Tumor-derived erythropoietin acts as an immunosuppressive switch in cancer immunity

Abstract

Successful cancer immunotherapy requires a patient to mount an effective immune response against tumors; however many cancers evade the body’s immune system. To investigate the basis for treatment failure, we examined spontaneous mouse models of hepatocellular carcinoma (HCC) with either an inflamed T-cell-rich or non-inflamed T-cell-deprived tumor microenvironment (TME). Our studies reveal that erythropoietin (EPO) secreted by tumor cells determines tumor immunotype. Tumor-derived EPO autonomously generates a non-inflamed TME by interacting with its cognate receptor EPOR on tumor-associated macrophages (TAMs). EPO signaling prompts TAMs to become immunoregulatory via NRF2-mediated heme depletion. Removing either tumor-derived EPO or EPOR on TAMs leads to an inflamed TME and tumor regression independent of genotype, due to augmented antitumor T-cell immunity. Thus, the EPO/EPOR axis functions as an immunosuppressive switch for anti-tumor immunity.

Endogenous antitumor immune responses in patients with cancer can result in “inflamed” tumors that contain substantial numbers of tumor reactive T cells. While these responses are relatively uncommon and rarely clinically sufficient (1, 2), a proportion of such patients experience significant benefits from immunotherapy using either anti-PD-1 or anti-PD-L1 antibodies. Targeting PD-1/PDL-1 functions by activating and expanding subsets of exhausted anti-tumor T cells located primarily in tumor sites and peripheral lymphoid organs (3, 4). Unfortunately, the majority of cancer patients have “non-inflamed” T-cell-restricted tumors and are resistant to checkpoint blockade immunotherapy (2). These non-inflamed tumors are often replete with immunosuppressive macrophages and neutrophils that hinder T cell priming, activation, and homing — critical processes for fostering antitumor immunity (5, 6). However, the mechanisms that determine the immune cell profile or immunotype of tumors remain poorly understood.

Genetic alterations within tumors can significantly influence their immune profiles. In particular, the tumor mutational burden (TMB), defined as the total number of mutations in the tumor genome, has been proposed as a biomarker due to its potential to generate immunogenic neoantigens that elicit T-cell responses. Early studies revealed positive correlations between TMB and anti-PD-1 responses across various cancer types (7), suggesting that a higher neoantigen load promotes CD8⁺ T-cell infiltration and enhances responsiveness to PD-1 inhibition. However, data from larger patient cohorts indicate that while cancer types such as melanoma, lung, and bladder exhibit strong associations between TMB, immunotypes, and responsiveness to anti-PD-1 immunotherapy, this correlation is surprisingly weak in others, including breast, liver, prostate, colon and glioma. These findings suggest that additional factors govern antitumor immune responses, underscoring the need for further investigation into the determinants of tumor immunotypes (8). Specific gene mutations can also impact chemokine and cytokine secretion patterns. For instance, TP53 deficiency increases CXCL17 expression, attracting neutrophils, whereas KEAP1 deficiency raises CXCL10 expression, recruiting T cells (9, 10). Although these mechanisms may contribute to shaping the tumor immunotype, their influence is, by definition, modest since most tumors are non-inflamed and myeloid-cell-rich (11). Therefore, other mechanisms must be responsible for the failure of patients to mount a more pronounced adaptive immune response.

Results

Non-inflamed tumors preferentially secrete EPO

To investigate why some tumors fail to mount a pronounced adaptive immune response, we developed a series of spontaneous tumor models, each with a defined inflamed or non-inflamed TME. We utilized genome-editing approaches, which combined the delivery of transposase-transposon vectors for Myc overexpression (Myc^OE) and CRISPR-Cas9 systems by hydrodynamic tail vein injection (HDTV) to mice (10), to generate various spontaneous hepatocellular carcinoma (HCC) models with specific tumor suppressor deficiencies that mimic different mutational subtypes of human HCC. By recapitulating these subtypes, including their immunotypes and responses to anti-PD-1 immunotherapy, this model offers a robust framework for investigating the development of the tumor immune landscape. In accord with their human HCC counterparts, Trp53^KO and Pten^KO tumors exhibited the traits associated with “non-inflamed” tumors, including low infiltration of immune cells, impaired activation and accumulation of CD8⁺ T cells, as well as resistance to PD-1 blockade. By contrast, “inflamed” Keap1^KO tumors were characterized by an abundance of tumor-infiltrating CD8⁺ effector memory T cells (T_EM) and were highly responsive to PD-1 blockade (Fig. 1A–B and fig. S1). Unexpectedly, splenomegaly and elevation of plasma EPO were exclusively observed in both non-inflamed tumor models (Fig. 1C), even in the absence of anemia (fig. S2) and regardless of tumor size. Hemoglobin (HGB), hematocrit (HCT) and red blood cell count (RBC) levels were slightly increased in mice with non-inflamed tumors, likely reflecting the erythropoietic effect of EPO (Fig. 1D). This raises the possibility that EPO, independent of tumor-associated anemia, may act as a mediator in creating an immunosuppressive or non-inflamed TME.

Fig. 1. Tumor-secreted EPO autonomously establishes a non-inflamed tumor microenvironment.

(A) Generation of a spontaneous HCC model by in vivo delivery of plasmids pCMV-SB13, pT3-EF1a-Myc and pX330-sgRNA targeting Trp53, Pten, Keap1 to mouse liver via HDTV. 5 weeks after HDTV, HCC tumors were harvested and the immune cell compositions were determined by flow cytometry (normal control, n = 5; experimental groups, n = 10–12/group of 2 independent experiments). (B) 2 weeks after HDTV, C57BL/6 WT mice were injected intraperitoneally (IP) with 2 mg/kg of αPD-1 (RMP1–14) or IgG Isotype control every 3 days (total 5 doses). Overall survival was measured (n = 7–8/group). (C) 5 weeks after HDTV, spleen weight and plasma EPO concentration of tumor-bearing mice were analyzed (normal control, n = 5; experimental groups, n = 6–19/group of 2 independent experiments). (D) 4 weeks after HDTV, blood samples were collected from tumor-bearing mice with comparable tumor burdens and analyzed for complete blood count (n = 5/group). The levels of RBC, HGB, and HCT are presented. (E-F) Correlation of EPO mRNA expression (EPO^high, upper quartile; EPO^low, lower quartile) with (E) 5-year overall survival and (F) immune composition in HCC patients (TCGA and LIRI-JR). (G-I) pX333 vector has two independent U6 promoters, which allow for the expression of two sgRNAs. pX333-sgTrp53-sgEpo was generated to render Trp53^KO tumors unable to secrete Epo. (G) 4 weeks after HDTV, blood samples were collected from tumor-bearing mice with comparable tumor burdens and measured for plasma EPO concentration (n = 6/group). 5 weeks after HDTV, Trp53^KO/Myc^OE and Trp53^KO/Epo^KO/Myc^OE HCC tumors were harvested for (H) size measurement and (I) intratumoral immune cell profiling (n = 6–7/group). (J-M) Regressive HCC model: C57BL/6 mice were orthotopically implanted with 3 × 10⁶ allogeneic Hepa1–6 cells. The tumors grew continuously for two weeks, and then spontaneously regressed, resulting in either complete or partial regression. Hepa1–6_EV (empty vector) and Hepa1–6_Epo^OE (EPO-overexpressing) tumors were harvested on Day 14 and Day 21 (normal control, n = 5; experimental groups, n = 8–11/group of 2 independent experiments) for (J) determination of tumor size and regression rate (CR: complete regression; PR: partial regression; NR: no regression), and (K-M) intratumoral immune cell profiling. (N) On days 14, 17 and 20 after orthotopic implantation (Hepa1–6_Epo^OE), mice were injected IP with 2 mg/kg of αCD25, αCTLA-4, αCCR8 or IgG control. Tumors were harvested on day 21 (n = 5/group). One-way ANOVA with Tukey’s multiple comparison test for (A), (C), (L right), (M right), (N); Log-rank (Mantel-Cox) test for (B), (E); Two-way ANOVA with Tukey’s multiple comparison test for (D); Unpaired t test for (G), (H), (I), (J left), (K), (L left), (M left); Fisher’s exact test for (J right); Mann-Whitney test for (F). In all panels, data are presented as violin plot with median and quartiles, or as mean ± SD. *P ≤ 0.05; **P ≤ 0.01; ***P ≤ 0.001, ****P ≤ 0.0001. Comparisons with P > 0.05 (ns, not significant) are not displayed.

