ABSTRACT
Regulatory T cells (Tregs) can impair anti-tumor immune responses and are associated with poor prognosis in multiple cancer types. Tregs in human tumors span diverse transcriptional states distinct from those of peripheral Tregs, but their contribution to tumor development remains unknown. Here, we used single cell RNA-Seq to longitudinally profile conventional CD4+ T cells (Tconv) and Tregs in a genetic mouse model of lung adenocarcinoma. Tissue-infiltrating and peripheral CD4+ T cells differed, highlighting divergent pathways of activation during tumorigenesis. Longitudinal shifts in Treg heterogeneity suggested increased terminal differentiation and stabilization of an effector phenotype over time. In particular, effector Tregs had enhanced expression of the interleukin 33 receptor ST2. Treg-specific deletion of ST2 reduced effector Tregs, increased infiltration of CD8+ T cells into tumors, and decreased tumor burden. Our study shows that ST2 plays a critical role in Treg-mediated immunosuppression in cancer, highlighting new potential paths for therapeutic intervention.
INTRODUCTION
The recent clinical success of immune checkpoint inhibitors in the treatment of non-small cell lung cancer (NSCLC) highlights how targeting mechanisms of immunosuppression in the tumor microenvironment may be an effective therapeutic strategy (Makkouk and Weiner, 2015; Soria et al., 2015). However, only a subset of patients responds to immune therapies, suggesting that an improved understanding of other immunosuppressive mechanisms is needed for effective treatment.
One major mechanism of immunosuppression is posed by CD4+ regulatory T cells (Tregs), which are thought to play a dominant role in impairing anti-tumor immune responses (Tanaka and Sakaguchi, 2017). Tregs are critical for maintaining peripheral immune tolerance and preventing autoimmunity (Josefowicz et al., 2012; Sakaguchi, 2011). Characterized by their expression of the transcription factor Foxp3, Tregs can inhibit adaptive immune responses through the production of inhibitory cytokines, direct killing of cells, competition with other T cell subsets for antigen or other substrates, and suppression of antigen presentation (Caridade et al., 2013; Savage et al., 2013; Vignali et al., 2008). Tregs are associated with poor prognosis in several cancers, including lung adenocarcinoma, which accounts for 40% of NSCLC (Fridman et al., 2012; Petersen et al., 2006; Shang et al., 2015; Shimizu et al., 2010; Suzuki et al., 2013). In mouse models, Treg depletion can enhance anti-tumor immunity (Bos et al., 2013; Joshi et al., 2015; Marabelle et al., 2013), and antibodies directed against CTLA-4 act in part by depleting Tregs in the tumor microenvironment (Simpson et al., 2013).
While curbing Treg function in tumors is an attractive therapeutic avenue, it is important to specifically target the tumor Treg cell population to avoid systemic and potentially lethal autoimmune reactions. Tregs have considerable phenotypic diversity, which may help inhibit tumor-associated cells in different settings. Functional diversity within tumor Treg populations may impact tumor immune responses, such that effector Tregs promote tumor growth (Green et al., 2017), whereas poorly immunosuppressive Tregs contribute to enhanced anti-tumor immunity (Overacre-Delgoffe et al., 2017; Saito et al., 2016). This functional diversity may be reflected in their transcriptional programs. Specific transcriptional profiles have been associated with Tregs in distinct tissues and inflammatory contexts, which are related to their tissue-resident functions (Arpaia et al., 2015; Burzyn et al., 2013; Cipolletta et al., 2012; Feuerer et al., 2009; Kolodin et al., 2015; Kuswanto et al., 2016). In human tumors, Tregs have a distinct program that may be shared across cancer types, and is associated with clinical outcome (De Simone et al., 2016; Magnuson et al., 2018; Plitas et al., 2016).
Inducible, autochthonous models of cancer are ideal for studying mechanisms of tumor tolerance because they recapitulate the longitudinal development of tumors and the immunosuppressive features of the endogenous tumor microenvironment better than transplanted, more “foreign”, tumors (Dranoff, 2011). Our group has previously developed a model of lung adenocarcinoma in which activation of oncogenic K-rasG12D and loss of Trp53 are driven by intratracheal delivery of a lentivirus expressing Cre recombinase (KP: LSL-KrasG12D, p53fl/fl) (DuPage et al., 2009; Jackson et al., 2005). By using lentivirus that also expresses firefly luciferase fused to chicken ovalbumin (Ova) and the antigenic peptide SIYRYYGL (Lenti-LucOS), we can program tumors to express known T cell antigens that can be used to monitor tumor-specific T cell responses (DuPage et al., 2011). Prior studies using this model have shown that T cell infiltration of Ova-expressing tumors delays tumor growth, but the number and activity of anti-tumor cytotoxic CD8+ T cells (CTLs) decline over time. The development of immune tolerance towards the tumor is partly due to the expansion of lung-resident Tregs that express various markers of effector activity and terminal differentiation (Joshi et al., 2015). Treg depletion results in massive infiltration of CD4+ and CD8+ T cells into the lungs, suggesting that Tregs actively suppress anti-tumor immune responses. Since Treg-depleted animals succumb to systemic autoimmunity, a strategy targeting features of lung tumor-specific Tregs is required to minimize self-directed cytotoxicity.
Here, we map the phenotypic diversity of CD4+ Tconv and Treg cells throughout tumor development in the KP model using scRNA-Seq. While Tconv subsets were stable over time, Treg heterogeneity changed with tumor progression. At early time points, Tregs were less differentiated and expressed genes associated with interferon signaling, while mice with advanced disease had a greater proportion of effector Tregs. Analyzing these data, we identified ST2 as a potential mediator of the accumulation of effector Tregs during tumor development. Indeed, Treg-specific ablation of ST2 increased CD8+ T cell infiltration of tumors and reduced tumor size while avoiding systemic autoimmunity. Our high-resolution characterization of Treg heterogeneity in the tumor microenvironment thus allows us to define refined and effective ways to target Treg function in cancer.
RESULTS
CD103 and KLRG1 mark an activated, heterogeneous population of lung tissue Tregs
We have previously demonstrated that tumor development in the KP model is associated with the expansion of lung-infiltrating Tregs, a large proportion of which express CD103 (integrin aE) and killer cell lectin-like receptor 1 (KLRG1), which have been associated with Treg effector activity and terminal differentiation, respectively (Beyersdorf et al., 2007; Cheng et al., 2012; Huehn et al., 2004; Lehmann et al., 2002; Sather et al., 2007). We characterized the heterogeneity of the Treg population in KP mice with advanced disease. While Treg cells in the draining lymph node (dLN) were predominantly CD103+KLRG1−(double-negative, DN) or CD103+KLRG1−(single-positive, SP), nearly 40% of lung Tregs from late-stage, tumor-bearing KP mice were CD103+KLRG1+ (double-positive, DP) (Figure 1A). DP Tregs in late-stage, tumor-bearing mice had increased expression of genes associated with enhanced Treg cell activity, including GITR, CD39, and PD-1, compared to SP and DN Tregs (Joshi et al., 2015). We therefore hypothesized that these Treg subsets may have distinct tissue and tumor-specific transcriptional programs.
