CXCL9:SPP1 macrophage polarity identifies a network of cellular programs that control human cancers

Abstract

Tumor microenvironments (TMEs) influence cancer progression but are complex and often differ between patients. Considering that microenvironment variations may reveal rules governing intratumoral cellular programs and disease outcome, we focused on tumor-to-tumor variation to examine 52 head and neck squamous cell carcinomas. We found that macrophage polarity, defined by CXCL9 and SPP1 (CS) expression but not by conventional M1 and M2 markers, had a noticeably strong prognostic association. CS macrophage polarity also identified a highly coordinated network of either pro- or anti-tumor variables, which involved each tumor-associated cell type, and was spatially organized. We extended these findings to other cancer indications. Overall, these results suggest that, despite their complexity, TMEs coordinate coherent responses that control human cancers and for which CS macrophage polarity is a relevant yet simple variable.

Introduction

TMEs are ecosystems composed of diverse stromal, immune, and cancer cells. The interactions between these components are incompletely understood, but some of them produce pro- or anti-tumor effects (1). It is therefore essential to decipher the components of these ecosystems and better understand some of their internal governance and rules of interaction. The results could lead to an improved understanding of the regulatory mechanisms that determine cancer progression, as well as resistance or response to treatment and, by extension, better define patient prognosis and treatment options. To date, many studies have described TMEs using low-resolution techniques, such as bulk RNA sequencing and immunohistochemistry. These data have allowed classification of TMEs based on broad criteria, such as the composition of the immune infiltrate and the character of the inflammatory response (2). For example, the abundance or phenotype of various immune cells within tumors, including T-cells (3–5), B-cells (6–8), dendritic cells (DCs) (9–11), and macrophages (12–14), may be associated with distinct clinical prognoses. More recently, single-cell RNA sequencing (scRNAseq) is increasingly used to define the TME, as it directly reveals its cellular heterogeneity without the need for predefined sets of markers for cell identification (15–20). Nevertheless, scRNAseq studies typically consider pooled cells from all patients as units of statistical replication, which prevents formal testing of tumor-to-tumor variation between patients. Here, we designed a population-oriented scRNAseq approach that considers tumors, not cells, as units of statistical replication. This approach allowed us to discover features that covary between individuals, and thus to better understand the complexity of TMEs and the rules governing their composition and relationship to disease outcome. We also performed histological analyses to validate the scRNAseq-based results and obtained spatial information on the variables of interest within the TME.

Results

Tumor-associated macrophages are major contributors to the clinical outcome of HNSCCs

We generated a scRNAseq dataset from 52 fresh HNSCC tumor tissues from 51 individual patients (MGH/MEE-HNSCC cohort) with a wide range of clinical characteristics, spanning the clinical spectrum, from primary tumors to locoregional recurrences to distant metastases, and without selection or enrichment for a particular cell type before sequencing (fig. S1, Table S1). We defined cell clusters from a total of 187,399 cell transcriptomes, using Seurat’s FindCluster and manually annotating them at different resolutions, from main cell compartments to minor cell states (fig. S2). This dataset comprehensively covered all three main cell compartments, namely tumor, stromal and immune cells, as well as all major cell types, namely tumor and epithelial cells, fibroblasts, endothelial cells, lymphocytes and myeloid cells (Fig. 1A, fig. S3, fig. S4). These cell types were reproducibly found in patients (fig. S5A), but their abundance varied greatly (fig. S5B). Epithelial cells, fibroblasts, endothelial cells, lymphocytes, and myeloid cell states contained a finite number of minor cell states, which were also reproducibly found in patients; this was in stark contrast to the tumor cell states from the same patients, which were largely patient-specific (fig. S5C).

Fig. 1. CXCL9:SPP1 polarity of tumor-associated macrophages is a major contributor to the clinical outcome of HNSCCs.

(A) Uniform manifold approximation and projection (UMAP) visualization of single-cell transcriptomic profiles combined from all 52 samples from the MGH/MEE-HNSCC cohort, highlighting separation of main cell lineages and showing finer clustering within each.

(B) Identification of genes that are dominantly expressed (with fold-change > 3 relative to the next most-expressing cell type; multiple-test testing p-value < 0.05 for patient-level comparison) in each major cell type, and based on scRNAseq data from the MGH/MEE-HNSCC cohort (left panel). Projection of these genes to bulk RNAseq data from public datasets (totalling n=886 HNSCC patients) shows their relative expression pattern together with associated cohort of origin, HPV status and molecular subtypes (middle panel). Pairwise correlations between the same genes (right panel). The patients were sorted according to hierarchical clustering based on the tumor-specific genes only.

(C) Comparison of three prognostic signatures, each derived using tumor, stromal and immune genes only. Multivariable Cox regression was used to obtain the hazard ratios (with Wald 95% confidence intervals, shown as horizontal bars, and p-values), based on the cross-validated prognostic scores derived using the GLMNET Cox model and applied to pairwise differences of expression of the genes.

(D) Similar analysis as in (C), but using genes dominant in each subdivision of the immune cell types.

(E) Kaplan-Meier plot of cross-validated macrophage prognostic score from (D) at 50% cutoff, showing the genes with opposite effects selected by the model.

(F) Scatter plot of CXCL9 and SPP1 expression in each of 16,292 macrophages, and a contingency table based on dichotomized expression (positive if UMI counts >3), with odds ratio and Fisher’s exact test p-value to indicate mutual exclusion (odds ratio < 1).

(G) Histological analysis with combined RNA-ISH/IF, using RNA probes for CXCL9 and SPP1 and antibody marker for macrophages (CD68) on 10 samples from the MGH/MEE-HNSCC cohort. Representative histology images (left) and quantification of TAM positive or not for CXCL9 and/or SPP1 (right) are shown.