EPO, a glycoprotein hormone primarily recognized for its role in stimulating red blood cell production, has recently been found to mediate other biological functions. For instance, EPO promotes efferocytosis in macrophages (12), a process of ingesting apoptotic cells and debris without triggering an immune response against self-antigens. This is vital for preventing excessive inflammation and autoimmune disease (13–15). Analysis of human HCC samples from The Cancer Genome Atlas (TCGA) and ICGC International Cancer Genome Consortium (ICGC) databases revealed that elevated EPO expression in tumors is associated with worsened survival, macrovascular invasion and poorer tumor differentiation by Edmondson-Steiner grading (Fig. 1E and fig. S3A). The elevated production of EPO in HCC patients may be attributed to the activation of hypoxia-inducible factors (HIFs), as evidenced by a positive correlation between EPO levels and common hypoxia-inducible genes identified in HCC (fig. S3B). Digital cytometry by CIBERSORTx further estimated that EPO overexpression correlates with higher frequencies of Tregs and M0 resting macrophages within the TME of HCC (Fig. 1F). Emerging evidence suggests that the M0 resting macrophages identified by CIBERSORTx represent an immunoregulatory macrophage population linked to poor prognosis in several cancers, including HCC (16–18). Beyond HCC, EPO is also produced by pancreatic ductile adenocarcinoma and prostate cancer (19, 20), and our analysis of the TCGA reveals that high EPO expression is associated with a poorer prognosis in other cancers such as kidney, breast, colorectal, and skin cancers (fig. S3C). These clinical findings, along with our observations in non-inflamed TMEs, prompted us to further investigate the underlying mechanism of this association.

Tumor-derived erythropoietin autonomously promotes an immunosuppressive tumor microenvironment

To directly examine the role of tumor-derived EPO in shaping the TME, we induced Trp53^KO HCC tumors that are unable to produce EPO, via HDTV delivery of a transient CRISPR-Cas9 vector that co-expresses two sgRNAs. Ablation of EPO in these non-inflamed tumors effectively abrogated the previously observed elevation of plasma EPO levels, suggesting that non-inflamed tumors are the primary source of EPO production in this model (Fig. 1G). Moreover, EPO ablation dramatically diminished both the tumor burden and incidence rate (Fig. 1H and fig. S4). This was accompanied by a marked increase in total leukocytes and effector CD8⁺ T cells (Fig. 1I), alongside a reduction in neutrophils and monocytes, indicating that EPO plays a critical role in promoting non-inflamed tumor development in HCC. To further study the impact of tumor-derived EPO, we used another tumor model in which the Hepa1–6 HCC cell line, derived from the C57L strain, was orthotopically implanted into C57BL/6 mice. The Hepa1–6 tumors exhibited growth for 14 days followed by subsequent spontaneous regression including complete tumor clearance in almost half the mice (21, 22). This regression was attributed to T cell immunity, as massive CD8⁺ T cell infiltration was observed in tumors in wild type C57BL/6 mice, while the tumors did not regress in Rag2-deficient mice on the same genetic background (Fig. 1J and fig. S5A–B), underlining the crucial role of T cells in tumor control. Notably, Hepa1–6 tumors that overexpressed EPO (Hepa1–6_Epo^OE) escaped regression and continued to grow, accompanied by a reduction in total leukocytes and a shift in the TME from T cell-rich to myeloid cell-rich (Fig. 1J–K). Furthermore, Hepa1–6_Epo^OE tumors displayed a lower proportion of activated CD8⁺CD69⁺ tissue-resident memory T cells (T_rm) but a higher proportion of CD4⁺ regulatory T cells (T_reg) (Fig. 1L–M). Interestingly, while EPO did not directly influence Treg polarization or activation, the higher Treg accumulation in Hepa1–6_Epo^OE tumors suggested an indirect mechanism (fig. S6A–B). Consistent with prior studies, tumor-infiltrating Tregs in this model preferentially expressed CCR8 and CTLA-4 (fig. S6C) (23, 24). Depleting Tregs using anti-CTLA4, anti-CCR8, or anti-CD25 antibodies beginning 14 days after tumor implantation led to the regression of Hepa1–6_Epo^OE tumors (Fig. 1N and fig. S6D–G). These results demonstrate that Tregs contribute to EPO-mediated immune suppression. Taken together, based on both loss-of-function and gain-of-function experiments, our data demonstrate that EPO autonomously shapes the immune compartment within the TME, thereby facilitating evasion from T cell-mediated tumor attack.

Immunoregulatory macrophages are the predominant EPOR⁺ populations in human and mouse HCC

Previous studies have suggested that EPO stimulates erythroid progenitors to induce tumor-promoting erythroid-differentiated myeloid cells (EDMCs) (25–27). In our spontaneous HCC models, we observed no significant difference in the proportion of EDMCs between inflamed and non-inflamed tumors (fig. S7A), suggesting that EDMCs are not a critical factor shaping the tumor microenvironment. Interestingly, we found that ablation of EPO secretion by tumor cells reduced EDMC accumulation, while EPO overexpression had the opposite effect (fig. S7B–C), indicating that EPO may contribute to EDMC accumulation. However, depletion of EDMCs or erythroid progenitors did not improve survival in non-inflamed HCC-bearing mice or enhance their responsiveness to PD-1 blockade therapy (fig. S7D), suggesting that the immunomodulatory effect of EPO on the TME is not primarily mediated by EDMCs.

Next, based on the hypothesis that EPOR⁺ immune cells account for the immune suppression instigated by EPO, we sought to identify which immune cell populations express EPOR in HCC. Studies of “non-inflamed” Trp53^KO HCC in EPOR-tdTomato reporter mice revealed that tumor-infiltrating EPOR⁺ leukocytes consisted mainly of F4/80⁺ macrophages, comprising three distinct subsets: CD11b^hi monocyte-derived macrophages (MDMs), CD11b^lo MDMs, and TIM-4⁺ macrophages (Fig. 2A). TIM-4 serves as a marker for Kupffer cells (KCs), liver-resident macrophages primarily originating from yolk-sac macrophages. In adults, monocytes can migrate from the bone marrow to the liver, giving rise to a minor proportion of KCs. Herein, regardless of their origin, we define TIM-4⁺ macrophages as KCs. Among the three EPOR⁺ macrophage subsets, CD11b^lo MDMs and KCs exhibited higher EPOR expression but lower MHCII levels (Fig. 2B). Ablation of EPO resulted in increased EPOR expression but reduced MHCII levels on macrophages (fig. S8A). Intriguingly, TIM-4⁺ KCs were seldom observed in “inflamed” Keap1^KO HCC, implying a potential role for these cells in EPO-induced suppression (fig. S8B) (28, 29). This notion gains further support from the Hepa1–6 orthotopic implantation model. While EPO derived from Hepa1–6_Epo^OE tumors did not alter the overall frequency of EPOR⁺ macrophages in liver, EPO expanded the TIM-4⁺ KC populations (Fig. 2C). The identity of KCs was further confirmed by their expression of CLEC4F and VSIG4 (fig. S8C). Additionally, EPO upregulated the levels of EPOR and downregulated the levels of MHCII on KCs (Fig. 2D). Using Ms4a3-EGFP reporter mice to trace monocyte-derived cells, we observed that while normal livers contained minimal EGFP⁺ KCs, EPO expression in tumors led to a dramatic increase in the proportion of EGFP⁺ KCs. This suggests that the EPO-driven expansion of KCs or KC-like cells may arise from the differentiation of monocytes or MDMs, rather than local proliferation of embryonic-derived KCs (Fig. 2E and fig. S8D).

Fig. 2. Kupffer cells and monocyte-derived macrophages are the dominant EPOR⁺ populations in hepatocellular carcinoma.

(A-B) 5 weeks after HDTV, Trp53^KO/Myc^OE tumors in EPOR-tdTomato reporter mice were harvested and subjected to flow cytometry analysis. (A) F4/80⁺EPOR⁺ macrophages, which consisted of CD11b^hi MDMs, CD11b^lo MDMs and TIM-4⁺ KCs, are the major EPOR⁺ populations in tumors, (B) and their EPOR and MHCII levels were examined (n = 4/group). (C-D) 2 weeks after orthotopic implantation (Hepa1–6_EV or -_Epo^OE), liver tissues of EPOR-tdTomato reporter mice were harvested and subjected to flow cytometry analysis. (C) The percentage of each F4/80⁺EPOR⁺ macrophage subset, and their expression levels of (D) EPOR and MHCII, were examined (n = 9/group of 2 independent experiments). (E) 2 weeks after orthotopic implantation (Hepa1–6_EV or -_Epo^OE), liver tissues of Ms4a3-EGFP reporter mice were harvested and the proportion of EGFP⁺ cells in KCs was determined (n = 3/group). (F-G) Fresh normal liver tissues were procured from deceased organ donors, while fresh tumors (T) and adjacent non-tumor liver tissues (NT) were obtained from surgical resection of HCC patients. Within 24 hours post-collection, samples were dissociated into single-cell suspensions. (F) Percentage of EPOR⁺CD68⁺ human macrophages and their CD14/CD163 levels in each tissue were determined by flow cytometry analysis. (G) The EPOR expression of EPOR⁺CD68⁺ human macrophages in T/NT pairs were examined. (H) Representative immunofluorescent images showing overlapping staining pattern of EPOR and CD68 in human HCC. (I) Correlation of EPO and EPOR mRNA expression in HCC patients (TCGA). (J) Correlation of the signature enrichment (SE) scores of two liver macrophage subsets (identified by MacParland et al.) with EPOR mRNA expression, in human HCC-NT (TCGA). (K) Fresh human liver tumors were obtained and dissociated into single cell suspensions. EPOR⁺ and EPOR⁻ macrophages (CD11b⁺CD14⁺) were isolated by FACS and subjected to bulk-RNA sequencing. (Left) The top five upregulated pathways (from 72 genes) enriched in EPOR⁺ macrophages relative to EPOR⁻ macrophages. (Right) Correlation between EPOR⁺ macrophage signature enrichment and cell type signature for five KC subsets (identified by Aizarani et al.). One-way ANOVA with Tukey’s multiple comparison test for (A), (B); Unpaired t test for (C), (D); Ratio paired t test for (G); Mann-Whitney test for (I). In all panels, data are presented as violin plot with median and quartiles, or as mean ± SD. *P ≤ 0.05; **P ≤ 0.01; ***P ≤ 0.001, ****P ≤ 0.0001. Comparisons with P > 0.05 (ns, not significant) are not displayed.

