Adenosine Signaling Is Prognostic for Cancer Outcome and Has Predictive Utility for Immunotherapeutic Response

The role of the immune system in controlling cancer is widely recognized (1). Therapeutically, this is evidenced by a number of recent drug approvals for immunotherapy agents that enhance endogenous antitumor immunity (reviewed by ref. 2) or target tumors directly (reviewed by ref. 3). Responses to immunotherapies are distinct from those seen from other targeted therapies in at least two respects. First, these responses are being observed in cancer indications of previously unmet need such as melanoma (4), lung adenocarcinoma (5), and hematologic malignancies (6). Second, the duration of response to immunotherapy appears to persist for longer in certain settings than those observed with targeted therapies (reviewed by ref. 7).

Despite the clinical success of immunotherapy, there are important questions regarding the initial or eventual failure to control disease, and the value of targeted versus broader approaches to tumor immunity. Current immunotherapies target specific molecules within the immune system, such as the checkpoint proteins PD1 and PDL1, and show responses in only a subset of patients with cancer in any given indication. Total mutational burden (TMB; ref. 8) and PDL1 protein (9) levels have been shown to correlate with immunotherapy response. However, only 30% of the responders are positive based upon these measures (10), suggesting that a more widespread response may be achieved by taking a broader approach; for example, by targeting both innate and adaptive antitumor immunity.

One example of a factor with broad immunosuppressive effects is the adenosine signaling axis (11), which has been shown to suppress natural killer (NK) and CD8⁺ T-cell cytolytic activity while enhancing suppressive macrophage and dendritic cell polarization as well as Treg and myeloid-derived suppressor cell (MDSC) proliferation (12). Beginning with landmark research by Ohta and colleagues (13), a series of preclinical studies (14–18) have been reported and clinical trials (19–21) initiated targeting adenosine signaling. In addition, preclinical evidence supports a role for adenosine axis antagonists in chimeric antigen receptor T-cell therapy (22), adoptive cell therapy (13), and cancer vaccines (23). Thus, targeting the adenosine axis may block a broadly relevant immunosuppressive pathway in cancer (24). Fong and colleagues have demonstrated tumor regressions in 24% of patients with renal cell cancer treated with an A2AR antagonist (25).

We sought to characterize the pan-cancer role of adenosine in human tumors, to test the hypothesis that adenosine signaling has a negative prognostic link within human tumors, and to identify segments of disease where this might be most pronounced. Ideally, a biomarker to address this question would be derived independently of any particular adenosine axis–targeting agent, be quantitatively measurable across various tumor indications, and test the hypothesis using existing public molecular databases. Here, we report the discovery and characterization of a gene expression signature that meets these criteria and which we use to define the landscape of adenosine signaling in human cancer with implications for existing immune checkpoint therapies and adenosine targeting agents in development.

To define a network of regulatory interactions for the A2AR receptor, we used two complementary datasets. Natural Language Processing (NLP) of abstracts and open-access full text from Medline and PubMed Central was performed by Biorelate Ltd (https://www.biorelate.com/) to broadly sweep as much of the literature as possible. In contrast, knowledge derived purely from manual curation in the Ingenuity Pathway Analysis tool database (Qiagen) was used to provide a deeper mining of full text articles from a smaller set of high-impact journals targeted by that resource.

Biorelate defines a causal (regulatory) interaction as a relationship between two entities (genes or proteins) where the subject (cause) entity has a directed edge with an object (theme) entity. Gene entity terms and their relationships from their in-house dictionaries were matched through their machine learning–named entity-recognition software, now incorporated within Biorelate Galactic AI. Protein entities from human, mouse, and rat were retained under the expectation that human data would be the most relevant, while mouse and rat would capture the majority of animal models used in biomedical research. Causal interactions were then collapsed such that all events containing the same pair of entities and the same interaction type were grouped. These groups were assigned a confidence score that was used to rank select events for manual verification.