(H) Cox regression analysis in bulk RNAseq cohorts (as in Fig 1B and 1E) comparing the prognostic impact of TAM abundance signature (derived from the pseudo-bulk profiles of the scRNAseq data using GLMNET classifiers) and the CXCL9:SPP1 (CS) ratio. Wald 95% confidence intervals and p-values were shown.

(I) Scatter plots showing lack of correlations between the macrophage signature score (Fig 1D-E) and common M1/M2 markers, and substantial correlations between the whole signature with individual expression of CXCL9 and SPP1, as well as the CS ratio. Spearman’s rank correlation is used and when significant a fitted red line is shown.

(J) Association between CS ratio and response to an anti-PD1 mAb-containing treatment regimen in 14 patients of the MGH/MEE-HNSCC cohort, with significant Fisher’s exact test.

We identified 1189 genes whose expression in the MGH/MEE-HNSCC cohort was dominant (i.e. at least 3-fold higher) in a given cellular compartment relative to all others (Fig. 1B, left, Table S2). We then used this gene list to interpret a collection of bulk mRNAseq data from 886 HNSCC patients obtained by pooling three independent cohorts (TCGA, LPZG, FHUT; Table S3). The expression heat-map of these genes in all 886 patients (Fig. 1B, middle, fig. S6A) showed mixing of the cohorts, indicating inter-cohort homogeneity. It also showed an association between human papilloma virus (HPV) status and molecular tumor subtype, as expected (21). Finally, pairwise correlation analysis showed that the TME compartment-dominant genes were co-enriched in their respective cellular compartments (Fig. 1B, right), indicating that what becomes apparent as co-regulation in bulk mRNA data may in fact largely be driven by heterogeneity in abundance of various TME cell types. Remarkably, tumor cell-gene intrinsic clustering was still associated with differential abundance of stromal and immune cell transcriptomes (fig. S6B), suggesting the existence of an orchestrated cellular communication between different TME compartments.

Using the information obtained from the analyses above, we next investigated the influence of the three main cellular compartments (tumor, stroma, immune) on clinical outcome. We applied supervised methods that are commonly used for bulk mRNAseq analysis, but considered only the cell-type dominant genes as the feature variables. To ensure internal normalization of gene expression in each cell type, we used a pairwise log ratio method (22) followed by construction of multi-pair signatures using an elastic network classifier (23) that was trained on overall patient survival. Univariate analysis showed that each major cellular compartment had a significant prognostic impact (fig. S7A). However, a multivariate analysis in which all three compartments compete to explain the patient’s clinical course showed that only the stromal (p=0.01) and immune (p=0.0003) signatures retained their prognostic ability (Fig. 1C), indicating a more direct association between these non-tumor components with disease progression.

Because immune cells include different subsets with distinct functions, we sought to determine which cells carried the most prognostic information. In univariate analyses, every major immune cell type except monocytes had a significant prognostic impact (fig. S7B); however, in multivariate analyses only mast cell (p=0.04) and macrophage (p=0.01) signatures retained independent effects (Fig. 1D). This showed that prognostic information from the different cell types was mostly redundant and suggested close coordination between these cells. We also illustrated the relevance of the macrophage-specific signature in Kaplan-Meier survival analyses (Fig. 1E).

Among the prognostic gene pairs identified in macrophages, we selected CXCL9 and SPP1 because the expression ratio of these two genes (CS) sufficiently captured the information in the entire signature (fig. S8A) and these two genes already have established and arguably opposing roles in cancer biology (24–26). The association of the other gene pairs with survival is shown in fig. S8A. CS tumor-associated macrophage (TAM) polarity was associated with overall survival when analyzed separately in the three independent HNSCC cohorts (fig. S8B,C); it also remained linked to disease outcome independently of CD274 expression (fig. S8D), which is frequently considered predictive of a patient’s clinical course (27). Furthermore, CS TAM polarity remained independent of tumor HPV status (fig. S9A-C) and disease stage (fig. S9D,E). Taken together, these data indicated that the TAM gene expression programs were particularly important in defining disease progression.

CXCL9 and SPP1 identify tumor-associated macrophage polarity in vivo

Quantitative analyses of scRNAseq (Fig. 1F) and histological (Fig. 1G) data both showed that expression of CXCL9 and SPP1 by macrophages was largely mutually exclusive at the single cell level. Also, by inferring the abundance of TAMs and their CS polarity in bulk RNAseq data (fig. S10), we found that CS polarity was prognostic, whereas TAM abundance was not (Fig. 1H). Patients with more SPP1^hi TAMs had poorer clinical outcomes, whereas patients with more CXCL9^hi TAMs had better ones and the CS ratio correlated strongly with our initial macrophage prognostic score; in contrast, TAM expression of M1 (ITGAX and CD80) or M2 (CD163 and MRC1) macrophage markers, which are widely used to define macrophages with anti- and pro-tumor functions, respectively, was not related to distinct clinical outcome in the same patients (Fig. 1I). These results supported the notion that M1 and M2 phenotypes define the possible extremes of polarized cells in vitro, but may not necessarily reflect the complexity of TAMs in vivo (28). Finally, CS TAM polarity was low pre-treatment in nine out of nine patients who did not respond to anti-PD1-containing treatment regimens, but high in three out of five patients who responded to treatment (Fig. 1J). This is additional evidence that CS TAM polarity alone may carry meaningful clinical information.

