11 cells (Figure 5B). Both these genes are prominent markers of NP3 pruriceptors in mouse (Figure 5B). However, not all known itch-related transcripts are expressed in H10 and H11 neurons and both classes of cells express genes that better define NP1 neurons in mice as well as other cell types (Figure 5B, Figure 2—figure supplement 2, Figure 5—figure supplement 2).

H10 cells are also distinguished from H11 and mouse pruriceptors by their prominent expression of the stretch-gated ion channel PIEZO2 (Figure 1C, Figure 2—figure supplement 2). The coexpression of itch-related transcripts and this low threshold mechanosensor hint that H10 neurons may be responsible for the familiar human sensation known as mechanical itch. However, their relationship to NP1-neurons revealed by co-clustering mouse and human data (Figure 3) and their expression of markers for various other cell types (Figure 2—figure supplement 2, Figure 5—figure supplement 2) possibly including nonpeptidergic cLTMRs suggest that their role in somatosensation may not be limited to itch alone.

A problem with single-cell sequencing approaches is the sparse nature of the data making it difficult to disentangle expression level from proportional representation in any cluster. This means that except for the most highly expressed genes, there is inherent ambiguity in interpreting the expression patterns. ISH provides an independent and more analog assessment of expression level that can help resolve this issue. Multiplexed ISH showed that SST divides the OSMR-positive cells into two intermingled types (Figure 5C) in keeping with the sequence data (Figure 5B) and the relative expression patterns of SCN10A and OSMR (Figure 5A). Moreover, the prediction that PIEZO2-OSMR coexpression should mark SST-negative neurons was also largely borne out by ISH (Figure 5C, Figure 5—figure supplement 1). However, ISH also shows that some neurons expressing lower levels of SST are PIEZO2-positive and that some OSMR-positive H10 cells, contain only a very low level of the mechanosensory channel (Figure 5C, Figure 5—figure supplement 1). Therefore, H10 and H11 are by no means homogeneous populations and may not be as functionally distinct as snRNA clustering suggests.

Transcriptomic analysis of human DRG neurons was carried out to establish similarities and differences between human somatosensory neurons and their counterparts in model organisms and to provide a resource. We chose a nuclear-based strategy because of its simplicity and quantitative nature relative to isolation of cells (Nguyen et al., 2019). All tissue was obtained from de-identified organ donors and was not preselected or otherwise restricted according to health conditions. We used DRGs from both male and female donors for the sequencing and ISH localization experiments. Randomization and blinding were not used because of the nature of our experiments. Similarly, before starting this study, we had no relevant information for setting sample size for snRNA sequencing from human DRG neurons. Therefore, we stopped data collection when we empirically determined that the cost of adding extra data outweighed the benefit of additional sequencing. In essence, numbers of sn-transcriptomes analyzed were limited by the availability of material and the difficulty of isolating human DRG nuclei with preservation of their transcriptome. We considered that the dataset would serve as a valuable and relatively comprehensive resource once including additional material from an individual preparation made only minor differences to the pattern of clustering we observed. Criteria for data exclusion followed standards in the field (see below) and sample sizes and numbers of replicates are also typical for this type of study and are described in the relevant experimental sections. Apart from the exclusions described for sn experiments, all data obtained were included in our study.

