Establishment of a treatment-naïve canine osteosarcoma reference database
To establish a treatment-naïve canine osteosarcoma reference, we completed single-cell RNA sequencing (scRNA-seq) on 6 dogs and collected data on a total of 35,310 cells. The average number of cells collected per tumor was 5,885 and on average each cell was sequenced to a depth of 72,649 reads per cell. All tumors were confirmed to be osteosarcoma by histology and histological subtyping was completed on each tumor. In total, the curated dataset consisted of 1 fibroblastic, 1 chondroblast, and 4 osteoblastic tumors, with one dog exhibiting radiographical evidence of lung metastasis (Table 1).
Table 1
Dog ID | Sex | Breed | Age (years) | Tumor location | Evidence of metastasis | Histological subtype |
Naïve 1 | FS | Mixed (Husky) | 8 | L proximal humerus | No | Osteoblastic |
Naïve 2 | MC | Catahoula | 11.5 | R distal femur | Yes | Osteoblastic |
Naïve 3 | MC | Labrador Retriever | 7.8 | L distal femur | No | Fibroblastic |
Naïve 4 | MC | Great Dane | 8 | R distal radius | No | Osteoblastic |
Naïve 5 | FS | Mixed | 11.3 | R distal radius | No | Chondroblastic |
Naïve 6 | FS | Catahoula | 8.4 | R distal radius | No | Osteoblastic |
Initial low-resolution cell type annotation revealed the presence of 7 major cell types consisting of T cells, B cells, tumor infiltrating monocytes (TIMs)/tumor associated macrophage (TAMs), dendritic cells (DCs), osteoclasts (OCs), tumor cells, cycling tumor cells, and an additional 3 minor cell populations consisting of neutrophils, mast cells, and endothelial cells (Fig. 1a). Evaluation of the dataset for evidence of batch effects indicated uniform distribution of cell types between biological replicates. The one exception was that naïve dog 6 had more neutrophils and T cells relative to the other study dogs (Fig. 1b). This skew might be a result of sampling bias in which necrotic tumor, blood, or bone marrow contamination was introduced during sampling. Subsequent analysis of cell type proportions revealed 42.3% of the dataset consisted of tumor cells or fibroblasts, 2.1% of the dataset was endothelial cells, and the remaining 55.6% was composed of immune cells (Fig. 1c).
Cell types were annotated based on expression of canonical markers, reference mapping using a human OS dataset, and gene set enrichment analysis (Fig. 1d). Cell type gene lists used by Liu et. al to define cell populations in human OS were applied using module scoring to provide further support for cell classifications (Supplemental Fig. 1a)12. These approaches consistently enabled the identification of T cells, B cells, osteoclasts, and endothelial cells. However, our high-level unsupervised clustering failed to distinguish between stromal fibroblasts and malignant osteoblasts. This unexpected observation may be in part due to the presence of a fibroblastic osteosarcoma tumor in our dataset and the broad expression of fibroblast markers (FAP, FBLN1) across all tumor cell clusters (Supplemental Fig. 1b).
Due to the inability to identify a distinct fibroblast population using feature expression, we applied CopyKAT to complete copy number variation (CNV) analysis to infer which cells exhibited aneuploidy based on their transcriptional properties (Fig. 1e)14. The analysis revealed that the majority of cells in the tumor/fibroblast cluster exhibited evidence of CNV aberrations with only a small subset of cells predicted to be diploid (Fig. 1e; purple arrow). The diploid cells were determined to represent a small cluster of fibroblasts which were investigated further through independent reclustering.
Dissection of the tumor and stromal populations reveals a distinct fibroblast cluster
Independent reclustering on cycling tumor cells and tumor/fibroblasts identified 10 distinct cell clusters which we defined as 4 cycling malignant osteoblasts clusters, 5 non-cycling malignant osteoblast clusters, and 1 fibroblast cluster (Fig. 2a). The defining features for each cluster were identified using a Wilcoxon Rank Sum test and the top 3–5 unique features were visualized using a heatmap and feature plots (Fig. 2b/c). Overall, the malignant osteoblasts exhibited a unique gene expression profile with collagen genes and alkaline phosphatase (ALPL) contributing to the gene signatures. We observed a small cluster of tumor cells (c9) that exhibited a gene expression pattern (OAS1, ISG15, OAS2) consistent with an interferon (IFN) response gene signature (Fig. 2b). This observation was further supported through completion of GSEA using Hallmarks gene set terms (Fig. 2d). Similar IFN signature enriched cells have been reported among immune cells, but the observation of such a cluster in a tumor population has not been previously reported in human OS studies12,13,15,16. Interpretation of GSEA further revealed that fibroblasts (c6) exhibited the most pronounced “epithelial-mesenchymal transition” (EMT) and “angiogenesis” signatures, which suggests the fibroblasts might play a role in promoting tumor growth. Additionally, GSEA supported the annotation of hypoxic osteoblasts (c4), as the cluster exhibited the strongest hypoxic transcriptomic signature.