We then filter the combined set of regulatory relationships to identify genes that are downstream of A2AR (154 genes; 136 from manual curation and 18 from NLP, with 13 detected in both), upregulated by A2AR (90 genes; 78 from manual curation and 12 from NLP), robustly expressed in human tumors, defined as having a median expression greater than the median expression of all genes (74 genes; 66 from manual curation and 8 from NLP), and, finally, by their presence on the Nanostring PanCancer Immune Profiling expression panel (14 genes; 10 from manual curation and 4 from NLP). This last step ensures that our signature retains maximum clinical utility given that the Nanostring panel is widely used to profile formalin-fixed, paraffin-embedded (FFPE) samples from clinical trials where whole transcriptome profiling is often unavailable. The 14 genes that meet these criteria and form the signature are: PPARG, CYBB, COL3A1, FOXP3, LAG3, APP, CD81, GPI, PTGS2, CASP1, FOS, MAPK1, MAPK3, and CREB1.

We score transcriptome data with the signature using Gene Set Variation Analysis [GSVA (26)]. This method is robust to outlying genes expressed at different orders of magnitude and generates scores amenable to downstream statistical interpretation. We observe a strong linear correlation between signatures representing immune processes/features and tumor purity in The Cancer Genome Atlas (TCGA; see Supplementary Fig. S1). To account for this bias, we adjust the signature scores for tumor purity by fitting a linear regression derived from all samples in TCGA versus tumor purity and then applying the following correction:

Corrected score = uncorrected score − (intercept + slope × purity)

Exome sequencing data from TCGA were processed as described in ref. (27). TCGA RNA sequencing (RNA-seq) data were described in ref. (28), and associated clinical data were taken from ref. (29). Copy-number variants made with GISTIC version 2.0.22 were obtained from the TCGA Firehose. Microsatellite instability (MSI) subtype information was obtained from ref. (30). Tumor purity data were obtained from ref. (31).

RNA-seq data from ADORA2A knockout NK cell lines generated in ref. (32) were obtained from the European Nucleotide Archive (PRJEB22631). Reads were aligned to the mouse genome (mm10) using HISAT2 (33), and expression levels were quantified using Salmon (34).

Published cohorts of immunotherapy-treated subjects with preprocessed gene expression profiles were obtained from refs. (35, 36) and scored with GSVA. The anti–CTLA-4 dataset (36) was generated with a custom nanostring panel that contained only 6 (CASP1, CD81, CYBB, LAG3, PARG, and PTGS2) of the 14 genes from our signature.

Survival analyses were performed using the Cox proportional hazards regression model as implemented in the Survival package from R (37). For the analysis presented in Fig. 2A and B, tumors were split into high (>75th), medium (25–75th), and low (<25th) based on quartiles. In all other survival analyses, adenosine signature scores were split on 0 with >0 high and <0 low.

Immune cell infiltrates were determined with a support vector regression (SVR) approach based on CIBERSORT (38) to define relative immune cell abundance. To study the association of adenosine with CD8⁺ T-cell infiltration, we consider CD8 high tumors to be greater than the median of CD8A expression across all samples. All other cell or cell-state signatures were scored using GSVA. NK-cell exhaustion was determined using expression of KIR3DL1, KIR3DL2, IL2RA, IL15RA, HAVCR2, and EOMES. Cytotoxicity was determined using the expression of NKG7, CST7, PRF1, GZMA, GZMB, and IFNG. CD8 exhaustion was determined using the signature provided in ref. (39). IFNG signaling was determined using the signature presented by ref. (40).

Genetic associations with adenosine signaling were studied for all genes with a mutation frequency >2% across the cohort being studied and for all copy-number variants. A linear model with tumor type, TMB, and MSI status as covariates was fit to the data, and ANOVA was used to test for significance. Effect sizes were computed as the Cohen's D effect size where the difference between means is normalized for the variance within the data. All P values were adjusted for multiple testing using the Benjamini–Hochberg procedure.

All animal studies were performed according to AstraZeneca Institutional Animal Care and Use Committee guidelines.

Transcriptional profiling data for the five syngeneic models shown in Fig. 1A were obtained from ref. (41). Tumor adenosine measurements from syngeneic models were performed as described in Goodwin and colleagues (42).

