Tumor ENPP1 (CD203a)/Haptoglobin Axis Exploits Myeloid-Derived Suppressor Cells to Promote Post-Radiotherapy Local Recurrence in Breast Cancer

Breast cancer is the second most common cause of cancer mortality in women worldwide (1). Although 60% of cases are diagnosed as localized disease, 10% of treated women will still develop locoregional recurrence or failure (LRF) in spite of state-of-the-art therapy, reaching 25% in triple-negative and HER2 subtypes (2, 3). In contrast to current treatments for the primary tumor, which include surgery, radiation, and chemotherapy, LRF poses a therapeutic dilemma, as treatment options are scant and are associated with a dismal prognosis. Within each subtype, the two major determinants of post-surgical and post-irradiation LRF are large tumor size and inappropriate resection margins (R1 and R2), among other minor clinical and pathologic factors (4). These clinical observations posit the emergence of quiescent tumor cells left outside the surgical tumor margins, which, if they overcome the constraints imposed by ionizing radiation (IR), could subsequently become reactivated by microenvironmental cues. These cells are found in specialized niches of the tumor microenvironment (TME) and may preserve phenotypic plasticity and resilience to genotoxic stress (5). Along these lines, the resected tumor bed (RTB) fluid obtained in patients treated only with surgery stimulates an aggressive phenotype in cultured tumor cells, whereas this effect is almost completely abrogated when irradiation is applied (6).

Breast tumors also release circulating tumor cells (CTC) from the primary tumor as an early event (7). CTC are detected in 10% to 30% of patients with nonmetastatic breast cancer during tumor evolution (8, 9). CTC status is becoming more commonly evaluated in clinical practice (10) because CTC also confer prognostic value. Treatment efficacy can also be assessed by monitoring CTC over time as a prognostic marker to estimate the risk of recurrence (11). Nevertheless, only a fraction of CTC are capable of surviving with potential to generate productive local or distant metastases (12). In animal models, CTC reinfiltrate an established tumor or spread to target organs, initiating secondary outgrowths (13). Based on this, a CTC-mediated mechanism of “tumor self-seeding” may also cooperate in local tumor recurrence. Yet surgical manipulation and IR perturbs the RTB to the extent that this milieu may also impose constraints on CTC engraftment and colonization. This is in part due to the altered immune landscape associated with a local inflammatory reaction. Indeed, several leukocyte subpopulations might efficiently promote the antitumor immune attack on incipient engrafted CTC. In contrast, other tumor-infiltrating leukocytes could exert protumorigenic effects, including polymorphonuclear myeloid-derived suppressor cells (PMN-MDSC) and monocytic MDSC (MON-MDSC). PMN-MDSC constitute a heterogeneous subpopulation that share many morphologic and phenotypic characteristics of neutrophils (14). They emerge for their salient proangiogenic role and immunosuppressive effects in a variety of tumors (15). Yet, the immune–CTC cross-talk and cellular and molecular underpinnings involved in LRF remain largely unexplored despite their suspected relevance for breast cancer and for a large variety of solid tumors.

To systematically dissect these events, we developed a new breast cancer model of LRF post-surgery and post-IR, in which we identified ENPP1 (ectonucleotide pyrophosphatase phosphodiesterase 1, or CD203a), a member of the non-classic adenosinergic CD38–ENPP1–CD73 pathway. ENPP1 catalyzes the hydrolysis of ATP to AMP and GTP to GMP, while generating inorganic pyrophosphates (16). ENPP1 also promotes the hydrolysis of cGAMP (17), a potent activator of the innate immune cyclic GMP–AMP synthase (cGAS)–stimulator of interferon genes (STING) pathway (18), which leads to strong immunosuppression and tumor progression (19). Extracellular release of ATP and AMP and their ENPP1-mediated hydrolysis products (20, 21) act as substrates to fuel adenosine (Ado) production by the classic adenosinergic CD39–CD73 pathway (22). These metabolites, specifically AMP and Ado, are known to tightly regulate immune responses acting on the purinergic (G-coupled) P1 receptors of immune cells (23, 24).

In this study, ENPP1^hi-expressing CTC were more fit to efficiently home and colonize the RTB during tumor recurrence. Engrafted tumor cells mediated an immunosuppressive TME characterized by the chemotactic infiltration of PMN-MDSC. Unexpectedly, tumor ENPP1 levels were comodulated with the proinflammatory protein haptoglobin (HP), an event mechanistically mediated by the autocrine signaling of extracellular ENPP1 hydrolysis products. HP was identified as a chemoattractant of MDSC and a novel unanticipated inducer of neutrophil extracellular trap (NET) formation. NET are extruded DNA meshes associated with cytotoxic enzymes released to the extracellular milieu that, aside from their antibacterial role, have been found to foster tumor progression and metastasis in mouse models (25). Conversely, genetic and pharmacologic ENPP1 inhibition or P1 receptor blockade abrogates PMN-MDSC recruitment. Moreover, inhibition of NET formation markedly extends relapse-free survival. More interestingly, a combination of ENPP1 pharmacologic blockade and standard radiotherapy treatment post-surgery eradicates LRF.

Furthermore, patients with ENPP1^hi or HP^hi triple-negative breast cancer (TNBC) were associated with a worse prognosis and with tumor recurrence. Recurrent tumors showed an increased NET formation as compared with primary paired tumors. Thus, the ENPP1/HP axis illuminates a novel mechanism involved in breast LRF.

To set up a valuable model of LRF, we selected two triple-negative aggressive syngeneic breast cancer cell lines, ANV5 and 4T1, to develop murine orthotopic tumors in syngeneic mice. In pilot experiments, we injected these luciferase-transduced tumor cells in a small group of immunosuppressed mice, to evaluate the quality of the surgical procedure by bioluminescence imaging (BLI). The experimental scheme is represented in Fig. 1A. We used a brachytherapy technique by implanting a catheter into the RTB at the time of surgery for inserting the IR source (Supplementary Fig. S1A–S1C). Briefly, ANV5 tumor cells were inoculated in the inguinal mammary gland of syngeneic mice. Tumors were resected at ∼300 ± 164 mm³ tumor volume. Mice were randomized to receive 15 Gy IR or no further treatment. In this group, the procedure for the placement of catheters was performed with no IR. No differences in tumor volume between groups were detected at surgery (Fig. 1B). Excised tumors were ink-embedded for surgical border-free assessment (Fig. 1B). The characteristics of treated mice are shown in Supplementary Table S1. No differences in the quality of surgical resection were detected by histologic evaluation (Supplementary Table S2). Interestingly, local control in the resected group only (Sg) was 44.8%, whereas in the resected and IR group it was 62.2% (Fig. 1C).

The extent of the surgical margin was a critical factor associated with LRF (Fig. 1D). Although no differences were detected in tumor size at surgery, the R1 subgroup obtained the greatest advantage from the use of radiation (Fig. 1E), a finding consistent with the clinical observations (Supplementary Fig. S2). Surgical margins and treatment arm were the main predictors of local control in a multivariate analysis (Supplementary Table S3). Hence, IR decreased the rate of local relapse and prolonged the time to LRF. Gross necropsy analysis did not detect development of macroscopic distant metastasis in this model.

In a similar approach using the murine breast cancer 4T1 cells, we found a trend toward a longer relapse-free survival rate in the locally irradiated group as compared with untreated animals that did not reach statistical significance, but we observed spontaneous animal death associated with aggressive lung metastasis preventing the evaluation of LRF.