In studies of human tissues, we confirmed the presence of EPOR⁺ cells in normal liver and HCC tissues (T), as well as in non-tumorous tissues (NT) adjacent to the tumors, by analyzing leukocytes from fresh tissues using flow cytometry and immunofluorescence imaging. The EPOR⁺ cells were predominantly CD11b⁺CD68⁺ macrophages, with a minor contribution from lymphocytes exhibiting low EPOR expression in the liver. The EPOR expression of macrophages showed a trend toward being higher in T compared to NT (Fig. 2F–H and fig. S9A–C). Overexpression of EPO mRNA also correlated with higher EPOR mRNA levels in human HCC (Fig. 2I). Furthermore, a previous study using single-cell RNA sequencing of human liver identified two distinct populations of macrophages, non-inflammatory KC-like macrophages and inflammatory macrophages (30). EPOR expression is positively associated with the signature enrichment scores of non-inflammatory KC-like macrophages (Fig. 2J and fig. S10).

To gain more insight into the nature of EPOR⁺ macrophages in human HCC, we procured fresh liver tumors from surgical resections of HCC patients. Subsequently, we used FACS to isolate EPOR⁺ macrophages (CD11b⁺CD14⁺) and EPOR⁻ macrophages and subjected them to bulk-RNA sequencing. We found 72 genes that were significantly enriched in EPOR⁺ macrophages, and these genes were related to the self-renewal and phagocytosis functions of KCs (Fig. 2K and fig. S11–S12). Moreover, among the five previously identified subsets of KCs (31), EPOR⁺ macrophages showed the highest correlation with LILRB5⁺ metabolic/immunoregulatory KCs (Fig. 2K). These macrophages expressed high levels of molecules such as VSIG4 and HMOX1, which are required for maintaining an intrahepatic tolerogenic and anti-inflammatory environment (31–33). This observation aligns with our mouse model data, indicating that macrophages expressing high levels of EPOR tend to display reduced MHCII expression, which is further downregulated by EPO. Consolidating our findings across both mouse and human studies, we propose that EPOR⁺ macrophages, potentially expanded and maintained by EPO, are the major EPOR⁺ populations in HCC and they have the characteristics of immunoregulatory KCs. This prompted us to investigate their role in shaping the TME.

EPO activation of EPOR⁺ macrophages drives a non-inflamed tumor immunotype and targeting the EPO receptor results in tumor regression

Based on these findings, we hypothesized that EPOR⁺ macrophages directly mediate the effects of EPO and are indispensable for preventing immunosurveillance in HCC. To test this hypothesis we generated LysM-driven Epor^−/− mice (Epor^ΔLysM) (34), in which mature myeloid cells, predominantly macrophages in liver and HCC, lack EPOR. In this context, the absence of EPOR in macrophages had a profound improvement on the survival of mice with non-inflamed tumors, while having no effect on those with hot tumors that secreted low amounts of EPO (Fig. 3A). In approximately 40–50% of Epor^ΔLysM mice (non-inflamed tumor models), we did not observe any tumors at the study endpoint. This raised the question of whether these mice never developed tumors at all or if tumors initially formed but subsequently regressed over time. To address this question, we used a modified transposon vector, which simultaneously encoded Myc and luciferase (Myc^OE-Luc⁺), allowing us to monitor tumor growth kinetics via bioluminescence imaging. Tumors began to grow from 2 weeks post-HDTV, and wild-type (WT) mice exhibited linear tumor growth. By contrast, Epor^ΔLysM mice displayed an oscillating pattern (Fig. 3B), indicating that in the absence of EPOR⁺ macrophages the tumors indeed developed but were likely subjected to active immunosurveillance. Keap1^KO tumors also displayed oscillating growth kinetics. Despite secreting significantly lower levels of EPO compared to non-inflamed tumors, administration of human recombinant EPO accelerated the growth of Keap1^KO tumors, resulting in linear growth kinetics similar to non-inflamed tumors (Fig. 3C). To confirm that tumor-derived EPO and macrophage-expressed EPOR are two indispensable factors for immunosuppression in non-inflamed HCC, we demonstrated that neither of the Epo-secreting tumors in Epor^ΔLysM mice nor Epo^KO tumors in WT mice could escape immunosurveillance and as a result, they failed to progress (Fig. 3D). Similarly, EPO-overexpressing Hepa1–6 tumors failed to evade T-cell surveillance and regressed in the orthotopic HCC model in which macrophages lacked EPOR (Fig. 3E and fig. S13). While our earlier finding showed that EPO alone could induce Treg polarization, here we demonstrated that this effect is indirect and that EPO-mediated Treg polarization in the liver is dependent on EPO/EPOR signaling in macrophages (fig. S14). Thus, not only is EPO-activation of EPOR on macrophages required for the development of non-inflamed immunosuppressed tumors, but removal of either one of these molecules results in an inflamed tumor, independent of tumor genotype.

Fig. 3. Removal of EPOR in macrophages promotes anti-tumor immunity and results in tumor regression.

(A-B) LysM^Cre;Epor^fl/fl (Epor^ΔLysM) mice, which develops Epor-deficiency in the mature myeloid cells (mainly macrophages), were generated. Spontaneous HCC with different tumor suppressor mutations were induced by HDTV. (A) Survival of HCC-bearing WT and Epor^ΔLysM mice (n = 7–10/group), and (B) tumor growth kinetics measured by luciferin-based bioluminescence imaging (n = 7–8/group). (C) 2 weeks after HDTV (Keap1^KO/Myc^OE-Luc⁺), mice were injected (IP) with either PBS or 50 IU of rHuEPO daily for 3 weeks and tumor growth kinetics were measured (n = 7–8/group). (D) Overall survival in WT mice and Epor^ΔLysM mice after HDTV (Trp53^KO/Myc^OE-Luc⁺; Epo^WT) or (Trp53^KO/Epo^KO/Myc^OE-Luc⁺; Epo^KO) (n = 8–9/group). (E) Orthotopically implanted Hepa1–6_Epo^OE tumors in WT and Epor^ΔLysM mice were harvested on day 21 for tumor size and CR rate measurement (n = 8–12/group of 2 independent experiments). (F-G) Liposomes containing 50 μg of siEpor or siNTC (non-targeted control) were administered intravenously (IV) to (F) Hepa1–6_Epo^OE (n = 6/group) or (G) Trp53^KO/Myc^OE (n = 8/group) HCC bearing mice, and tumor burden was measured. Log-rank (Mantel-Cox) test for (A), (D); Unpaired t test for (C), (E left), (F), (G); Fisher’s exact test for (E right). In all panels, data are presented as violin plot with median and quartiles. *P ≤ 0.05; **P ≤ 0.01; ***P ≤ 0.001, ****P ≤ 0.0001. Comparisons with P > 0.05 (ns, not significant) are not displayed.

To further validate our results and determine if inhibition of EPOR expression results in tumor regression after tumors are established, we employed another method to reduce EPOR expression in macrophages. We administered liposomes loaded with siRNA targeting EPOR to tumor-bearing mice and confirmed that they were mainly taken up by macrophages (fig. S15). This strategy led to tumor regression similar to that seen in mice with macrophage-specific deletion of EPOR, indicating that established tumors are dependent on EPOR signaling in macrophages and regress when such signaling is diminished (Fig. 3F–G). Collectively, our data strongly support the notion that non-inflamed HCC relies on EPO production to evade immune surveillance, with the macrophage EPO/EPOR axis being the primary immunosuppressive mechanism.