For the in vivo treatment study shown in Fig. 1B, MC38 cells were confirmed free of mycoplasma and mouse pathogens by PCR as part of a rodent pathogen testing panel (IMPACT, IDEXX Bioresearch). Thawed cells were cultured in DMEM supplemented with 10% heat-inactivated FBS and 1% l-glutamine (Sigma Aldrich) at 37°C in a humidified incubator maintained at 5% CO₂. Cell counts were performed prior to implantation by Countess Cell Counter (Invitrogen). For subcutaneous implants, 5 × 10⁻⁵ MC38 cells/mouse were resuspended in sterile PBS and injected s.c. into the right flanks of 4- to 6-week-old female C57BL/6 mice (Charles River Labs) in a total volume of 0.1 mL/mouse.

Mice were randomized into treatment groups at a starting tumor volume of 50 to 90 mm³. AZD4635 nanosuspension formulation (Aptuit, Verona) was reconstituted in sterile water and dosed orally b.i.d. at 50 mg/kg. Tumor volume and body weight were measured twice weekly after randomization. Growth rate was calculated as the slope of a linear model fit to the percent change in tumor volume from day 0 over time.

The first-in-human trial, NCT02740985, was conducted to assess safety, pharmacokinetic (PK), and pharmacodynamic activity of AZD4635 as monotherapy and in combination with durvalumab in patients with treatment refractory solid tumors. Predose and on-treatment tumor biopsies within 3 weeks were collected from 7 subjects who were treated with AZD4635 monotherapy. Six subjects (ovarian, sarcoma, pancreatic, colon, myxofibrosarcoma, and cholangiocarcinoma) were treated with 75 mg, whereas the single remaining subject (breast) was treated with 100 mg AZD4635 PO in a once-daily dosing regimen, thus all were at or below the recommended phase 2 dose of 100 mg QD, showed no significant differences in PK, and none of the subjects in this cohort had a RECIST response to AZD4635.

Total RNA was extracted from tumor tissue macrodissected from 5-mm-thick FFPE sections using the miRNeasy FFPE Kit (QIAGEN). RNA integrity and quantity were assessed on the TapeStation 2200 using the RNA ScreenTape System (Agilent). Manufacturer's recommended protocols were followed.

The RNA was subsequently analyzed for gene expression using the NanoString nCounter FLEX Analysis System and the commercially available 770-gene, human PanCancer Immune Profiling Panel (NanoString). Following the manufacturer's standard XT CodeSet Gene Expression Assays protocol, 25 to 100 ng RNA was hybridized with Capture and Reporter probes at 65°C for 22 hours. Posthybridization sample processing on the Prep Station using the high-sensitivity setting was followed by data collection on the Digital Analyzer scanning at 555 fields of view. Preprocessing of the raw count data, which included background subtraction of the negative control probes, positive control normalization, and housekeeping gene normalization, was performed in the nSolver 4.0 (NanoString) software using the geometric means and default parameters. All samples included in downstream analyses fell within the default nSolver QC parameters.

It is challenging to measure tumor adenosine levels in a high-throughput manner, so we sought to create a gene expression signature that would recapitulate adenosine signaling and allow us to study the wealth of existing data from large collections of tumor transcriptomes. It has previously been shown that causal, or regulatory, protein/gene interaction knowledge is a powerful substrate for the interpretation of transcriptomic data (43–45). We thus sought to compile a regulatory network for the adenosine signaling pathway. Both NLP, as described previously (46–48), and manually extracted knowledge (49) were used to define a network of interactions between the A2AR receptor and downstream entities. Of the four adenosine receptors, A2A was selected as the basis of our study given that A1 and A3 function to increase cAMP rather than decrease it, which is necessary for immune cell suppression (50), whereas A2B has considerably lower affinity for adenosine (50). We focused on regulatory interactions where there was evidence that A2AR increased expression of the downstream entity in the primary scientific literature. We found 172 genes that have been reported to be regulated by A2AR, 90 of which were reported to be positively regulated by A2AR signaling activity. We applied additional filters to ensure the genes are robustly expressed in human tumors and present on a widely used clinical transcriptomics assay. Our final signature consisted of 14 genes whose concordant activity is indicative of adenosine signaling: PPARG, CYBB, COL3A1, FOXP3, LAG3, APP, CD81, GPI, PTGS2, CASP1, FOS, MAPK1, MAPK3, and CREB1.

