Enhanced antitumor efficacy of combined targeting of adenosine A receptor and PD-1 is mediated via multiple effects on different cell populations within tumor microenvironment.

Introduction
Immune checkpoint inhibitors (ICIs) have become a cornerstone of immunotherapy for patients with NSCLC. However, the 5-year survival rates for patients undergoing ICI-based therapies in advanced stages of the disease remain low, around 15–16% [1–3]. While this represents a significant improvement over chemotherapy alone, further advancements are essential for improving long-term outcomes. A deeper understanding of how tumors evade immunotherapy is crucial. It is increasingly clear that the tumor microenvironment (TME) is an important part of this resistance, and the mechanisms underlying the role of the TME remain poorly understood.
A promising target for overcoming ICI resistance is adenosine, an immunosuppressive metabolite that accumulates in the TME [4]. Extracellular adenosine (eADO) is primarily generated through the breakdown of ATP by ectoenzymes CD39 and CD73, along with the processing of cyclic AMP (cAMP) and NAD+ by CD38 and CD157 [5]. Adenosine exerts its effects through interactions with adenosine receptors (A1R, A2AR, A2BR, A3R) on immune and tumor cells [6]. A2BR has a lower affinity for adenosine compared to other receptors and remains inactive under normal physiological conditions, but becomes active in tumors where adenosine levels are high, thus representing an attractive therapeutic target in the TME [6]. In preclinical and early clinical studies, targeting adenosine receptors has shown promise in enhancing anti-tumor immunity and undermining the resistance to ICIs [7–9]. Therapeutic agents targeting CD39 and CD73 are being developed to target this pathway and enhance anti-tumor immune responses [10]. CD38 and CD157 are also being targeted in clinical trials [5].
TGFβ signaling is also an important player in the TME. Adenosine has been shown to induce TGFβ in both cancer and myeloid cells [11, 12]. TGFβ contributes to carcinogenesis through multiple mechanisms, including immunosuppression and the promotion of cancer-associated fibroblasts (CAFs) to produce extracellular matrix (ECM) components [13]. TGFβ signaling also promotes PD-L1 expression in tumors, enhancing T cell exhaustion [12]. Furthermore, TGFβ signaling can induce CD73 expression, thereby increasing eADO levels and establishing a feedback loop that exacerbates immune suppression [14, 15]. Disrupting TGFβ signaling in myeloid cells can decrease CD39 and CD73 expression, resulting in enhanced T cell infiltration [10, 14]. However, inhibiting TGFβ signaling in normal tissues poses significant challenges due to its critical role in tissue homeostasis [16, 17].
In this study, we assessed the effects of PBF1129 (an A2B receptor antagonist), an anti-PD-1 antibody, and their combination in the highly immunosuppressive LLC and CMT167 murine models [7, 18] and explored the mechanisms underlying the efficacy of combined therapy using single-cell RNA sequencing (scRNA-seq) to gain insights into impact of combination treatment in NSCLC. We generated cell–cell interaction networks in the TME highlighting how malignant and other cells interact with each other thus providing insights into the mechanisms of action of the drugs and their combination [19–21]. This high-resolution approach provides a more in depth understanding of the mechanisms by which adenosine receptor antagonism enhances ICI efficacy, potentially through increasing Gzmb in M1-like macrophages and inhibition of T cell exhaustion. Our findings also reveal that TGFβ signaling plays a central role in solid tumor progression, and that blockade of A2BR signaling in combination with ICI significantly reduces TGFβ signaling, supporting this strategy to overcome IO resistance in NSCLC.

Materials and methods

Mouse models
Lewis lung carcinoma cells (LLC, 1 × 105 cells) or CMT167 cells (3 × 105 cells) were subcutaneously injected on the flank of C57BL/6 female mice. Vehicle or PBF1129 (100 mg/kg) were given i.p. every other day in both models. IgG (8 mg/kg) or anti PD-1 (8 mg/kg) were given i.p. two times a week in the LLC model, three times a week in the CMT167 model. 10 mice in each group and all mice were subjected to tumors until the experiment ended. Treatment was maintained for 2–3 weeks. Tumors were measured every other day and tumor growth was monitored. Whole tumors of the LLC model and CMT167 models were harvested and sliced either snap frozen or fixed in formalin immediately.

Single-cell RNA sequencing analysis
The snap frozen tumors were put in a -80 freezer for long-term storage. Four samples with the median size of each group were chosen for single-cell RNA sequencing. Due to the small tumor size due to the efficacy of the combination treatment in LLC model, we either combined three individual tumors or two individual tumors as one sample for analysis. Therefore, we forwarded 16 samples of each model for sequencing. Each sample had a minimum weight of at least 15 mg. Cell preparation followed the instructions for tissue fixation and dissociation for chromium fixed RNA profiling from 10X Genomics (Supplementary Method). We used the Nexcelom Bioscience Cellometer® Auto 2000 counter to do the quantification of the 16 samples in both models by using ViaStain AOPI staining solution (#CS2-0106-5 ml) and SD100 Cellometer®cell counting chambers (#CHT4-SD100-002). All samples were normalized to the limiting sample with the lowest total cell number in hybridization and the same number of cells were pooled. All the samples were qualified to proceed to the preparation of scRNA-seq libraries and followed by processing for scRNA-seq library construction using 10X Genomics and sequenced using Novaseq6000 platforms. Raw scRNA-seq data were aligned to mouse reference and analyzed by using Cell Ranger pipeline (10X Genomics). Details of quality control including the removal of cellular debris and doublets were found in Supplementary Methods. Following quality control, a total of 68,388 cells of the LLC model and 99,889 cells of the CMT167 model were retained for the downstream analysis. The Uniform Manifold Approximation and Projection (UMAP) was used for visualization of clusters by using the unsupervised clustering analysis Seurat. Differentially expressed genes (DEGs) for each cell types and treatment comparisons were identified using the FindMarkers function in Seurat R package. To identify specific pathways signatures, AddModuleScore in the Seurat R package was applied. The enriched pathways were analyzed through the Gene Set Enrichment Analysis (GSEA) tool and by using ShinyGO 0.82 (https://bioinformatics.sdstate.edu/go/) or Enrichr (https://maayanlab.cloud/Enrichr/). We defined cell types by integrating the top-ranked DEGs and their global distribution (Supplementary Method). To evaluate cell–cell interactions and ligand-receptor pairs, significant cellular interactions and pairs were identified by using CellChat as previously described (Supplementary Method). NicheNet is a method to predict ligand-target genes between interacting cells (Supplementary Methods). To study the cell cycle states, Cell Cyle Score was applied (Supplementary Methods). Monocle3 was applied to construct trajectories and the DEGs of the clusters were identified (Supplementary Methods). To study signature of the pathways, PROGENy analysis was used (Supplementary Methods).