Next, we evaluated whether CXCL9 and/or SPP1 expression was restricted to discrete, minor monocyte or TAM states. To this end, we first used unsupervised clustering of scRNAseq data to reveal the complexity of monocytes and macrophages in HNSCCs. We found stereotyped gene expression programs that defined a total of four conserved monocyte states and seven conserved TAM states (fig. S11A), which were reproducibly found in patients (fig. S11B). Each state was named after the gene that it expressed dominantly over all other monocyte and TAM states (fig. S11C, Table S4). Monocytes mostly expressed low levels of CXCL9 and SPP1. Among TAMs, Mac_APOE, Mac_CXCL10, and Mac_F13A1 preferentially expressed CXCL9, and Mac_SPP1 and Mac_MT1H preferentially expressed SPP1 (fig. S11D); however, all states of TAMs expressed both transcripts at variable levels (these findings are further discussed below and presented in Fig. 2F and fig. S19-20).

Fig. 2. CXCL9:SPP1 tumor-associated macrophage polarity determines the broader TME.

(A) Patient ranking according to the CS TAM ratio.

(B) Cell counts of major cell types. Patients were ranked from lowest to highest CS TAM ratio. Significant correlation based on Spearman’s rank correlation with CS TAM is indicated with a red line. Each dot represents the value for one sample with the dot size being representative of the cell number contributing to that value.

(C) Quantitative immunofluorescence on 23 samples showing correlation between CS TAM ratio and CD8⁺ T-cell abundance in both tumor nest and stroma.

(D) Correlations between T, B and dendritic cell abundance (relative to TAM counts) and common M1/M2 markers, CXCL9, SPP1 and CS ratio. Spearman rank correlation is used, and when significant, a fitted red line is shown.

(E) Cell-type- and patient-specific expression of example genes, plotted according to patient order defined by CS ratio. The left panels show genes that are positively correlated with CS ratio, and the right panels show negatively correlated genes.

(F) Association between CS TAM polarity and the expression of CXCL9 (left), and SPP1 (right) in minor TAM states.

(G) Heat-map summarizing the correlation analysis (similar to Fig 2E) when applied to 195 cytokines in all cell types and indicating those that are significant in at least one cell type (FDR < 0.05).

(H) Heat-map of gene set enrichment analysis results, applied separately to each cell type and taking as the input the correlation with CS ratio of Fig 2F, showing significant MSigDB hallmark gene sets.

(I) Circos plots showing well-known ligand-receptor pairs from curated databases (as compiled in the NicheNet software package) under the requirement that either the ligand or the receptor (or both) is correlated with CS ratio, in addition to both being expressed. Arrows are pointing from the ligand towards the receptors. Only selected example pairs are labeled and highlighted for CXCL9 network. Abbreviations: Tu: tumor, T: T-cells, B: B-cells, N: neutrophils, MC: mast cells, DC: dendritic cells, Mo/Mø: monocyte/macrophages, EC endothelial cells, F fibroblasts.

**** p < 0.0001, *** p < 0.001, ** p <0.01, * < 0.05. P-values are adjusted for the number of cell types. ECs: endothelial cells, DCs: dendritic cells.

We also examined whether any of the observed monocyte or TAM states corresponded to M1 or M2 cells (fig. S11E). To this end, we used M0, M1 and M2 signatures of in vitro-polarized cells and involving a total of 596 genes identified from public dataset GSE158094 (29). We found that SPP1^hi and CXCL9^hi TAMs all expressed elevated M0, M1 and M2 signatures compared to monocytes and Mac_MARK4. The only substantial differences were higher M2 vs M1 signatures in Mac_SPP1 and higher M1 vs M2 signatures in Mac_CXCL10. These results align with previous findings suggesting that the phenotypes of human TAMs only partially overlap with those of M1 and M2 markers (18, 28, 30, 31). Therefore, the study of CXCL9 and SPP1 expression by TAMs should be more relevant to capture the polarity of these cells in vivo and their association with other TME components and disease progression.

CXCL9/SPP1 tumor-associated macrophage polarity defines the broader HNSCC microenvironment

For subsequent analyses, we used the CS ratio to assess TAM polarity in vivo (Fig. 2A). Furthermore, we examined tumor-to-tumor variation of TAM polarity, considering that this variation may reveal rules governing the cellular and molecular composition of the TME. Specifically, we classified the 52 samples according to CS TAM polarity and performed correlation analyses between this variable and other TME variables, including the abundance of different cell types and their gene expression profiles. For the latter, the main unit of analysis was the expression of a gene averaged over all cells from the same sample and cell state (fig. S12). The resulting value is independent of cell abundance, unlike “pseudo-bulk” sums that are typically used (32). In doing so, we were able to detect co-regulation of gene expression between cell types, which indicates the existence of a coordinated communication network in the TME, as detailed below.

CS TAM polarity was positively associated with increased tumor infiltration by three major immune cell types, namely T-cells, B-cells and DCs (Fig. 2B, fig. S13). Operationally dividing T-cells into well-defined transcriptional states (33) revealed that all of them contributed to the association between CS TAM polarity and overall T-cell abundance (fig. S14A,B). Histological analyses on tissue sections (n=23) derived from the same tumors that were analyzed by scRNAseq confirmed this finding for various CD4⁺ and CD8⁺ T-cell populations (Fig. 2C, fig. S15). Similarly, both plasma and non-plasma B-cells contributed to the association between CS TAM polarity and overall B-cell abundance (fig. S16A). On the other hand, operationally dividing intratumoral DCs into well-defined states (34) showed that only the abundance of type 1 conventional DCs (cDC1_XCR1) was associated with CS TAM polarity (fig. S16B). Also, CS TAM polarity did not associate with the abundance of other stromal or immune cell types, including TAMs themselves (Fig. 2B). Therefore, the quality of TAMs, not their quantity, was related to the intratumoral abundance of three cell types, namely T-cells, B-cells and DCs, all of which have been linked to antitumor immunity (2).