Ablation of EPOR in macrophages provokes a T-cell-mediated anti-tumor immune response that synergizes with anti-PD-1 immunotherapy

We then investigated how EPOR⁺ macrophages shape the TME. In Trp53^KO tumors, knockout of EPOR in macrophages led to substantial increases in absolute leukocyte infiltration and the percentage of effector T cells and macrophages with elevated MHCII expression (Fig. 4A). This implied a transformation from a non-inflamed to an inflamed HCC, further supporting our hypothesis that EPOR⁺ macrophages are essential for creating an immunosuppressive TME. While antitumor immunity was restored and some tumors regressed spontaneously in Epor^ΔLysM mice, depletion of CD8⁺ T cells dramatically shortened the survival of these mice and there was no sign of regression, indicating that the antitumor immunity unleashed by ablation of EPOR in macrophages is CD8⁺ T cell-dependent (Fig. 4B and fig. S16).

Fig. 4. Removal of EPOR in macrophages promotes T cell activation and augments the response to anti-PD-1 immunotherapy.

(A) Trp53^KO/Myc^OE tumors of WT and Epor^ΔLysM mice were harvested 5 weeks after HDTV. Total leukocyte infiltration per gram of tumor, frequency of different immune populations and functional status of TAMs (CD11b⁺F4/80⁺CD64⁺) were analyzed (n = 5/group). (B) 2 weeks after HDTV (Trp53^KO/Myc^OE), C57BL/6 WT or Epor^ΔLysM mice were injected (IP) with 4 mg/kg of an anti-CD8 mAb (YTS 169.4) or IgG control twice during the first week and then once weekly (total 6 doses). Overall survival was measured (n= 6–8/group). (C-D) Trp53^KO/Myc^OE HCC tumors of WT and Epor^ΔLysM mice were harvested 5 weeks after HDTV. Tumor-infiltrating CD8⁺ T cells were profiled using spectral flow cytometry with 20 markers. PhenoGraph analysis of the CD8⁺ T cell population delineated 20 clusters (excluding the dead cell population), with 8 out of 20 differentially expressed markers (DEMs) contributing to the cluster definition. (C) UMAP plot displaying the distribution of CD8⁺ T cells in WT and Epor^ΔLysM groups (n = 6/group; 2500 CD8⁺ T cells/sample). (D) Distribution percentage of each cluster along with the expression levels of 8 DEMs. (E) 2 weeks after HDTV (Trp53^KO/Myc^OE-Luc⁺), C57BL/6 WT and Epor^ΔLysM mice were injected (IP) with 2 mg/kg of αPD-1 (RMP1–14) or IgG Isotype control every 3 days (total 5 doses). Tumor growth was monitored using luciferin-based bioluminescence imaging. Tumor growth kinetics and overall survival were measured (n = 7/group). Gray area refers to the threshold level of background noise. (F) 1.5 weeks after HDTV (Trp53^KO/Myc^OE), C57BL/6 Epor^ΔLysM-ERT2 mice were injected (IP) with 75 mg/kg tamoxifen or corn oil every 3 days (total 8 doses), in addition to the αPD-1 regimen. Overall survival was measured (n = 7–8/group) (G) 2 weeks after HDTV (Trp53^KO/Myc^OE). C57BL/6 WT mice were injected (IP) with 20 mg/kg mEPOR-Fc or PBS weekly (total 4 doses), in addition to the αPD-1 regimen. Overall survival was measured (n = 5–6/group). Log-rank (Mantel-Cox) test for (B), (E), (F), (G); Unpaired t test for (A), (D). In all panels, data are presented as violin plot with median and quartiles. *P ≤ 0.05; **P ≤ 0.01; ***P ≤ 0.001, ****P ≤ 0.0001. Comparisons with P > 0.05 (ns, not significant) are not displayed.

In light of these findings, we utilized high-dimensional spectral flow cytometry to comprehensively characterize the tumor-infiltrating CD8⁺ T cells in both WT and Epor^ΔLysM mice. The antibody panel consisted of 20 markers (including viability dye) indicative of the activation, exhaustion, differentiation, memory and functional status of CD8⁺ T cells. The tumor-infiltrating CD8⁺ T populations were analyzed with UMAP, a dimensionality reduction method, and the overlaid UMAP plot showed a dramatic subpopulation shift in Epor^ΔLysM mice (Fig. 4C). We identified 20 unsupervised clusters of CD8⁺ T cells (excluding the dead cell population) with an optimized data-derived stratifying algorithm (Phenograph). Among the 20 clusters, 8 markers (CD44, CD62L, KLRG1, TNFα, IFNg, PD-1, TIGIT, EOMES) critically defined the clusters. Intriguingly, ablation of EPOR in macrophages led to an expansion of early activated and TNFα-producing effector memory CD8⁺ T cell subsets (cluster 4,6,8) (Fig. 4D and fig. S17–S19), while the CD8⁺ T cells in WT mice were either terminally exhausted (cluster 2,17) or CD62L⁺ TNFα⁻ bystander T cells (35, 36). These results suggest that EPOR acts as an immunosuppression switch in macrophages and triggering EPOR signaling stops macrophages from activating tumor-killing CD8⁺ T cells. Epor^ΔLysM mice showed restoration of T cell immunity and expansion of PD-1⁺TNFα⁺ effector memory CD8⁺ T cells, which have been reported to critically control tumor growth and respond to PD-1 blockade (37). Consistent with these data, in the Trp53^KO model, which does not respond to anti-PD-1 monotherapy, the combination of ablation of EPOR in macrophages and anti-PD-1 led to a profound synergistic antitumor effect, with complete tumor regression in all mice (Fig. 4E). Similar synergy was observed in established non-inflamed tumors treated with tamoxifen-induced EPOR deletion in macrophages (Fig. 4F) or pharmacological inhibition of EPO/EPOR signaling using a recombinant mouse EPOR-Fc chimera protein (mEPOR-Fc) as a decoy receptor to neutralize EPO (Fig. 4G). In combination, our data indicate EPO/EPOR signaling in macrophages suppresses T cell activation and plays a central role in generating a non-inflamed TME, and that targeting this pathway in macrophages can transform non-inflamed HCC into T cell-inflamed and immunotherapy responsive HCC.

EPO transforms macrophages into immunoregulatory cells via EPOR

Our results demonstrate that EPO programs macrophages via EPOR to promote immunosuppression and hinder antitumor T cell immunity. In both mice and humans, macrophages undergo differentiation into KCs or adopt KC-like characteristics in response to EPO. Furthermore, in HCC-bearing mice, EPO stimulation leads to a reduction in the expression of inflammatory and antigen-presenting proteins such as MHCII, TNFα, and IL-1β in KCs, accompanied by an increase in regulatory proteins including CD31, MERTK, and CD206 (Fig. 5A). Therefore, to gain insight into how EPO modulates the functions of macrophages in HCC, we isolated TIM-4⁺ macrophages (KCs) from mice bearing Hepa1–6_EV (KC_EV) and Hepa1–6_Epo^OE (KC_Epo) and subjected them to bulk-RNA sequencing. Principal component analysis of these data showed that EPO/EPOR signaling induces a distinct transcriptomic profile in KCs associated with downregulation of inflammatory responses and secretion of cytokines important for T cell activation and recruitment, including Il1b and Cxcl10 (Fig. 5B–C). Among the differentially expressed genes, we identified 593 upregulated and 319 downregulated genes in KC_Epo compared to KC_EV (Log₂ fold change > 1.0, P-adjusted value < 0.05; Fig. 5D and fig. S20A–B). Pathway enrichment analysis revealed that the top 50 upregulated genes in KC_Epo were involved in scavenger receptor-mediated ligand binding and uptake, histone trimethylation, resolving acute inflammation, and cellular iron metabolism, while the top 50 downregulated genes in KC_Epo were involved in antigen presentation, inflammation, response to external stimuli, protein folding, and cell activation (Fig. 5E and fig. S20C). Importantly, most of these changes induced by EPO in KCs were also observed in MDMs and were abolished by knockout of EPOR in macrophages (fig. S21), confirming that the gene regulation is conserved in different populations of macrophages in HCC, and occurs via the EPO/EPOR signaling axis.

Fig. 5. Kupffer cells and monocyte-derived macrophages develop a regulatory phenotype upon activation of the EPO/EPOR axis.

(A-E) KCs were isolated by FACS from adjacent liver tissues of Hepa1–6_EV (KC_EV) and Hepa1–6_Epo^OE (KC_Epo) tumors at day 18 post-implantation. KC_EV and KC_Epo were analyzed with flow cytometry and subjected to bulk-RNAseq (n = 3 independent replicates). (A) Protein levels of functional markers in KC_EV and KC_Epo (n = 6/groups of 2 independent experiments). Data are presented as mean ± SD. Unpaired t test, **P ≤ 0.01; ***P ≤ 0.001, ****P ≤ 0.0001. (B) Principal component analysis of gene expression profiles, (C) Geneset enrichment analysis for inflammatory response gene signature and heatmap for core proinflammatory cytokines enriched in KC_EV, (D) volcano plots of differentially expressed genes in KC_EV and KC_Epo. (E) The top five upregulated and downregulated pathways enriched in KC_Epo relative to KC_EV. (F) Correlation between EPO and KC_Epo signatures in human HCC (TCGA). (G-I) Gene set enrichment analysis to compare the gene expression profiles of KC_EV and KC_EPO. (G) KC_EPO profile was enriched for the human LILRB5⁺ immunoregulatory KC gene signature (Geneset_{ILIRB5_KCs}). (H) Core genes enriched in KC_Epo for the Geneset_{ILIRB5_KCs} and (I) potential transcriptional factors that regulate these core genes (analyzed by ChEA3).