Using microfluidics CTC detection technology (Parsortix), we evaluated the number of CTC pre- and post-surgery in the 4T1 and ANV5 models (Supplementary Fig. S3A). After bilateral orthotopic injection of 4T1 cells transduced with a GFP reporter gene, tumors were resected. Interestingly, the number of CTC in blood was markedly increased four hours after surgery compared with one day before surgery (Supplementary Fig. S3B and S3C). In the ANV5 model, an increase in the CTC number was detected post-surgery, although the number of CTC was very low in comparison with the 4T1 model (Supplementary Fig. S3B and S3C). Thus, CTC are commonly detected during primary tumor growth in both models. A few hours after tumor resection, the CTC number increases, suggesting a possible tumor-release effect during the surgical procedure.

To investigate the causal link between CTC and LRF, we orthotopically inoculated naïve ANV5 cells in the mammary gland (day 0). At the time of tumor resection (day 17), the contralateral mammary gland was partially excised as a sham-surgery control. Subsequently, luciferase-GFP–transduced ANV5 cells were intracardiacally (i.c.) inoculated to mimic CTC (Fig. 1F). BLI assessed in vivo during follow-up revealed an increased signal in the RTB (ipsilateral mammary gland) that increased over time as compared with the contralateral mammary gland (Fig. 1G), although there was also an initial BLI detected in the contralateral mammary gland that disappeared over time. Interestingly, BLI was obliterated in IR mice. More importantly, immunohistochemical (IHC) analysis of recurrent tumors revealed the presence of several patches of GFP⁺ cells within the tumor mass. This approach enabled us to detect CTC derived from i.c. inoculation that engrafted in the RTB and contributed to the formation of recurrent tumors. These findings were further substantiated by assessing the fluorescently labeled cells GFP⁺ within the tumor (Fig. 1H). Similar results were obtained in the 4T1 model (Supplementary Fig. S3D and S3E). We also noted a strong signal emanating from the thoracic cavity, a finding consistent with the mediastinal contamination after i.c. inoculation and with the emergence of pulmonary metastases, an event profusely described in this model. These data thus suggest the proclivity of CTC to engraft in the RTB and contribute to the formation of the recurrent tumor mass.

Next, we sought to compare the efficacy of two irradiation regimens using single dose (SD) of 15 Gy versus fractionated dose (FD) of four fractions of 6.2 Gy equi-effective IR doses in both models, ANV5 or 4T1 cells (Supplementary Fig. S3F and S3G).

To evaluate whether this IR effect was due to the efficient eradication of engrafted CTC, we assessed the number of CTC homing and engrafting to the RTB with different FD (Fig. 1I). Briefly, after tumor resection, i.c. inoculation of ANV5-GFP⁺ was performed to mimic CTC. One group was 2FD-IR (2 × 6.2 Gy) and the other was 4FD-IR (4 × 6.2 Gy). The control group was sham-IR. At sacrifice, single-cell suspensions derived from total minced mammary glands were cultured ex vivo. Interestingly, a high number of clones derived from GFP⁺ single cells were isolated from sham-IR animals. In contrast, few clones were detected in any of the 2FD group and none in the 4FD group, indicating efficient IR-mediated eradication of GFP⁺ CTC-engrafted cells (Fig. 1I). We hypothesized that CTC could still engraft several days post-IR. We evaluated the number of CTC post-resection after 2FD or 4FD treatment by microfluidics in mice injected with 4T1-GFP⁺ cells in two mammary glands per mouse. After surgery, one group of animals was irradiated with an SD two days after surgery, the second group was FD-IR, and control mice were sham-IR. Blood samples were collected four days after tumor IR and seven days post-surgery (Fig. 1J). A slight decrease in the number of CTC was observed after SD (Fig. 1J) and a substantial decrease in the number of CTC was observed in the FD group. However, a small number of CTC were still detected after FD, indicating that these CTC could still be competent to engraft and contribute to LRF. Furthermore, local IR eradicates CTC engrafting in the RTB, and the IR schedule also has an impact on the number of CTC. Overall, CTC represent a determining factor contributing to the emergence of tumor recurrence. Interestingly, CTC after FD or SD were competent to give rise to local recurrence in the ANV5 model (using a similar approach as in Fig. 1F), an event associated with a diminished number of CTC colonizing the RTB (Supplementary Fig. S3H–S3K).

As engraftment represents a highly inefficient process, we obtained a marked relapsed phenotype by successive cycles of enrichment. Briefly, we isolated cell subpopulations from recurrent non-IR and IR tumors in the ANV5 model (Fig. 2A). Relapse-free survival curves showed a shorter latency time of recurrence for the cell lines 720Cy2Rx and 701Cy2Rx derived from second cycles in IR conditions, indicating an enrichment in the LRF phenotype (Fig. 2B).

To identify the determinants of the acquisition of increased competencies for LRF, we performed a transcriptomic analysis using RNA sequencing (RNA-seq) comparing parental ANV5 cells with second cycle IR cell subpopulations (Fig. 2C). Clustering analysis represents a selection of differentially expressed genes, one of which encodes for Enpp1 (19). We also performed a transcriptomic analysis comparing ANV5 with the second cycle isolated cells from IR and non-IR recurrent tumors (Supplementary Fig. S4A and S4B), a signature that was further validated in non-IR and IR tumors (Fig. 2D). Cells showed higher protein expression levels over increased cycles in ANV5 as compared with parental cells (Fig. 2E). Similar results were validated in 4T1 cells following an identical strategy (Supplementary Fig. S4C and S4D). Thus, the acquisition of an LRF-enriched phenotype is mediated by a common gene transcriptomic program, which is conserved irrespective of radiation conditions.

Based on previous findings, we reasoned that CTC could be enriched in LRF gene signature. To test this hypothesis, we isolated CTC from the blood of incipient orthotopically implanted 4T1-derived tumors, which release high numbers of CTC. These cells, called “CTC-out,” were expanded ex vivo and GFP transduced (Fig. 2F). These CTC-out cells showed similar levels of Enpp1 to the control 4T1 cells (Fig. 2G). Subsequently, in resected mice carrying 4T1 unlabeled mammary tumors, previously isolated CTC-out were i.c. inoculated with GFP⁺ CTC-out cells (n = 10 mice per each CTC-out subpopulation). After recurrence, tumors were digested and GFP⁺ cells isolated (called “CTC-in”). To further enrich this phenotype, recovered cells were subjected to another cycle of i.c. inoculation. Enpp1 expression levels were found elevated in all isolated CTC-in subpopulations as compared with control and CTC-out cells (Fig. 2G and H). To extend these results to the ANV5 model, we had to perform several cycles after i.c. inoculation of GFP⁺ ANV5 cells in mice previously tumor resected to obtain CTC-in cells, as we were unsuccessful at isolating CTC-out. Interestingly, Enpp1 gene-expression levels in CTC-in of different cycles consistently showed an enrichment in Enpp1 in recurrent tumors (Fig. 2G and H). Other components including Srpx2 and Epas1 showed no consistent expression levels in CTC-in cells (Supplementary Fig. S5A and S5B). Interestingly, varied Enpp1 expression levels were found in cancer cell lines derived from different tumors (Supplementary Fig. S5C).