To identify such a program, we bred KP mice to Foxp3 reporter mice to facilitate isolation and manipulation of Tregs from tumor-bearing mice. Using a previously-described method (Anderson et al., 2012), mice were injected with antibody prior to sacrifice to label intravascular cells and distinguish tissue-infiltrating populations. We profiled DP, SP, and DN Tregs isolated from the lungs of tumor-bearing KP-Foxp3RFP mice at 20 weeks post infection (p.i.) with Lenti-LucOS by bulk RNA-Seq (Figure 1A, Methods). We also profiled SP and DN Tregs from matching mediastinal lymph nodes (msLNs) and DN Tregs from the spleen of one tumor-bearing mouse for comparison.
The most significant distinction in the data by Independent Component Analysis (ICA) was between lung-infiltrating and peripheral Tregs (Figure S1A). A 284 gene signature strongly distinguished lung-infiltrating Tregs (“KPLungTR signature genes”, Methods, Figure 1B, Table S1), which we confirmed by quantitative RT-PCR (qPCR) of Pparg1, Nr4a1, Areg, and Gata1expression (Figure S1B). This KPLungTR signature was enriched for signatures of other tissue Tregs, including Tregs in visceral adipose tissue (VAT), colonic lamina propria, and wounded muscle (Figure S1C, Table S2). Genes upregulated in the KPLungTR signature also included activation, differentiation, and growth factor signaling genes (Figure S1D), consistent with prior reports that Tregs promote tissue repair (Arpaia et al., 2015; Burzyn et al., 2013). Notably, the signature was enriched for orthologs of genes induced in human colorectal cancer (CRC) and NSCLC-associated Tregs (De Simone et al., 2016) (Figure S1E), suggesting that lung Tregs in human cancer and the KP model have a common “tissue Treg” phenotype.
Several lines of evidence further suggest that the DP population is activated. First, genes upregulated and downregulated transiently in activated Tregs were differentially expressed in DP vs. DN Tregs (Figure S1F), which may reflect antigen exposure of this Treg population in the tumor microenvironment (van der Veeken et al., 2016). Second, genes upregulated in DP Tregs vs. all other Tregs in tumor-bearing lungs (Methods, Figure 1C) were associated with T cell activation and putative Treg effector functions (e.g., Nr4a1, Cd69, Il1rl1, Areg, Srgn, and Fgl2). Notably, Cxcr3, which has been associated with a T-bet+ Treg phenotype specialized to counter Th1 inflammation (Koch et al., 2009; Levine et al., 2017), was downregulated in DP Tregs vs. SP and DN Tregs (Figure 1C). The DP Treg phenotype may thus represent an effector cell state different from Cxcr3+T-bet+ Tregs.
While the DP subset of lung Tregs may be particularly active and an attractive target for immunotherapy, PD-1 and CD69 expression across DN, SP, and DP Tregs revealed considerable heterogeneity within each subset (Figure 1D). In particular, 52% of DP Tregs expressed PD-1 and 68% expressed CD69. We thus turned to more fully characterize the variation within Tregs in the tumor microenvironment.
scRNA-Seq reveals heterogeneity within tumor-associated CD4+ Tconv cells
We sought to characterize patterns of heterogeneity in tumor-associated CD4+ T cells over time to contextualize the diversity of Treg responses in relation to their Foxp3-CD4+ T cell (Tconv) counterparts. By scRNA-seq we profiled 1,254 Tconv and 1,679 Tregs sorted from the lungs and msLN of non-tumor bearing KP-Foxp3GFP mice and tumor-bearing mice at weeks 5, 8, 12, and 20 after tumor induction with Lenti-LucOS (Figure 2A,~4 mice per timepoint).
The tissue-specific expression program partitioned into genes shared by lung infiltrating Tconv and Tregs, and genes uniquely upregulated in each (Figure 2B, Table S3). For example, lung-infiltrating Tregs expressed high levels of Il1rl1, Cxcr4, Areg, and Klrg1, while Tconv cells expressed Cd44, Ccr4 and Itgb1 (Figure 2B). Genes from the KPLungTR signature and from a recently described trajectory of tissue-resident Tregs (Miragaia et al. 2017) were both differentially expressed in the scRNA-seq profiles (Figure S2A).
Both the lung and msLN cells spanned a phenotypic continuum, with the lung cells showing particular diversity (Figure 2C, S2B, DC1 p < 10−13; DC2 p < 10−16, Levene’s test). The spectrum of cell states was apparent when scoring for the expression of lung Tconv or Treg signatures, and when cells were arranged along diffusion components that describe their tissue-specific expression program (Figure 2C). Both Tregs and Tconv in the msLN expressed genes associated with a naive or central memory phenotype, including Lef1, Sell, and Ccr7 (Figure 2B, S2C). Conversely, cells were more activated in the lung (Figure 2B). Subsets of lung Tconv and Treg cells that scored highly for the msLN signature also expressed genes associated with TCR signaling, including Nr4a1 and Junb, suggesting that they may be recently activated (Figure 2C, S2C). Lung-infiltrating Tconv and Treg cells that scored highly for the respective lung signature may represent cells that were more tissue-adapted or localized to a particular region of the lung.
Lung Tconv subsets remain in stable proportions throughout tumor development
Lung Tconv subsets expressed programs associated with different CD4+ T cell subsets, including naïve T, Th17, Th1, Th9 and NKT17 cells (Methods, Figure 2D-E), whose proportions remained largely stable over time. Within Th1 cells, a subset expressed Eomes and Gzmk, which may reflect cytolytic function, and Cxcr3 and Ccr5, which promote antigen-specific CD4+ T cell recruitment to lungs during respiratory virus infection (Kohlmeier et al., 2009) (Figure S2D). Some of the Th17-like cells expressed Zbtb16, a marker for NKT cells, and also scored highly for a gene module that includes genes associated with natural killer T17 (NKT17) cells, such as Blk and Gpr114 (Figure S2E) (Engel et al., 2016). Furthermore, these cells had lower expression of CD4 than other Tconv (Figure S2F) and did not express TCR chains associated with γδ T cells (Table S5). We found little evidence of Th2-like cells, despite their role in lung inflammation in other settings (Walker and McKenzie, 2018), but did observe a small population of Th9-like cells expressing Il9r, Il4, and Il1rl1, which have been implicated in driving anti-tumor immune responses (Végran et al., 2015) (Figure S2G). Finally, we identified a population that scored highly for both the Th1 and the Th17 modules. We validated the presence of cells expressing both RORγt and T-bet (Figure 2F); such cells have been described as a plastic, Th17-derived population in other pathogenic states (Lee et al., 2012, 2009; Wang et al., 2014). The overall expression of the gene modules associated with these Tconv subsets showed subtle variation over time by scRNA-Seq (Figure S2H-I), but the relative cell proportions measured by flow cytometry remained stable during tumor development (Figure 2E-F).
A RORγt+ Treg population is present throughout tumor development and may have shared clonal origin with Th17 Tconv cells
Lung-infiltrating Tregs expressed several gene modules with similar features to those in transcriptional signatures of previously-described Treg subsets (Figure S3A). For example, Module 18 includes genes that characterize a resting, or central, Treg (rTreg) phenotype, such as Sell, Ccr7, and Tcf7 (Campbell, 2015; Li and Rudensky, 2016), whereas Module 13 identified a Treg population expressing Rorc and Il17a (Figure 3A, S3A), reminiscent of Th17-like Tregs (Tr17), a subset with immunosuppressive activity directed at Th17 responses (Kim et al., 2017). We validated this population by flow cytometry and found that RORγt+ Tregs comprise roughly 10% of lung-infiltrating Tregs throughout tumor progression (Figure 3B). The Tr17-like cells represented a distinct state among lung Tregs and the expression of Tr17-associated genes was inversely correlated with the expression of genes previously identified in lung-resident Tregs, including KLRG1 (Figure 3C-D). Additionally, whereas Ccr6 expression within the Tconv was restricted to Th17 cells (Figure 2E), Ccr6 was expressed in multiple Treg subsets (Figure S3B), consistent with previous findings (Yamazaki et al., 2008), which may result in the localization of different Treg subsets to common sites in the lung.