To determine if the macrophages in human HCC exhibit similar regulation by EPO, we first defined the top 50 significantly upregulated genes in KC_Epo as the KC_Epo signature. Analysis of human HCC samples in the TCGA database revealed a positive correlation between the expression level of EPO and this KC_Epo signature (Fig. 5F). Further analysis showed that KC_EV resembled human CD1C⁺ antigen-presenting KCs, while KC_Epo resembled human ILIRB5⁺ metabolic/immunoregulatory KCs (Fig. 5G and fig. S22). As mentioned previously, the expression profile of EPOR⁺ liver macrophages strongly aligned with ILIRB5⁺ KCs. Collectively, these results indicate that the EPO/EPOR axis regulates the functions of liver macrophages in human HCC and likely promotes the expansion of the immunoregulatory KC subset in HCC, as it does in mice.

NRF2 is a critical downstream mediator of the EPO/EPOR axis in macrophages

To further interrogate the downstream mediators of the EPO/EPOR pathway of immune regulation, we identified 34 genes critically enriched in KC_Epo for human ILIRB5⁺ KCs using geneset enrichment analysis (GSEA) (Fig. 5H). Subsequent ChIP-X enrichment analysis showed that general transcription factors governing hematopoiesis and stem cell biology, such as MECOM, GATA1, GATA2, and NANOG, can regulate these genes (Fig. 5I). Notably, many of these genes, particularly those involved in iron metabolism and antioxidation, are also under the control of the transcriptional factor nuclear factor erythroid 2-related factor 2 (NFE2L2 or NRF2) (Fig. 5I). Furthermore, pathway analysis utilizing the KEGG database revealed the KEAP1-NRF2 signaling pathway as the most enriched and significant in EPO-stimulated KCs (fig. S23). A previous study demonstrated EPO-induced nuclear translocation of NRF2 in neuroblastoma cells (38), but its impact on NRF2 signaling in macrophages remains unexplored.

Our studies show that EPO stimulation also induces NRF2 nuclear localization in KCs of HCC-bearing mice (Fig. 6A), suggesting a role for EPO in modulating NRF2 signaling within macrophages and implicating iron metabolism and antioxidation pathways in EPO-induced immune regulation. Heme oxygenase-1 (HMOX1) is the most significantly upregulated gene in EPO-stimulated KCs. HMOX1 is an enzyme that catalyzes the degradation of heme, a component of hemoglobin, into iron, carbon monoxide and biliverdin (39). Heme is essential for macrophages to generate pro-inflammatory cytokines and reactive oxygen species, but its depletion shifts macrophages to anti-inflammatory phenotypes (40). Along with HMOX1, EPO/EPOR activation also upregulated several genes involved in iron recycling and antioxidant production in KCs and MDMs (Fig. 6B–D). Most of these genes are well-known targets of NRF2 regulation, and the absence of EPO/EPOR signaling results in a significant reduction in their expression levels (Fig. 6E). Therefore, we hypothesized that activation of EPOR induces heme depletion and antioxidation in liver macrophages via regulation of NRF2. In accord with this hypothesis, EPO stimulation lowered the heme levels in KCs in the Hepa1–6 orthotopic model, as well as human monocyte-derived macrophages (Fig. 6F and fig. S24). Moreover, EPOR deletion increased the intracellular heme levels in intratumoral macrophages in Trp53^KO and Pten^KO non-inflamed HCC models (Fig. 6G). Additionally, EPO stimulation temporarily elevated intracellular free iron levels, likely due to heme breakdown, an effect reversed by EPOR deletion (fig. S25). To further demonstrate the role of NRF2 in EPO-mediated macrophage modulation, we generated LysM-driven Nrf2^−/− mice, which lack Nrf2 expression in macrophages (Nrf2^ΔLysM). We found that Hepa1–6_Epo^OE tumors could not escape T cell immunosurveillance (Fig. 6H–I), and in Nrf2^ΔLysM mice, non-inflamed HCC progressed much more slowly and was associated with an increase in CD8⁺ T_EM infiltration (Fig. 6J–K). In combination, these data reveal that EPO/EPOR signaling modulates macrophage polarization and immunosuppression in HCC by regulating heme metabolism and NRF2 activation.

Fig. 6. NRF2 activation and heme deprivation are crucial for EPO-mediated immune suppression.

(A) The nucleus-to-cytoplasm ratio of NRF2 signal in KC_EV and KC_Epo (n = 6 independent replicates). (B-E) KCs and MDMs were isolated from Hepa1–6_EV and Hepa1–6_Epo^OE HCC-bearing WT and Epor-deficient mice (n = 3 independent replicates). mRNA expression of iron metabolism-associated genes and antioxidant genes were examined in (B-C) KC_EV and KC_Epo, (D) MDM_EV and MDM_Epo, (E) KC_Epo with Epor deficiency. (F-G) Intracellular heme level in (F) KCs isolated from Hepa1–6 tumors and (G) TAMs (both KCs and MDMs) isolated from cold HCC tumors of WT and Epor^ΔLysM mice. (H-K) LysM^Cre;Nrf2^fl/fl (Nrf2^ΔLysM) mice, in which Nrf2-deficiency develops in mature myeloid cells, mainly macrophages, were generated. (H) Orthotopically implanted Hepa1–6_Epo^OE tumors in WT and Nrf2^ΔLysM mice were harvested on Day 18, and tumor size and CR rate were measured (n = 6/group). (I) Representative H&E and CD8-stained images of partially regressed tumors in Nrf2^ΔLysM mice. (J) Survival of HCC bearing WT and Nrf2^ΔLysM mice after HDTV (Trp53^KO/Myc^OE) (n = 6–8/group). (K) 5 weeks after HDTV (Trp53^KO/Myc^OE ), HCC tumors of WT and Nrf2^ΔLysM mice were harvested, and the percentages of the indicated T cell populations were determined (n = 4/group). Log-rank (Mantel-Cox) test for (J); Unpaired t test for (A), (B), (C), (D), (E), (F), (H left), (K); Fisher’s exact test for (H right). In all panels, data are presented as violin plot with median and quartiles, or as mean ± SD. *P ≤ 0.05; **P ≤ 0.01; ***P ≤ 0.001, ****P ≤ 0.0001. Comparisons with P > 0.05 (ns, not significant) are not displayed.

Concluding remarks

Although the association between EPO expression in tumors and poor prognosis in cancer patients was reported more than a decade ago, initial investigations primarily targeted its impact on cancer cell proliferation and tumor angiogenesis, overlooking its effects on immune cells within the TME. (41). Our study demonstrates that EPO autonomously determines the tumor immunotype and promotes an immunosuppressive TME through activation of EPOR signaling in TAMs. Not only is EPO activation of EPOR on TAMs sufficient for the development of non-inflamed immunosuppressed tumors, but removal of either tumor derived EPO or EPOR from TAMs results in an inflamed tumor, independent of tumor genotype. These findings likely explain numerous studies that have shown that recombinant EPO, when administered to cancer patients for the treatment of anemia, promotes tumor progression and reduces survival in a wide range of cancers (41).

The upregulation of EPO in HCC patients is likely due, at least in part, to HIF activation. Indeed, EPO is one of the earliest proteins known to be induced by hypoxia, and the search for the transcription factor that regulates EPO resulted in the purification and identification of HIF-1 (42). Beyond hypoxia, the loss of TP53 or PTEN, two frequently mutated or deleted tumor suppressor genes in HCC, further promotes HIF stabilization and downstream gene expression (43, 44). Our HDTV models consistently demonstrated elevated EPO secretion in Trp53^KO and Pten^KO tumors, reinforcing the association between TP53/PTEN loss and heightened EPO levels. Notably, eliminating EPO in non-inflamed tumors triggers tumor regression, while enforced overexpression of EPO by tumor cells or administration of EPO transforms an inflamed tumor into a non-inflamed one, regardless of genotype. EPO's indispensable role in the creation and maintenance of non-inflamed HCC has implications for other tumor types as well, given the prevalence of hypoxia and TP53 mutations in solid cancers. Since KCs and MDMs constitute the predominant EPOR⁺ cells in HCC, it is not surprising that the EPO/EPOR axis modulates the TME through macrophages. For other cancer types, different EPOR⁺ immune cells, such as subsets of dendritic cells and B cells (45), may also contribute to EPO-induced immunomodulation, warranting further investigation into their functions.