To confirm the identification of fibroblasts, we used module scoring with a human fibroblast gene list17. This analysis confirmed Cluster 6 exhibited the strongest fibroblast gene signature (Supplemental Fig. 2a). We then completed differential gene expression (DGE) analysis contrasting fibroblasts (c6) and non-hypoxic osteoblasts (c0, c1, and c2) to better define the canine fibroblast gene signature (Fig. 2e; Supplemental data 1). While key fibroblast markers such as FAP and ACTA2 were identified, the top features consisted of SFRP2 and PRSS23 which were recently reported to be associated with a fibroblast population involved in wound healing18. To conclude our analysis of tumor cells, we sought to further investigate the transcriptomic signature of hypoxic osteoblasts (c4) by contrasting with non-hypoxic osteoblasts (c0, c1, and c2). Few differentially expressed genes were identified, suggesting cell types are similar, but subsequent pathway analysis identified enrichment of “response to oxygen levels” to be a top enriched pathway, suggesting that the tumor cells were indeed hypoxic (Fig. 2f, Supplemental data 1, Supplemental Fig. 2b). In summary, we were able to resolve a population of fibroblasts through completion of independent reclustering, as well as define the transcriptional heterogeneity within malignant osteoblasts.
Independent reclustering reveals a population of CXCL13+ follicular helper CD4 T cells
To ensure we captured all biologically relevant T cell populations, we completed independent reclustering, which led to the identification of 10 transcriptomically distinct clusters: 3 CD8 T cell, 4 CD4 T cell, 1 NK cell, and 2 mixed CD4/CD8 T cell clusters (Fig. 3a/b). Next, we interrogated T cell subtypes using an approach that has been applied in human breast cancer and OS scRNA-seq datasets to describe T cell populations13,19. We modified the gene lists used in previous applications to include signatures for cycling T cells, NK cells, and IFN-signature T cells that we recently established in circulating canine leukocytes20. Overall, this approach proved to be consistent with annotations assigned using canonical markers (Fig. 3c). Although the gene signatures were definitive for naïve CD4 T cells and cytotoxic CD8 T cells, other gene signature scores provided weaker support for their corresponding cell types. For instance, regulatory T cells (Tregs) and follicular helper CD4 T cells (CD4fh) both exhibited moderate enrichment for exhausted and costimulatory terms, with minimal distinction between the two T cell types. The analysis also revealed the presence of a T cell cluster with an IFN gene signature, a population that has been reported to be hypersensitive to stimulation16.
After identifying each T cell subset, we completed pseudobulk conversion and DGE analysis to further establish the transcriptomic signatures of Tregs and CD4fh. Comparisons between Tregs (c3) and activated CD4 T cells (c2) revealed overexpression of IL21R, TNFRSF4, and TNFRSF18, with CTLA4 being the most definitive marker of Tregs (Fig. 3d, Supplemental data 1). When repeating this analysis on CD4fh (c5) cells, we identified CXCL13, IL4I1, and TMEM176A to be defining features (Fig. 3e, Supplemental data 1). Although intratumoral CD4fh cells exhibited a distinct exhaustion profile (PDCD1, TOX, TOX2, IL4I1), they also displayed a gene signature consistent with follicular helper T cells (CXCL13, IL21, CD70)21,22. A similar population of CXCL13+ CD4fh T cells has been identified in multiple human tumors and the cell type has been implicated in tertiary lymphoid follicle formation and modification of intratumoral adaptive immune responses23–25. Our analysis confirms expression of CTLA4/FOXP3 on Tregs and CXCL13/IL21 on CD4fh is conserved across species, while also providing complete gene signatures for the canine T cell subtypes17.