Remarkably, shared clonotypes between Treg and Tconv cells were predominantly Tr17-like and Th17-like cells, respectively. Specifically, based on paired-chain T cell receptor (TCR) sequences of profiled cells (Figure S3B, Methods, Table S5), 12 TCR clonotypes were shared across Treg and Tconv cells. Indeed, dedicated TCR profiling of Tregs and Tconv from KP mice with advanced disease showed that ~5% of Treg clones were shared with Tconv on average in advanced disease (Figure S3C). Of the 19 Tregs and 20 Tconv cells belonging to the 12 TCR clonotypes shared between Tconv and Treg, the Treg cells were predominantly of the Tr17-like phenotype (13 of 19 Tregs had a z-score > 1.5 in the Tr17-like Module, hypergeometric p-value < 10−5, Figure 3F, S3D). The Tconv cells were also predominantly of the Th17 phenotype, although this was not a significant enrichment . 67 out of 178 identified Tconv clones were of the Th17 phenotype (hypergeometric p = 0.68), of which 8 were clonotypes shared with Tregs, (Figure 3F). Thus, Tr17 differentiation may reflect a shared clonal origin with Th17 cells.
An effector-like Treg phenotype becomes predominant during tumor development
In contrast to Tr17-like cells, where a program was expressed by a fixed proportion of cells during tumor development, other Treg programs changed in prominence throughout tumor development (Figure 4A). For example, there was decreased expression of Modules 1, 3, 8, and 9, which mark cycling cells, after 8 weeks (Figure 4A), corresponding to a decline in Ki67 expression on Tregs (Figure 4B). Two other programs also changed over time, reflecting an interferon response and a T effector program (Figure 4A).
The interferon program (“IFNstim_TR”) was characterized by the expression of Modules 6 and 23 (Figure 4C), which included many interferon-stimulated genes (ISGs) downstream of either type I or II interferon (IFN) signaling, including Stat1, guanylate binding protein genes (GBPs), type I interferon-specific genes (e.g., oligoadenylate synthetase family members), and IFNγ-specific genes (e.g., Irf1, Irf9) (Der et al., 1998). 28 genes from the IFNstim_TR program were significantly downregulated by Tregs during tumor progression (Figure S4B). IFNγ promotes a Tbet+CXCR3+ Th1-like Treg cell population that can suppress Th1 responses (Hall et al., 2012; Koch et al., 2009, 2012). Neither Cxcr3 nor Tbx21 are IFNstim_TR genes, but IFNstim_TR expression was correlated with Tbx21 expression (Figure S4C). Moreover, the program was enriched for genes expressed by lymphoid tissue Tregs and genes downregulated in DP Tregs (Figure S4D), which include Cxcr3. IFNstim_TR expression may thus reflect recent arrival to the lung, consistent with its presence early in tumor development.
The T effector program (“Eff_TR”) was characterized by the expression of Modules 12 and 21 (Figure 4C), which were enriched for genes in the DP signature (p-value ≤ 10−25, Figure S4E) and genes upregulated in Tregs from mouse non-lymphoid tissues and human breast cancer, NSCLC, and CRC (De Simone et al., 2016; Guo et al., 2018; Magnuson et al., 2018; Miragaia et al., 2017; Plitas et al., 2016; Zheng et al., 2017) (Figure S4D), confirming the distinct expression profile we had previously identified in the DP Treg subpopulation.
The interferon and effector programs represented independent phenotypes of Tregs within each timepoint but followed opposite patterns over time: expression of IFNstim_TR genes was highest in cells from week 5 and declined thereafter, while expression of Eff_TR genes increased and remained elevated (Figure 4A, D). This temporal transition was also highlighted when testing for individual temporally varying genes: Cxcr3 expression decreased with time, and Pdcd1 and Lilrb4 (Module 21) increased in expression during tumor development (Figure S4F), consistent with down-regulation of Cxcr3 in DP Treg cells (Figure 1D). More generally, Eff_TR genes were upregulated in DP Tregs compared to DN Tregs in mice with late-stage tumor burden, whereas IFNstim_TR genes were significantly downregulated (Figure S4G). We confirmed that protein levels of Cxcr3 decreased, and proteins encoded by Eff_TR genes, including CD85k, CD69, CXCR6, PD-1 and ST2, increased during tumor progression (Figure 4E).
Taken together, our data suggest that tumor progression may be associated with a shift from a Treg cell phenotype specialized for responding to Th1 inflammation to an effector Treg cell population. In particular, we hypothesized that the strong immunosuppression associated with the late-stage tumor environment may be a result of the emergence and stabilization of cells with the Eff_TR phenotype.
ST2 is upregulated on effector Tregs in mice bearing advanced lung tumors
We reasoned that Il1rl1, an Eff_TR gene that encodes the interleukin 33 (IL-33) receptor ST2, may highlight a pathway that could be targeted to alter longitudinal changes in Treg cell phenotype and prevent the accumulation of effector Tregs in advanced tumors. First, Il1rl1/ ST2 levels tracked with the effector Treg phenotype; Il1rl1 is a member of Module 21 in the Eff_TR program and ST2 was most highly expressed in DP lung Tregs (Figure 5A), and its expression in Tregs increased during tumor development (Figure 4D). Moreover, ST2 was expressed by ~40% of lung Tregs vs. ~10% of Tregs in the msLN, and <5% of Tconv cells in the lung in late-stage tumor-bearing mice (Figure 5B). Second, Treg cells from tumor-bearing KP, LucOS-infected mice expressed both the membrane-bound and soluble isoforms of ST2 (Figure 5C); soluble ST2 (sST2) is thought to diminish ST2 signaling through sequestration of IL-33, the only known ligand of ST2 and an alarmin that recruits immune cells to sites of tissue damage (Cayrol and Girard, 2014). Finally, IL-33 was highly expressed in normal lung, and in early and late lung adenocarcinomas in the KP model (Figure 5D). In normal lung, IL-33 was predominantly expressed on surfactant protein C (SPC)-expressing type II epithelial cells (Figure S5). We thus hypothesized that ST2 may be a critical mediator of Treg cell function in the lung tumor environment.
Recombinant IL-33 treatment increases effector Tregs in tumor-bearing lungs
To determine the effect of IL-33 on the immune microenvironment of tumors, we administered recombinant mouse IL-33 (rIL-33) intratracheally to tumor-bearing KP, Lenti-LucOS-infected mice (Figure 6A). Consistent with prior reports (Kondo et al., 2008; Schmitz et al., 2005), rIL-33 induced significant inflammatory infiltration and epithelial thickening in tumors and throughout the lung (Figure 6B). rIL-33-treated mice had greater numbers of eosinophils (Figure 6C) and CD4+ and CD8+ T cells per lung weight (Figure 6D), although the proportion of tumor-specific, SIINFEKL tetramer-positive cells among CD8+ T cells was unchanged (Figure 6E). We observed similar inflammation in non-tumor bearing wild-type mice treated with rIL-33 (data not shown). CD4+ T cells in rIL-33-treated mice had an increased proportion of Tregs (Figure 6F), of which 64% were DP compared to 34% in PBS-treated controls, with proportionally fewer SP and DN Tregs (Figure 6G). rIL-33 treatment of ST2-deficient mice failed to elicit the same change in the proportion of Tregs, which was similar to that of untreated, wild-type mice (Figure S6). Taken together, rIL-33 administration is sufficient to drive both a major increase in the lung Treg population in general, and to promote an increase in effector Tregs cells in particular.