Next, we investigated whether Enpp1 levels could endow cells with an advantageous function for homing and engraftment at the RTB. We overexpressed Enpp1 (OE-Enpp1) in ANV5-GFP⁺ cells. Cells were i.c. inoculated one day after resection of ANV5-derived tumor (Fig. 2I). Interestingly, mice inoculated with OE-Enpp1-GFP⁺ cells showed an increased number of single cell–derived colonies obtained from the resected mammary glands at day 6 post-inoculation with no effect on cell growth kinetics in vitro (Fig. 2I). We also found that Enpp1^hi cells were endowed with increased radioresistance (Supplementary Fig. S5D). Thus, Enpp^hi CTC showed an enhanced ability to engraft in the RTB.

We assessed the functional contribution of Enpp1 in LRF using Enpp1 (OE-Enpp1) and parental ANV5 cells (control; Fig. 3A). After tumor resection, local recurrence was evaluated by histology. Tumor volume was matched between groups before surgery. A diminished recurrence-free survival of OE-Enpp1–derived tumors was detected over control mice (Fig. 3B and C).

We also silenced Enpp1 (shEnpp1) in 720Cy2Rx cells (Fig. 3D). Tumor volume was matched between groups before surgery. Interestingly, Kaplan–Meier curves of relapse-free survival showed an extended time to LRF in shEnpp1-derived tumors with a survival rate around 70% (Fig. 3E). In the LRF model, similar tumor volumes were reached in both groups before surgery (Fig. 3F).

As Enpp1 has been categorized as an immunotransmitter, we evaluated whether Enpp1 levels could modulate tumor-associated immune infiltrate. First, we estimated the relative levels of the Enpp1 diesterase activity in parental ANV5, 720Cy2Rx, and these cells transduced with two different shRNAs targeting Enpp1 (Supplementary Fig. S5E), and we assessed the immune landscape after orthotopic injection (Supplementary Fig. S6A and S6B).

At day of sacrifice, tumors derived from shEnpp1 cells showed a significant decrease in tumor volume (Fig. 3G). Moreover, tumors derived from Enpp1-silenced cells showed a consistent decrease in total myeloid cells and PMN-MDSC (CD11b⁺Ly6C^loLy6G⁺) and CD4⁺ regulatory T cells (Treg), whereas CD8⁺ T cells, including effector subsets, were unaffected (Supplementary Fig. S7A and S7B).

A converse experiment was performed using forced expression of Enpp1 (OE-Enpp1) in ANV5 cells. A significant increase in orthotopic tumor volume was observed (Fig. 3H). Interestingly, a significant increase in total myeloid cells CD11⁺ with an increase in PMN-MDSC (P < 0.01) was detected in OE-Enpp1–derived tumors, whereas no changes were detected in Tregs (Fig. 3H). CD8⁺ and CD4⁺ effector cells were unaffected. By contrast, we found a diminished number of natural killer (NK) and dendritic cells infiltrating OE-Enpp1 tumors (Supplementary Fig. S7C). The infiltration kinetics of PMN-MDSC in the RTB were measured at early time points (days 1, 6, and 8) post-injection in mice (n = 8 per time point) injected with OE-Enpp1 cells (Fig. 3I). Interestingly, at days 6 and 8 post-injection when the tumors were barely palpable, increasing numbers of infiltrated PMN-MDSC were already detected. Conversely, Enpp1-silenced tumors showed a significantly decreased number of infiltrated PMN-MDSC at day 13 (Fig. 3J).

Thus, a steady infiltration kinetics of PMN-MDSC concurrent with tumor cell colonization of the RTB was detected from early phases of tumor growth. Moreover, modulation of Enpp1-tumor levels regulated the immune landscape by consistently altering PMN-MDSC tumor infiltration.

To discriminate the functional role of PMN-MDSC infiltration, we depleted PMN-MDSC with an anti-Gr1⁺ antibody in an orthotopic model of tumor growth using OE-Enpp1 and control ANV5 cells. At day 15, we observed an expected increase in tumor volume in OE-Enpp1–derived tumors until sacrifice. Interestingly, anti-Gr1–treated animals inoculated with OE-Enpp1 cells showed a dramatic decrease in tumor volume, reaching a similar size to tumors derived from control mice (P < 0.0001; Fig. 3K). No changes in tumor volume were detected in control animals treated with anti-Gr1 antibody. Blood counts of neutrophils were significantly diminished at sacrifice (two days after last depletion) in the Gr1⁺-treated OE-Enpp1 group (Fig. 3L), whereas no significant changes were detected in the number of lymphocytes. IHC analysis showed a decreased number of tumor-infiltrating Gr1⁺ cells in OE-Enpp1 tumors treated with anti-Gr1 in correlation with tumor volumes (Fig. 3M).

Next, we performed an LRF assay with OE-Enpp1 and control cells. Tumors were resected at similar tumor volume (Fig. 3N). As expected, OE-Enpp1–injected animals showed an accelerated LRF rate. Interestingly, the LRF rate diminished in the anti-Gr1–treated group injected with OE-Enpp1 cells, whereas anti-Gr1 treatment had no effect in mice injected with control cells (Fig. 3N). IHC analysis revealed a decrease in the Gr1⁺ infiltrate in Gr1-depleted tumors overexpressing Enpp1 as compared with vehicle-treated OE-Enpp1–derived tumors (Fig. 3O). These findings indicate that Gr1⁺ cell subpopulations were associated with Enpp1 tumor expression levels. The Gr1⁺ cell subpopulations played a significant immune-mediated role in orthotopic tumor growth and in LRF-free survival rate.

Next, we interrogated the efficacy of pharmacologic blockade of Enpp1 using a small-molecule Enpp1 inhibitor (ENPP1i; CM3163), whose activity was tested in vitro (Supplementary Fig. S7D; ref. 19). In an orthotopic experiment using a 4T1 subpopulation (374Cy2NoRx cells) that expresses high Enpp1 levels, tumor volume of treated mice decreased in comparison with vehicle-treated mice (Supplementary Fig. S7E). In orthotopic tumors inoculated with 720Cy2Rx cells, a significant decrease in PMN-MDSC in Enpp1i-treated animals was found in an independent experiment performed where Enpp1i also decreased tumor volume. We also detected an increase in NK, CD8⁺ T effector, and dendritic cells and a decrease in tumor-associated macrophages and M2 macrophages in Enpp1-treated mice (Fig. 4A). Thus, Enpp1 inhibition strongly remodels the tumor-associated immune landscape, mounting a marked antitumor effector response.

In an LRF assay using the Enpp^hi ANV5 subpopulation (720Cy2Rx), Enpp1i diminished the LRF rate and extended relapse-free survival (Fig. 4B). Abrogation of other cell subpopulations including CD8⁺ T or NK cells in this assay did not significantly influence relapse-free survival. IHC analysis showed a decreased number of Gr1⁺ leukocytes infiltrating the tumors in Enpp1i-treated mice (Fig. 4B). Interestingly, the STING pathway was also partially mediating Enpp1 function on tumor growth in vivo, as codepletion of STING (C-176) in control or silenced Enpp1 cells led to increased tumor growth kinetics (Fig. 4C). Furthermore, codepletion of CD39 in Enpp1-silenced tumor cells led to a significant increase in LRF-free survival rate (Fig. 4D), an effect presumably mediated by the immune compartment as ANV5 cells were devoid of CD39.