Following initial cell classification, we determined the cellular composition of each cell type as a percentage of all immune cells and of all cells in each sample (Fig. 3f, Supplemental table 1). This analysis revealed exhausted CD8 T cells (CD8ex) and effector CD8 T cells (CD8eff) to be among the most abundant populations, along with activated CD4 T cells and Tregs. We then curated a heatmap of defining features predicted to be expressed on the cell surface with the objective of identifying potential cell markers to be used in alternative cell identification approaches, such as flow cytometry (Fig. 4g, Supplemental Fig. 3, Supplemental data 2)26. With the caveat that transcript presence does not always correlate with protein expression, the analysis suggested that TNFRSF4 (OX-40), TNFSF8 (CD153), and TMEM140 may represent valuable surface markers for further investigation of canine Tregs, CD4fh, and TIFN−sig, respectively. Together, the relative cellular percentages and potential surface markers provide a foundation for further functional study of the cell types identified in our transcriptomic analysis.
Mature regulatory dendritic cells are present in canine OS and are predicted to modulate T cell mediated immunity
Five dendritic cell (DC) subtypes were identified when completing independent reclustering on FLT3+ cells. The subtypes identified included conventional DC2s (cDC2; c0), mature regulatory DCs (c1; mregDC), cDC1s (c2), plasmacytoid DCs (c3; pDC), and precursor DCs (c4; preDC) (Fig. 4a). Key features used to assign cell type identities included DNASE1L3 (cDC1), CCR7/IL4I1 (mregDCs), CD1C (cDC2), IL3RA (preDC), and IGKC (pDC) (Fig. 4b)27,28. The population of canine preDCs closely resembled a recently redefined human preDC cell type that exhibits a tendency to cluster with pDCs when investigated using scRNA-seq27. Of note, we previously identified cDC2, cDC1, preDC, and pDC cell types in canine peripheral blood, however mregDCs (c1) were not observed, suggesting a potential tissue specificity20. The identification of mregDCs, also reported as migratory (mig) DCs, is of note as this cell type is predicted to modulate T cell responses29,30. Thus, we provide evidence that a key immune regulatory cell type is present in canine OS.
We next used hierarchical clustering and Toll-like receptor expression to investigate differences between preDCs and pDCs. Hierarchical clustering indicated preDCs are closely related to myeloid cDC2s and cDC1s, suggesting a myeloid lineage, while the pDCs were located on a unique clade, suggesting a lymphoid origin (Fig. 4c). In humans, pDCs are reported to exhibit high expression of TLR9 and TLR7, which we identified to be highly expressed on canine pDCs (Fig. 4d)31. To ensure that none of the DC populations were of B cell origin, we evaluated MS4A1 (CD20) expression and found it to be minimally expressed (Supplemental Fig. 4a). We then used SCENIC to predict active regulons in each DC subtype (Supplemental Fig. 4b). This analysis revealed RUNX2, a master regulator of pDC development, to be enriched in both pDCs and preDCs32. Overall, these findings suggest canine preDCs are closely related to the recently defined plasmacytoid-like human preDCs27.
To confirm mregDCs exhibited a mature, immune regulatory transcriptomic signature, we used module scoring with gene lists previously applied to investigate human DC subtypes29,33. This analysis revealed that mregDCs had a marked enrichment for migration, regulatory, and maturation associated gene signatures (Fig. 4e). Subsequent, DGE analysis of canine mregDCs (c1) relative to cDC2s (c0) revealed a distinct mregDC signature of CCR7, IL4I1, CCL19, and FSCN1 with substantial overlap to the human mregDC transcriptional program (Fig. 4f, Supplemental data 1)29. With the precedent that mregDCs interact with intratumoral T cells to shape adaptive immune responses in humans, we wanted to determine whether a similar interaction might occur between mregDC and T cells in dogs23,24. We used CellChat to evaluate interactions between mregDCs and T/NK cells34. This analysis revealed enriched PD-1/PD-L1 and CTLA4/CD80 interactions between mregDCs and CD4 Tregs, Tfh cells, and naïve T cells (Fig. 4g). In summary, we present the transcriptomic signature of canine mregDCs and provide evidence of intratumoral interactions between canine mregDCs and T cells.