We found that EPO expands TIM-4⁺ macrophages, by promoting the differentiation of MDMs into KCs or KC-like cells. Tissue-resident macrophages have been largely overlooked in cancer research, but emerging evidence suggests that these resident macrophages are the major source of TAMs, generating a pro-tumorigenic niche for early tumor initiation and progression, as well as later recurrence and metastasis (46, 47). In addition, KCs and other tissue-resident macrophages, such as peritoneal macrophages and red pulp macrophages, are capable of cross-presentation, which enables them to directly modulate CD8⁺ T cells via MHCI-associated antigen presentation (28, 48, 49). Under steady state conditions, KCs promote the suppressive activity of Tregs and inhibit CD8⁺ T cell activation within the liver. However, under pathological conditions such as chronic liver injury, fibrosis or cancer, KCs may lose their tolerogenic phenotype and secrete proinflammatory cytokines (50, 51). Our study highlights the role of EPO/EPOR signaling acting as a switch to prevent the proinflammatory activation of KCs, and the dominant role of this mechanism in the generation and maintenance of a non-inflamed TME in HCC.

Our data also demonstrate that the EPOR⁺ macrophages isolated from human HCC are enriched in genes associated with KC functions, and more specifically, with those identified in the LILRB5⁺ KC subset (31). LILRB5⁺ KCs are an immunoregulatory subpopulation responsible for phagocytosis and iron metabolism (31), which are the upregulated features we observed in EPO-stimulated KCs in mice, prompting us to investigate the potential crosstalk involving heme depletion as a crucial mechanism for EPO-induced immunosuppression. Macrophages, which play a critical role in iron metabolism, are very sensitive to the intracellular level of heme-iron (40). Heme depletion, mediated by the NRF2/HMOX1 axis prevents macrophages from polarizing into pro-inflammatory cells (52, 53). On this basis NRF2 would appear to be a promising target for manipulating TAMs in cancer settings. However, NRF2 is a ubiquitous transcription factor, and its untargeted inhibition can cause broad effects (54). Our data show that EPOR signaling is one of the upstream regulators of NRF2 and selective removal of EPOR on TAMs can alter NRF2 signaling in these cells and dramatically impact antitumor immunity.

In this study, we utilized spontaneous preclinical models of HCC with different tumor suppressor mutations. Based on differences in the frequency, functional profile and location of CD8⁺ T cells in these tumors, each model is clearly distinguished by a non-inflamed or inflamed tumor immunotype, which mimic human HCC immunotypes and predict anti-PD-1/PD-L1 antibody responses (10). By deleting EPOR in macrophages or EPO in tumor cells, we successfully converted a non-inflamed tumor into an inflamed tumor with increased TNFα-producing CD8⁺ T_EM infiltration, leading to tumor regression in some mice, and complete cure of all mice when combined with anti-PD-1 immunotherapy. Similar synergistic effects were observed when treating established non-inflamed tumors with tamoxifen-induced EPOR deletion in macrophages or pharmacological inhibition of EPO/EPOR signaling. Additionally, macrophage-targeted EPOR siRNA treatment in mice with established tumors demonstrated similar therapeutic efficacy. Since the siRNA we used only partially reduced EPOR RNA expression on macrophages, this suggests that partial blockade or removal of the receptor is sufficient for a therapeutic effect. Taken together, these findings reveal a novel but reversible mechanism to explain the failure of immune surveillance in HCC. While our study relied on murine models of HCC, which limits generalization, the association between high EPO expression and poor prognosis across various solid malignancies suggests that similar mechanisms could contribute to the non-inflamed immunotype and resistance to anti-PD-1 immunotherapy in other tumor types. On this basis, targeting the EPO/EPOR axis may have application for the treatment of solid tumors beyond HCC.

Materials and Methods

Human Tissues

The use of human samples was approved by the Stanford University Institutional Review Board (IRB, 6304). Healthy human liver, HCC and nontumorous liver tissue samples were obtained from the Stanford Diabetes Research Center (SDRC), Donor Network West (DNW) and the Stanford Tissue Bank. Human blood samples were obtained from the Stanford Blood Center. All human samples were de-identified and exempted (exemption 4). Informed consent was obtained from patients.

Mice

C57BL/6J (Stock # 000664), B6.129P2-Lyz2^tm1(cre)Ifo/J (LysM^Cre/Cre; Stock # 004781), B6.Cg-Rag2^tm1.1Cgn/J (Stock # 008449), C57BL/6-Nfe2l2^tm1.1Sred/SbisJ (Nrf2^fl/fl; Stock # 025433), C57BL/6J-Ms4a3^em2(cre)Fgnx/J (Ms4a3^Cre/^Cre; Stock # 036382), B6.Cg-Tg(Gt(ROSA)26Sor-EGFP)I1Able/J (ROSA26-EFGP; Stock # 007897), B6.129P2(FVB)-Lyz2^{tm1(cre/ERT2)Grtn}/J (LysM-ERT2; Stock # 031674) were acquired from Jackson Labs (JAX). Epor^fl/fl mice were a gift from Dr. Hong Wu (UCLA) (55). EpoR^tdTomato mice were generated by XiuLi An (New York Blood Center)(56). LysM^Cre/Cre mice were crossbred with Nrf2^fl/fl and Epor^fl/fl mice to generate LysM^Cre;Nrf2^fl/fl (Nrf2^ΔLysM) mice and LysM^Cre;Epor^fl/fl (Epor^ΔLysM) mice, respectively. Ms4a3^Cre/^Cre mice were crossbred with ROSA26-EFGP to generate Ms4a3^Cre/^Cre; ROSA26-EFGP (Ms4a3^EGFP) mice. LysM-ERT2 mice were crossbred with Epor^fl/fl mice to generate LysM-Cre; Epor^fl/fl (Epor^ΔLysM-ERT2) mice. All mice were housed and bred in an American Association for the Accreditation of Laboratory Animal Care–accredited animal facility at Stanford University and maintained in specific pathogen-free conditions. All animal studies were approved by the Institutional Animal Care and Use Committee at Stanford University (APLAC, 17466). Orthotopic implantation studies were performed in mice between 5 to 7 weeks of age, while HDTV studies were performed between 8 to 10 weeks of age.

Cell Lines

The Hepa1–6 murine HCC line was acquired from ATCC (CRL-1830) and the 293FT line was acquired from Thermo Fisher Scientific. Cells were grown in Dulbecco’s Modified Eagle Medium (DMEM) supplemented with 4mM L-glutamine, 10% Fetal Bovine Serum (FBS), and 1% Penicillin Streptomycin. Cells were routinely tested for mycoplasma by PCR and all tests were negative. No additional authentication was performed.

Generation of Epo-overexpressing Hepa1–6 Cells

Epo-ORF (open reading frame) was inserted into lentiviral vector pLVX-EF1a-IRES-Puro by using EcoRI and BamHI restriction sites. pLVX-EF1a-IRES-Puro was a gift from Nathan E. Reticker-Flynn (Stanford University). Lentiviral particles were packaged in 293FT cells using the pCMV-VSVG and psPAX2 envelope and packaging plasmids. psPAX2 was a gift from Didier Trono (Addgene plasmid # 12260). pCMV-VSV-G was a gift from Bob Weinberg (Addgene plasmid # 8454). HEK-293T cells were transfected with Lipofectamine 2000 (Thermo Fisher Scientific) in DMEM. Viral supernatant was collected after 24 hours and filtered through a 0.45μm filter and added to Hepa1–6 cells in the presence of 8 μg/ml polybrene (Sigma Aldrich). Transduced Hepa1–6 cells were selected in puromycin (Invivogen) at a concentration of 1 μg/ml. Conditioned media were collected from selected cells, and Epo-overexpression efficiency was confirmed by ELISA.

Tumor Models

For orthotopic implantation, 3 × 10⁶ Hepa1–6 cells were resuspended in 15 μL of Matrigel Basement Membrane Matrix (Corning) and injected into the left lobes of the livers of the mice. For hydrodynamic tail vein injection (HDTV) model, sterile saline/plasmid mix, with a total volume corresponding to 10% of body weight, was injected into the lateral tail vein over 8 to 10 seconds. The sterile saline/plasmid mix contained a total of 20 μg of pX330-sgTrp53/Pten/Keap1 or pX333-sgTrp53-sgEpo, 10 μg of pT3-EF1a-Myc or pT3-EF1a-Myc-IRES-Luciferase, and 2 μg of pCMV-SB13. pT3-EF1a-Myc and pCMV-SB13 were gifts from Carmen Chak-Lui Wong (The University of Hong Kong) and generated by Scott Lowe (MSKCC). pX330 was a gift from Feng Zhang (Addgene plasmid # 42230). pX333 was a gift from Andrea Ventura (Addgene plasmid # 64073). sgRNAs were inserted into pX330 and pX333 by using BbsI and BsaI restriction sites. The sequences of the sgRNAs used in this study are provided in Table S1. Plasmids for HDTV were prepared using PureLink Endotoxin-Free Maxi Plasmid Purification Kit (Thermo Fisher Scientific). To monitor the growth of luciferase-expressing tumors, mice were administered 100 mg/kg D-luciferin (Gold Biotechnology) through intraperitoneal injection prior to bioluminescence imaging using an IVIS Spectrum imager (Caliper Life Sciences). At endpoints, tumors were measured with a caliper and calculated with the formula length × width × depth × 0.52 (cubic millimeters). Tumors and adjacent non-tumorous liver tissues were harvested for histology analysis or dissociated into single cell suspension for downstream analysis.