DRG recovery was reviewed by the University of Cincinnati IRB #00003152; Study ID: 2015-5302, title Human dorsal root ganglia and was exempted. Lumbar L4 and L5 DRGs were recovered from donors withing 90 min of cross-clamp (Valtcheva et al., 2016). For ISH and IHC, DRGs were immersion fixed in 4% paraformaldehyde in phosphate-buffered saline (PBS) overnight, cryoprotected in 30% sucrose and were frozen in Optimal Cutting Temperature compound (Tissue Tek). For RNA sequencing, human DRGs immediately were cut into 1–2 mm pieces and stored in RNA-later (Thermo Fisher, Cat# AM7021). Excess RNA-later was removed and the tissue was frozen on dry ice and stored at −80°C. Nuclei were isolated from each donor separately as described previously (Nguyen et al., 2019) with minor modification. Briefly, the tissues were homogenized with a Spectrum Bessman tissue pulverizer (Fisher Scientific, CAT# 08-418-3) in liquid nitrogen. The sample was then transferred to a Dounce homogenizer (Fisher Scientific, Cat# 357538) in 1 ml of freshly prepared ice-cold homogenization buffer (250 mM sucrose, 25 mM KCl, 5 mM MgCl₂, 10 mM Tris, pH 8.0, 1 µM DTT, 0.1% Triton X-100 [vol/vol]). To lyse cells and preserve nuclei homogenization used 5 strokes with the ‘loose’ pestle (A) and 15 strokes with the ‘tight’ pestle (B). The homogenate was filtered through a 40 µm cell strainer (Thermo Fisher, cat# 08-771-1), was transferred to low bind microfuge tubes (Sorenson BioScience, cat# 11700) and centrifuged at 800 g for 8 min at 4°C. The supernatant was removed, the pellet gently resuspended in 1 ml of PBS with 1% bovine serum Albumin (BSA) and SUPERaseIn RNase Inhibitor (0.2 U/µl; Thermo Fisher, Cat#AM2696) and incubated on ice for 10 min.

Neuronal nuclei selection was performed by incubating the sample with a rabbit polyclonal anti-NeuN antibody (Millipore, cat#ABN78) at 1:4000 dilution with rotation at 4°C for 30 min. The sample was then washed with 1 ml of PBS with 1% BSA and SUPERaseIn RNase Inhibitor and centrifuged at 800 × g for 8 min at 4°C. The resulting pellet was resuspended in 80 µl of PBS, 0.5% BSA, and 2 mM EDTA. 20 µl of anti-rabbit IgG microbeads (Miltenyi biotec, cat# 130-048-602) were added to the sample followed by a 20-min incubation at 4°C. Nuclei with attached microbeads were isolated using an LS column (Miltenyi Biotec, cat# 130-042-401) according to the manufacturer’s instruction. The neuronal nuclei enriched eluate was centrifuged at 500 × g for 10 min, 4°C. The supernatant was discarded, and the pellet was resuspended in 1.5 ml of PBS with 1% BSA. To disrupt any clumped nuclei, the sample was homogenized on ice with an Ultra-Turrax homogenizer (setting 1) for 30 s. An aliquot was then stained with trypan blue and the nuclei were counted using a hemocytometer. The nuclei were pelleted at 800 × g, 8 min at 4°C and resuspended in an appropriate volume for ×10 chromium capture. A second count was performed to confirm nuclei concentration and for visual inspection of nuclei quality. Note that prior to sequencing no check was made that this selection procedure was unbiased but we have previously used the approach to purify mouse trigeminal ganglion neuron nuclei (Nguyen et al., 2019) and have demonstrated highly quantitative recovery of all neuronal types (von Buchholtz et al., 2020). IHC analysis (Figure 1—figure supplement 1) supports the use of NeuN enrichment as a means to purify human DRG neurons and together with ISH quantitation (Figure 2—figure supplement 1) substantiates the relatively unbiased purification of various classes of neuronal nuclei using this method.