Treg-specific ST2 is required for the increase in effector Tregs during tumor progression
To test whether ST2 signaling on Tregs was necessary for the development of a robust effector Treg cell response in tumors, we studied the effects of Treg-specific Il1rl1 deletion. We used a modified version of the KP model wherein FlpO recombinase drives expression of oncogenic K-ras and loss of p53 (KPfrt: FSF-KrasG12D, p53frt/frt), which allowed us to use the Cre-lox system to study Treg-specific Il1rl1 deletion. We crossed KPfrt mice to Foxp3YFP-Cre and Il1rl1fl/fl mice to model lung adenocarcinoma development in the setting of Treg-specific ST2 deficiency (Figure 7A). We infected the mice with a lentivirus expressing FlpO recombinase and GFP fused to Ova and SIYRGYYL (FlpO-GFP-OS) in order to induce tumors that would express the same strong T cell antigens as those in the Lenti-LucOS model.
Early-stage KPfrt, Foxp3YFP-Cre, Il1rl1fl/fl mice did not differ from KPfrt, Foxp3YFP-Cre mice in the fraction of CD4+ T cells that were Tconv or Treg cells, but late in tumor progression there was a slight reduction in the proportion of Treg cells (Figure 7B and S7A), a significantly lower proportion of DP Tregs, and a higher proportion of SP cells (Figure 7C). Expression profiles of DP, SP, and DN Tregs from KPfrt, Foxp3YFP-Cre, Il1rl1fl/fl and KPfrt, Foxp3YFP-Cre control mice identified an expression signature lower in ST2-deficient vs. wild-type Tregs, where it was highest among wild-type DP Tregs (Figure 7D, S7D). The signature was enriched for KPLungTR and DP signature genes, including Dgat2, Furin and Nfkbia, as well as for genes upregulated by Tregs in human NSCLC (Figure 7E, S7B, C). ST2-deficient Tregs also showed higher expression of some genes, including Itgb1, Il10, Klf6, and Fos (Figure 7E), suggesting that they may adopt alternative phenotypes. Taken together, our data supports the hypothesis that ST2 regulates the accumulation of effector Tregs in the tumor microenvironment over time by promoting the expression of DP signature genes.
Treg-specific ST2 ablation leads to increased CD8+ T cell infiltration and a reduction in tumor burden
Finally, we found that tumors from KPfrt, Foxp3YFP-Cre, Il1rl1fl/fl mice had over 50% higher CD8+ T cell infiltration than tumors from control mice by immunohistochemistry (Figure 7E). KPfrt, Foxp3YFP-Cre, Il1rl1fl/fl mice also had a significantly lower total tumor burden and lower average tumor size compared to control mice (Figure 7F,G), suggesting that greater CD8+ T cell infiltration of tumors may result in better inhibition of tumor growth. Overall, our studies suggest that Treg-specific inhibition of ST2 signaling may result in a less immunosuppressive tumor microenvironment characterized by increased anti-tumor CD8 T cell activity and lower tumor burden.
DISCUSSION
To identify specific features of Tregs in the tumor microenvironment that can be targeted therapeutically without adversely affecting Tregs in other tissues, we profiled Tconv and Tregs longitudinally in a mouse model of lung adenocarcinoma by scRNA-Seq. We show that Treg diversity undergoes temporal shifts that would be missed in analyses of bulk populations or at a single timepoint. Leveraging these dynamic changes, we identified IL-33 as a critical mediator of effector Treg function in tumors. Although previous scRNAseq studies have defined signatures of Treg exhaustion and activation in human cancer (Guo et al., 2018; Zheng et al., 2017), our study is the first to effectively impair tumor growth by characterizing and perturbing a major pathway responsible for the development of transcriptionally-distinct subsets of Tregs in tumor-bearing lungs.
IL-33 has been shown to promote tumorigenesis through the recruitment of Tregs and other cells in transplant and xenograft models of breast and lung cancer (Jovanovic et al., 2014; Wang et al., 2016, 2017), and mice with Treg-specific ST2 deficiency have impaired growth of a transplantable tumor model (Magnuson et al., 2018). Here, we show in a genetically-defined, autochthonous mouse model of lung adenocarcinoma that loss of Treg-specific ST2 function is sufficient to impair tumor development without provoking systemic autoimmunity. Several therapeutic antibodies directed against ST2 and IL-33 are in preclinical development for the treatment of allergy and asthma. Our data point to the potential value of disrupting ST2 signaling in cancer.
Although ST2-deficient Tregs have been reported to be equally immunosuppressive as their wild-type counterparts in vitro (Schiering et al., 2014), our results suggest that in vitro suppression assays may fail to capture the full spectrum of Treg effector activity in vivo. We observed a slight reduction in lung Treg cell numbers as a result of Treg-specific ST2 deficiency, which may be related to reports that IL-33 can stimulate TCR-independent expansion of Tregs (Arpaia et al., 2015; Kolodin et al., 2015). DP Tregs from KPfrt, Foxp3YFP-Cre, Il1rl1fl/fl mice had lower expression of Eff_TR genes compared to wild-type DP Tregs, suggesting that ST2 may promote the maintenance of the effector Treg phenotype. Indeed, IL-33 has been shown to increase expression of Foxp3 and GATA-3 (Kolodin et al., 2015; Vasanthakumar et al., 2015), transcription factors integral for Treg terminal differentiation. Taken together, ST2-deficient Tregs may adopt an alternate functional state due to loss of IL-33 signaling.
Tregs across multiple tumor types likely have a common transcriptional program that is closely related to that of healthy tissue Tregs (Magnuson et al., 2018). Indeed, the effector Tregs in the KP model express a program similar to that of Tregs in several human cancers (De Simone et al., 2016; Guo et al., 2018), including a TNFRSF9+ Treg population in human NSCLC (Zheng et al., 2017). This similarity may be due to the fact that clinically-detectable tumors are likely to have convergent strategies for evading immune destruction by recruiting highly suppressive Tregs. Tumor-bearing lungs from KP mice also harbor Th1-like CXCR3+ Tregs, which express IFN-stimulated genes and peak early in tumor development, following an opposite temporal pattern from the Eff_TR program. CXCR3 directs Tregs to sites of Th1 inflammation (Koch et al., 2009), which may explain the prominence of the IFNstim_TR program during early tumorigenesis, when CD8+ T cell infiltration of tumors and IFN signaling are most robust (DuPage et al., 2011). Cxcr3 may mark recently-arrived Tregs that have distinct functions from effector Tregs, and temporal shifts in IFNstim_TR and Eff_TR gene expression may reflect Treg adaptation to the tumor microenvironment over time. Alternatively, the decline in Cxcr3+ Tregs during tumor development may reflect cellular turnover and/or the outgrowth of an alternate subset of Tregs due to reduced IFN, and availability of IL-33 ligand. Several reports have described an IFN signature or a distinct population of CXCR3+ Tregs in human tumors, although their functional significance is not well-defined (Halim et al., 2017; Johdi et al., 2017; Redjimi et al., 2012).