To confirm the validity and specificity of our signature, we quantified the intratumoral levels of adenosine in five murine syngeneic models for which we also have transcriptional profiles (41). We find a significant correlation (r² = 0.92, P = 0.018) between measured intratumor adenosine concentrations and adenosine signaling as captured using our signature (Fig. 1A). We next assessed whether the adenosine signature tracked with inhibition of the adenosine receptor in vivo within the MC38 syngeneic model using AZD4635, an A2AR selective small-molecule currently in clinical development (15, 51). We find that our signature correlates (r² = −0.62, P = 0.001) with reduced growth rate after A2AR inhibition (Fig. 1B). Furthermore, knockout of the A2AR receptor (32) abrogated adenosine signaling signature scores in CD11b⁺ CD27⁻ NK cells (Fig. 1E). A key biological effect of adenosine within human tumors is to suppress immune cell activity (32). In concordance with this, the adenosine signature scores have a significant association with NK-cell (r² = 0.4, P < 2.2e⁻¹⁶ and OR = 3.1, P < 2.2e⁻¹⁶, Fig. 1C) and CD8⁺ T-cell (r² = 0.6, P < 2.2e⁻¹⁶ and OR = 7.8, P < 2.2e⁻¹⁶, Fig. 1D) exhaustion marker expression in TCGA. Finally, 7 patients with a variety of solid tumors were treated once daily with AZD4635 in a phase IA study (NCT02740985) to assess pharmacodynamic changes in signature scores within humans. Adenosine signaling scores were reduced in 5 of the 7 (70%) patients, 4 of which also had concordant increases in gene expression signatures of cytolytic activity and IFNG signaling. Taken together, these data demonstrate that our proposed signature is a useful surrogate for adenosine signaling activity when studying bulk transcriptomes of human and mouse tumors.

Having established that our signature captures adenosine signaling activity within tumors, we next explored the association of adenosine signaling with disease outcomes. Adenosine suppresses a functional antitumor response and so we would expect tumors with a high adenosine drive to be more aggressive and have reduced survival. To confirm this, we used our signature scores to compare survival in tumors with high adenosine signaling with tumors with low adenosine signaling across all cancers in TCGA. However, before doing so, we studied the potential for tumor purity to bias our scores across large datasets. We observed that low purity trended with greater adenosine signature scores (Supplementary Fig. S1A). We therefore established a normalization of signature scores for tumor purity (Supplementary Fig. S1B and Methods) to remove this bias from further studies of human tumors in TCGA.

Adenosine signaling high tumors were defined as the upper quartile of signature scores across all samples, and likewise adenosine low consisted of the lower quartile. We find that high levels of adenosine signaling associate with significantly worse overall survival (OS; HR = 0.6, Cox PH P < 2.2e⁻¹⁶) and progression-free survival (PFS; HR = 0.77, Cox PH P = 0.0000006) in a pan-cancer model (Fig. 2A and B). This association remains if the data are split by tertiles (OS HR = 0.75, Cox PH P = 0.000000006; PFS HR = 0.83, Cox PH P = 0.000025) or on the median (OS HR = 0.81, Cox PH P = 0.0000002; PFS HR = 0.86, Cox PH P = 0.00007).

Considerable progress has been made in the characterization of the tumor microenvironment from the perspective of immune cell infiltration. However, it remains unclear why some apparently “hot” tumors with an otherwise adequate infiltration of immune cells do not appear to mount an effective antitumor response. We first assessed the relationship of adenosine signaling to immune cell infiltrates inferred from bulk RNA-seq in TCGA using a support vector regression approach based upon the CIBERSORT algorithm (38). There are no strong associations, but we observe weak negative correlations with activated NK-cell and T follicular helper cell scores and a positive correlation with resting NK-cell and macrophage scores (Supplementary Fig. S3A). We therefore studied the ability of adenosine to modulate the activity of existing immune infiltrates by studying only tumors with a high level of CD8⁺ T-cell infiltration, defined as greater than the median of CD8A expression across all samples. We found a dramatic survival deficit in tumors that are both CD8 high and adenosine high versus tumors that are CD8 high but adenosine low, for both OS (HR = 0.47, Cox PH P < 2.2e⁻¹⁶) and PFS (HR = 0.65, Cox PH P = 0.0000002; Fig. 2C and D). Further, the survival deficit between adenosine high and low tumors is reduced or ablated in CD8 low tumors (OS Cox PH P = 0.001, PFS Cox PH P = 0.05).