10x-chromium capture and library generation were performed according to the manufacturer’s instructions using v3 chemistry kits. Next generation sequencing was performed using Illumina sequencers. 10x chromium data were mapped using CellRanger to a pre-mRNA modified human genome (GRCh38.v25.premRNA). Data analysis used the Seurat V3 packages developed by the Satija lab and followed standard procedures for co-clustering (Stuart et al., 2019). For sn-RNA sequencing experiments cell filtering was performed as follows: outliers were identified and removed based on the number of expressed genes (500–10,000 retained) and mitochondrial proportion (<10% retained). Normalization and variance stabilization used regularized negative binomial regression (sctransform). After initial co-clustering of data from the different preparations, non-neuronal cell clusters were identified by their gene expression profiles. Clusters not expressing high levels of neuronal or somatosensory genes like SNAP25, SCN9A, SCN10A, PIEZO2, NEFH, etc. but instead expressing elevated levels of markers of non-neuronal cells including PRP1, MBP, QKI, LPAR1, and APOE were tagged as non-neuronal and were removed to allow reclustering of ‘purified’ human DRG neurons. A total of 1837 human DRG neuronal nuclei were included in the analysis (Table 1). The mean number of genes detected per nucleus was 2839 (range 501–9652), with a standard deviation of 1917. Doublet detection was performed on the individual datasets using DoubletFinder (McGinnis et al., 2019). For the clustering shown in the main figures the small number of potential doublets (Figure 1—figure supplement 2E) were not removed; principal components (PCs) were determined from integrated assay data and PCs 1–16 were used both for UMAP display of the data and for determining clusters. The resolution for clustering used relatively low stringency (2.0) and closely related clusters without distinguishing markers were merged. Changes in display clustering parameters and in the cutoffs for data inclusion/exclusion as well as leaving out nuclei from any single preparation made differences in how the data were represented graphically and the number of clusters identified but not to the main conclusions (Figure 1—figure supplement 3). All the different transcriptomically related neuron types described here could still be readily discerned in UMAP analysis of expression data. Identification of markers used the FindMarkers and FindAllMarkers functions of Seurat using default settings and with markers limited by a minimal level of expression in the positive cluster set using (min.pct = 0–0.3) and the difference in expression between positive and negative clusters set using (min.dif.pct = 0–0.3).

For analysis of the mouse, a random subset of data from sn-RNA sequencing of DRGs from wild type mice were extracted from data deposited by the Woolf lab (Renthal et al., 2020) using the R-function: sample. The data were filtered according to gene count (400–12,000 retained) and mitochondrial DNA (<1% retained) leaving 6895 DRG nuclei that were clustered using standard methods (Stuart et al., 2019). For the data that are displayed in most figures, PCs 1–20 were used for UMAP display with resolution for clustering set at 3.5; closely related clusters were merged. For Figure 3—figure supplement 1, PCs 1–30 were used for UMAP display with resolution for clustering set at 1.6; related clusters were not merged. The expression patterns that are described for genes in mice can also be checked in the outstanding and easy to search single-cell analysis provided by Sharma et al., 2020.

Co-clustering of mouse and human data used methods described by the Satija lab (Stuart et al., 2019). Briefly, we capitalized gene names in both mouse and human DRG-neuron data and then limited the datasets to genes that were shared between them. Mouse and human data were individually normalized and subjected to variance stabilization (sctransform) and were then combined using default parameters for integration (dims = 15). After integration, data were scaled and 30 PCs were calculated, used for UMAP display and clustering. Clustering resolution was set at 0.5 and cells were color coded by transferring their identity in the original clustering of human or mouse data. We experimented using different numbers of mouse neurons and found broadly similar results when approx. 1800, 3500, or 7000 mouse nuclei were used. However, although the relationships shown in Figure 3 could be discerned using the lower numbers of mouse nuclei the mouse neurons were less organized (with some of the mouse clusters splitting). When we used substantially greater numbers of mouse neurons from the full Renthal dataset (Renthal et al., 2020), the mouse neurons dominated the clustering and only co-clustering of large diameter neurons across species was observed with the other human and mouse nuclei clustering separately from each other. In order to quantify the similarity/dissimilarity between a given mouse cluster and any human cluster, the KL divergence between their distributions in 30-dimensional continuous space (PCs 1–30 from the co-clustering) was estimated as described previously (Perez-Cruz, 2008) using R (https://gist.github.com/lars-von-buchholtz/636f542ce8d93d5a14ae52a6c538ced5; von Buchholtz, 2021a). The natural logarithm of the KL divergences for each mouse/human pairing was plotted as a heatmap in R.