Longitudinal profiling in the KP model provides a window into the natural history of effector Treg activity that is challenging to achieve using patient samples. While Tr17-like, CXCR3+, and effector Treg populations have been described previously in human tumors, we have shown that these states exist simultaneously, and their relative proportions vary with tumor development. Future studies may help elucidate the contribution of each distinct Treg subset to tumor immune responses. While Treg transcriptional heterogeneity may pose a challenge for efforts to target tumor Treg activity, we show that loss of Treg-specific ST2 signaling can alter Treg composition and ultimately impact tumor growth. Our study provides proof of concept that pathways that control Treg diversity, maturation, and function may be useful targets for future therapies.
EXPERIMENTAL METHODS
Mice
KP, KPfrt, Foxp3GFP, Foxp3RFP, Foxp3GFP/DTR, Il1rl1−/− and Il1rl1fl/fl mice have been previously described (Bettelli et al., 2006; Chen et al., 2015; DuPage et al., 2011; Kim et al., 2007; Townsend et al., 2000; Wan and Flavell, 2005; Young et al., 2011). Both male and female mice were used for all experiments, and mice were gender and age-matched within experiments. Experimental and control mice were co-housed whenever appropriate. All studies were performed under an animal protocol approved by the Massachusetts Institute of Technology (MIT) Committee on Animal Care. Mice were assessed for morbidity according to MIT Division of Comparative Medicine guidelines and humanely sacrificed prior to natural expiration.
For in vivo labelling of circulating immune cells, anti-CD4-PE (eBioscience, RM4-4, 1:400) and anti-CD8β-PE (eBioscience, 1:400) were diluted in PBS and administered by IV injection 5 minutes before harvest (Anderson et al., 2012). Alternatively, anti-CD45-PE-CF594 (30-F11, BD Biosciences, 1:200) was also used for intravascular labeling and was administered 2 minutes before sacrifice.
For rIL-33 treatment studies, 200ng of recombinant mouse IL-33 (BioLegend) was diluted in 50 mL of PBS and administered intratracheally to mice as described previously (Li et al., 2014). Control mice received PBS only.
Lentiviral production and tumor induction
The lentiviral backbone Lenti-LucOS has been described previously (DuPage et al., 2011). Lentiviral plasmids and packaging vectors were prepared using endo-free maxiprep kits (Qiagen). The pGK::GFP-LucOS::EFS::FlpO lentiviral plasmid was cloned using Gibson assembly (Akama-Garren et al., 2016; Gibson et al., 2009). Briefly, GFP-OS was created as a protein fusion of GFP and ovalbumin257-383, which includes the SIINFEKL and AAHAEINEA epitopes, and SIYRYYGL antigen. Lentiviral plasmids and packaging vectors were prepared using endo-free maxiprep kits (Qiagen). Lentiviruses were produced by co-transfection of 293FS* cells with Lenti-LucOS or FlpO-GFP-OS, psPAX2 (gag/pol), and VSV-G vectors at a 4:2:1 ratio with Mirus TransIT LT1 (Mirus Bio, LLC). Virus-containing supernatant was collected 48 and 72h after transfection and filtered through 0.45mm filters before concentration by ultracentrifugation (25,000 RPM for 2 hours with low decel). Virus was then resuspended in 1:1 Opti-MEM (Gibco) - HBSS. Aliquots of virus were stored at −80°C and titered using the GreenGo 3TZ cell line (Sánchez-Rivera et al., 2014).
For tumor induction, mice between 8-15 weeks of age received 2.5 × 104 PFU of Lenti-LucOS or 4.5 × 104 PFU of FlpO-GFP-OS intratracheally as described previously (DuPage et al., 2009).
Tissue isolation and preparation of single cell suspensions
After sacrifice, lungs were placed in 2.5mL collagenase/DNAse buffer (Joshi et al., 2015) in gentleMACS C tubes (Miltenyi) and processed using program m_impTumor_01.01. Lungs were then incubated at 37°C for 30 minutes with gentle agitation. The tissue suspension was filtered through a 100 μm cell strainer and centrifuged at 1700 RPM for 10 minutes. Red blood cell lysis was performed by incubation with ACK Lysis Buffer (Life Technologies) for 3 minutes. Samples were filtered and centrifuged again, followed by resuspension in RPMI 1640 (VWR) supplemented with 1% heat-inactivated FBS and 1X penicillin-streptomycin (Gibco), and 1X L-glutamine (Gibco).
Spleens and lymph nodes were dissociated using the frosted ends of microscope slides into RPMI 1640 supplemented with 1% heat-inactivated FBS and 1X penicillin-streptomycin (Gibco), and 1X L-glutamine (Gibco). Spleen cell suspensions were spun down at 1500 RPM for 5 minutes, and red blood cell lysis with ACK Lysis Buffer was performed for 5 minutes. Cells were filtered through 40 μm nylon mesh and, after centrifugation, resuspended in supplemented RPMI 1640. Lymph node suspensions were filtered through a 40 μm nylon mesh, spun down at 1500 RPM for 5 minutes, and resuspended in supplemented RPMI 1640.
For ex vivo T cell stimulation experiments to detect intracellular cytokines, 0.5 × 105 cells were plated in a 96-well U-bottom plate (BD Biosciences) in RPMI 1640 (VWR) supplemented with 10% heat-inactivated FBS, 1X penicillin-streptomycin (Gibco), 1X L-glutamine (Gibco), 1X HEPES (Gibco), 1X GlutaMAX (Gibco), 1mM sodium pyruvate (Thermo Fisher), 1X MEM non-essential amino acids (Sigma), 50μM beta-mercaptoethanol (Gibco), 1X Cell Stimulation Cocktail (eBioscience), 1X monensin (BioLegend), and 1X brefeldin A (BioLegend). Cells were incubated in a tissue culture incubator at 37°C with 5% CO2 for 4 hours.
Staining for flow cytometric analysis
Approximately 0.5-1 × 106 cells were stained for 15-30 minutes at 4°C in 96-well U-bottom plates (BD Biosciences) with directly conjugated antibodies (Table S8). SIINFEKL-Kb tetramer was prepared using streptavidin-APC (Prozyme) and SIINFEKL-Kb monomer from the NIH Tetramer Core.
After staining, cells were fixed with Cytofix/ Cytoperm Buffer (BD). Samples that were destined for Foxp3 or other transcription factor staining were fixed with the Foxp3 Transcription Factor Staining Buffer Kit (eBioscience). Intracellular cytokine and transcription factor staining were performed right before analysis using either the BD Perm/Wash Buffer (BD) or the Foxp3 Transcription Factor Staining Buffer Kit (eBioscience); staining was performed for 45 minutes at 4°C. Analysis was performed on an LSR II (BD) with 405, 488, 561, and 635 lasers. Data analysis was performed using FlowJo software.
Isolation of Treg populations for bulk RNA-Seq
For sequencing of LucOS-infected, KP, Foxp3-RFP mice: 100-200 DP, SP, and DN Treg cells were sorted into Buffer TCL (Qiagen) plus 1% b-mercaptoethanol using a MoFlo Astrios cell sorter. cDNA was prepared by the SMART-Seq2 protocol (Picelli et al., 2013) with the following modifications: RNA was purified using 2.2X RNAclean SPRI beads (Beckman Coulter) without final elution, after which beads were air-dried and immediately resuspended with water and oligoDT for annealing, and 18 cycles of preamplification were used for cDNA. cDNA was then mechanically sheared and prepared into sequencing libraries using the Thru-Plex-FD Kit (Rubicon Genomics). Sequencing was performed on an Illumina HiSeq 2000 instrument to obtain 50 nt paired-end reads.