Our scRNAseq data further showed that M1 or M2 markers, in contrast to CXCL9 and SPP1, were not strongly associated with T-cell, B-cell or DC abundance in the same tumors (Fig. 2D). Histological staining of CD8⁺ T-cells in tumors confirmed these findings (fig. S17). Overall, these data confirmed the existence of a coordination between CS TAM polarity and adaptive immune responses in human HNSCC; they also indicated the limited link between traditional M1 and M2 markers and such immune responses.

We extended our analyses of tumor-to-tumor variation in CS TAM polarity by investigating its connection to gene expression in all TME cell types. Initially, we assessed CXCL9 and found that not only TAMs but also tumor cells, fibroblasts, endothelial cells, monocytes and DCs could express it, despite at lower levels (Fig. 2E). Importantly, all these cell types did so in a coordinated manner with TAMs. For example, patients with higher CXCL9 expression in TAMs also had higher CXCL9 expression in these other cell types. In a like manner, tumor cells and monocytes could express SPP1 at lower levels; they did so in a coordinated manner with TAMs, and opposite to CXCL9 expression (Fig. 2E). This harmonized gene expression across cell types was independent of key tumor and patient characteristics, namely HPV/EBV status, nature of specimen obtained, anatomic site of tumor origin, sex, and smoking (see also fig. S18). Together, these results suggested that CS TAM polarity could broadly inform on related activities of other cells of the TME.

All conserved minor macrophage states expressed both CXCL9 and SPP1; they did so at varying levels but in a coordinated manner (Fig. 2F, fig. S19). For example, Mac_CXCL10 cells expressed CXCL9 at the highest level on average, but all other minor macrophage states expressed CXCL9 to some degree. The level of CXCL9 expression in Mac_CXCL10 cells varied between patients, but the inter-patient variation was preserved in all other minor macrophage states. Analysis of other minor cell states indicated that the results obtained for macrophages were broadly generalizable. For example, among monocytes, Mo_LYPD2 and Mo_CCL3L1 expressed CXCL9 in a manner that was coordinated with each other and with that of macrophages; similarly, Mo_CD300LB and Mo_CCL3L1 expressed SPP1 in a coordinated manner (fig. S20A). The same findings applied to DCs: cDC1_XCR1, DC3_LAMP3 and cDC2/MoDC_CLEC10A all expressed CXCL9 in a manner that was coordinated with that of macrophages (fig. S20B). Minor states that did not follow this coordination rule were generally those in which the gene of interest was never expressed; for example, Mo_PADI4, Mo_CD300LB, cDC2_CD1A and pDC_LILRA4 largely lacked CXCL9 expression in the 52 tumor samples analyzed. We concluded that CXCL9 and SPP1 were primarily expressed in macrophages, with expression of both genes coordinated across macrophage states and, more broadly, across various immune and nonimmune cells in the TME.

Next we investigated whether CS TAM polarity was co-dependent on that of other transcripts. To this end, we again studied all TME cell types. We found that 2,320 cell type-gene combinations (including 1,812 unique genes) were significantly associated with CS TAM polarity (FDR < 0.05; Table S5, fig. S21). The expression of a few of these genes, such as STAT1, was globally shared among all immune and nonimmune cell types (Fig. 2E). The coordinated expression of STAT1 in different cell types suggests the existence of coordinated programs that can occur across the entire TME and correlate with CS TAM polarity; yet cell type-specific programs correlating with CS TAM may be directly or indirectly linked to STAT1 activity, and possibly influenced by STAT1-independent pathways. The expression of other genes was instead coordinated with CS TAM polarity only in certain cell types, such as CXCR6 in T-cells and SLC2A1 (GLUT1) in tumor cells and macrophages (Fig. 2E). In addition, CS TAM polarity was directly associated with the expression level of a large set of cytokines and cytokine receptors, either positively (i.e. genes expressed more strongly in CS^hi tumors, such as CXCL9, CXCL10, CXCL11, CXCL16, and IL7) or negatively (i.e. genes expressed more strongly in CS^lo tumors, such as CXCL5, CXCL8, IL1A, IL1B and IL1RN) (Fig. 2G). The abundance of other genes, such as IFNG, CCL5, and CXCL13, was indirectly correlated with CS TAM polarity, as it depended on the abundance of the cell state expressing this gene (fig. S22). Finally, some genes were coordinated with CS TAM polarity only in certain minor cell types; such as IL12B in DC_LAMP3 but no other DC state (fig. S23). Overall, this coordination of expression of many genes, all connected to CS TAM polarity, occurred simultaneously in virtually all TME cell types.

To determine whether the genes associated with CS TAM polarity globally contributed to distinct biological processes, we performed gene set enrichment analyses in each cell type separately. This allowed us to reveal not only which gene expression programs were associated with CS TAM polarity, but also in which cell types these programs were active (Fig. 2H). CS^hi TAM-associated pathways included immune-related IFNα and IFNγ signaling in almost all cell types; IL-6/JAK/STAT3 signaling in tumor cells, fibroblasts, and endothelial cells; and complement activity in tumor cells, fibroblasts, and macrophages. In contrast, CS^lo TAM-associated pathways included epithelial-mesenchymal transition, hypoxia, and mTOR signaling in almost all cell types; glycolysis and angiogenesis in tumor cells, myeloid cells and T-cells; TNFα signaling in myeloid cells and T-cells; and TGFβ signaling in monocytes, mast cells and T-cells. Taken together, these data linked CS TAM polarity to a variety of immune and nonimmune cell programs, all of which have been independently proposed to regulate tumor growth in a positive or negative manner, and which we reveal here to be deeply interconnected in human tumors.