Macrophage transcriptomic states supports a spectrum of cell types
Due to the transcriptional overlap between tumor associated macrophages (TAMs) and osteoclasts (OCs), we analyzed these two cell types in the same UMAP space. In doing so, our analysis highlighted the relatedness of OCs and TAMs which would have been overlooked if analyzed independently. Through independent reclustering we identified 8 transcriptomically distinct macrophage/monocyte populations which were annotated using modified nomenclature derived from Ma et al (Fig. 5a-c)35. Activated TAMs (c0, TAM_ACT) and intermediate TAMs (c1, TAM_INT) did not fit into any of the macrophage subtypes presented in Ma et al, so they were instead annotated based on an activated signature (CD5L, CD40, CD80) and an intermediate polarization signature, respectively. Tumor infiltrating monocytes (TIMs) were divided into two populations based on CD4 expression, a division of monocytes unique to dogs20,36. Unsupervised clustering divided lipid-associated (LA-) TAMs into two subclusters defined by either C1QChi expression (c3) or SPP2hi expression (c2). To better define the distinctions between the two LA-TAM populations we completed pseudobulk-based DGE analysis (Fig. 5d, Supplemental data 1). The analysis revealed IL2RA, CXCL10, and SERPING1 as key markers of C1QChi LA-TAMs, while ENO1, LGALS3, and RBP4 defined SPP2hi LA-TAMs. Based on the analysis, C1QChi LA-TAM appear to most closely resemble the definitions of human LA-TAMS provided by Ma et al.
In addition to the recently proposed TAM nomenclature, we used module scoring with pro- and anti-inflammatory gene lists to investigate the macrophage populations in a more traditional dichotomy (Supplemental table 2)37. We identified the C1QChi LA-TAM (c3) cluster to have the strongest anti-inflammatory transcriptomic signature while CD4+ monocytes (c11) exhibited the most prominent pro-inflammatory transcriptomic signature (Fig. 5e). To further investigate the signatures of Clusters 11 and 3 we completed pseudobulk-based DGE analysis (Fig. 5f, Supplemental data 1). The genes upregulated in Cluster 11 exhibited overlap with the predefined gene set that was used to identify the cluster as pro-inflammatory, while also revealing IL-1B, S100A12, LTF, and VCAN as defining features. DGE analysis of the anti-inflammatory cluster (c3) exhibited less overlap with the gene list originally used to identify the cluster as anti-inflammatory (with MRC1 the only overlapping feature), but analysis further revealed APOE, IGF1, and complement receptors C1QA/B/C as enriched markers. The top features identified when contrasting Clusters 11 and 3 were then used to generate a heatmap to evaluate how the expression of these features varied across all macrophage clusters (Fig. 5g). Findings from the analysis suggested that there is a spectrum of macrophage phenotypes, which is consistent with human macrophage literature38. As such, we next sought to better define the heterogeneity of the macrophage populations without relying on predefined cell type gene signatures.
Gene set enrichment analysis was used to provide further insights into the inferred functional capacity of each macrophage subtype (Fig. 5h). Cell clusters 4, 7, and 11 clustered together based on pathway enrichment scores suggesting the three transcriptionally distinct clusters have similar underlying gene signatures. LA-TAMs (c2, c3) and intermediate TAMs (c1) exhibited the strongest scavenger receptor associated activation, suggesting a mature macrophage population with immune suppressive properties39. Several terms were identified suggesting that both SPP2hi LA-TAMS and intermediate TAMs preferentially utilize oxidative phosphorylation and mitochondrial metabolic pathways. C1QChi LA-TAMs had a distinct profile suggestive of lipid and polysaccharide metabolism. Lastly, GSEA confirmed c10 to be consistent with IFN-TAMs based on strong enrichment of IFN signaling associated terms. In summary, we described the transcriptional profiles of macrophages in the canine OS TME which provides a foundation for further investigation of the functional relevance of each cell type.