For comparison of wild-type and ST2-deficient Tregs and CD8+ T cells: 100-200 DP, SP, and DN Tregs or SIINFEKL-tetramer-positive and negative CD8+ T cells were sorted and cDNA was prepared with 14 cycles of preamplification. Nextera library preparation was performed as previously described (Picelli et al., 2013) and sequencing was performed with 50 × 25 paired end reads using two kits on the NextSeq500 5 instrument.
Single-cell sorting of Tconv and Treg populations for RNA sequencing
Tconv (DAPIneg, i.v. neg, Thy1.2+CD4+Foxp3-GFPneg) and Treg (DAPIneg, i.v. neg, Thy1.2+CD4+Foxp3-GFP-positive) cells were single-cell sorted into Buffer TCL (Qiagen) plus 1% B-mercaptoethanol in 96-well plates using a MoFlo Astrios cell sorter. Each plate had 30-100 cell population well and an empty well as controls. Following sorting, plates were spun down for 1” at 2000 RPM and frozen immediately at −80C.
Preparation of scRNAseq libraries
Plates were thawed and RNA was purified using 2.2X RNAclean SPRI beads (Beckman Coulter) without final elution (Shalek et al., 2013). SMART-seq2 and Nextera library preparation was performed as previously described (Picelli et al., 2013), with some modifications as described in a previous study (Singer et al., 2017). Plates were pooled into 384 single-cell libraries, and sequenced 50 × 25 paired end reads using a single kit on the NextSeq500 5 instrument.
Quantitative PCR for validation of RNA-Seq experiments
Quantitative PCR was performed using various primer sets (Table S5). 1ng of cDNA generated using SMART-Seq2 was included in a reaction with 1μL of each primer (2μM stock) and 5μL of KAPA SYBR Fast LightCycler 480 (KAPA Biosystems). Cp values were measured using a LightCycler 480 Real-Time PCR System (Roche). Relative fold-change in expression values were calculated using the following formula: 2(ΔCp(Sample) - ΔCp(Spleen)), where (ΔCp(Sample) = Sample CpGene of Interest - Sample CpGAPDH, and ΔCp(Spleen) = Spleen CpGene of Interest - Spleen CpGAPDH.
Population-level TCR Beta chain sequencing and analysis
For bulk TCR beta chain sequencing, T cells were sorted directly into 250μl RNAprotect buffer (Qiagen), spun down for 1 minute at 2000 RPM, and immediately frozen at −80°C. Samples were sent to iRepertoire (Huntsville, AL) for library preparation and sequencing. TCR sequences were analyzed and compared with VDJtools software (Shugay et al., 2015).
Immunohistochemistry (IHC) and immunofluorescence staining
Lung lobes and spleens allocated for IHC and IF were perfused with 4% paraformaldehyde in PBS and fixed overnight at 4°C. Lung lobes and/ or spleen were transferred to histology cassettes and stored in 70% ethanol until paraffin embedding and sectioning (KI Histology Facility). H&E stains were performed by the core facility using standard methods.
For IHC, 5 μm unstained slides were dewaxed, boiled in citrate buffer (1 g NaOH, 2.1 g citric acid in 1L H2O, pH 6), for 5 minutes at 125°C in a decloaking chamber (Biocare Medical), washed with 3X with 0.1% Tween-20 (Sigma) in TBS, and blocked and stained in Sequenza slide racks (Thermo Fisher). Slides were blocked with Dual Endogenous Peroxidase and Alkaline Phosphatase Block (Dako) and then with 2.5% Horse Serum (Vector Labs). Slides were incubated in primary antibody overnight, following by washing and incubation in HRP-polymer-conjugated secondary antibodies (ImmPRESS HRP mouse-adsorbed anti-rat and anti-goat, Vector Laboratories). Slides were developed with ImmPACT DAB (Vector Laboratories). Primary antibodies used were goat anti-IL-33 (R&D, AF3626) and rat anti-CD8a (Thermo Fisher, 4SM16). Stains were counterstained with hematoxylin using standard methods before dehydrating and mounting.
After fixation, lung lobes and spleen allocated for IF were perfused with 30% sucrose in PBS for cryoprotection for 6-8h at 4°C. Tissues were then perfused with 30% optimum cutting temperature (O.C.T.) compound (Tissue-Tek) in PBS and frozen in 100% O.C.T in cryomolds on dry ice. 6μm sections were cut using a CryoStar NX70 cryostat (Thermo), and air-dried for 60-90 minutes at room temperature. Sections were incubated in ice-cold acetone (Sigma) for 10 minutes at −20°C and then washed 3 x 5 minutes with PBS. Samples were permeabilized with 0.1% Triton-X-100 (Sigma) in PBS followed by blocking with 0.5% PNB in PBS (Perkin Elmer). Primary antibodies were incubated overnight. Primary antibodies used were rabbit anti-prosurfactant protein C (SPC) (Millipore, AB3786, 1:500) and goat anti-IL-33 (R&D, AF3626, 1:200). After washing 3 x 5 minutes, samples were incubated in species-specific secondary antibodies conjugated to Alexa Fluor 568 and Alexa Fluor 488, respectively, at 1:500. Sections were then fixed in 1% PFA and mounted using Vectashield mounting media with DAPI (Vector Laboratories).
Immunohistochemistry and immunofluorescence tissue section images were acquired using a Nikon 80 Eclipse 80i fluorescence microscope using 10x and 20x objectives and an attached Andor camera. Stained IHC slides were scanned using the Aperio ScanScope AT2 at 20X magnification.
COMPUTATIONAL ANALYSIS
Bulk RNA-seq data processing and signature analyses
Bulk RNA-Seq reads that passed quality metrics were mapped to the annotated UCSC mm9 mouse genome build (http://genome.ucsc.edu/) using RSEM (v1.2.12) (http://deweylab.github.io/RSEM/) (Li and Dewey, 2011) using RSEM’s default Bowtie (v1.0.1) alignment program (Langmead et al., 2009). Expected read counts estimated from RSEM were upper-quartile normalized to a count of 1000 per sample(Bullard et al., 2010). Genes with normalized counts less than an upper-quartile threshold of 20 across all samples were considered lowly expressed and excluded from further analyses. The dataset was log2 transformed before subsequent analysis.
Unsupervised clustering of samples was performed using a Pearson correlation-based pairwise distance measure.
Signature analyses between bulk Treg cell populations were performed using a blind source separation methodology based on ICA (Hyvärinen and Oja, 2000), using the R implementation of the core JADE algorithm (Joint Approximate Diagonalization of Eigenmatrices)(Biton et al., 2014; Nordhausen et al., 2014; Rutledge and Jouan-Rimbaud Bouveresse, 2013) along with custom R utilities. Multi-sample signatures were visualized using relative signature profile boxplots (Li et al., 2018). Signature correlation scores (z-scores) for each gene are included in Tables S1 and S7. Heat maps were generated using the Heatplus package in R.
Gene Set Enrichment Analysis (GSEA)
Gene set enrichment analyses were carried out using the pre-ranked mode in GSEA with standardized signature correlation scores for the KPLungTR signature and default settings using gene-sets from MsigDB v5.1 (Subramanian et al., 2005) and a custom immunologic signatures library of gene sets (Table S2) added to version 4.0 of the MSigDB immunologic collection (c7). Normalized Enrichment Score (NES), p-values and FDR for the custom gene-sets were calculated in the context of the combined c7 v4.0 MSigDB collection.