Histology
We generated Lewis lung carcinoma (LLC) model tumors by subcutaneous injection of 1 × 105 cells on the flank of C57BL/6 female mice.
Tumors were fixed in formalin from control and combination treatments for both LLC and CMT167 models and were embedded in paraffin. Paraffin-embedded tissues were sectioned and placed on slides. Deparaffinization was conducted by melting the paraffin on the slide in a moat at 60 °C for 30 min and dipping the slides in xylene and subsequently passing them through the series of reducing ethanol gradients (100, 95, 70%) for 10 min.
For picrosirius red staining, the picrosirius red stain (Polysciences, #24901) was applied to the deparaffinized sections and hydrated with distilled water and rinsed well. According to the instructions, the slides were placed in Solution A (Phosphomolybdic acid) for 2 min followed by rinsing with distilled water, Solution B (Picrosirius Red F3BA Stain) for 60 min and Solution C (0.1 N Hydrochloride acid) for 2 min subsequently, followed by 70% ethanol for 45 s and then dehydrated and mounted with coverslips. Fiji ImageJ was used to analyze collagen deposition in these sections.
For Cd31 immunohistochemical staining, the deparaffinized sections were rehydrated with running cold tap water for 5 min followed by antigen retrieval for 40 min at 95 °C using antigen retrieval reagent (Enzo, #ENZ-ACC113, pH9) and then washed for 5 min by using PBS (Sigma, #79383-1L). The slides underwent peroxidase and alkaline blocking (BLOXALL, #SP-6000–100) for 10 min at 25 °C, followed by 2.5% horse serum (ThermoFisher, #16050130) blocking for 20 min at 25 °C and incubation with the 3 μg/ml goat primary antibody Cd31 (R&D, #AF3628) overnight at 4° C followed by washing. The concentration of Cd31 antibody was optimized at first. Secondary detection was performed (Cell Signaling Technology, SignalStain® Boost IHC Detection Reagent (HRP, Goat), #63707S) on the section for 30 min followed by washing. All washing cycles used PBS (Sigma, #79383-1L) for 5 min. Utilizing DAB (SignalStain®DAB substrate kit, #8059S) as a chromogen within a minute. The slides were washed with PBS (Sigma, #79383-1L) for 5 min followed by running cold tap water for 5 min and a positive control slide and isotype control slide (R&D, #AB-108-C) used as the reference. Hematoxylin staining (Thermo Fisher Scientific, #22–220-100) of cell nucleus was performed for 2 min, followed by washing the slides with running cold tap water for 10 min. Dehydration was conducted by dipping the slides in a series of increasing ethanol gradients (70, 95, 100%) for 5 min. Mounting medium was used with a coverslip. Fiji ImageJ was used to analyze Cd31 expression in these sections.

Module score
The AddModuleScore function from the Seurat in R package was applied and the genes sets were represented in violin plots. Statistical analysis by using ggviolin plot function stat_compare_means in the R package.

Statistical analysis
The tumor growth curves of both models which were shown as an average size of each treatment group with the mean and SEM and linear mixed models were used to assess the associations among the treatment groups over time. The distributions of cell type proportion and collagen/Cd31 expression were tested using the Shapiro–wilk test, collagen/Cd31 expression indicated they were non-normally distributed. Therefore, the non-parametric Mann–Whitney test or parametric unpaired t-test was used to compare the cell type proportion and the non-parametric Mann–Whitney test or Kruskal–Wallis test was used to compare collagen/Cd31 expression on the slide sections between the paired groups (control vs PBF1129, control vs anti PD-1, control vs combination treatment, anti PD-1 vs combination treatment, anti PD-1 vs PBF1129, PBF1129 vs combination treatment). Comparisons of the subcluster proportions of M2-like macrophages were adjusted for multiple testing using Holm-Bonferroni correction. The comparison between gene expression in paired groups were analyzed using Wilcoxon rank-sum test, with a p value of less than 0.05 as an indication of significance. (*p < 0.05, **p < 0.01, ***p < 0.001, ****p < 0.0001).

Results

The combination of an A2BR blockade with an immune checkpoint inhibitor significantly improved antitumor efficacy in two murine NSCLC models
To determine whether the combination of the A2BR antagonist PBF1129, and an anti-PD-1 antibody was more effective than either agent alone, mice were injected subcutaneously with Lewis lung carcinoma (LLC) cells or CMT167 cells and treated as indicated in Fig. 1A. As shown in individual mice (n = 10) (Supplementary Fig. 1A-B) and the average tumor volumes (Fig. 1B, C), the combination treatment was significantly more effective than the vehicle and either monotherapy alone in the LLC model. In fact, for the LLC model the monotherapies were not significantly different than vehicle. In the CMT167 model, both monotherapies and the combination were significantly more effective than the vehicle. While the combination was trending to be more effective than either monotherapy, a significant difference was observed between the combination and PBF1129 treatment in the time frame of the experiment. While the drug combination significantly reduced tumor volume in both models, the effect was greater in the LLC model. These results indicate that A2BR inhibition enhances the efficacy of an ICI.