We next studied the adenosine signaling profile of each tumor type from TCGA individually. All tumor types exhibit a wide range of adenosine signaling levels, and all have some individuals with high adenosine signaling (Fig. 3A). Kidney renal clear cell carcinoma (KIRC) has the highest levels of adenosine signaling on average across all tumor types, whereas thymoma (THYM) has the lowest (Fig. 3A). Consistent with this observation, adenosine is known to play an important role within the kidney where it regulates a variety of physiologic functions and is present at significant extracellular concentrations (52). Interestingly, adenosine also plays a role in the thymus, regulating the thymocyte selection process (53).

Concordantly reduced OS and PFS in adenosine high tumors are seen in 13 individual diseases (Fig. 3C), with four having an HR < 0.7 for both survival measures: uveal melanoma (UVM, OS HR = 0.08, PFS HR = 0.38), cervical (CESC, OS HR = 0.70, PFS HR = 0.69), pancreatic (PAAD, OS HR = 0.74, PFS HR = 0.68), and thyroid (THCA, OS HR = 0.75, PFS HR = 0.52). However, UVM (HR = 0.08, P = 0.016) is the only case where OS is statistically significant for an individual tumor type (Fig. 3B). Similarly, glioblastoma (GBM, HR = 0.66, P = 0.02), thyroid carcinoma (THCA, HR = 0.52, P = 0.03), and UVM (HR = 0.37, P = 0.05) are the only diseases where adenosine signaling is statistically associated with worse PFS. Interestingly, diffuse large B-cell lymphoma (DLBCL) tumors with high levels of adenosine signaling are associated with improved PFS (DLBC, HR = 5.19, P = 0.02), although notably not concordant with OS.

Adenosine signaling is not correlated with TMB at a pan-cancer level (r² = 0.02); however, MSI high tumors have significantly higher levels of adenosine signaling (Fig. 4C, P = 5e⁻¹⁶). We therefore derived a linear model that incorporated MSI as a covariate with which to identify single-nucleotide variants (SNV) associated with adenosine signaling. Our analysis identifies 23 mutated genes that associate with adenosine signaling (at q < 0.1) when all samples are considered in a pan-cancer model: 9 with enhanced adenosine signaling and 14 with reduced adenosine signaling (Fig. 4A and Supplementary Table S1).

Six adenosine-associated genes have an established role in cancer pathogenesis, being members of the cancer gene census (54, 55), including VHL, ACVR2A, FIP1L1, and NSD1 which all correlate with increased adenosine signaling, and GATA3 and STK11 that associate with reduced adenosine signaling (Supplementary Fig. S3).

VHL has the largest effect size and is thought to be an E3 ubiquitin ligase that suppresses HIF1a expression. Consequently, VHL loss-of-function mutations lead to constitutive expression of HIF1a which upregulates CD73 and CD39, thereby enhancing the production of adenosine (56). This previously described mechanism gives us further confidence in the relevance of our signature.

GATA3 is an important transcription factor associated with breast cancer and as a key regulator of CD4⁺ T-cell development with some evidence to suggest its activity is regulated by adenosine in other settings (57).

The tumor suppressor STK11 has recently been shown to drive primary resistance to checkpoint inhibition (58), and the negative association with adenosine signaling identified here most likely reflects the immunologically cold/excluded tumor microenvironment for which an immunosuppressive phenotype has not been activated. This raises the interesting possibility that the other negatively associated genetic segments might also exhibit resistance to immunotherapy. Notably, the most significantly associated genetic mutations are in NPRL3 which is part of the GATOR1 complex, which, like LKB1 via AMPK, feeds into the mTOR signaling pathway (59, 60).