Network representations of GSEA results were generated using EnrichmentMap (http://www.baderlab.org/Software/EnrichmentMap) for Cytoscape v3.3.0 (http://www.cytoscape.org).
Identification of DP signature
To identify a signature separating CD103+KLRG1+ lung Tregs from other populations we applied ICA to the data prior to log transformation, which allowed us to detect signatures with lower amplitudes of gene expression changes. We detected a signature separating CD103+KLRG1+ lung Tregs from other populations. Genes in this signature with |z-score| > 3 were selected for downstream analysis (75 up-regulated and 31 down-regulated genes). An additional expression level filter was implemented to narrow the list of genes of interest. For upregulated genes, expression in all CD103+KLRG1+ lung Treg samples had to be greater than all but a maximum of 3 other samples (3 out of a total 8 other samples). A similar filtering scheme was employed in the other direction for down-regulated genes. This yielded a total of 43 up-regulated and 2 down-regulated genes in CD103+KLRG1+ lung Tregs (Table S1). This set of genes was used to illustrate gene expression level changes in a heatmap (Figure 1D).
Filtering of genes differentially-expressed in ST2-deficient Tregs
A signature distinguishing ST2-deficient Tregs from wild-type Tregs was identified through ICA (Table S7). To identify particular genes of interest, signature genes (|z-score| > 3) were filtered to include only genes that had an absolute fold change exceeding 1.5x within any of the CD103+KLRG1+ (DP), CD103+KLRG1−(SP), CD103−KLRG1−(DN) sample types between wild-type and ST2-deficient Tregs. These gene lists were then filtered to retain only those genes that appeared in at least two of the three sample types (i.e. up/down-regulated in wild-type or ST2-deficient in at least two of DP/DN/SP comparisons). Genes with opposite directionality across the three sample types (n=5 genes) were dropped. Expression levels of the resulting curated set of 14 genes were visualized using a row-normalized heatmap (Figure 7D).
Pre-processing of SMART-Seq2 scRNA-seq data
BAM files were converted to de-multiplexed FASTQs using the Illumina-provided Bcl2Fastq software package v2.17.1.14. Paired-end reads were mapped to the UCSC mm10 mouse transcriptome using Bowtie with parameters ‘-n 0 -m 10’, which allows alignment of sequences with zero mismatches and allows for multi-mapping of a maximum of ten times.
Expression levels of genes were quantified using TPM values calculated by RSEM v1.2.8 in paired-end mode. For each cell, the number of detected genes (TPM > 0) was calculated and cells with less than 600 or more than 4,000 genes detected were excluded as well as cells that had a mapping rate to the transcriptome below 15%. To further remove potential doublets (mostly of B cells and epithelial cells), we calculated the sum log2(TPM+1) over Cd79a, Cd19, Lyz1, Lyz2 and Sftpc, and excluded any cell that scored higher than 3. We retained only genes expressed above log2TPM of 3 in at least five cells in the whole dataset.
Since we could not sort for Treg for two of the mice (#336 and #338), we had to infer which cells are Tregs from these data. To this end, we trained a random forest classifier for mice for which we have sorted both Tconv and Tregs, using the train function from the caret package in R, based on the expression of the following genes: Foxp3, Ikzf2, Areg, Il1rl1, Folr4, Wls, Tnfrsf9, Klrg1, Il2ra, Dusp4, Ctla4, Neb, Itgb1, and Cd40lg. The labeled data was partitioned into training and test sets. The model has a sensitivity and specificity above 90% in cross validation. We then applied the classifier on the unlabeled data and cells with a probability above 0.6 to be either Tconv or Treg were given the corresponding label. The remaining 4% of cells were discarded as unambiguous.
Identifying tissue-specific gene programs for Treg and Tconv
To identify genes that are differentially expressed between lung and msLN in Treg and/or Tconv, we performed a regression analysis. We focused on the proportion of cells expressing a gene, and hence on logistic regression. We performed logistic regression using the bayesglm function from the arm package in R, including only those mice (# 338, #3642, #3839, #3889) for which we had matched cells from both lung and msLN, as well as for Treg and Tconv, and excluding all genes expressed in >95% or <5% of cells in lung and msLN. We ran the logistic regression with expression data binarized at a log2(TPM+1) of 2 and using the following full model: gene expression ~ genes detected + batch effect + tissue versus a reduced model: gene expression ~ genes detected + batch effect. We corrected for multiple hypothesis by computing an FDR of the likelihood ratio test p-value, and retained genes as differentially expressed between lung and msLN with P< 10−5 and an |coefficient| > 2.
Comparing the extent of cell heterogeneity between lung and msLN
Diffusion components were calculated on a gene expression matrix limited to genes that were differentially expressed between lung and msLN using the DiffusionMap function from the destiny package in R (Angerer et al., 2016) with a k of 30 and a local sigma. In order to be able to compare the variance in distributions in diffusion component 1 and 2 between lung and msLN Treg/Tconv, we downsampled the cells from the lung to the (lower) numbers of cells from the msLN. To test for significant differences in variance in the distributions of lung and msLN Treg/Tconv, we used Levene’s test for the equality of variances on the distributions of the coefficients of the downsampled cells in each of diffusion components 1 and 2.
Identifying gene modules and their time dependence
Gene modules were identified using PAGODA using the scde R package version 2.6.0. (Fan et al., 2016) on the counts table from RSEM after cleaning the data using the clean.counts function (min.lib.size=600,min.detected=5). The knn.error.model function was run using a k of 30, which is much lower than default, but yields statistically indistinguishable results from the default k (# cells / 4). We then ran the pagoda.varnorm to normalize gene expression variance, and the pagoda.subtract.aspect function to control for sequencing depth which then allowed us to run pagoda.gene.clusters which identifies de-novo correlated genes in the dataset. We forced PAGODA to return 100 modules. We identified modules with a significance z.score above 1.96. We removed several highly significant newly identified gene modules consisting of paralog groups with high expression correlation, likely because of multimapping of reads.
Mean module expression was calculated by averaging over the genes in each module of the centered and scaled gene expression table and transforming to a z-score over 1,000 randomly selected gene sets with matched mean-variance patterns. As an initial step, all genes were binned into 10 bins based on their mean expression across cells, and into 10 (separate) bins based on their variance of expression across cells. Given a gene signature (e.g. list of genes in a module), a cell-specific signature score was computed for each cell as follows: First, 1,000 random gene lists were generated, where each instance of a random gene-list was generated by sampling (with replacement) for each gene in the gene-list a gene that is equivalent to it with respect to the mean and variance bins it was placed in. Then, the sum of gene expression in the given cell was computed for all gene-lists (given the 1,000 random lists generated) and the z-score of the original gene-list for the generated 1,000 sample distribution is returned, as in (Singer et al., 2017).
Another module of highly correlated genes identified by PAGODA showed no biological relevance based on gene annotation, but was associated with cells processed on specific dates, suggested they reflect a contamination or batch effect. We scored each cell for this module with the above described method for scoring cells for gene signatures. When testing for differential gene expression over tumor development (described below), we included this batch effect score as a covariate in the regression analysis to control for genes that are correlated with it.
To test if a module’s expression changes over the course of tumor development, we estimated a linear model for each module and compared with a likelihood ratio test a full model: module.activity ~ detected genes + time point to a reduced model: module.activity ~ detected genes. For the time point covariate, healthy lung was taken as reference. We corrected the likelihood ratio test p-values for multiple hypotheses for the number of modules using the p.adjust function computing the false discovery rate in the stats package.