ScRNA-seq reveals the cellular composition of the TME
To determine the cellular composition and transcriptional profile of each tumor, we employed fixed cell single-cell RNA sequencing. Individual cells were isolated from snap-frozen tumor tissue. For each treatment group, we selected four samples. Due to the strong response in LLC tumors treated with combination therapy, some samples had to be pooled to obtain a sufficient number of cells for analysis. Sample sizes were not a limiting factor in any of the CMT167 treatment groups. For LLC treatment groups, a total of 68,388 cells were studied. For CMT167 treatment groups, we obtained 99,889 cells (Supplementary Table 1). We performed unsupervised clustering analysis and visualized cell populations using UMAP. We classified tumor-associated cell clusters into eight broad cell types in the LLC model: Malignant cells, M2-like macrophages, Inflammatory cancer-associated fibroblasts (iCAFs), Endothelial cells, M1-like macrophages, T cells, Neutrophils and Myofibroblast-like cancer-associated fibroblasts (myCAFs). Cell type classification was based on the results of the singleR package and the expression profiles of specific marker genes (Fig. 2A, Supplementary Fig. 2A, Supplementary Table 1). A similar classification approach identified seven distinct cell types in the CMT167 model (Fig. 2B, Supplementary Fig. 2B, Supplementary Table 1): Malignant cells, M2-like macrophages, iCAFs, Endothelial cells, M1-like macrophages, T cells and myCAFs. In the LLC model, the M2-like macrophage population was significantly decreased in the combination therapy group compared to the PD-1 inhibitor-treated group, a reduced myCAFs population was observed comparing combination therapy to the control group (Fig. 2C). Other cell populations, excluding malignant cells, showed a general reduction in the combination group relative to the control group, though the differences were not statistically significant. In the CMT167 model, the most notable change was a significant increase in T cell numbers in the anti-PD-1 antibody and combination treatment groups compared to the control group (Fig. 2D). The proportions of all other cell types remained largely unchanged.

Impact of combination therapy on ECM remodeling, immune pathways, and apoptosis in malignant cells and T cells
Expression of the cytokeratin tumor marker genes Krt8 and Krt18 were used to identify malignant cells in both mouse models (Supplementary Fig. 2A-B). In the LLC model, we analyzed transcriptional changes in malignant cells and conducted a differential expression genes (DEGs) analysis across the four treatment groups (Fig. 3A, right). Comparing the control and combination therapy groups for malignant cells, we observed significantly higher expression of ECM-related genes in the control group, including Fn1, Col4a1, Col6a1, Col3a1, and Col4a2 (Fig. 3A, top left). Combination therapy resulted in significant downregulation of some of these genes, such as Col6a3 and Col3a1, along with other genes associated with ECM organization, including Emilin1 and Chpf (Fig. 3A, bottom left). mt-Nd6 was the top upregulated gene with a lowest adjusted p value found in the combination treatment of malignant cells in the LLC model and it is related to the production of ATP in the tricarboxylic acid cycle (TCA cycle) (Fig. 3A, bottom left). ECM organization, collagen formation, and carbohydrate metabolism were significantly suppressed in the combination therapy group compared to the control (Fig. 3B, left). Given the important role of TGFβ in ECM remodeling[13, 17], we analyzed the malignant cell combination versus control groups for downregulated genes in the ECM organization pathway that overlapped with those in the TGFβ receptor signaling complex. We found that expression of Tgfb1, Tgfb3, Ltbp2, Ltbp4, Fbn1, Furin, Itgb1, Itgb3, and Itgb8 was reduced in the combination therapy group, suggesting that TGFβ may play a crucial role in tumor progression and be mechanistically important for the efficacy of the combination therapy in the LLC model (Fig. 3B, right, Supplementary Table 2). We also observed that the ECM organization pathway was significantly decreased in malignant cells when comparing combination treatment to either PBF1129 treatment (Fig. 3C, up) or anti PD-1 treatment (Fig. 3C, bottom) in the LLC model.
In the CMT167 model, we determined the transcriptional signature across treatment groups for malignant cells (Fig. 3D). In the combination group we observed upregulated DEGs related to antigen presentation such as H2-K1 and Cd74 when compared to control. Upregulated DEGs such as mt-Nd4 and mt-Nd5 were also observed in the malignant cells of the combination treatment in the CMT167 model (Fig. 3D). Collagen biosynthesis and modifying enzymes, collagen formation, and ECM organization pathways were downregulated in malignant cells of the combination therapy group when compared to control (Fig. 3E, top). The upregulated DEGs in the combination therapy group were predominantly in immune response and antigen presentation pathways (Fig. 3E, bottom). Similar to the LLC model, combination treatment suppressed genes linked to TGFβ signaling in malignant cells and included Itgb3, Bmp2, Bmp4, and Ltbp4 (Fig. 3F, Supplementary Table 2). When comparing the combination treatment to either monotherapy, the pathways related to collagen formation were significantly downregulated (Supplementary Fig. 3A-C). To validate the ECM changes determined by our scRNA-seq analysis, specifically changes in collagen, we used picrosirius red staining to visualize collagen fibers in both LLC and CMT167 models. We found that the amount of collagen was significantly decreased in the combination treatment samples for both the LLC and CMT167 models (Fig. 3G). Several upregulated genes in the CMT167 model were associated with antigen processing and presentation, particularly for major histocompatibility complex (MHC) (Fig. 3H, left). These included genes involved in antigen processing such as proteasome components (Psmb8, Psmb10, Psme1), transporters (Tap1, Tap2), transmembrane proteins (Tapbp, Cybb) and genes involved in antigen presentation H2-K1, H2-Q6, H2-Q4, H2-M3 and B2m (Fig. 3H, left). Similarly, in the LLC model, H2-K1, H2-Q6, H2-Q4, and H2-M3 genes, which are involved in MHC presentation to T cells, were also increased in the combination therapy group (Fig. 3H, right). Antigen presentation scores, based on the expression of these genes, were significantly increased in both models (Fig. 3I, Supplementary Table 2). To evaluate the efficacy of formation of the immunological synapse between tumor and T cells, we analyzed the T cell activation score, TCR signaling score, and co-stimulation score via the CD28 family [22]. Compared to the control group, these scores were generally decreased in the combination therapy group, with the exception of the CD28 co-stimulation score in the CMT167 model, which was significantly increased in anti–PD-1-treated samples compared to control (Supplementary Fig. 3D-F, Supplementary Table 2). We assessed the expression of Bcl-2 family members to investigate apoptosis in T cells of both models. In the LLC model, combination therapy induced apoptosis, as indicated by decreased expression of anti-apoptotic markers (Mcl1, Bcl2l1, Bcl2) and increased expression of pro-apoptotic markers (Bid, Bax). In the CMT167 model, Mcl1 expression was notably reduced (Supplementary Fig. 3G).
To investigate the impact of treatments on the key processes promoting carcinoma progression such as epithelial-mesenchymal transition (EMT), inflammation, and angiogenesis [23–25], we assessed the respective scores in malignant cells from both models (Fig. 3J-L, Supplementary Table 2). These scores were significantly reduced in the PBF1129 monotherapy and combination therapy groups. We also evaluated the apoptosis score in malignant cells and observed enhanced apoptosis score in the combination treatment of both models (Fig. 3M, Supplementary Table 2). To validate the angiogenesis score, we performed immunohistochemistry with an anti-CD31 antibody, which recognizes endothelial cells. Consistently, we found that Cd31 staining in the tumors was significantly decreased in the combination group for both the LLC and CMT167 models (Fig. 3N).