Preparation of Single Cell Suspension from Tissues

Tissues (including human/mouse normal livers, HCC tissues and adjacent non-tumorous tissues) were collected in ice-cold PBS. All human tissues underwent processing within 24 hours following collection, whereas mouse tissues were processed promptly after harvesting. Tissues were subjected to mincing, digestion with Liberase TM (Roche) and DnaseI (Sigma Aldrich) RPMI-1640, and dissociation with gentleMACS Dissociator (Miltenyi Biotec) for 20 minutes at 37°C. Dissociated samples were filtered through a 70 μm cell strainer and ACK lysis buffer was used to lyse the red blood cells. Cells were resuspended in PBS supplemented with 2% FBS and 2mM EDTA (FACS Buffer, FB) for fluorescence-activated cell sorting (FACS), magnetic cell separation or flow cytometry analysis.

Flow Cytometry Analysis and Cell Sorting

Cells were stained for viability using LIVE/DEAD Fixable Blue Dead Cell Stain (Thermo Fisher Scientific) on ice for 30 minutes. Cells were treated with Fc Block (BD Biosciences, 2.4G2) at room temperature for 10 minutes. Surface staining was performed for 30 minutes at 4°C prior to fixation. For intracellular staining, cells were fixed and permeabilized using the eBioscience FoxP3 Fixation/Permeabilization kit (Thermo Fisher Scientific). Intracellular staining was performed at room temperature, and cells were washed twice in permeabilization buffer prior to transfer back into FB. The following anti-mouse antibodies were used for flow cytometric analysis or FACS: CD45 (BioLegend, 30-F11), CD11b (BioLegend, M1/70), Ly-6C (BioLegend, HK1.4), Ly-6G (BioLegend, 1A8), CD11c (BioLegend, N418), CD86 (BD Biosciences, GL1), CD8a (BD Biosciences, 53–6.7), CD40 (BD Biosciences, 3/23), F4/80 (BioLegend, BM8), CX₃CR1 (BioLegend, SA011F11), I-A/I-E (BioLegend, M5/114.15.2), CCR2 (BioLegend, SA203G11), CD103 (BioLegend, 2E7), CD3 (BioLegend, 17A2), FoxP3 (eBioscience, FJK-16s), CD25 (eBioscience, eBio7D4), CD44 (BioLegend, IM7), CD62L (BioLegend, MEL-14), CD4 (BioLegend, RM4–5), CD8b (BioLegend, YTS156.7.7), CD69 (BioLegend, H1.2F3), PD-1 (BD Biosciences, 29F.1A12), NK1.1 (BioLegend, PK136), KLRG1 (BioLegend, 2F1/KLRG1), CD64 (BD Biosciences, X54–5/7.1), CD19 (BioLegend, 1D3/CD19), Granzyme B (BioLegend, GB11), CTLA-4 (BioLegend, UC10–4B9), TIGIT (BD Biosciences, 1G9), TIM-3 (BioLegend, RMT3–23), TCF1/7 (BD Biosciences, S33–966), EOMES (BD Biosciences, X4–83), CD107a (BD Biosciences, 1D4B), IFNg (BioLegend, XMG1.2), Ki67 (BioLegend, 16A8), TNFa (BD Biosciences, MP6-XT22), TIM-4 (BD Biosciences, RMT4–54), CD71 (BioLegend, RI7217), Ter-119 (BioLegend, TER-119), CLEC4F (BioLegend, 3E3F9), VSIG4 (Invitrogen, NLA14), Gr-1 (BioLegend, RB6–8C5), Helios (BioLegend, 22F6). The following anti-human antibodies were used for flow cytometric analysis or FACS: CD45 (BioLegend, HI30), CD68 (BioLegend, Y1/82A), CD14 (BioLegend, M5F2), CD163 (BioLegend, GHI/61), CD131 (BioLegend, 3D7), HLA-DR (BioLegend, L243), CD11b (BioLegend, LM2), EPOR (R&D Systems, 38409), CD4 (BioLegend, SK3), CD3 (BioLegend, SK7), CD8 (BioLegend, SK1), CD19 (BioLegend, HIB19), CD56 (BioLegend, QA17A16). For traditional flow cytometry analysis, samples were run on LSRFortessa cytometer (Becton Dickinson) and analyzed using FlowJo V10 software (TreeStar). Absolute counts were determined by using SPHERO AccuCount Fluorescent Particles (Spherotech), according to the manufacturer’s guidelines. For spectral or high-dimensional flow cytometry analysis, samples were run on Cytek Aurora and analyzed using FlowJo V10 software. The FlowJo plugin Downsample was used to select 2500 gated live CD3⁺CD8⁺ T cells from each sample. All samples were concatenated prior to the downstream analysis. The FlowJo plugin uniform manifold approximation and projection (UMAP) was used for dimensionality reduction to visualize Flowjo high-parameter data sets in 2-dimensional space, and the plugin PhenoGraph was used to group data into different unsupervised clusters. Heatmap was generated using the heatmap2 and flowCore packages in R. For FACS, samples were run on FACSAria II (Becton Dickinson). Mouse Kupffer cells (KCs) and monocyte-derived macrophages were pre-enriched by magnetic cell separation against F4/80 (Miltenyi Biotec) prior to FACS.

RNA Extraction and NGS Library Preparation and RT-qPCR

RNA was extracted using RNEasy Plus mini kits (Qiagen). Library preparation and sequencing were performed by MedGenome, Inc. (Foster City, CA). For mouse KCs in Figure 5, libraries were prepared using the Illumina TruSeq stranded mRNA kit. For human EPOR⁺ and EPOR⁻ macrophages, libraries were prepared using the Takara SMART-Seq v4 Ultra low Input RNA kit. Libraries were checked for size using the Agilent DNA 100 kit on a BioAnalyzer (Agilent) and quantified by qPCR using the KAPA Library Quantification Kit for Illumina. Libraries were sequenced on NovaSeq 6000 Flow Cells (S4) at 2 × 100 cycles paired-end to a depth of approximately 20M paired reads per sample. To validate some bulk-RNAseq results, RNA was converted to cDNA with the iScript cDNA synthesis kit (BioRad). qRT-PCR amplification was performed using PowerUp SYBR Green Master Mix (Thermo Fisher Scientific) with specific primers. The sequences of the primers used in this study are provided in Table S2.

RNA Sequencing Analysis

Raw sequencing reads were filtered, adapters trimmed using fastq-mcf program (v1.05), and data quality check was performed using FastQC (v0.11.8). Contamination removal was performed using Bowtie2 (v2.5.1)(57). The paired-end reads were aligned to the reference Human or Mouse genome downloaded from the Ensembl database, and the alignment was performed using STAR (2.7.3a)(58). The raw read counts were estimated using HTSeq (v0.11.2)(59). Additionally, the aligned reads were used for estimating expression of the genes using cufflinks (v2.2.1). The expression values were reported in FPKM (Fragments per kilobase per million) units for each gene. Differential gene expression analysis was performed using DESeq2 (60). Counts were transformed and normalized using a regularized log transformation. Principal component analysis (PCA) was performed using the R statistics package and the PCA plots were calculated using all genes differentially expressed between the samples. Volcano plots were generated using ggplot2. Differentially expressed genes in blue were defined as genes with an adjusted p-value < 0.05 and |fold change| > 2. Heatmaps were generated using the heatmap2 package, which clusters based upon Euclidean distances. Gene set enrichment analysis (GSEA) was performed using GSEA v4.3 software (61, 62). Pathway enrichment analysis and cell type analysis were performed using Metascape.

Immunohistochemical (IHC) analysis

Prior to paraffin embedding, tissues were fixed in 10% buffered formalin and washed with 75% ethanol. Heated paraffin sections were dewaxed in xylene followed by ethanol gradation. Antigens were retrieved in 1mM EDTA buffer (pH 7.8) by boiling for 15 minutes. Following blocking endogenous enzymes with 3% hydrogen peroxide and preventing non-specific binding with 2% bovine serum albumin (BSA), slides were stained with primary antibodies overnight at 4°C. For horseradish peroxidase (HRP) immunodetection, slides were stained with EnVision+ System- HRP Labelled Polymer (Agilent) and developed with 3,3’-Diaminobenzidine tetrahydrochloride (Sigma-Aldrich). Slides were counterstained with hematoxylin. For immunofluorescent (IF) detection, primary antibodies were conjugated with fluorescent dyes using Zenon Antibody Labeling Kits (Thermo Fisher Scientific) or DyLight Conjugation Kits (Abcam) prior to staining. Slides were counterstained with 4',6-diamidino-2-phenylindole (DAPI). Slides were imaged on a Keyence BZ-X810 microscope. The following antibodies were used for IHC: anti-mouse CD8a (Cell Signaling, D4W2Z), anti-mouse FoxP3 (Cell Signaling, D6O8R), anti-human CD68 (Cell Signaling, D4B9C), anti-human EPOR (ImmunEdge, in-house).