To evaluate whether Enpp1i therapy could benefit the current standard of care (surgery and fractionated IR), Enpp1i treatment was used with or without IR after surgical resection of tumors. As expected, FD treatment improved the LRF-free survival as compared with the non-IR group. Interestingly, FD treatment with Enpp1i further extended the LRF-free survival, reaching a remarkable improvement in local control, and resulted in only 1 of 15 mice with LRF (Fig. 4E). Thus, these findings indicate that the combination of IR with Enpp1i reached a superior benefit in terms of obliterating LRF.

Next, we tested whether tumor cells with different Enpp1 levels could differently modulate PMN-MDSC migration and function. Murine PMN-MDSC were purified from the spleen of tumor-bearing mice, reaching 97% purity. We also obtained PMN-MDSC differentiated from murine bone marrow after ex vivo culture, reaching 60% purity. In both cases, flow cytometry analysis revealed the expression of C–C chemokine receptor type 2 (CCR2), also reportedly expressed on human neutrophils (ref. 26; Supplementary Fig. S8A and Methods). Using a Boyden chamber assay, we observed a significant increase in the migration/invasion of PMN-MDSC placed in the upper compartment when incubated with OE-Enpp1 ANV5 tumor cells in the lower chamber. This effect was abrogated with an Enpp1i. Consistently, PMN-MDSC had a significant decrease in chemoattraction when Enpp1-silenced 720Cy2Rx cells (shEnpp1a and shEnpp1b) were placed in the lower chamber as compared with control cells, suggesting a chemotactic effect associated with Enpp1 levels (Fig. 5A).

To identify mediators of this Enpp1-dependent effect on MDSC invasiveness, we performed RNA-seq in OE-Enpp1 versus control in ANV5 and 4T1 cells. We selected 146 common differentially overexpressed genes that were coherent between cell lines with P < 0.05 and log-fold change >0. A similar approach was used to identify downregulated genes in two Enpp1-silenced cells (720Cy2Rx and 374Cy2NoRx cells) with two independent shRNAs (Fig. 5B). By crossing upregulated genes in OE-Enpp1 cell lines with downregulated genes in silenced cell lines, we identified a coherent 16-gene signature associated with Enpp1 levels (Fig. 5B). The most significant gene associated with Enpp1 levels was Hp, an acute secreted inflammatory component known for binding free hemoglobin (Fig. 5C).

Consistent with Enpp1 levels, Hp expression levels were validated in cell lysates (Supplementary Fig. S8B) and in the conditioned media (CM; Fig. 5D). Hp expression was also diminished when cells were incubated with Enpp1i (Supplementary Fig. S8C), indicating an Enpp1-mediated comodulation of Hp levels. Based on these findings, we reasoned that a mechanistic link between Enpp1 and Hp could be mediated by Enpp1 activity or its product metabolites as a result of the adenosinergic pathway. As Ado is the common downstream metabolite to the noncanonical and canonical pathways, we postulated that high levels of Ado could mediate the coregulation of Hp. As Ado stimulates intracellular cyclic AMP levels through Gs-coupled P1 receptor subtypes A2aR and A2bR, we used forskolin, an activator of adenylyl cyclase, or 8-Br-cAMP, as cAMP mimetics. Incubation strongly induced Hp secretion in the CM of ANV5 (Fig. 5E). Secretion was inhibited by coincubation with a competitive antagonist of P1 receptors (SCH-58261). Moreover, NECA (5′-N-ethylcarboxamidoadenosine), a nonselective Ado receptor agonist, also stimulated Hp secretion, and this effect was markedly reduced by the Ado receptor antagonist. In Enpp1-silenced cells, Hp levels were rescued upon agonist stimulation, and these effects were reduced by coincubation with the Ado receptor antagonist (Fig. 5E). Similar effects were observed in 4T1 cells and its derivative (Supplementary Fig. S8D). Concordant with Enpp1 effects, purified Hp per se increased the migration chemotactic index (MCI) of murine PMN-MDSC in a dose-dependent manner, an effect abrogated by the use of a CCR2 inhibitor (RS-102895; Fig. 5F). Furthermore, CM of Hp-silenced cells led to diminished MCI of PMN-MDSC cells (Fig. 5G). CM of ANV5 parental cells treated with cAMP mimetics or the Ado receptor agonist increased the MCI of myeloid cells, effects that were abrogated by coincubation with a P1 receptor antagonist (Fig. 5G). Similar results were obtained in 4T1 cells (Supplementary Fig. S8E). These findings indicate that the Enpp1 hydrolysis product AMP, and its derivative, Ado, stimulate the secretion of Hp to the extracellular milieu in breast cancer cell lines, an effect mediated by purinergic P1 receptors. In turn, Hp induces the migration and invasion of PMN-MDSC in vitro, an effect mediated by CCR2.

As other proinflammatory factors such as C5a (a downstream effector of complement activation) are strong inducers of NET formation, an extrusion primarily composed of DNA complexed with proteins, we inquired whether Hp as an acute phase component could also promote NET formation. As expected, a significant number of NET were observed with C5a, as compared with unstimulated cells (Supplementary Fig. S8F). Interestingly, upon incubation with Hp, we observed prominent NET formation, a process that was again strongly diminished by the NET inhibitor, Pad4 inhibitor, and using 720Cy2Rx cells with Hp-silenced levels (Fig. 6A and B).

Based on this finding, we also investigated the ability of PMN-MDSC to induce NET formation in vitro upon incubation with tumor cells with different Enpp1 levels. To test this, ANV5 cells were cocultured with mouse PMN-MDSC. Interestingly, OE-Enpp1 tumor cells showed a significant induction of NET that was further abrogated with an Enpp1i and A2aRi (SCH-58261; Fig. 6C and D). Similarly, silenced levels of Enpp1 led to a significant decrease in NET formation. Taken together, these findings indicate that tumor Enpp1 levels differentially modulate neutrophil chemotactic invasiveness and NET formation in vitro. To further substantiate these findings, we assessed the relevance of Hp in LRF; mice inoculated with Hp-silenced cells showed an almost significant improvement in LRF-free survival as compared with control mice (P = 0.07). Of note, tumor volume and resection margins at surgery were significantly different between groups (Supplementary Fig. S8G), suggesting that Hp might be an important component involved in LRF possibly mediating NET formation in vivo. More importantly, LRF of animals treated with NET inhibitors (PAD4i or DNase) improved LRF-free survival rate, indicating that abrogation of NET formation strongly influences relapse-free survival (Fig. 6E).

To investigate the clinical relevance of the identified LRF gene signature, we used the in silico tool Kaplan–Meier plotter in a cohort of patients with TNBC. Patients were segregated according to the computed gene-expression levels. Patients with TNBC with high levels of Enpp1 had significantly shorter recurrence-free survival (RFS; P = 0.0011; Fig. 7A). Interestingly, in a cohort of 405 patients with TNBC, high levels of HP were associated with a diminished RFS (P = 0.0093; Fig. 7A). In contrast, gene-expression levels of other components of the LRF signature did not show the expected correlation with RFS in those patients (Supplementary Fig. S9A). Of note, copy number of ENPP1 is frequently amplified in breast cancer cohorts (Supplementary Fig. S9B).