Higher anti-tumor activity of M1 macrophages and distinct transcriptional profiles associated with various M2-like macrophage phenotypes were observed with PBF1129, anti PD-1 and combination treatments
Tumor associated macrophages (TAMs) are one of the cell components in the TME, which are composed of two distinct subtypes: M1-like TAMs and M2-like TAMs. Given the important role of M1-like macrophages in anti-tumor immunity, we examined their Ifng and Gzmb expression. Gzmb expression was significantly upregulated in the combination therapy group in both models, indicating enhanced cytotoxic capacity (Supplementary Fig. 4A). These findings suggest M1-like macrophages maintain robust cytotoxic activity, contributing to sustained anti-tumor effects.
M2-like TAMs are associated with malignant metastasis, invasion and treatment resistance [26]. A recent synthesis of multiple single-cell studies has proposed a consensus model of TAM diversity [27, 28]. In our dataset, we observed the presence of some of these TAM subtypes and performed reclustering of M2-like macrophages in the LLC model to examine treatment-related changes in their composition. Three distinct M2-like macrophage subclusters were identified: immune-regulatory (Reg-M2-like), proangiogenic (Angio-M2-like), proliferating (Prolif-M2-like) (Fig. 4A, left). Among these, Prolif-M2-like and Undefined clusters were present in low numbers and did not show significant changes across treatment groups (Fig. 4A, right). Based on gene expression profiles [27, 28], Reg-M2-like were characterized by high expression of Trem2, Apoe, Arg1, and Tgfb1, upregulated genes associated with antigen presentation as H2-K1, Cd74 and enhanced expression of established Reg-M2-like markers [29–31], including C1qa, Pf4, Mrc1, and Adgre1 (Supplementary Fig. 4B). Angio-M2-like [27, 32], which are associated with angiogenesis, CAF interactions, and ECM proteolysis, were marked by significant expression of Vcan and Vegfa in the LLC model (Supplementary Fig. 4C). Consistent with the human Angio-TAM gene signature, which includes THBS1, IL1B, and VEGFA, we observed increased expression of Thbs3, Vegfd, and Il1r1 in our dataset (Supplementary Fig. 4C). Additional genes involved in cell–cell interactions and TGFβ signaling [33, 34], such as Sema4c, Fgfr1, and Ltbp4, were also enriched in this cluster (Supplementary Fig. 4C). Prolif-M2-like displayed proliferative properties [35], characterized by high expression of Top2a and Cdk1, as well as a predominance of cells in the G2/S phase based on cell cycle scoring (Supplementary Fig. 4D). Due to the low number of cells in the Undefined cluster, we performed pathway enrichment analysis based on upregulated DEGs. These pathways were primarily related to signal transduction, molecular transport, and cell motility, suggesting a potential migratory phenotype, though further validation is needed (Supplementary Fig. 4E). To functionally validate the subtypes, we analyzed pathway-specific gene signatures across M2-like macrophage subtypes in the LLC model (Supplementary Fig. 4F). Angio-M2-like macrophages showed enhanced TGFβ, VEGF, and hypoxia signaling, while Prolif-M2-like macrophages exhibited reduced p53 signaling relative to other subtypes. TGFβ signaling is known to reduce inflammation and protect tissues from damage, whereas downregulated p53 signaling supports active cell division [36, 37]. Trajectory analysis was used to characterize the development of cell types[38] and was applied to M2-like macrophages subtypes in the LLC model to further support our hypothesis (Supplementary Fig. 4G). We found that Lyz2, a myeloid marker commonly used in Lyz2-Cre models to target macrophages, was more highly expressed in Reg-M2-like, underscoring their immunoregulatory identity. Additionally, the high expression of Col3a1, an ECM component, in Angio-M2-like aligned with their matrix remodeling function. Elevated Top2a expression in Prolif-M2-like confirmed their proliferative status.
Importantly, the proportion of Reg-M2-like significantly decreased in the combination group compared to the anti–PD-1 and control groups, while Angio-M2-like increased significantly (Fig. 4A). Based on these results, we focused our downstream analysis primarily on Reg-M2-like and Angio-M2-like in the LLC model. Analysis of DEGs revealed distinct expression patterns and diversity within each treatment group (Fig. 4B). Based on the DEGs in Fig. 4B, we performed a pathway analysis to distinguish the differences between Reg-M2-like and Angio-M2-like. A Venn diagram illustrating suppressed genes in the combination group from Fig. 4B showed distinct downregulated genes in Reg-M2-like (236 genes) and Angio-M2-like macrophages (159 genes), with 153 genes shared between both subtypes (Fig. 4C). Distinct downregulated pathways in Reg-M2-like were primarily related to TGFβ signaling and immune responses, while ECM organization and collagen formation pathways were reduced in Angio-M2-like (Fig. 4C) and were observed in shared 153 genes (Supplementary Fig. 4H). Analysis of common downregulated genes revealed suppression of Fn1, Apoe, Col3a1, Lrp1, C1qb, Mpeg1, Stab1, Lgmn, App, and C1qa in both macrophage subtypes (Supplementary Table 3). Receptor tyrosine kinase (RTK) signaling was also suppressed in both subtypes, with Fn1, Col3a1, Atp6v0b, Spp1, and Ctsd among the top downregulated genes (Supplementary Table 3). Upregulated DEGs in the combination therapy group included mt-Nd6, Gbp4, Gbp8, Pdcd1lg2, Gbp2, and Ligp1, which were common to both macrophage subtypes (Supplementary Fig. 4I). In Reg-M2-like, genes associated with interferon signaling (G0s2, Zbp1, Acod1, Gbp2b, Hdc, Phf20l1) were significantly increased in the combination group (Supplementary Fig. 4I). In Angio-M2-like, upregulated genes were enriched in RNA metabolism and interleukin signaling (Myc, Hmox1, Hsp90b1, Il1b, Cxcl1, Cxcl2) (Supplementary Table 3). Pathway analysis showed weakened TGFβ signaling in Reg-M2-like with significant reductions in expression of Tgfb1, Pmepa1, Tgfbr3, Tgfbr1, Itgav, and Cdk8 (Supplementary Table 3, Fig. 4C). Notably, in the volcano plot, Reg-M2-like macrophages exhibited increased expression of Lat2, Zbp1, Fyb, and Ccl7 in the combination group compared to control, while Angio-M2-like macrophages upregulated genes such as Gbp8, Fosl1, suggesting chemokine and cytokine production (Fig. 4D). Both subtypes demonstrated downregulated DEGs such as Apoe, C1qa, Lgmn and Mpeg1 with combination treatment (Fig. 4D). The UMAPs and violin plot demonstrated the decreased expression of Tgfb1 in Reg-M2-like macrophages when comparing the combination treatment to control in LLC model (Fig. 4E). Interestingly, the decreased expression of Tgfb1 also occurred with PBF1129 as a single agent.
In the CMT167 model, reclustering analysis identified eight distinct M2-like macrophage subpopulations based on DEGs profiles (Supplementary Fig. 4 J). We categorized three major subpopulations: immune regulatory M2-like macrophages (Reg-M2-like), proangiogenic M2-like macrophages (Angio-M2-like), and proliferative M2-like macrophages (Prolif-M2-like) (Fig. 4F). Using pathway-responsive gene signatures and trajectory analysis, we further characterized each cluster and found a pattern similar to that observed in the LLC model (Supplementary Fig. 4 K, L). Key genes highly expressed in Reg-M2-like included C1qa, Trem2, Tgfbi, and Cd74 (Fig. 4G). Notably, unlike the LLC model, the proportion of Reg-M2-like was significantly increased in the combination therapy group compared to other treatments (Fig. 4F). However, like the LLC model, Tgfb1 expression was significantly reduced following combination treatment (Supplementary Fig. 4 M). Additionally, Tgfbi expression was significantly decreased in the LLC model but remained unchanged in the CMT167 model (Supplementary Fig. 4N). Interestingly, MHC genes such as Ciita, H2-Aa, H2-Q6, H2-Ab1, H2-K1, and H2-Eb1 were upregulated in Reg-M2-like in the CMT167 model, while MHC genes including H2-DMa, Cd244a, and Cd300lf were downregulated in the LLC model (Supplementary Fig. 4O, P). Vcan and Vegfa remained the upregulated DEGs in Angio-M2-like macrophages in the CMT167 model (Fig. 4H). Among the top downregulated genes in Reg-M2-like macrophages (combination treatment vs. control) were those associated with ECM organization, such as Mmp9, Mmp12, Spp1, Ctsl, Col1a1, Fn1, and Emilin2. Additionally, decreased expression of Gsr and Lpl was noted, implicating reduced metabolic activity. Notably, Lpl interacts with Apoe and facilitates lipid binding and uptake, which was highly downregulated in Reg-M2-like macrophages with combination treatment (Fig. 4I). In contrast, top upregulated genes in Reg-M2-like macrophages after combination therapy included immune checkpoint genes Cd274, Pdcd1lg2 and Gbp2, a marker of interferon signaling. Enhanced expression of H2-Aa and Ciita indicated activation of antigen presentation pathways (Fig. 4I). Consistently, the downregulated genes in Reg-M2-like macrophages were enriched in the top 10 pathways, most of which were associated with ECM organization while upregulated genes were almost enriched in immune system such as PD-1 signaling (Fig. 4J).
These findings show that combination therapy suppresses TGFβ signaling in Reg-M2-like and inhibits ECM organization in Angio-M2-like, while simultaneously promoting interferon signaling in both subtypes. The activation of antigen presentation pathways was observed in the CMT167 model, but not in the LLC model. Furthermore, anti-tumor activity may be mediated by Angio-M2-like macrophages through interferon signaling and chemokine production. We performed a pseudotime analysis in Reg-M2 and Angio-M2 in both models and demonstrated the gene expression by using Monocle3 (Supplementary Fig. 4Q-R) to generate hypotheses about how these features evolve over time.