Analysis of osteoclasts reveals four transcriptomically distinct populations
Using the same UMAP space, we next shifted our focus to further characterize osteoclast heterogeneity. Consistent with human and murine reports using single-cell RNA sequencing to characterize OCs, we identified 4 transcriptiomically distinct OC populations12,13,40. The cycling OCs (c5/c8) in our canine OS dataset likely correspond to previously reported pre/progenitor OCs, while the mature OCs (c6) are consistent with previous reports (Fig. 6a/b). CD320+ OCs (transcobalamin receptor expressing OCs, c9) have not been described in macrophage or osteoclast clusters from human and mouse tissues and may represent a canine specific cell type, or more likely a previously unresolved OC subtype (Fig. 6a/b). Due to the similarity of OCs and macrophages we completed hierarchical clustering to confirm the unsupervised clustering results (Fig. 6c). The secondary analysis was consistent with unsupervised clustering and further suggested Clusters 5, 6, 8, and 9 are distinct from the macrophage clusters.
To confirm the mature OC classification and provide a canine specific transcriptomic signature, we completed DGE analysis. When comparing mature OCs (c6) to macrophages (c0, c1, c2, c3) we identified canine mature OCs to be defined by ATP6V1C1, CD84, HYAL1, and CAMTA2 expression, with subsequent GSEA analysis suggesting an association with bone resorption and remodeling (Fig. 6d, Supplemental data 1, Supplemental Fig. 5a).
We next completed DGE analysis contrasting CD320+ OCs (c9) with macrophages and mature OCs. (Supplemental data 1, Supplemental Fig. 5b/c). By evaluating the intersection of the differentially expressed genes we determined CD320+ OCs are defined by HMGA1, TNIP3, and CD320 expression (Fig. 6e). The analysis also provided further evidence that c9 is an OC cluster based on TNFRSF11A (RANK) enrichment when contrasted with macrophage, but not when contrasted to mature OCs41. Lastly, we used SCENIC’s regulon specificity scoring to better define the transcription factors active in mature OCs and CD320+ OCs. We identified ZEB1 and NFATC1, known regulators of OC development, to be enriched in mature OCs, while TCF4, IRF5, and TP53 were enriched in CD320+ OCs (Fig. 6f/g)42,43. Together this suggests, CD320+ OCs are a distinct population from mature OCs and may represent macrophage-like OC precursors.
Transcript abundance of widely used immunohistochemistry macrophage markers exhibit distinct specificity to myeloid cells
In contrast to other tumor types, there have been multiple reports in humans and dogs suggesting that increased TAM infiltrates in OS are correlated with reduced metastasis rates and increased patient survival8,9. Despite these reports, other groups completing similar analysis have concluded that increased macrophage infiltrates have a negative impact on OS clinical outcomes44. Given the conflicting nature of previous reports we sought to employ our dataset to investigate which cell types express the transcript of the prototypical macrophage markers used for IHC analysis in these previous studies. To complete this analysis, we profiled TIMs, TAMs, DCs, and OCs for the expression of widely used canine (MSR1 aka CD204 and AIF1 aka Iba1) and human (CD163 and CD68) macrophage markers (Fig. 7a). With the caveat that this analysis is limited to transcript abundance and does not evaluate protein expression, we found that CD163 transcript expression was the most specific for macrophages. CD68 expression was detected in TIMs, TAMs, and OCs, with a remarkably high expression levels in mature OCs. The expression of CD68 on mature OCs is consistent with human literature45. AIF1 (Iba1) was the most non-specific marker with diffuse expression across all cell types, except for mature OCs. Lastly, CD204 (MSR1) was determined to be largely specific to TAMs, but the expression also extended to CD320+ OCs and CD4+ monocytes. To investigate the translational relevance of this finding, we evaluated expression of the markers in human OS (Supplemental Fig. 6). We observed similar expression patterns, with marked variability in specificity of each marker, suggesting the variability is conserved across species.
Given the degree of heterogeneity within the myeloid compartment in the OS TME, we used a Wilcoxon Rank Sum test to identify features that define each cell type, then selected for features predicted to be expressed on the cell surface (Fig. 7b, Supplemental Fig. 7, Supplemental data 3). Overall, the analysis suggested there is substantial overlap in expression of most features. Despite the overlap, we were able to identify candidate markers which include ADAM28 for LA-TAM_C1QChi, TNFSF13B for IFN-TAMs, and CD84 for mature OCs. Lastly, we calculated the relative percentages of each cell type to further facilitate cell identification (Fig. 7c, Supplemental table 3). Together, the data presented here act as a foundation to further investigate the role of myeloid cells in OS biology.