To identify possible cellular interactions involved in this broad program of co-regulation associated with TAMs, we used the well-curated subset of ligand:receptor pairs obtained from the NicheNet database (35), and only considered those for which the expression of the ligand and/or the receptor correlated with that of CXCL9 or SPP1 in TAMs. In addition, we manually annotated these predicted interactions a priori as putatively immuno-activating, immuno-inhibiting, or unknown, based on the nature of the role of the ligand or receptor reported in the literature (Table S6). This analysis identified multiple ligand:receptor interactions associated with CXCL9^hi TAMs, many of which were putatively immuno-activating and did not involve TAMs (Fig. 2I, left). In particular, we found predicted interactions between DCs and T-cells through the ligand:receptor pairs IL12B:IL12RB1, CXCL16:CXCR6, XCL1:XCR1, and CXCL9:CXCR3. All of these molecular interactions involving tumor-associated DCs have been shown to promote local antitumor T-cell functions in experimental mouse models (25, 36–39). We also found predicted interactions between TAMs and T-cells through the CXCL9:CXCR3 pair. The ability of macrophage-derived CXCL9 to stimulate antitumor T immunity has been shown in mice (40). CXCL9 expression in TAMs was also associated with putatively immuno-inhibiting CD274:PDCD1 (PD-L1:PD1) interactions between tumor cells and T-cells. This is consistent with our findings that tumors with CS^hi TAMs were enriched in IFNγ signaling and that CD274 is an IFNγ inducible gene (41). Overall, the CXCL9-associated interaction network was strongly associated with inflammatory reactions and antitumor immunity. The interaction network associated with SPP1^hi TAMs was more limited than that associated with CXCL9^hi TAMs and focused less on immune cell interactions (Fig. 2I, right). SPP1 was expressed in TAMs and tumor cells, whereas known SPP1 receptors (integrins and CD44) were found in a variety of cell types, such as fibroblasts and endothelial cells, consistent with the role of SPP1 in metastasis and angiogenesis (42). Since SPP1 is an extracellular matrix protein, activation of the SPP1-associated interaction network in tumors may reflect a shift from tissue inflammation to tissue repair.

To refine the above results, we further investigated interactions at the level of minor cell states (fig. S24). We found that interactions of some ligand:receptor pairs, such as CXCL9:CXCR3, CXCL16:CXCR6, and CD40LG:CD40, were rather pleiotropic because they involved many different minor cell states (with interesting new details, such as some minor DC states potentially acting as both sender and receiver in CXCL9:CXCR3 interactions). In contrast, other ligand:receptor pairs, such as XCL1:XCR1 and IL12B:IL12RB1, connected only certain minor cellular states. For instance, XCL1:XCR1 interactions selectively involved cDC1s as a sensor of the chemokine XCL1, which is known to recruit cDC1s to the TME and promote immune control of cancer (38). Similarly, IL12B:IL12RB1 interactions selectively involved DC_LAMP3 DCs (previously identified as DC3 (18, 34), mregDC (20), and LAMP3⁺ DC (17)) as producers of IL-12, which can locally stimulate antitumor immunity (36).

In conclusion, these data indicate that CS TAM polarity identifies critical cellular and molecular TME programs; these programs have been independently recognized for their ability to regulate tumor growth, and we show here they are all connected and identifiable by CS TAM polarity.

Regulation of CXCL9/SPP1 tumor-associated macrophage polarity

One possible explanation for the largely mutually exclusive expression of CXCL9 and SPP1 in TAMs could be that the tissue niche provides discrete cell-extrinsic cues that promote the generation of CXCL9⁺ or SPP1⁺ TAMs. Thus we investigated first whether these cells distributed differently in tumors. Histological analyses, combining RNA in situ hybridization (RNA-ISH) with immunofluorescence (IF), not only confirmed the CS TAM polarity of tumors initially identified by scRNAseq (fig. S25), but also showed that TAMs accumulated often as clusters of either CXCL9⁺ or SPP1⁺ cells (Fig. 3A,B, fig. S26A-C). The distribution of CXCL9⁺ and SPP1⁺ TAMs in the tumor stroma and tumor nests varied among patients, although there were increased accumulations of CXCL9⁺ or SPP1⁺ TAMs at the interface between the tumor stroma and tumor nests in CS^hi or CS^lo patients, respectively (fig. S26D,E). In contrast, CXCL9^– SPP1^– TAMs were more found in the tumor stroma and further away from the interface between stroma and nest (fig. S26E). Additional histological analyses showed that tumor cells expressing CXCL9 or SPP1 were more likely to be found at the vicinity of CXCL9⁺ or SPP1⁺ TAMs, respectively (Fig. 3C,D, fig. S26F). These data align with the identification of a coordinated regulation of CXCL9 and SPP1 expression between different cell types in the same tumors (Fig. 2E); they also suggest that local cues may foster the acquisition of CXCL9⁺ or SPP1⁺ phenotypes.

Fig. 3. IFNγ and hypoxia respectively increase and decrease CXCL9:SPP1 tumor-associated macrophage polarity.

(A) “Pseudo-image” of summarized biomarker status per cell from a combined RNA-ISH and IF histology from a HNSCC tissue sample. Each one of the 650,340 cells is plotted as a dot located at the center of the original cell image, colored according to categorized biomarker combinations. Two magnified regions highlighting concentration of CXCL9⁺ or SPP1⁺ TAMs are shown to the right.

(B) Quantitative analysis comparing the density of the CXCL9⁺ TAMs (top row) and SPP1⁺ TAMs (bottom) in the neighborhood surrounding every SPP1⁺ or CXCL9⁺ TAM cell (horizontal axis). The neighborhood is defined as four layers of cells according to Delaunay triangulation (with approximately ~40 µm distance to the outmost layer and a total of ~70 cells per neighborhood). The percentage of the cell type of interest is compared by t-test (summarized by the fold change and p-value). The box plots show Tukey’s lower and upper hinges, the quartiles and the median. Data from 6 samples were analyzed in the same manner.