To further study this association, we constructed a breast cancer tissue microarray (TMA) containing 54 paired specimens from 27 HER2⁺ and TNBC patients with primary breast cancer who developed LRF (Supplementary Table S4). IHC analysis of ENPP1 in a TMA of primary versus recurrent breast cancer human tumors obtained from the same patient showed a significant increase of ENPP1 in recurrent tumors as compared with that in paired primary tumors (P = 0.0324; Fig. 7B). Furthermore, we evaluated the density occupied by NETs in the same human cohort including naïve primary tumors and relapsed tumors post-surgery and post-IR. Interestingly, the NET-occupied area, identified by costaining with CD15, myeloperoxidase (MPO), and citrullinated histone H3 (H3Cit), was significantly increased in relapsed tumors as compared with their respective primary tumor NET levels at diagnosis (P = 0.013; Fig. 7C–E). Of note, the density of neutrophil infiltration correlated with the area of NET, possibly indicating a relationship between neutrophil tumor infiltration and NETosis

.

LRF post-surgery and post-radiotherapy poses a major challenge in the management of patients with breast cancer. By modeling this event, we demonstrate a conserved mechanism by which Enpp1^hi CTC are endowed with increased fitness for homing and colonization of the recurrent tumor bed and subsequent tumor relapse, in part for their ability to recruit PMN-MDSC. Infiltration by this immunosuppressive myeloid subpopulation is enabled by Enpp1 hydrolysis metabolites acting in an autocrine/paracrine loop to trigger Hp tumor secretion. Interestingly, its chemoattractant activity promotes this myeloid intrusion and generates robust NET formation. Accordingly, blocking Enpp1 or NET formation or eradicating engrafted CTC by IR diminishes CTC colonization, leading to a reduction in the LRF rate. The most salient finding was an almost complete LRF abrogation after Enpp1 blockade was added to standard fractionated IR post-surgery. This synergism is substantiated by the antitumor immune response elicited by Enpp1i acting in the immune infiltrate, its tumor radiosensitizing effect, and by the more radiosensitive immune infiltrate by eradicating neutrophil infiltration and NET formation (27, 28). Beyond the cytotoxic effect of FD, the number of residual cells and intervals between fractions may affect LRF (6). FD has also shown a significant antitumor benefit leading to the production of IFNβ required for the function of dendritic cells, which in turn primed tumor-specific CD8⁺ T cells, boosting immune-mediated antitumor responses and synergizing with checkpoint inhibitors (29). These results underscore the potential of this combination to boost current therapeutic options in post-surgical IR and strongly substantiates a proof-of-concept for the future development of clinical trials.

This report highlights a unique role of Enpp1 in tumor relapse, specifically in CTC colonizing the RTB. Enpp1 confers further advantages during early events of recurrence, in which Enpp1^hi cells are more prone for efficient engraftment, survival, and growth in the RTB.

This Enpp1^hi CTC phenotype arose independently of IR and conferred an increased fitness for secondary outgrowths in the RTB. Tumor-intrinsic factors leading to Enpp1^hi phenotype are unknown but could be prompted by tumor–stromal interactions or other contextual factors of the tumor milieu at the primary tumor and can be further enriched during relapse. Of note, a strong Enpp1 expression signal was also detected in the stromal compartment of some infiltrating breast cancer tumors, indicating a potential tumor–stromal cross-talk (30).

Several mechanisms were also unpredictably exploited by Enpp1^hi tumor cells to mediate LRF. First, Enpp1 levels were unexpectedly comodulated with the tumor-secreted Hp. This protein is a well-known component of the acute inflammatory response, aside from its ability to sequester free hemoglobin released into the bloodstream to prevent hemoglobin-induced oxidative damage (31). This previously uncovered association links its proinflammatory function with the antitumor immune response and posits the Enpp1/Hp axis as a new mechanism by which tumor cells subdue myeloid cells to boost LRF. We showed that AMP by an autocrine/paracrine signaling via Ado receptors in the tumor strongly stimulated Hp secretion, establishing a mechanistic link between Enpp1 and Hp through purinergic P1 receptors. Ado is also increased by IR, which also upregulates other Ado-generating enzymes such as CD38 (32, 33). Ado levels are also dependent on the activity of the CD39/CD73 pathway expressed by cancer cells, immune cells, and the vasculature. Indeed, CD39 blockade in Enpp1-silenced cells led to a significant reduction in the LRF rate, indicating that a blockade of the remnant activation of the adenosinergic pathway, especially in immune-infiltrating cells, could be necessary to reach a more salient effect. Although this point requires further investigation, several clinical trials targeting CD73/Ado receptors are currently ongoing.

Besides its role in the adenosinergic pathway, tumor ENPP1 also hydrolyzes the extracellular cGAMP, a danger signal of the STING pathway, to AMP and GMP, and this degradation prevents detection by the immune system. We cannot discard the possibility that other extracellular purines or pyrimidines in the TME could also stimulate tumor Hp. Moreover, IR could exacerbate these findings as this treatment augments extracellular cGAMP. Importantly, accumulation of cGAMP by ENPP1 inhibition activates the cGAS–STING pathway, leading to an antitumor immune attack, which was reversed using a STING inhibitor, resulting in tumor growth in Enpp1-silenced cells. Indeed, STING activation in noncancer cells leads to increased numbers of cross-presenting dendritic cells and subsequent T-cell activation and NK-cell responses (34). Similarly, Enpp1i treatment led to an increased tumor-associated dendritic CD11c⁺ MHC-CII⁺ population and NK infiltration. Although cGAS–STING activation by emerging agonists could reduce tumor relapse, we could anticipate that concurrent Enpp1 blockade could reach more noticeable antitumor effects because Enpp1i boosted a strong immune landscape remodeling with an additional reduction of myeloid cells.

Of note, Enpp1 inhibition was more pronounced than Enpp1-silencing effects, as it may also affect most of the immune subpopulations expressing Enpp1 (35). Our findings were also in line with previous studies showing that PMN-MDSC inhibit the activity and function of other myeloid subpopulations and NK cells (36). Consequently, ENPP1 levels were associated with poor prognosis in a subset of patients with breast cancer. These observations support the potential benefit of combining immune-checkpoint inhibitors with ENPP1 inhibition to obtain a superior antitumor immune response.

Second, Hp released in the inflammatory milieu constitutes a novel factor orchestrating myeloid infiltration and NET formation. This dual function exerted by Hp on PMN-MDSC might be dependent on its concentration gradient and the duration of the stimulus, as myeloid intrusion predates NET formation (37). Hp-induced myeloid migration was supported by early observations where Hp-specific binding to neutrophils was shown to modulate effector functions (38) and subsequent findings (26). However, we cannot discard the possibility of a more complex interaction between Hp and PMN-MDSC that may involve other unidentified players. The PMN-MDSC infiltration kinetics detected from incipient nonpalpable tumors indicated a sustained effect (Fig. 3I and J). Interestingly, NET formation was required for LRF because its abrogation was associated with a significantly better LRF-free survival, underscoring a novel role of NET formation in LRF. It is tempting to speculate that NET could participate in the accumulated trapping of CTC in the RTB, further increasing the LRF rate, and eventually initiate the development of secondary outgrowths at local or distant locations (Fig. 7F). Reportedly, NET participate in the capture of metastatic emboli during colonization (39). Furthermore, NET could also interfere with antitumor immune cytotoxicity, creating a tumor-protective shielding mechanism (40). As NET formation also primed the premetastatic niche and promoted tumor dissemination in a lung metastasis model of breast cancer (41), inhibition of NET hampered the progression of established tumors and impaired metastasis (40). Concordantly, an increased NET area was detected in relapsed breast cancer specimens in our cohort (42). Thus, we could anticipate that several NET inhibitors currently tested in the clinic, including PAD4 antagonists, IL8-neutralizing antibodies (43), and CXCR2 inhibitors, could also affect LRF. Similarly, given the central role of ENPP1 in myeloid infiltration, NET formation, and tumor progression, it is not surprising that ENPP1i have emerged as potent immunostimulatory agents (44). In fact, eradicating MDSC infiltration by blocking CXCR2 also impaired NET formation and sensitized tumors to immune-checkpoint inhibitors (45). Besides Hp, other factors comodulated with Enpp1 could act as chemoattractant factors for PMN-MDSC boosting the LRF phenotype. For instance, Vegfa, a component of the Enpp1-comodulated gene signature, was shown to trigger MDSC infiltration and impair T-cell function, and could favor tumor growth in more advanced stages by enhancing tumor angiogenesis. A limitation of our model was that the short time of latency did not recapitulate a more complex clinical scenario emerging >1 year after IR, in which signals emanating from healed RTB, derived from persistent TGFβ activation, could foster the recruitment of reactivated dormant or disseminated tumor cells (46).