T cell population as signal receiver in the cell–cell interactions in the TME
To explore the global intercellular communication within the tumor microenvironment (TME), we used CellChat, a tool designed to quantitatively infer and analyze intercellular communication networks from scRNA-seq data[19]. CellChat outputs patterns of signaling between cell groups, considering either outgoing signals (cells as sender/ligand source) or incoming signals (cells as receiver/receptor source)[19].
In the LLC model, we observed reduced communication between malignant cells (sender) and T cells (receiver) as well as between M2-like macrophages (sender) and T cells (receiver), in both the PBF1129 and combination treatment groups compared to control (Supplementary Fig. 5A, left). We further examined interactions involving M1-like macrophages or neutrophils as sender targeting T cells (Supplementary Fig. 5A, left). Unlike the patterns seen with malignant cells and M2-like macrophages, the interactions between M1-like macrophages or neutrophils and T cells were largely preserved in the combination group (Supplementary Fig. 5A, left), and no significant changes in ligand–receptor pairs were detected. We performed a similar analysis in the CMT167 model, examining the crosstalk between malignant cells, M2-like macrophages, and M1-like macrophages (sender) targeting T cells (receiver) (Supplementary Fig. 5A, right). Consistent with the LLC model, cell–cell interactions between malignant cells and T cells were reduced in the combination group compared to anti–PD-1 treatment. A similar reduction was seen in M2-like macrophage-T cell communication. However, signaling between M1-like macrophages and T cells remained intact (Supplementary Fig. 5A, right). To understand the molecular mechanisms underlying these changes, we analyzed significantly upregulated and downregulated ligand–receptor pairs in T cells (receiver) comparing PBF1129 treatment to control in both models (Fig. 5A, B). In both LLC and CMT167, collagens–CD44 signaling was reduced in PBF1129 treatment. In the LLC model, integrin signaling was also downregulated in T cells (Fig. 5A). Previous studies have shown that CD44 promotes Treg function by enhancing FOXP3 and TGF expression[39], and integrin ablation can impair TGFβ signaling in Tregs, leading to increased cytotoxic T cell activation [40]. These results suggest that TGFβ signaling is reduced in T cells under PBF1129 treatment in both models. Despite the overall reduction in TGFβ-associated pathways, we identified upregulated signaling in T cells under PBF1129 treatment, including Thbs3–Cd47 and Fgf7–Fgfr1 in the LLC model (Fig. 5A), and Lama4–Cd44 in the CMT167 model (Fig. 5B). CD47, which delivers a “don’t eat me” signal, has been associated with T cell antitumor immunity and is a potential target in cancer immunotherapy [41, 42].
To assess the effect of combination therapy, we compared increased and decreased signaling in T cells versus control, CD44 and integrin associated signaling were downregulated in T cells in both models (Fig. 5C, D). Consistent with PBF1129 monotherapy, Fgf7–Fgfr1 signaling remained enhanced in T cells under combination therapy in the LLC model (Fig. 5C). Additionally, Sema4c–Plxnb2 signaling, another pathway upregulated in T cells under combination treatment, was identified (Fig. 5C). Previous studies have shown that FGFR1 signaling can suppress TGFβ signaling and that the Sema4c–Plxnb2 axis is essential for optimal T cell activation[43, 44]. Interestingly, Jam2–(Itga3 + Itgb1) and Jam2–(Itgav + Itgb1) signaling were upregulated specifically in the CMT167 model in combination treatment between endothelial cells and T cells (Fig. 5D). Furthermore, Cd274–Pdcd1 signaling between M1/M2-like macrophages and T cells was increased (Fig. 5D).
Finally, we investigated differentially regulated ligand–receptor pairs with defined sender cell types, comparing either combination or PBF1129 treatment to control (Supplementary Fig. 5C–L). Remarkably, combination therapy resulted in a general reduction of signaling events, particularly integrin signaling, across nearly all cell types. In contrast, Fgf7–Fgfr1 and Sema4c–Plxnb2 signaling pathways were consistently upregulated not only in T cells but also in other cell types in the LLC model under combination treatment (Supplementary Fig. 5C–L).