(C) Similar plot to Fig 3A, adding CK⁺ tumor cells that expresses CXCL9 or SPP1.

(D) Similar analysis to Fig 3B, but comparing the percentage of CXCL9⁺ or SPP1⁺ tumor cells in the neighborhood surrounding CXCL9⁺ or SPP1⁺ TAMs.

(E) Histology image from a HNSCC sample, with antibody markers for TAMs (CD68) and T-cells (CD3), as well as RNA-ISH probes for CXCL9 and IFNG. Areas with different CXCL9⁺ TAMs densities are magnified, showing enrichment of IFNG⁺ CD3⁺ cells around CXCL9⁺ TAMs.

(F) Quantitative analysis of the histological images as in Fig 3E from 5 patients with CS^hi score, showing consistent enrichment of IFNG⁺ T-cells in areas with high CXCL9⁺ TAM density.

(G) Percentage of CD3⁺ cells among IFNG⁺ cells.

(H) RNA-ISH-IF analysis of IFNG and hypoxia marker GLUT1 in the context of SPP1⁺ and SPP1^– TAMs.

(I) Quantitative analysis comparing GLUT1 occurrence in the neighborhood of SPP1⁺ versus SPP1^– TAMs, in samples from 3 patients (with the same method as that for Fig 3B and 3D).

(J) Impact of IFNγ and hypoxia on the expression of CXCL9 and SPP1 by THP-1 macrophage-like cells. Cells were exposed to 40 ng/ml of IFNγ or 1 % O₂ for 24 hours. Expression of CXCL9 and SPP1 was measured by RT-qPCR and the combined CS ratio was calculated. Shown are technical triplicates representative of one out of two independent experiments.

(K) Impact of IFNγ or hypoxia on the expression of CXCL9 and SPP1 by TAMs obtained from two primary human HNSCC samples (HUG-01 and HUG-02). The cells were obtained by enzymatic digestion of fresh biopsies immediately exposed to 40 ng/ml of IFNγ or 1 % O_2. TAMs were FACS-enriched after 72 hours for analysis of CXCL9 and SPP1 content by RT-qPCR.

To examine whether distinct environmental signals might foster the production of CXCL9⁺ or SPP1⁺ TAMs, we characterized cellular and molecular neighbors of these two cell populations. We found that CD3⁺ and CXCR3⁺ cells were enriched near CXCL9⁺ TAMs (fig. S27A,B), which validated the ligand:receptor analysis results suggesting CXCR3:CXCL9 interactions between T-cells and macrophages (Fig. 2I, fig. S24A), and aligned with previous findings in mice linking CXCR3:CXCL9 signaling to intratumoral antitumor immunity (39). Furthermore, the neighborhood of CXCL9⁺ TAMs was enriched of IFNG⁺ T-cells (Fig 3E,F), which represented the vast majority of IFNG-producing cells (Fig. 3G, fig. S22, fig. S28). Conversely, the neighborhood of SPP1⁺ TAMs was enriched in GLUT1⁺ cells (Fig. 3H,I) indicating hypoxic areas (43). Taken together, the data indicate that CXCL9⁺ and SPP1⁺ TAMs exist in distinct neighborhoods, which contain factors that are linked to tumor control (T- cells, IFN-γ) and tumor progression (hypoxia), respectively, in broad agreement with the results of non-spatial scRNAseq analysis (Fig. 2E,G).

Given the physical proximity between CXCL9⁺ TAMs and CXCR3⁺ IFNG⁺ T-cells in vivo, and the ability of IFNγ to stimulate CXCL9 expression in other cancer settings (25), we sought to test in vitro whether exposing macrophages to IFNγ alone would promote the acquisition of a CXCL9⁺ phenotype. Conversely, given the physical proximity between SPP1⁺ TAMs and hypoxic areas identified by GLUT1 positivity, and the association between CS^lo TAM polarity and hypoxia signatures, we tested whether hypoxia would promote the acquisition of an SPP1⁺ phenotype in macrophages in vitro. To test these hypotheses, we first used THP-1 cells differentiated into macrophage-like cells (Fig. 3J). We found that exposure to IFNγ promoted CXCL9 expression, whereas hypoxic conditions promoted SPP1 expression; consequently, IFNγ increased the CS macrophage ratio, whereas hypoxic conditions decreased it. To extend these analyses to TAMs, we generated single cell dissociations of primary human HNSCCs biopsies (Fig. 3K). Again, we found that exposure to IFNγ promoted CXCL9 expression and increased the CS macrophage ratio, whereas hypoxic conditions promoted SPP1 expression and decreased the CS macrophage ratio. These results indicate that certain microenvironmental factors, namely IFNγ and hypoxia, can respectively regulate CXCL9 and SPP1 expression in TAMs. These results also help explain the spatial segregation of CXCL9⁺ and SPP1⁺ TAMs in tumors.

Relevance of CXCL9/ SPP1 tumor-associated macrophage polarity in a variety of cancers

Using a pan-cancer approach exploiting bulk RNAseq data from >10,000 patients (TCGA pooled data), we found that CS TAM polarity was strongly associated with overall survival in this combined cohort (fig. S29A). This association was particularly notable for patients with liver, ovarian, cervical, thyroid, or bladder cancer, as well as melanoma and sarcoma, in addition to head and neck cancer as shown above (fig. S29B). These data confirmed that CS TAM polarity alone carries significant clinical information; they also suggested that the relevance of this polarity to understanding the inner workings of the TME extends to many cancer indications.