We found 55 SNVs associated with adenosine within an individual tumor type (q < 0.05, Fig. 4B and Supplementary Table S2), comprising 25 from kidney renal papillary cell carcinoma, 23 from breast cancer, 3 from kidney renal clear cell carcinoma, and 1 each from lung adenocarcinoma (STK11), prostate adenocarcinoma (RABL6), stomach adenocarcinoma (TP53), and head and neck squamous cell carcinoma (BRD7). Seven of these associations feature cancer census genes: TP53 in BRCA and STAD, GATA3 in BRCA, CDH1 in BRCA, VHL in KIRC, FIP1L1 in KIRP, and STK11 in LUAD (Supplementary Fig. S4).

Somatic copy-number alterations (SCNA) are also associated with adenosine signaling. Note that 124 SCNA are significant (q < 0.05) with 11 having an effect size greater than 0.5 (Table 1 and Supplementary Fig. S5). This includes a deletion on chromosome 3 which removes VHL and replicates the observation seen with SNVs.

TGFBR2 and ACVR2A mutations are among the most significant associations with adenosine levels in a pan-cancer model even after correction for MSI status. Both are members of the TGFβ superfamily encoding the TGFβ receptor and the structurally related activin growth factor receptor, respectively. TGFβ signaling has a complex and highly context-dependent association with cancer biology. As a tumor suppressor, TGFβ mutation promotes tumorigenesis, but its loss has also been shown to increase chemokine signaling resulting in infiltration of myeloid-derived suppressor cells which themselves produce TGFβ and eventually drive immunosuppression, thereby promoting tumor growth (61). Our result raises the possibility that this suppression is driven largely through the adenosine axis.

To further explore this relationship, we conducted a deeper study of the association between adenosine and TGFβ. Thorsson and colleagues (62) defined six primary immune subtypes of cancer including a TGFβ-dominant group, cluster 6 (“C6”). We find that adenosine signaling is significantly higher in this group compared with the other five immune subtypes (Fig. 5A). We further expanded our analysis to include the 43 members of the TGFβ superfamily (63) and find that mutations in any of these genes are associated with a higher level of adenosine signaling (Fig. 5B). Finally, tumors that are both adenosine high and mutant in a TGFβ superfamily member have worse OS compared with tumors that are adenosine low and TGFβ wild-type (HR = 0.43, P < 2.2e⁻¹⁶), or those that are either TGFβ mutant (HR = 0.74) or adenosine high (HR = 0.72; Fig. 5C).

To test the clinical utility of the signature and the extent to which adenosine affects immune checkpoint therapy, we studied cohorts of patients treated with checkpoint inhibitors. Prat and colleagues generated gene expression profiles of 65 patients from a variety of solid tumors that were treated with anti-PD1 therapy (35). Chen and colleagues profiled 53 patients with metastatic melanoma that were treated with anti–CTLA-4 therapy (36). We find that responders to immune checkpoint therapy, as classified by their best overall response, have lower levels of baseline tumor adenosine signaling than do patients which progress on both anti-PD1 therapy (Fig. 6A) and anti–CTLA-4 therapy (Fig. 6C). We used logistic regression to model the probability of a patient being a complete responder (CR), partial responder (PR), or having stable disease (SD) versus progressive disease (PD) in these cohorts. A signature score just below 0 (−0.01368) equates to a 50% probability of being a responder, and a signature score of −0.4 equates to a 75% probability of being a responder (Supplementary Fig. S7).

In the anti–CTLA-4 dataset, only 6 genes from our 14-gene signature are present on the panel used. To study the effect this might have we scored the anti-PD1 dataset with the same 6 genes. The overall trend of results is retained, but the sensitivity of the signature is reduced (PD vs. SD 6 gene P = 0.072 vs. 14 gene P = 0.076, and PD vs. PR/CR 6 gene P = 0.13 vs. 14 gene P = 0.0027).