Our new LRF model also integrates the mechanistic role of CTC in tumor recurrence. The presence of CTC frequently detected as an early event in patients with breast cancer with the increased number of CTC during surgical manipulation was also concordant in our model and supports the notion that they make a significant contribution to LRF (47). At early stages after resection, CTC were homing to the ipsi- and contralateral mammary glands to a similar extent, but only those CTC engrafting in the RTB yielded productive recurrent tumors over time. The emergence of CTC clusters within recurrent tumors suggests the requirement of cell–cell interactions as a potential mechanism involved in the competency of the “fertile soil” with fit Enpp1^hi CTC for proficient local tumor relapse.

In line with our results, fractionated IR was advantageous mainly in patients with CTC positivity (47). Indeed, the selective obliteration of CTC prevents metastasis and extends survival (48). Beyond the cytotoxic effect of FD, the number of residual cells and intervals between fractions could affect LRF (6). However, Enpp1 emerged as a common mechanism impairing antitumor immune response in IR and non-IR conditions. Thus, our model opens new avenues to explore other previously defined roles of PMN-MDSC, including radioresistance (49, 50), angiogenesis (51), and metastatic activity (52).

Another inherent limitation of our CTC modeling approach was the impossibility to discriminate whether the unlabeled recurrent tumor cells were derived from unlabeled CTC (released from the primary tumor) or arose from the local infiltration of recalcitrant cells beyond resection margins. Our strategy could also be translated to other tumors such as sarcomas, head and neck carcinomas, and other neoplasms to reveal novel mechanistic insights where LRF represents an unmet clinical need.

In summary, our study unveils novel links with patient prognosis. The role of the Enpp1/Hp axis, the relevance of CTC, tumor-secreted Hp inductive role in PMN-MDSC infiltration, and NET formation by the adenosinergic pathway emerge as a novel players uncovered in LRF. The identification of new mediators paves the way for introducing novel post-radiation therapies to prevent LRF and discloses new vulnerabilities for drug development. This represents a promising avenue for therapeutic development, with wide implications in clinical practice for the treatment of breast cancer and other tumors (53).

ANV5 cells were originally derived from relapsed tumors after subcutaneous inoculation of mouse mammary carcinoma cells into nontransgenic FVB/N mice (a kind gift of K.L. Knutson; ref. 54). ANV5 cells were cultured in RPMI 1640 supplemented with 10 mmol/L HEPES, 1% GlutaMAX, and 1 mmol/L sodium pyruvate. The 4T1 cells were obtained from ATCC. Cells were tested monthly for Mycoplasma infection using MycoAlert. Cells were authenticated by PCR assessment.

Bulk tumors were minced into small fragments and digested in RPMI medium containing collagenase (1 mg/mL) and DNase (50 µg/mL) for 30 minutes at 37°C. Subsequently, fragments were mechanically disaggregated and filtered through a 70-µm cell strainer. Cells were collected by centrifugation, resuspended in complete medium, and cultured. We obtained different cell subpopulations from recurrent tumors in IR and non-IR animals.

Cells were transduced with pSIN-puro-GFP lentivirus for cell tracking. To isolate CTC-out cells, mice were orthotopically inoculated in the mammary gland with pSIN-puro-GFP–transduced 4T1 cells. Blood extraction was performed by intracardiac puncture under anesthesia in EDTA Microvette 500K3E tubes (Sarstedt). Cells were cultured in complete medium and selected with puromycin.

For CTC-in, cells were inoculated in the left cardiac ventricle in six-week female immune-compromised mice (Rag2tm1.1Flv Il2rgtm1.1Flv/J). Tumors were disaggregated to single-cell suspension, and after selection with 4 to 6 µg/mL of puromycin, CTC-in cells were cultured until single cell–derived colonies were detected. In homing experiments, cells were fixed and stained with crystal violet for 15 minutes. Colonies were counted using an automated macro developed in Fiji software (ImageJ, RRID:SCR_003070).

All experiments were executed in compliance with institutional guidelines and regulations and after approval from the Local Animal Ethics Committee (160-14 and 114-19). Sex did not represent a biological variable in this study. All experiments were performed in female mice. Cells were resuspended in PBS:Matrigel (BD Biosciences) at 1:1 ratio. Fifty microliters containing 2 × 10⁴ ANV5 or 4T1 cells (RRID:CVCL_0125) in Matrigel at 1:1 volume (BD Bioscience) were injected through a 1-cm-long incision into the fourth and fifth inguinal mammary fat pads of six- to seven-week-old female syngeneic FVB/N mice or BALB/c (RRID:IMSR_ORNL:BALB/cRl) mice (Envigo) for 4T1 cells. After inoculation, the skin was sutured with Vicryl 6.0. When tumors reached the indicated volume, tumors were resected and catheters were implanted until radiation (Fig. 1A). Tumors were usually detected by palpation two weeks after tumor cell inoculation. Orthotopic tumors were resected en bloc. Surgery aimed to completely remove all visible gross tumor. The tumor in the inguinal mammary gland was grasped between the thumb and the index finger for immobilization, and a small incision, 1–2 mm deep and 10–15 mm long, was created in the skin overlying the tumor. The tissues surrounding the index lesion (radial tumor margins) were dissected to allow mobilization. The tumor was then gently pulled out through the skin opening with a mini-forceps, and the tumor was removed from its base (deep tumor margin). The whole surface of the tumor specimen was ink-marked for pathologic assessment (Fig. 1A). A summary of the surgical margins of all experiments is included in Supplementary Table S5.

Cells were expanded ex vivo (first cycle) and reinoculated in the mammary glands of another cohort of mice, n = 8–10 per each cell subpopulation. After local recurrence, tumors were surgically resected, and recurrent cells were isolated and expanded in vitro (second cycle). Isolated first and second cycle cell subpopulations were compared to assess their “local recurrence” phenotype by evaluating the time to LRF.