To widen our analyses to scRNAseq data, we interrogated CS TAM polarity in an independent HNSCC cohort (n=18 patients (44)), as well as in non-small cell lung cancer (NSCLC; n=37 patients (45)) and colorectal cancer (CRC; n=62 patients (46)) cohorts. We chose these cohorts because they included a sufficiently large number of patients to examine tumor-to-tumor variations. For each cohort, we ranked patients according to CS TAM polarity, and then analyzed various cell types in the TME. We found in all three newly analyzed datasets that CXCL9 expression in TAMs was inversely correlated with SPP1 expression in the same cells (Fig. 4A), recapitulating our initial results obtained with the MGH/MEE-HNSCC cohort (Fig. 1F). Therefore, the expression of CXCL9 and SPP1 may be used to define TAM polarity in different cancer types. In marked contrast, monocytes and macrophages found in adjacent normal tissue were all largely CXCL9^– and SPP1^– (Fig. 4B), further suggesting that the acquisition of a CXCL9⁺ or SPP1⁺ phenotype requires environmental signals that may be absent in normal tissues.

Fig. 4. CXCL9:SPP1 tumor-associated macrophage polarity identifies a coordinated network of cellular programs in a wide range of human cancers.

(A) Scatter plot of CXCL9 and SPP1 expression in single TAMs from three independent scRNAseq datasets, including HNSCC (UPit cohort), lung and colorectal cancer patients. The number of patients for each cohort is indicated above. The inserts show contingency tables based on dichotomized expression (positive if UMI counts >3), with odds ratio and Fisher’s exact test p-value to indicate mutual exclusion (odds ratio < 1).

(B) CXCL9 and SPP1 expression in macrophages or monocytes, from tumor or adjacent normal tissue, in colorectal cancer. Healthy tissue is highlighted in green.

(C) Comparison between cell-type specific gene expression correlations with CS ratio in the MGH/MEE-HNSCC cohort (fig S21 and Table S5) and the corresponding correlations in three independent scRNAseq datasets, quantified by the “correlation of correlations” (p-value by Spearman’s rank correlation test).

(D) Rank correlation between CS TAM polarity and relative abundance of T, B and dendritic cells (Fig 2D), on the three independent scRNAseq datasets. P-values are indicated to the right of the forest plot and the 95% confidence intervals are shown as horizontal bars. The totals are summarized by random-effect meta-analysis.

(E) Spatial transcriptomics analysis of CosMx lung cancer data, showing spatial separation of CXCL9⁺ versus SPP1⁺ TAMs density in the same tumor sample (left panel), quantified by the same analysis as in Fig 3B. Enrichment of SLC2A1⁺ cells around SPP1⁺ TAMs and INFG⁺ T-cells around CXCL9⁺ TAMs are also observed (middle and right panel, respectively).

(F) Assessment of overall survival prognostic impact of bulk RNAseq CS ratio in HNSCC cohorts (as in Fig 1B-E), simultaneously with an EMT signature as competing explanatory variable in Cox regression analysis (HR: hazard ratios, Wald 95% confidence intervals are shown as horizontal bars).

(G) As in (F), applied to TCGA pan-cancer collection.

(H) As in (F), on HNSCC cohorts comparing CS ratio pairwise to single-gene expression of IFNG and CD8A, as well as the effector cells and Th1 signatures from Bagaev’s TME signature collection.

(I) As in (H) applied to TCGA pan-cancer collection.

We extended our analysis to all cell types and found that in addition to TAMs, tumor cells and fibroblasts could also express CXCL9, although at lower levels. Moreover, these cells did so in a manner coordinated with TAMs (fig. S30). Tumor cells and monocytes could also express SPP1, and did so in a way that was consistent with TAMs, and inversely with CXCL9 expression (fig. S30). We also noted that fibroblasts, and to a lesser extent DCs, mast cells, T-cells, and B-cells, expressed SPP1 in a manner coordinated with TAMs in the different cohorts analyzed, suggesting a possible extension of the synchronized expression of CXCL9 and SPP1 in the TME to these cells. We next investigated whether CS TAM polarity was co-dependent with that of other transcripts in the same patient cohorts presented above. We analyzed each cell type separately to reveal genes in each that might be coordinately expressed with CS polarity in TAMs; this approach identified the existence of CS TAM co-regulatory programs in virtually all cell types and in all cohorts (fig. S31). By performing a “correlation of correlations” analysis, we confirmed that the strong association between CS TAM and cell type-specific gene expression initially identified in the MGH/MEE-HNSCC cohort was largely replicated in the other three cohorts (Fig. 4C). In addition, CS TAM polarity correlated with the abundance of T-cells, B-cells and DCs in these cohorts. The correlations were particularly significant for T-cells and DCs in the CRC cohort, which contained a larger number of patients, and for all three cell types in all cohorts combined (Fig. 4D). Thus, in all cancer indications that we were able to analyze, CS TAM polarity was linked to (i) the abundance of antitumor cell types, and (ii) critical gene expression profiles in virtually all immune and nonimmune cells of the TME.

To examine the spatial distribution of CXCL9⁺ or SPP1⁺ TAMs in another cancer indication, we considered single cell spatial transcriptomics of tissue sections obtained from lung adenocarcinoma patients (47). We found that TAMs could accumulate as clusters of either CXCL9⁺ or SPP1⁺ cells (Fig. 4E). These data were confirmed with two other approaches using the CellCharter algorithm (48): first, unsupervised identification of spatial clusters centered on TAMs indicated two major hubs, which were enriched in CXCL9⁺ and SPP1⁺ TAMs, respectively (fig. S32A); second, neighborhood enrichment analysis revealed that SPP1⁺ TAMs were depleted near CXCL9⁺ TAMs and vice versa (fig. S32B). Furthermore, SPP1⁺ TAMs were more surrounded by SLC2A1⁺ (GLUT1) cells, whereas CXCL9⁺ TAMs had more CXCR3⁺ and IFNG⁺ T-cells near them (Fig. 4E, fig. S33). These data indicate that, irrespective of the cancer indication, CXCL9⁺ and SPP1⁺ TAMs may exist in distinct neighborhoods, which contain factors that are linked to tumor control (T-cells, IFNγ) and tumor progression (hypoxia), respectively.