Cell-cell interaction analysis indicates TAMs are involved in immune regulatory pathways
Following cell identification through independent reclustering of major cell types, we evaluated the cell-cell interaction networks using CellChat. Between the 41 cell types included in the analysis, we identified a total of 15,405 inferred interactions across 61 signaling networks. The number of interactions and the predicted interaction strength of incoming (express receptor) versus outgoing (express ligand) signals were used to infer the activity of cells within the TME (Fig. 8a, Supplemental Fig. 8a). The top three cell types predicted to have the strongest interactions were fibroblasts, mature OCs, and endothelial cells. We next categorized the significantly enriched networks as “immune specific”, “immune related”, and “non-immune” to investigate if certain cell types were more active in a subset of networks (Fig. 8b, Supplemental table 4). We found that malignant osteoblasts and stromal cells were largely predicted to be involved in “non-immune” interactions, while “immune specific” interactions were largely confined to TAMs and DCs with strong outgoing interactions.
By subsetting on immune cells and evaluating interactions of known immune regulatory pathways we identified mregDCs and IFN-TAMs to have the most interactions, while activated (CD5L+) macrophages and C1QChi LA-TAMs were predicted to have the strongest outgoing signals (Fig. 8c). It was further predicted that follicular helper and regulatory CD4 T cells make up the populations receiving most of the signals originating from myeloid cells. When evaluating the PD-L1 network, we identified mregDCs, TIMs, and IFN-TAMs to have the highest expression of PD-L1 and were predicted to interact with Tfh, Tregs, and exhausted CD8 T cells (Fig. 8d, Supplemental Fig. 8b). The CD80 and CD86 networks involved a larger portion of myeloid cells, with all CD4 T cells predicted to be influenced by the interactions (Fig. 8e, Supplemental Fig. 8c/d). Overall, activated TAMs, IFN-TAMs, and C1QChi LA-TAMs are predicted to be key contributors to the suppression of T cell mediated immunity.
Comparison of human and canine scRNA-seq OS datasets reveal a high degree of similarity in cell type gene signatures between species
Lastly, we obtained 6 publicly available treatment-naïve human OS scRNA-seq samples to complete a cross-species analysis (GSE162454)12. The two datasets were integrated using a SCTransform workflow which is reported to overcome genome annotation differences between species46. Hierarchical clustering of SCT normalized data revealed a high degree of similarities between species, with major clades containing similar cell types based on pre-integration annotations (Fig. 9a). All canine lymphocyte subtypes paired 1:1 with their human counterpart, as did endothelial cells and fibroblasts. Discrepancies between species included the placement of mast cells, which clustered into separate clades. Overall, macrophages clustered in the same clade, but due to differences in annotation levels, many cell types did not pair off into terminal clades.
To further compare transcriptional programs across species we used an analysis approach adopted from Scheyltjens et al47. Briefly, the approach used DGE analysis between two cell populations in each species, then signing the adjusted P value to determine if transcriptomic signatures were conserved. When contrasting fibroblasts and endothelial cells, we found substantial overlap in gene expression patterns with key endothelial cell markers (PLVAP, CD34, and PECAM1) enriched in both species (Fig. 9b, Supplemental data 4). Top features conserved in fibroblasts included VCAN, COL6A1, and LUM, while key features such as FAP and ACTA2 were also conserved. Interesting discrepancies included the expression of HYAL2 and NOTCH as defining features in human endothelial cells, but nonsignificant in canine endothelial cells.
Completion of the same analysis on plasmacytoid DCs and cDC2s revealed TCF4 to be enriched in pDCs and BATF expression enriched in cDC2s, which is consistent with human literature (Fig. 9c, Supplemental data 5)28. An intriguing distinction between species included the high expression of GZMB and PTGDS (prostaglandin D2 synthase) in human pDCs, but not in canine pDCs. Lastly, we applied the same approach to compare mature OCs with TIMs (Fig. 9d, Supplemental data 6). As expected, mature osteoclasts were defined by CSTK, ACP5, and ATP6V0D2 expression, while monocytes in both species were defined by CXCL8, OSM, and LYZ expression. Notable differences included canine monocytes exhibiting high expression of SLAMF9 and PLBD1, while human monocytes had high S100A8 and HCST expression. In summary, we present a comprehensive comparison of human and canine OS cell types, which suggests a high degree of consistency in cell type gene signatures across the two species, however we also present evidence of distinct transcriptional programs in pDCs, mast cells, and monocytes.