Combination treatment disrupts interactions of T cells with M2 macrophages and stromal cells mediated by TGFβ signaling
Given the central role of macrophages in mediating TGFβ signaling and ECM remodeling [45], we next explored how combination therapy affects broader cell–cell interactions in the TME, particularly the crosstalk between macrophages, CAFs, and T cells. Focusing on TGFβ ligand-receptor pairs, we found strong evidence that TGFβ plays a crucial role in immune cell-immune cell and immune cell-stromal cell communication. In terms of the three TGFβ ligand genes (Tgfβ1, Tgfβ2, and Tgfβ3), Tgfβ1 was highly expressed by M2-like macrophages among the cell types and decreased in the combination treatment group compared to control in both models (Fig. 6A, B).
We further examined the expression of TGFβ receptors (Tgfbr1, Tgfbr2, Acvr1, Acvr1b, Acvr1c) across different cell types. In the LLC model, combination therapy led to a decrease in Tgfbr1 expression in T cells and CAFs (iCAFs and myCAFs) (Fig. 6C). However, certain receptors were upregulated, including Tgfbr2 in CAFs and Acvr1b in T cells. Acvr1 expression was induced in T cells, iCAFs, and myCAFs (Fig. 6C). In the CMT167 model, combination treatment resulted in decreased Tgfbr2 expression across all receiving cells, whereas Acvr1 and Acvr1b were upregulated. Tgfbr1 expression was reduced in T cells and myCAFs, but was enhanced in iCAFs with combination treatment (Fig. 6D). Additionally, other ligands corresponding to TGFβ receptors were identified in different cell types when comparing combination treatment to control (Supplementary Fig. 6A-B). In general, with a few exceptions, expression of most of the ligands was reduced in the combination treatment group when compared to control.
To infer the effects of sender-cell ligands on receiver-cell target gene expression, we employed NicheNet analysis[46], which is a tool to predict the affected downstream target genes in receiver-cell that correspond to the anticipated ligands from sender-cell. In the LLC model, the top 10 ligands when iCAFs, myCAFs and T cells as the receiver and were demonstrated in the dot plots (Fig. 6E, left panel). When comparing control to combination treatment, Tgfb1 ligand expression from M2-like macrophages (sender) had a strong influence on the downstream genes of Tgfb1 in T cells, iCAFs, and myCAFs (receiver) (Fig. 6E, right panel). Similarly, in the CMT167 model, Tgfb1 ligand expression in M2-like macrophages was predicted to mediate target gene expression changes in T cells (Fig. 6F). Additionally, when iCAFs and myCAFs were examined as receiver-cell in the CMT167 model, Il1b and Il1a emerged as key ligands that most affected gene expression (Supplementary Fig. 6C). These cytokines were highly expressed in M1/M2-like macrophages (Supplementary Fig. 6C), indicating their potential role in macrophage-CAF interactions.
In line with these findings of lower Tgfb1 expression in the combination treatment group, we observed a reduction in the expression of exhaustion-related genes, leading to a significantly decreased T cell exhaustion score in the combination treatment compared to control in the LLC model (Fig. 6G). In the CMT167 model, while the exhaustion score was not significantly different between control and combination groups, Lag3, Cd244a, Cd160, and Ctla4 expression was reduced in the combination group compared to control, and the exhaustion score was significantly lower in the combination group compared to anti-PD-1 antibody treatment (Fig. 6G). Using NicheNet analysis, we identified target genes influenced by Tgfb1 expression in T cells and CAFs in the LLC model. These genes were enriched in pathways related to ECM organization, collagen formation, and RTK signaling (Fig. 6H, top). Within the ECM organization pathway, genes such as Sparc, Fn1, Col5a1, Col4a2, Col4a1, and Col1a1 were downregulated in iCAFs and myCAFs following combination treatment (Fig. 6H, middle). In T cells, the influenced ECM organization pathway was still observed (Fig. 6H, bottom left) and ECM-related genes such as Vcan, Tnc, Sdc2, Fn1, Eln, Col5a1, and Col1a1 were upregulated (Fig. 6H, bottom right). A similar pathway enrichment pattern was observed in the CMT167 model, with ECM-related genes as Col4a2, Col4a1, and Col1a1 downregulated in T cells after combination treatment (Fig. 6I). We also examined the target genes influenced by Tgfb1 enriched in RTK signaling. We observed that the genes within RTK signaling overlapped with genes in ECM organization, such as Fn1, Col5a1, Col4a2, Col4a1, and Col1a1 in CAFs and T cells of the LLC model and Thbs1, Col4a2, Col4a1, and Col1a1 in T cells of the CMT167 model (Supplementary Fig. 6D-E). These data suggest that combination therapy disrupts TGFβ-mediated interactions between M2-like macrophages and immune cells or stroma cells in the TME, leading to reduced T cell exhaustion and ECM remodeling. While both LLC and CMT167 models exhibited reductions in TGFβ signaling, subtle differences in receptor expression patterns and ECM-related gene expression suggest that the therapy exerts context-specific effects on stromal and immune cell interactions.