There is also a highly significant association between adenosine signaling at baseline and PFS on anti-PD1 therapy (Fig. 6B; HR = 0.29, Cox PH P = 0.00012). Interestingly, expression of the gene encoding PDL1 (CD274), which is highly correlated with PDL1 IHC measurements (35), does not associate with PFS in the same dataset (HR = 0.8, Cox PH P = 0.47). Furthermore, combining adenosine and CD274 expression does not enhance the ability to predict immunotherapy response beyond adenosine alone (Fig. 6B and Supplementary Fig. S6). These results would suggest that baseline levels of adenosine are an important correlate of response to immunotherapy and that our signature might complement PDL1 as a marker in this regard for the existing checkpoint inhibitors and potentially as A2AR inhibitors progress through the clinic.

We sought to characterize the pan-cancer role of adenosine in human tumors, to test the hypothesis that adenosine signaling has a negative prognostic link within human tumors, and to identify segments of disease where this might be most pronounced. We therefore report the derivation and validation of a gene expression signature that recapitulates adenosine signaling and with which we address these questions.

Our data show that adenosine signaling levels vary across the tumor types of TCGA and that this is associated with suppression of antitumor immunity in tumors where an otherwise adequate CD8⁺ T-cell infiltrate is present. This observation has important implications for the way in which we model and identify functional antitumor immune responses. Significant progress has been made in the identification of immune infiltrates alone that associate with outcomes (e.g., the Immunoscore; ref. 64), yet our results suggest that orthogonal measures of immunosuppressive effectors could enhance these measures.

We also identify genetic segments of disease that associate with higher adenosine signaling including MSI tumors and specific genetic variation in TGFβ. These mutations have potential as markers for adenosine-targeted therapies and are consistent with the concept that adenosine signaling acts to suppress the inflammatory response to highly immunogenic tumors (65). The relationship between adenosine signaling and TGFβ associates adenosine with fibroblast biology and reflects early clinical data from the anti-CD73 monoclonal antibody Oleclumab (ASCO) in pancreatic cancer, an indication known to be rich in cancer-associated fibroblasts (66).

Finally, we demonstrate that adenosine correlates with survival across tumors of all types and within specific indications. Furthermore, baseline adenosine signaling scores appear to predict response to immune checkpoint therapies. We show that this is independent of PDL1 expression and that adenosine signaling does not correlate with TMB. Further investigation is therefore warranted to establish the potential utility of adenosine signaling as an additional measure to identify patients likely to respond to checkpoint therapy. In addition, these findings support the development of adenosine targeting agents for use in combination with existing checkpoint inhibitors. However, it requires further clinical validation using an agent targeting the adenosine axis to be considered predictive of response to this class of agents as opposed to prognostic in the presence of intact adenosine signaling. As our signature has been derived independently of any specific molecular agent targeting the adenosine pathway, it may have utility across a broad spectrum of candidate drugs currently in development.

Our results present several unexpected findings. First, the CT26 mouse model and MSI high human tumors are sensitive to immune checkpoint inhibitors, while the MC38 mouse model has shown variable responsiveness to immune checkpoint inhibition, yet we find both associated with high adenosine signaling (Figs. 1A and 4C). In addition, not all tumor types with high adenosine signaling on average appear to exhibit a survival deficit. Further, although reduced adenosine signaling enriches for responders to checkpoint inhibition, not all adenosine low patients respond and vice versa. Immune checkpoint inhibitor sensitivity is likely determined by many factors in addition to adenosine. For example, the presence of CD8⁺ T cells, expression of PDL1, and high TMB are all associated with checkpoint response (67, 68). It is also likely that an immune infiltration/response must occur prior to a state of adenosine-mediated repression. As such, adenosine is another factor that contributes to the balance between those that induce antitumor immunity and those that are immunosuppressive. Further works to characterize the tipping point in this balance are necessary.