To further assess the relevance of CS polarity, we examined whether its association with clinical outcome was dependent, or not, of cancer-associated variables. These analyses used bulk RNAseq data from 33 pooled cancer entities (n=10,429 patients), from whom the clinical outcome was known. Univariate analyses identified that both cancer stage and the CS ratio were associated with clinical outcome (fig. S34A). Furthermore, bivariate analysis showed that both variables remained significantly associated with clinical outcome (fig. S34B). These data indicate that CS polarity was independent of cancer stage across cancer types, thereby extending our initial findings in HNSCC (fig. S9D,E). Using a similar approach, we found that CS polarity was independent of known EMT markers linked to poor prognosis (49), both in HNSCC (Fig. 4F, fig. S35A) and in combined cancer indications (Fig. 4G, fig. S35B). Similarly, we compared how CS polarity, interferon and effector T-cell signatures (49) were associated with clinical outcome. We also considered IFNG expression on its own, as well as of CD8A expression as a proxy for a CTL response. Univariate analyses identified that each variable was associated with a better clinical outcome in HNSCC, with the exception of IFNG alone whose hazard ratio was just not significant (fig. S36A). However, bivariate analyses revealed that the interferon and T-cell signatures were all competed by CS polarity (Fig. 4H), suggesting that the latter is superior and sufficient to identify at least some TIL processes, e.g., those involving IFNγ responses. We extended these findings to all combined cancer indications (Fig. 4I, fig. S36B), further indicating that CS TAM polarity alone carries meaningful clinical information in multiple cancer types.

Concluding remarks

Through a focused, clinically-driven approach to single-cell analysis of HNSCCs, we addressed a major difficulty in human disease research, namely the very different manifestation of the same condition from patient to patient, and gained biological insight on TAM-related transcriptional co-regulation across cell types in the TME. Specifically, we show that cancer patient disease outcome is closely linked to CS TAM polarization, and that the latter, although independent of cell differentiation, exhibits plasticity and is strongly linked to the polarization of other immune, stromal, and tumor cells within the TME. CS TAM polarity identifies TME-wide coordination of pro- and anti-tumor pathways. The broad impact of our findings is highlighted in the identified relevance of this variable across multiple solid cancers.

One of the conclusions of this study is that the CS axis is not restricted to any particular TAM state. Therefore, instead of differentiating into highly polarized states, all sublineages of TAMs can activate the CS axis more or less strongly. This is consistent with the memoryless phenotypic plasticity that macrophages exhibit in vitro in response to different stimuli (29), and suggests that pathways associated with CS regulation can be transiently activated in many cell types in response to orchestrated extracellular signaling networks independent of intrinsic lineage differentiation programs. Defining whether the CS ratio can be used as a bona fide prognostic or predictive marker goes beyond the scope of this study; nevertheless, we have provided here a simple testable hypothesis involving the ratio of two genes, which may be measured by histology or qPCR analysis. We show that axes orthogonal to CS also exist in TAMs, for example those related to tissue-specific functions, and/or in other cell types; these additional axes require investigation and might be combined with CS to define non-redundant TME variables. In any case, the results presented here reveal that CS TAM polarity identifies a large co-regulated network of cellular states that control human cancers, and is thus a simple, but critical, determinant of the TME.

Our results have implications for how TAMs should be studied, particularly in the context of single-cell transcriptomics. First, for the analysis to have clinical relevance, it should focus on the expression of certain genes and the presence of certain signaling pathways in TAMs, rather than solely on the identification of TAM states and their abundance. Second, TAM activity should be considered at the patient level, as it is an important factor to explain transcriptomic heterogeneity as well as clinical outcome. Third, the patient-level approach also permits the study of co-regulation of various genes in different cell types, and intercellular communication in the TME. These considerations have implications not only for better understanding of the TME at the time of cancer diagnosis, but also for studying the efficacy of TAM-targeting therapies or other cancer treatments that may influence or be influenced by TAMs.

Considering the TME as an ecosystem of interdependent elements, perturbation of any of them should have circular effects. For example, our results showing IFNγ-induced CXCL9 (and hypoxia-induced SPP1) expression by macrophages suggest that CS TAM polarity is a consequence of signals from the TME. However, the resulting TAM phenotypes could contribute to the amplification of TME features, as indicated by the acquisition of distinct hallmark gene programs in TAMs in CS^hi and CS^lo tumors. Thus, TAM polarization is likely to be both a cause and a consequence of the TME composition, consistent with the notion that macrophages are plastic cells that tailor their responses to their environment (50). By extension, the entire network associated with CS TAM polarity could be induced or disrupted by multiple therapeutic means. Overall, insights gained from this work could help dissect the TME in various indications, and have broad implications for our understanding of cancer biology and the development of precision medicine.

Data and Materials Availability:

Data produced through the NIH/NCI P01 CA240239 grant will be shared in a manner consistent with data-sharing under the NIH Genomic Data Sharing Policy (NOT-OD-14-124). Raw counts of scRNA data are available from Gene Expression Omnibus with accession number GSE234933. Human data will be available for secondary research only after investigators have obtained approval from NIH to use the requested data for a particular project. Further information and requests for resources and reagents should be directed to M.J.P. and S.I.P.