After tumor excision, a leader Angiocath 16GA was subcutaneously implanted over the RTB to allow the insertion of a standard 6 Fr afterloading catheter CT/MR Flexible Implant Tube 6F (SL, 30 cm, Elekta), passed backward toward the former skin entry point, and then pulled outside and fixed with a stopper (plastic button). The next day, mice were anesthetized to allow the insertion of the 4 Fr catheter coupled to the afterloader (Fig. 1A). Immediately after last irradiation dose, the catheter was pulled out for the rest of the experimental period. Brachytherapy treatment was used with predefined CT-based standards to deliver an SD of 15 Gy to the CTVD₉₀ (minimum dose received by 90% or the CTV). Standards were generated previously using several implanted mice for CT image acquisition at 1 mm intervals. For non-IR conditions, surgery and the catheter placement were performed as in IR animals. The standard CTV created was 20 mm long in the axis of the catheter with a diameter of 10 mm. In fractionated irradiation, four fractions of 6.2 Gy each, twice a day, at least six hours apart were delivered over two consecutive days to reach a total dose of 24.8 Gy. SD irradiation of 15 Gy and fractionated irradiation of 24.8 Gy (6.2 Gy × 4) are equi-effective as per the linear-quadratic formulation.

Gene silencing was performed by lentiviral infection using shRNA in pLKO.1 vector (RRID:Addgene_52920) as previously described (55). For gene overexpression, mouse cDNA encoding Enpp1 was cloned into pBABE-Neo as described (RRID:Addgene_1767; ref. 56).

We performed experiments using PMN-MDSC obtained by two different procedures:

1. Purification of spleen-derived PMN-MDSC was performed by immunomagnetic positive selection from mice carrying 21-day Lewis lung cell carcinoma tumors with Ly6G microbead kit (Miltenyi RRID: AB_2895065) according to the manufacturer's recommendations.

2. Differentiated PMN-MDSC obtained from murine bone marrow were generated as previously described (57). Briefly, mononuclear precursors were obtained from bone marrow of 8- to 12-week-old female C57BL/6J mice (RRID:IMSR_JAX:000664) and cultured in complete medium for 24 hours. Nonadherent cells were recovered 24 hours after extraction and reseeded. These cells were incubated for 7 days with 50% CM obtained from 3-day culture of 8 × 10⁵ genetically modified Lewis lung carcinoma cells and 50% complete medium. Adherent cells were harvested with PBS with 1 mmol/L EDTA and were used for flow cytometry analyses.

The upper wells of 8-µm-pore-size Boyden chambers were coated with MDSC stained with Hoechst (1:1,000 dilution, Thermo Fisher Scientific) at a density of 10³ cells/µL embedded in hydrogel composed by type 1 collagen (2 mg/mL, BD Bioscience) and Matrigel poor in growth factors (2 mg/mL, BD Bioscience) as previously described (58). HP was used at a concentration of 0.02 to 0.5 mg/mL based on the findings that murine serum levels are between 0.5 and 1.5 mg/mL in normal conditions. After 24 hours, cells were recovered from the lower chamber and were quantified by flow cytometry using a CytoFLEX (Beckman Coulter) and CyteExpert software. Experiments were repeated three times.

In vitro–differentiated murine MDSC (1 × 10⁶ cells) stained with Hoechst (20 µmol/L, Thermo Fisher) and calcein (1 mg/mL diluted 1:2,000, Invitrogen) were cocultured with the indicated tumor cells (ANV5, 4T1, or derivatives), previously stained with orange CMRA (1 mg/mL, diluted 1:2,000, Invitrogen) and Hoechst. The coculture was embedded at 10³ cells/µL in hydrogels. After 48 hours (for mouse cells) and 20 hours (for human cells), images were captured by a confocal microscope LSM 800 laser-scanning (Carl Zeiss) with Plan-apochromat-63×. Images in Z were acquired with Zen 2.3 software (Carl Zeiss) and were visualized with Volocity 3D (RRID:SCR_002668, Thermo Fisher Scientific). NET area and the percentage of NETs were calculated using the plugin DANA for ImageJ (developed by Dr. Miriam Shelef, University of Wisconsin–Madison, Madison, WI). At least 60 cells per experimental condition were analyzed (59). Experiments were repeated three times.

A day before the assay, cells were seeded at 5 × 10⁴ cells per well in 96-well plates. The next day, culture media were replaced by 50 µL of the buffer assay (50 mmol/L Tris, 250 mmol/L NaCl, pH 9.5). Subsequently, 50 µL of the assay buffer containing the reagent thymidine 5′-monophosphate p-nitrophenyl ester (Sigma, #T4510) at 10 mmol/L diluted in deionized water was added in each well. Absorbance was measured at 405 nm over time. Measurements were taken after 60-minute reaction.

Evaluation of NETs was performed using an Automated Quantitative Pathology Imaging System (Vectra Polaris, Akoya Biosciences). Tissue imaging and spectral unmixing were performed using inForm software (version 2.4.8, Akoya Biosciences). Image analysis was then performed using the open-source digital pathology software QuPath version 0.2.3 and ImageJ Software, an open-source Java-based image processing software. Measurement workflow and scoring have been previously described (42). In short, neutrophils were identified by the costaining with CD15 and MPO. NETs were detected based on the costaining with CD15, MPO, and H3Cit. An ImageJ software plugin was developed and validated to accurately calculate the percentage area of neutrophils and NETs for each TMA spot (42).

We used Rag2^−/−/Il2Rγ ^−/− in-house bred mice (Rag2tm1.1Flv Il2rgtm1.1Flv/J, Jackson Labs, RRID:IMSR_JAX:014593) to avoid immune recognition in preliminary experiments to improve the surgical procedure. Cells were transduced with triple modality (GFP-luciferase-Δ45HSV1-tk; refs. 55, 56) or pSIN-puro-GFP lentiviruses. BLI evaluation was performed immediately after surgery under anesthesia after inoculation of 50 µL of 15 mg/mL D-Luciferin into the retro-orbital plexus for assessment of residual disease as well as during follow-up for detection of local relapse or distant spread. The date of local relapse was registered as the date of the first positive BLI test post-surgery or the date of the first palpable recurrence post-surgery in those cases with weak or negative bioluminescence. Images were captured immediately after injection during one minute with PhotonIMAGER imaging system (Biospace Lab) and analyzed using M3Vision software (Biospace Lab). Photon flux was calculated for each mouse using a region of interest for each mammary gland inoculated. All BLI signals were normalized with values from day 0.

Mammary tumors were formalin-fixed for 24 hours and maintained in 70% ethanol, dehydrated, and paraffin-embedded according to standard protocols conducted at the Morphology Core Facility at CIMA. Five-micrometer sections were stained in hematoxylin and eosin (H&E). Ink-marked margins were evaluated for each tumor by an experienced pathologist. Resections were categorized into R0 (clear margins in all directions), R1 (focally involved surgical margin), and R2 (extensively involved surgical margins). For IHC, the anti-GFP antibody (ab6556) was used in serial sections. For TMA, anti-Enpp1 antibody was used at 1:4,000 dilution (ab40003). The detection of Gr1⁺ cells was performed with an anti-Gr1 antibody (BioLegend, ref. 108402) at 1:5,000 dilution. IHC staining was performed using EnVision (Dako) and revealed using diaminobenzidine (DAB). Whole mount sections were acquired and digitized using Leica Aperio CS2 system. Image analysis was performed with Fiji software (RRID:SCR_002285).