Cryosections from human DRGs (from different donors than those used for sequencing) were cut at 20 µm and used for ISH with the RNAscope HiPlex Assay (Advanced Cell Diagnostics) following the manufacturer’s instructions. The following probes were used: NEFH (cat# 448141); TRPM8 (cat# 543121); PIEZO2 (cat# 449951); SCN10A (cat# 406291); NTRK2 (cat# 402621); TAC1 (cat# 310711); OSMR (cat# 537121); SST (cat# 310591); TRPV1 (cat# 415381); and PVALB (cat# 422181). We also used RNAscope Multiplex Fluorescent Assay (Advanced Cell Diagnostics) for localization of OSMR (cat# 537121); NPPB (cat# 448511); HRH1 (cat# 416501); and SST (cat# 310591).

Confocal microscopy (5 µm spaced optical sections for HiPlex and 1 µm spaced optical sections for Multiplex Assays) was performed with a Nikon C2 Eclipse Ti (Nikon) using a ×40 objective. All confocal images shown are collapsed (maximum projection) stacks. HiPlex images were aligned and adjusted for brightness and contrast in ImageJ as previously described (von Buchholtz et al., 2020). Diagnostic probe combinations were used on at least three sections from at least two different individuals with qualitatively similar results. Overall, we used sections from ganglia from five different donors, however, the signal intensity of all probes varied between the individual ganglia making identification of positive signal risky in some cases. Strongest signals were observed for sections from an 18-year-old male donor and 23- and 35-year-old female donors. All images displayed here, and our analysis including cell counts were only from sections of ganglia isolated from these three individuals. When individual channels are displayed in the supplements, the strongest autofluorescence signals have been selected and superimposed using photoshop to help focus attention on ISH signal.

IHC was carried out using 20 µm sections of human DRGs using standard procedures. Primary antibodies to NeuN (Millipore) and β3 tubulin (Proteintech) were used at 1:1000 and 1:500 dilution, respectively, and were visualized using appropriate secondary antibodies conjugated to fluorescent reporters (Jackson Immuno Research, 1:1000). Sections were counterstained with 4′,6-diamidino-2-phenylindole (DAPI) and/or green NeuroTrace (Fisher Scientific). Confocal images were acquired with a Nikon C2 Eclipse Ti (Nikon) using a ×40 objective with optical sections 5 µm apart. Images are maximum projection (collapsed) stacks of individual optical sections; consistency of NeuN staining was assessed using sections from two individuals; images were processed using Adobe Photoshop CC to adjust brightness, contrast, and set channel color for display.

Animal experiments were carried out in strict accordance with the US National Institutes of Health (NIH) guidelines for the care and use of laboratory animals and were approved by the National Institute of Dental and Craniofacial Research animal care and use committee (protocol #20-1041). Male and female mice were used for all experiments but were not analyzed separately. Mice were C57BL/6NCrl and were 6 weeks or older. Cryosections from fresh frozen mouse lumbar DRGs in OCT (Tissue-Tek) were cut at 10 µm and used for ISH with RNAscope Multiplex Fluorescent Assay (Advanced Cell Diagnostics) according to the manufacturer’s instructions. Confocal images were acquired with a Nikon C2 Eclipse Ti (Nikon) using a ×10 objective at 1 µm optical section. Images are maximum projection (collapsed) stacks of 10 individual optical sections; consistency of staining was assessed using multiple sections from at least three mice as is considered standard; images were processed using Adobe Photoshop CC to adjust brightness, contrast, and set channel color for display.

In order to quantify spatial clustering of cell types, neurons in ISH images (for humans, one male, one female and for mice several sections from different DRGs and animals) were manually outlined and annotated as NEFH-only or SCN10A-only; cells expressing both genes were not counted. Centroid coordinates of these cells and their distances were analyzed in Python 3.7. The nearest neighbors were identified based on Euclidean distance (Scikit-Learn package) and the percentage of NEFH and SCN10A cells in each neighborhood of size 1–40 cells was calculated. Statistical significance between NEFH- and SCN10A-surrounding neighborhoods was determined using a one-tailed Mann–Whitney U-test (Scipy Stats package).