Preparation of siRNA-loaded liposomes

FITC-labelled siNTC (sense: 5'-CGUUAAUCGCGUAUAAUACGCGUdAdT-3'; antisense: 5'-ACGCGUAUUAUACGCGAUUAACGdAdT-3'-FlTC) and siEpor (sense: 5'-CCCAGAGAGCGAGUUUGAGGGUCdTdC-3'; antisense: 5'-GACCCUCAAACUCGCUCUCUGGGdTdC-3’-FlTC) were synthesized by Dharmacon (Horizon Discovery). siRNA was dissolved in 10 mM sodium citrate (0.45 mg/mL). 1,2-Distearoyl-sn-glycero-3-phosphocholine (DSPC), 1,2-dipalmitoyl-sn-glycero-3-phosphocholine (DPPC), cholesterol and DSPE-PEG-NH2 (MW:3400) at a weight ratio of 1:1:0.5:0.4 were then dissolved in chloroform and the lipid mixture was evaporated to produce dry liposome powder. Dry liposome powder and siRNA (1:40 w/w) were resuspended in UltraPure DNase/RNase-Free Distilled Water (Thermo Fisher Scientific) and sonicated to form siRNA-loaded liposomes of desired molecular size. The molecular size and siRNA encapsulation efficiency of liposomes were determined as previously described (63). The physical properties of our siRNA-loaded liposomes are shown in Extended Data 11.

Measurement of EPO, heme and intracellular free iron

To assess plasma EPO levels in mice with HCC, blood was obtained via retro-orbital bleeding before sacrifice. The collected blood was centrifuged at 1500 × g for 15 minutes at 4°C to isolate the plasma layer. To verify the overexpression efficacy of EPO in Hepa1–6_Epo^OE cells, conditioned media from both Hepa1–6_EV and Hepa1–6_Epo^OE cells were gathered. Plasma EPO levels were determined using the Mouse Erythropoietin/EPO Quantikine ELISA Kit (R&D Systems). To assess the intracellular heme levels in macrophages, mouse Kupffer cells (KCs) and monocyte-derived macrophages were purified with magnetic cell separation against F4/80 (Miltenyi Biotec) and FACS. The purified macrophages were lysed, and intracellular heme levels were determined using a Heme Assay Kit (Sigma-Aldrich). All colorimetric test results were recorded using a Victor X4 Multilabel Reader (Perkin Elmer). Intracellular free iron levels of tumor macrophages and bone-marrow-derived macrophages were determined using FerroOrange (Cell Signaling Technology) with flow cytometry analysis.

NRF2 subcellular localization

Mouse KCs were purified using magnetic cell separation followed by FACS. Purified KCs were concentrated onto microscope slides using StatSpin Cytofuge 2 cytocentrifuge (Beckman Coulter). Slides were fixed with 4% paraformaldehyde (PFA) and permeabilized with PBS containing 0.2% Triton X-100. Slides were blocked with 1% BSA, 22.52 mg/mL glycine in PBST (PBS+ 0.1% Tween 20). Rabbit anti-human/mouse NRF2 antibody was conjugated to FITC using Zenon Antibody Labeling Kit (Thermo Fisher Scientific). Slides were stained with anti-NRF2 antibody overnight at 4°C and counterstained with DAPI. Slides were imaged on a Zeiss LSM 700 confocal microscope. The subcellular localization of NRF2 in the cell nuclei or cytoplasm was defined on the basis of colocalization with DAPI. The NRF2 signal was quantified using the colocalization tool in ZEN Blue software (Zeiss).

Immune cell depletion

C57BL/6 wild-type mice were orthotopically implanted in the liver with 3 × 10⁶ Hepa1–6_Epo^OE cells. On Days 14, 17, and 20 after implantation, 2 mg/kg αCD25 (BioXCell, PC-61.5.3), αCTLA-4 (BioXCell, 9H10), αCCR8 (BioLegend, SA214G2) or IgG control (BioXCell, hamster polyclonal) was injected intraperitoneally to deplete Tregs. Tumors and spleens were collected on Day 21 for further analysis. For CD4⁺ and CD8⁺ T cell depletion studies, 2 weeks after HDTV, mice received intraperitoneal injections of 4 mg/kg of αCD8 (BioXCell, YTS169.4), αCD4 (BioXCell, GK1.5) or IgG control (BioXCell, LTF-2) twice for the first week and then once weekly (total 6 doses). Overall survival was measured. For EDMC depletion studies, 2 weeks after HDTV, mice were given IP injections of 2 mg/kg of αTer-119 (TER-119) or IgG Isotype control every 3 days (total 9 doses). Overall survival was measured.

PD-1 blockade

Following administration of plasmid cocktails via HDTV, mice received intraperitoneal injections of 2 mg/kg of αPD-1 (BioXCell, RMP1–14) or IgG Isotype control (BioXCell, 2A3) every 3 days (total 5 doses). Tumor growth kinetics were assessed using luciferin-based bioluminescence imaging, and overall survival was determined.

EPO administration

To examine the effects of EPO on Keap1^KO tumor progression, C57BL/6 wild-type mice were injected with plasmid cocktail (20 μg of pX330-sgKeap1, 10 μg of pT3-EF1a-Myc-IRES-Luciferase, and 2 μg of pCMV-SB13) via HDTV. 2 weeks after HDTV, mice received intraperitoneal injections of either PBS or 50 IU of rHuEPO (PeproTech) daily for 3 weeks, and tumor growth kinetics were monitored using luciferin-based bioluminescence imaging.

Induction of EPOR ablation in macrophages by tamoxifen

Tamoxifen (Sigma‐Aldrich) was dissolved in corn oil at a concentration of 20 mg/ml by shaking overnight at 37°C. 1.5 weeks after HDTV C57BL/6 EporΔLysM-ERT2 mice were injected IP with 75 mg/kg tamoxifen or corn oil every 3 days (total 8 doses).

In vitro studies of human macrophages

Fresh human PBMCs were isolated from whole blood using SepMate PBMC Isolation Tubes (StemCell Technologies). PBMCs were cultured in ImmunoCultTM-SF Macrophage Differentiation Medium supplemented with 50 ng/ml of human recombinant M-CSF (PeproTech). On Day 6, macrophages were treated with 10 ng/ml of human recombinant IL-4 (PeproTech) and 100 ng/ml of rHuEPO. On Day 8, macrophages were harvested, and the intracellular heme level was measured.

In vitro studies of mouse macrophages

Fresh mouse bone marrow cells (BMCs) were isolated from tibia and femur. BMCs were cultured in complete RPMI Medium supplemented with 50 ng/ml of mouse recombinant M-CSF (PeproTech). On Day 6, macrophages were treated with 100 ng/ml of rHuEPO. On Day 8, macrophages were harvested, and the intracellular heme level was measured.

In vitro Treg polarization

Fresh mouse CD4⁺ T cells were isolated from spleens using magnetic cell separation against CD4 (Miltenyi Biotec). CD4⁺ T cells were cultured in complete RPMI Medium supplemented with Dynabeads^® Mouse T-Activator CD3/CD28 (Thermo Fisher Scientific) according to manufacturer’s recommendation, 5 ng/ml of recombinant mouse IL-2 (PeproTech), 5 ng/ml of recombinant human TGF-β1 (PeproTech) and 100 ng/ml of rHuEPO. On Day 5, cells were harvested for Treg analysis.

Statistical Analyses

All statistical tests were performed with Prism (GraphPad) and are described as follows except where otherwise noted in figure legends. For all figures, *P < 0.05, **P < 0.01, ***P < 0.001, ****P < 0.0001. For most data, replicates are shown as individual points, and where summary data is presented, the number of replicates (n) is noted in the figure legends. All comparisons involving more than two groups utilize tests that adjust for multiple comparisons. Comparisons of two conditions were evaluated using two-tailed unpaired t test unless otherwise noted. Comparisons of three or more conditions were performed using One-Way ANOVA with Tukey’s multiple comparison test, with a single pooled variance. Comparisons of categorical variables were performed with Fisher’s Exact test. Comparisons of survival curves were performed with Log-rank test.

Data and Materials Availability :

Bulk RNA-seq data have been deposited to GEO (GSE290606). Tabulated data underlying the figures are provided in data file S1. Requests for plasmids, EPOR antibodies and transgenic mouse lines can be sent to the corresponding author, E.G.E. Any outgoing material shared with external beneficiaries will be based on a material transfer agreement between Stanford University and the other institution/company. All data needed to evaluate the conclusions in the paper are present in the paper or the Supplementary Materials.