Discussion
In this study, we demonstrated that combination therapy of PBF1129 and PD-1 inhibitor significantly improved the antitumor efficacy compared to anti-PD-1 alone and is associated with consistent reductions in TGFβ signaling, which alters cell–cell interactions within the tumor microenvironment (TME). These findings highlight the therapeutic potential of targeting adenosine signaling in NSCLC and define a candidate mechanism.
Decreased TGFβ signaling in malignant cells alleviates TGFβ-driven epithelial-to-mesenchymal transition (EMT) and weakens immune evasion by attenuating the expression of integrins, such as Itgb3 [47]. The reduction in TGFβ signaling genes in malignant cells may also alter interactions with other cells in the TME to shape tumor immunity. We examined Tgfb1 expression in both models and found it significantly reduced in malignant cells comparing combination treatment to either PBF1129 treatment or control (data not shown). In both models, gene expression in pathways related to ECM organization or collagen formation were also downregulated in malignant cells when comparing combination treatment to either PBF1129 treatment or control. ECM proteins, such as collagen, laminin and fibronectin are known to promote migration and progression of cancer [48–52]. These changes were correlated with decreases in TGFβ signaling [13]. Evidence shows that other signaling pathways observed to change with therapy may also interact with TGFβ signaling to mediate decreased matrix deposition, such as MAPK, Wnt, PAR2/PAR1, etc. [53–56]. Reduced angiogenesis and EMT scores, along with diminished ECM pathways, suggest that combination treatment may hold promise in suppressing metastasis, and providing a possible mechanistic explanation for the previously observed therapeutic efficacy of combination of PBF1129 with anti-VEGF antibody [7].
With the combination treatment, we observed upregulation of several mitochondrial encoded genes for the respiratory electron transport chain (ETC) in the malignant cell population for both models. In the LLC model we saw an increase in mt-Nd6 expression, and in the CMT167 model we observed increased expression of mt-Nd4 and mt-Nd5. The increased mitochondrial ETC genes we observed in malignant cells correlated to immune response have previously been shown to correspond to alterations in oxidative phosphorylation and ROS and MHC molecules in antigen presentation [57–59]. MHC expression was shown to be enhanced in the combination treatment in both models, suggesting that antigen presentation and T cell cytotoxicity were promoted. Increased antigen presentation not only prevents immune evasion by malignant cells but also enhances the efficacy of immunotherapy [60]. Furthermore, we identified upregulated DEGs of malignant cells enriched in antigen presentation pathway in the CMT167 model. Different molecular, immunological and tumor microenvironment features of the two models used likely account for pronounced T cell infiltration in the CMT167 model, and our current models provide the foundation in other murine and human tumors [18, 61]. The reasons why the monotherapies were not significantly different than vehicle may be due in part to the lower immunotherapy sensitivity and increased expression of Adora2b and Cd73 under anti PD-1 or PBF1129 in the T cells of LLC model (Supplementary Fig. 7). Together, these observations indicate that LLC tumors represent a more fibrotic and immune-excluded state characterized by elevated EMT, angiogenesis, and ECM deposition, whereas CMT167 tumors exhibit a more immune-inflamed and immunogenic microenvironment with greater baseline T cell infiltration.
Using CellChat, software used to analyze cell–cell interactions, we observed that when T cells acted as the receiver and interacted with other cell types, both Cd44 and integrin signaling were significantly downregulated in the combination treatment across both models. Since Cd44 and integrin receptors are known to activate TGFβ signaling, a decrease in these signaling pathways is consistent with decreased TGFβ signaling in T cells in both models [62, 63]. Whether comparing PBF1129 treatment to control or combination treatment to control, we found increased Fgf7-Fgfr1 signaling in T cells as well as nearly all cell types interacting with iCAFs in the LLC model. Increased Fgf7 expression in iCAFs was also observed with combination treatment (data not shown). Evidence suggests that TGFβ signaling is deactivated in high Fgf7 expressing CAFs, further validating our previous observation [43]. TGFβ can inhibit Fgf7 expression [64], so higher Fgf7 may be due to the reduction of TGFβ. We also identified increased Thbs3-Cd47 signaling in CAFs of the LLC model and increased Spp1-(Itga5 + Itgb1) signaling in CAFs of the CMT167 model under PBF1129 treatment, while these signals were diminished with the combination treatment in both an autocrine and a paracrine manner. CD47 and integrins may mediate ECM production in CAFs through TGFβ signaling, as supported by related studies [63, 65]. Collagen, one of the ECM proteins, was significantly decreased in iCAFs under combination treatment. This indicates that while PBF1129 may increase ECM expression to some extent, the combination therapy appears to be more effective in tumor suppression than PBF1129 monotherapy due to decreased ECM. Future time-course studies would be needed to determine if this is a causal relationship [66]. We also observed inactivation of RTK signaling in CAFs and T cells in the combination treatment, corresponding to reduced Tgfb1 ligand expression. Gene analysis within the RTK pathway revealed extensive crosstalk with ECM components, suggesting that ECM genes may act as upstream regulators or mediators of RTK signaling. Discoidin domain RTKs (DDRs), a subclass of RTKs, interact with ECM molecules such as collagens [67], while integrins share common oncogenic signaling pathways with RTKs [68, 69]. This suggests that combining PBF1129 with an RTK inhibitor could be a promising therapeutic strategy, with one of the mechanisms being due to decreased ECM. This approach might be particularly useful for cancer patients harboring driver gene mutations, which is consistent with the observation that tumor suppression was observed using an A2BR inhibitor in combination with erlotinib in a mouse model [7].