The signature described here may be limited by the decision to only include genes within a commercially available RNA expression panel. This decision ensures the translatability of the signature to ongoing clinical studies as well as direct comparison with other reported gene expression systems (69, 70). However, expansion and further development of the signature using a broader panel of transcripts could enhance the sensitivity of the signature. A second area which falls outside of the scope of this report but may well be relevant here involves other branches of adenosine biology beyond A2AR. For example, inflammatory signaling through ATP (71) or other nodes of the larger adenine nucleotide signaling axis (72). Finally, a group from Corvus Pharmaceuticals has taken an orthogonal approach to generating an adenosine-related gene signature. Here, the authors identified genes upregulated by NECA, an adenosine analog, and suppressed by CPI-444, an A2AR antagonist (69). The two signatures have just one gene in common (PTGS2) which may reflect the compound-specific nature of the CPI-444 signature.

In conclusion, the expression signature reported here is a useful clinical tool that establishes the adenosine pathway as a key mediator of a successful antitumor immune response, demonstrates the prognostic potential for existing immunotherapies of quantifying adenosine drive, and informs patient selection strategies for adenosine pathway modulators currently in development.

B. Sidders, P. Zhang, D.L. Russell, A. Borodovsky, J. Armenia, J.R. Dry, and K. Sachsenmeier are employees/paid consultants for AstraZeneca. T.M. Bauer is an employee/paid consultant for Pfizer, Blueprint Medicines, Exelixis, Bayer, LOXO Oncology, and Guardant Health, and reports receiving speakers bureau honoraria from L&M Healthcare. G.S. Falchook is an employee/paid consultant for Fujifilm and EMD Serono, reports receiving commercial research grants from 3-V Biosciences, Abbisko, Abbvie, ADC therapeutics, Aileron, American Society of Clinical Oncology, Amgen, ARMO, AstraZeneca, BeiGene, Bioatla, Biothera, Celldex, Celgene, Ciclomed, Curegenix, Curis, Cyteir, Daiichi, DelMar, eFFECTOR, Eli Lilly, EMD Serono, Epizyme, Exelixis, Fujifilm, Genmab, GlaxoSmithKline, Hutchison MediPharma, Igntya, Jacobio, Jounce, Kolltan, Loxo, Medimmune, Millennium, Merck, miRNA therapeutics, National Institutes of Health, Novartis, OncoMed, Oncothyreon, Precision Oncology, Prelude, Regeneron, Rgenix, Ribon, Strategia, Syndax, Taiho, Takeda, Tesaro, Tocagen, Turning point therapeutics, U.T. MD Anderson Cancer Center, Vegenics, and Xencor, speakers bureau honoraria from Total Health Conferencing, and other remuneration from Wolters Kluwer, Bristol-Myers Squibb, EMD Serono, Fujifilm, Milennium, and Sarah Cannon Research Institute. M. Merchant is an employee/paid consultant for AstraZeneca and Epizyme. G. Pouliot is an employee/paid consultant for, reports receiving commercial research grants from, and holds ownership interest (including patents) in AstraZeneca. J.C. Barrett and R. Woessner are employees/paid consultants for and hold ownership interest (including patents) in AstraZeneca. No potential conflicts of interest were disclosed by the other authors.

Conception and design: B. Sidders, M. Merchant, J.C. Barrett, J.R. Dry, R. Woessner, K. Sachsenmeier

Development of methodology: B. Sidders, K. Goodwin, M. Merchant, J.C. Barrett, J.R. Dry, K. Sachsenmeier

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): K. Goodwin, G. O'Connor, A. Borodovsky, J. Armenia, J.C. Bendell, T.M. Bauer, M.R. Patel, G.S. Falchook, M. Merchant, J.R. Dry

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): B. Sidders, P. Zhang, J. Armenia, R. McEwen, B. Linghu, M.R. Patel, M. Merchant, G. Pouliot, J.C. Barrett, J.R. Dry, R. Woessner, K. Sachsenmeier

Writing, review, and/or revision of the manuscript: B. Sidders, P. Zhang, D.L. Russell, J. Armenia, B. Linghu, J.C. Bendell, T.M. Bauer, M.R. Patel, M. Merchant, G. Pouliot, J.R. Dry, R. Woessner, K. Sachsenmeier

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): D.L. Russell, R. McEwen, R. Woessner, K. Sachsenmeier

Study supervision: M.R. Patel, M. Merchant, G. Pouliot, J.R. Dry, R. Woessner

This study was funded by AstraZeneca Ltd.

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.