Dimensionality reduction using diffusion component analysis
Diffusion components were calculated on a gene expression matrix limited to genes from modules of interest: modules 1,4,5,14,15 and 21 for Tconv, and modules 1,3,6,8,9,12,13,18,21,23 and 26 for Treg. Gene expression was scaled for Tregs only across all cells. Diffusion components were calculated using the DiffusionMap function from the destiny package in R (Angerer et al., 2016) with a k of 30 and a local sigma. Significant diffusion components identified by the elbow in the eigenvalues were further used for dimensionality reduction to two dimensions. The eigenvectors of the significant diffusion components were imported into gephi 0.9.2 and a force directed layout using forceatlas 2 was run until it converged to get a two dimensional embedding.
Testing for differential gene expression during tumor development
To test whether individual genes change in gene expression over the course of tumor growth, we performed a two-step regression analysis. We focused on the proportion of cells expressing a gene, and hence on logistic regression. We performed logistic regression using the bayesglm function from the arm package in R. Because gender is often confounded with a particular time point in our experiment, we did not include it as a covariate in the model, but did remove all Y chromosome genes from analysis. We also excluded all genes expressed in >95% or <5% of cells in each mouse. We ran the logistic regression with expression data binarized at a log2(TPM+1) of 2 and using the following full model: gene expression ~ genes detected + batch effect + week p.i. (healthy lung as reference) versus a reduced model: gene expression ~ genes detected + batch effect. We identified a threshold for significance by the elbow method, identifying the peak of the second derivative of the ordered fdr distribution of the likelihood ratio test for each time point. To remove significant genes whose signal was driven by only one mouse, we performed another logistic regression using a mixed effect model, accounting for mouse variability: To this end, we added to the significant genes 1,000 randomly selected genes that were non-significant by the initial test to serve as background genes, and performed a mixed effect logistic regression using the glmer function of the lme4 package in R, with the model gene expression ~ tmp + (1|mouse), allowing the intercept to vary by mouse. We combined the elbow method above and the background genes to select an FDR cutoff for significance of 0.01. A gene was classified as significantly varying during tumor development if it passed this FDR cutoff in at least one time point.
T cell receptor (TCR) reconstruction and clonotype calling
TCR were reconstructed using Tracer (Stubbington et al., 2016), run in short read mode with the following settings ‘--inchworm_only=T --trinity_kmer_length=17’. To call shared clonotypes between Treg and Tconv cells, we required all cells of a clone to have identical productive TCRA and TCRB.
Comparison of bulk and scRNA-seq signatures to published signatures
Lists of differentially expressed genes in human cancer Tregs, mouse tissue Tregs, Tr17 cells from mice, and mouse activated Tregs (Table S4) were collected either from the supplementary tables of the relevant publications, or generously provided by the authors upon request (De Simone et al., 2016; Kim et al., 2017; Miragaia et al., 2017; Plitas et al., 2016; van der Veeken et al., 2016).
Population-level TCR Beta chain sequencing and analysis
Analysis of IHC Images
QuPath software was used to annotate tumor and lobe areas (Bankhead et al., 2017). CD8-stained images were standardized to a common set of stain vector parameters. CD8+ cell detection was performed using the PositiveCellDetection plugin with the following parameters: runPlugin(‘qupath.imagej.detect.nuclei.PositiveCellDetection’, ‘{“detectionImageBrightfield”: “Optical density sum”, “requestedPixelSizeMicrons”: 0.5, “backgroundRadiusMicrons”: 8.0, “medianRadiusMicrons”: 0.0, “sigmaMicrons”: 1.5, “minAreaMicrons”: 7.0,“maxAreaMicrons”: 125.0, “threshold”: 0.3, “maxBackground”: 2.0, “watershedPostProcess”: true, “excludeDAB”: false, “cellExpansionMicrons”: 2.0, “includeNuclei”: false,“smoothBoundaries”: false, “makeMeasurements”: true, “thresholdCompartment”: “Cytoplasm: DAB OD max”, “thresholdPositive1”: 0.7, “thresholdPositive2”: 0.4, “thresholdPositive3”: 0.6, “singleThreshold”: true}’);
Scored cells were normalized to tumor area.
Additional statistical analyses
Unpaired, two-tailed Student’s t tests, Mann-Whitney tests, Tukey’s multiple comparisons tests, and Sidak’s multiple comparisons tests were used for all statistical comparisons using GraphPad Prism software.
AUTHOR CONTRIBUTIONS
A.L., R.H.H., D.C., J.M.S., C.G.R., M.B., L.C., A.R., and T.J. designed the study; A.L., D.C., and J.M.S. performed all of the mouse experiments and collection of samples for RNA-Seq in the laboratory of T.E.J.; R.H. performed all computational analysis of RNA-Seq data in the lab of A.R., with help from A.B.; C.D. and M.H. provided technical assistance; C.G.R. conducted TCR repertoire analyses in the laboratory of M.B.; L.C., O.C.S., J.Y.K., and M.C. performed scRNA-seq in the laboratory of A.R., under guidance and supervision from O.R.R.; P.S.R. assisted with cell sorting. A.L., R.H.H., D.C., J.M.S., A.R., and T.J. wrote the manuscript with input from other authors.
DISCLOSURES
T.J. receives research support from the J&J Lung Cancer Initiative. T.J. is a member of the Board of Directors of Amgen and Thermo Fisher Scientific and an equity holder in both companies. He is co-Founder and Scientific Advisory Board member of Dragonfly Therapeutics, a co-founder of T2 Biosystems, and a Scientific Advisory Board member of SQZ Biotech; he is an equity holder in all three companies. His laboratory currently receives funding from the Johnson & Johnson Lung Cancer Initiative and Calico.
ACKNOWLEDGEMENTS
We thank N. Joshi, N. Marjanovic, R. Satija, D. Gennert, K. Thai, and C. Jin for thoughtful discussions,and technical advice; S. Riesenfeld for computational advice; J. Park, J. Wilson and N. Cheng for technical assistance with animal experiments; S. Levine at the MIT BioMicro Center for sequencing support; C. Otis and S. Saldi in the Broad Flow Cytometry Core for sorting assistance; G. Paradis in the Koch Institute Flow Cytometry Facility for technical advice on flow cytometry; K. Cormier and C. Condon from the Hope Babette Tang (1983) Histology Facility for histology assistance; L. Gaffney for artwork and advice on figures; A. Rudensky for Foxp3DTR-GFP mice; A. Sharpe for Foxp3GFP mice; D. Artis for Il1rl1−/− mice; D. Mathis for Il1rl1fl/fl mice; K. Anderson, J. Teixeira, M. Magendantz, and K. Yee for administrative and logistical support.
This work was supported by the Howard Hughes Medical Institute (T.J. and A.R.), Margaret A. Cunningham Immune Mechanisms in Cancer Research Fellowship Award (A.L.), David H. Koch Graduate Fellowship Fund (A.L.), NCI Cancer Center Support Grant P30-CA1405, an Advanced Medical Research Foundation grant (D.C.), and by the Klarman Cell Observatory at the Broad Institute. A.L. is supported by T32GM007753 from the National Institute of General Medical Sciences. A.R. is an Investigator of the Howard Hughes Medical Institute, SAB member for Thermo Fisher and Syros Pharmaceuticals, and a consultant for Driver Group. T.J. is a Howard Hughes Medical Institute Investigator, David H. Koch Professor of Biology, and a Daniel K. Ludwig Scholar.