TGFβ signaling is also known to suppress immune responses through multiple mechanisms [36, 70, 71]. However, complete inhibition of TGFβ signaling disrupts adaptive immunity, posing a challenge for developing effective anti-TGFβ therapies in cancer treatment [72]. Demonstrating that TGFβ pathway re-activation rescues tumor growth would formally prove its role in the observed effects, but is difficult to robustly accomplish experimentally. The fact that the A2BR seems to be active only in the tumor microenvironment due to its high IC50 and the high extracellular adenosine concentrations specifically observed there, may be a way of specifically targeting TGFβ signaling there and not elsewhere. Despite the potential challenges in targeting TGFβ, the importance of the adenosine-TGFβ axis has been recognized and several therapies targeting both pathways have been developed and are currently in clinical trials. An anti-CD73-TGFβ-trap bifunctional antibody has been tested in a phase I first-in-human study involving patients with advanced solid tumors. This study informs potential therapeutic interventions in cancer treatment and suggests the promising anti-tumor effects of bispecific antibodies targeting A2BR-TGFβ in future drug development [73]. Targeting A2AR via CRISPR/Cas9 significantly enhanced in vivo efficacy and increased IFNγ production in A2AR-edited CAR T cells [74]. Additionally, dual A2AR and A2BR antagonism has been shown to enhance immune responses by activating T cells and promoting antigen-presenting cell function [75]. The deletion of A2BR in T cells exerts anti-tumor effects by adoptive transfer but there are few data on the role of A2BR in host T cells of TME on immune responses [76]. This could be addressed using PD-1 deficient hosts or CD8 + T cells depletion in future studies.
Adora2b was expressed at a higher level than Adora2a among all cell components, including M2-like macrophages in both models. In addition, A2BR is only active in high adenosine environments like that found in the TME, suggesting that it is A2BR rather than A2AR that plays an important role in TGFβ signaling in M2-like macrophages, especially Reg-M2-like macrophages. M2-like macrophages produce extracellular matrix proteins and TGFβ to enhance tumor cell invasion [77], and decreased ECM pathway activity with combination therapy may contribute to reduced interactions with other cell components, facilitating T cell migration to execute anti-tumor functions [49]. In the LLC model, combination treatment also led to a reduction in T cell exhaustion scores, possibly due to decreased TGFβ1 signaling and reduced ECM deposition by M2-like macrophages. Both decreased exhaustion score and increased pro-apoptotic score in T cells under combination treatment suggest different subsets and activation states of T cells, however, the limited numbers of identified T cells limited our ability to analyze this apparent paradox. This limitation may be solved by multiple time point analysis in the evolution of tumor growth under treatments in future studies and our study provides the foundation for future validation and examination of dynamics of cellular and molecular alterations of such as TGFβ suppression and ECM remodeling in response to A2BR inhibition. The reduced T cell exhaustion score under PBF1129 treatment in LLC model was consistent with the phenotype in human NSCLC peripheral blood mononuclear cells (PBMCs) samples after treatment with PBF-1129 [78]. More detailed evaluation of post-treatment human tumor samples would be valuable to show whether reduced TGFβ signaling and ECM remodeling is observed. We found increased DEGs in Reg-M2-like macrophages associated with interferon signaling [79, 80] after combination treatment, indicating Reg-M2-like macrophages may exert anti-tumor function by secreting cytotoxic cytokines. We also observed enhanced processes of RNA splicing or RNA metabolism in Angio-M2-like macrophages in the CMT167 model in combination treatment. Evidence indicates that cytokines can influence RNA splicing patterns, which play a role in immune response and limiting inflammation [81]. This is consistent with our observation of decreased inflammation in malignant cells and enhanced interferon signaling in Angio-M2-like macrophages in combination treatment of LLC model, while significant increased Il1b derived from Angio-M2-like macrophages indicate chronic inflammation which may drive tumor progression [82]. Decreased Reg-M2 macrophages in the LLC model but increased in the CMT167 model is perplexing, but the shifts observed involving Reg-M2 and Angio-M2 macrophages may indicate that Reg-M2 macrophages play an important role in controlling and balancing the inflammatory response to exert anti-tumor effects but not induce immunopathology under combination treatment [83]. A direct molecular link between A2BR blockade and regulation of TGFβ in macrophages is hard to define. Our data show that it is likely mediated by Reg-M2 and Angio-M2 macrophages, sustaining and prolonging the active immune response. Montalbán del Barrio et al. demonstrated that M2-like TAMs would increase T cell proliferation in response to CD39/CD73 inhibitors in a co-culture system, which supports the hypothesized role of Reg-M2 in our study and provides an avenue for further studies [84].
Our study provides insights into the therapeutic potential of adenosine A2B receptor blockade (PBF1129) combined with anti-PD-1 therapy in NSCLC. It would be supportive in future studies to show that the A2BR knockout (Adora2b-/-) phenocopies PBF1129. Since PBF1129 is a truly selective A2B adenosine receptor antagonist over other adenosine receptors, a role for compensatory activation of A2AR cannot be ruled out, but other studies have shown this not to be the case [7]. It should also be noted that this study was conducted exclusively in female mice, future studies are warranted to determine if the identified different cell populations within the tumor microenvironment is different in males. Our study has implications on clinical strategies in human NSCLC, but also underscores the effects of inter-tumoral heterogeneity and further work will be required to explore the efficacy of this combination in all human NSCLC subsets. By utilizing scRNA-seq, we elucidated mechanisms underlying the efficacy of this combination, including modulation of TGFβ signaling, ECM interactions, and immune responses. These findings support the further development of combination strategies to enhance immune checkpoint blockade therapy.

Supplementary Information
Below is the link to the electronic supplementary material.