The optimization and validation workflow of the multiplex immunolabeling protocol development has been previously described (42). A five-color multiplex quantitative immunofluorescence assay for formalin-fixed, paraffin-embedded tissue human specimens was implemented for simultaneous detection of the granulocyte marker CD15, the neutrophil marker MPO, H3Cit, cytokeratin (CK), and DAPI. Briefly, TMA sections were deparaffinized and subjected to sequential rounds of antibody staining. Antigen retrieval was performed using DAKO PT-Link heat-induced antigen retrieval with low pH (pH = 6) or high pH (pH = 9) target retrieval solution (DAKO). Antibodies were CD15 (mouse monoclonal, clone Carb-3, 1:100, Agilent, product # IR062, RRID:AB_10838679), MPO (rabbit monoclonal, clone E1E7I, isotype IgG, 1:2,000, Cell Signaling, product # 14569S, RRID:AB_2798516), anti-histone H3 citrullinated (citrulline R2 + R8 + R17, H3Cit) antibody (rabbit polyclonal, 1:100, Abcam, product # ab5103, RRID:AB_304752), and CK wide spectrum screening (1:100, Agilent, product # Z0622). Tyramide signal amplification with fluorophores Opal 620 and Opal 690 was used for visualization of CD15 and H3Cit, respectively. Cy3 Tyramide Plus (1:100) and Alexa488 were used for visualization of MPO and CK, respectively. Thereafter, nuclei were counterstained with spectral DAPI (Akoya Bioscience). Sections were then mounted with Faramount Aqueous Mounting Medium (Dako).

Cells were lysed in RIPA-buffer [1% Nonident P-40, 50 mmol/L Tris-HCl (pH 7.4), and 150 mmol/L NaCl] supplemented with protease inhibitors (Sigma). Protein lysates were subjected to 6% SDS-PAGE, transferred to nitrocellulose membrane (Bio-Rad), and incubated with anti-Enpp1 (PA5-17097, Invitrogen), anti-Epas1 (sc-13596, Santa Cruz, RRID:AB_627525), anti-Srpx2 (Ab01584, Abcam), anti-Apcdd1 (sc-84694, Santa Cruz, RRID:AB_2057638), anti-Hp (ab256454, Abcam), and anti-tubulin antibody (T4026, Sigma-Aldrich). Bands were developed with enhanced chemiluminescence system (Amersham Bioscience).

Blood extracted by cardiac puncture under anesthesia was recovered into sterile cell-free DNA BCT, EDTA-glass tubes (Streck) and analyzed by Parsortix Technology (Angle). Blood parameters were assessed by Hemavet 950 counter (Drew Scientific Inc).

Tumors were dissected, cut into small pieces, and incubated with collagenase/DNase buffer (as previously described) for 30 minutes and mechanically disaggregated through 70-µm cell strainer. Erythrocytes were lysed (red blood cell lysis buffer, Sigma), and cell suspensions were harvested with Percoll 35% (GE Healthcare). One million cells were treated with CD16/CD32 (Mouse BD Fc Block, 1:200, clone 2.4G2, BD Bioscience) and stained on ice with the following fluorochrome-conjugated antibodies: anti-mouse CD45 (APC-H7, 1:500, clone 30-F11, BioLegend), CD11b (BUV661, 1:160 clone M1/70, BD Bioscience), CD11c (BUV395, 1:80, clone HL3, BD Bioscience), Ly6G (AF700, 1:200, clone 1A8, BioLegend), Ly6C (BV510, 1:200, clone HK1.4, BioLegend), F4/80 (BV605, 1:80, clone BM8, BioLegend), MHC-C2 (BV650, 1:80, clone M5/114.15.2, BioLegend), CD38 (PerCP/Cy5,5, 1:160, clone 90, BioLegend), and CD8 (BUV395, 1:200, clone 53–6.7, BD Bioscience), CD44 (BV510, 1:200, clone IM7, BioLegend), CD62 L (AF700, 1:800, clone MEL-14, BioLegend), CD4 (BUV496, 1:400, clone GK1.5, BD Bioscience), CD25 (PerCP/Cy5,5, 1:50, clone PC61, BioLegend), CD19 (BUV661, 1:200 clone 1D3, BD Bioscience), NKp46 (CD335; APC, 1:20, clone 29A1.4, BioLegend), LAG3 (BV650, 1:400, clone C9B7W, BioLegend), GITR (FITC, 1:100, clone DTA-1, BioLegend), and PD-1 (BV605, 1:80, clone 29F.1A12, BioLegend) diluted in FACS buffer (PBS, 0.24 mmol/L EDTA, and 5% serum FetalClone). For intracellular staining, cells were permeabilized after surface staining with eBioscience Fixation/Permeabilization Kit for 15 minutes and intracellularly stained with GzB (BV421, 1:20, clone GB11, BioLegend), FoxP3 (PECy7, 1:320, clone 3G3, Abcam) and CD206 (PECy7, 1:40, clone C068C2, BioLegend). Dead cells were excluded using PromoFluor-840 NIR Maleimide (PromoCell). Cells were acquired using a CytoFLEX LX flow cytometer (Beckman Coulter) and analyzed with CytExpert software (RRID:SCR_017217). Gating strategy of flow cytometry analysis is shown in Supplementary Fig. S5.

Normalization of microarray data (Affymetrix Mouse Gene 2.0 ST) was performed with robust multiarray average algorithm (RMA; ref. 60). After quality assessment and outlier detection with R/Bioconductor, a filtering process was carried out to eliminate low expression probe sets (61). Applying the criterion of an expression value greater than 16 in at least 50% of the samples of one of the experimental conditions, 18,311 probe sets were selected for statistical analysis. Linear models for microarray data (LIMMA; ref. 62) were used to identify the probe sets that showed significant differential expression between experimental conditions. Hierarchical clustering of microarray data was performed using R (63).

Quality of the samples was verified using FastQC (RRID:SCR_014583) software https://www.bioinformatics.babraham.ac.uk/projects/fastqc/. The alignment of reads to the mouse genome (mm10) was performed using STAR (ref. 64; RRID:SCR_004463). Subsequently, gene-expression quantification using read counts of exonic gene regions was carried out with feature Counts (65), and the gene annotation reference was Gencode M25 (66). Differential expression statistical analysis was performed using R/Bioconductor.

First, gene-expression data were normalized with edgeR (ref. 67; RRID:SCR_012802) and voom (62). After quality assessment and outlier detection using R/Bioconductor (RRID:SCR_006442), a filtering process was performed. Genes with read counts lower than 6 in more than the 50% of the samples of all the studied conditions were considered as not expressed in the experiment under study. LIMMA (RRID:SCR_010943) were used to identify the genes with significant differential expression between experimental conditions. Genes were selected as differentially expressed using a B > 0 cutoff. Further functional and clustering analyses and graphical representations were performed using R/Bioconductor. The selection of coherent changes caused by Enpp1 silencing or overexpression was based in a P value cutoff P < 0.05 in all the experiments and coherent logFC changes between experiments. The clustering analysis of these genes was performed with transformed expression data on the same scale making comparable different experiments.

In silico analysis was performed with patients with breast cancer included in the Kaplan–Meier plotter (http://kmplot.com/analysis/), a web-based tool that uses transcriptomic data.

Selection was performed from the archives of the Clínica Universidad de Navarra (CUN) from 1995 to 2017. Inclusion criteria included all available archival material examined for integrity and representative areas of H&E-stained tumors selected by an expert pathologist. We included only TNBC and HER2⁺ subtypes from patients who were surgically resected and irradiated. We excluded patients for whom material was unavailable or deteriorated. The protocol was approved by the Research Ethics Committee of the University of Navarra (reference 2019.180 y el 2019.181), and written informed consent was obtained from all living patients or the closest relative in the case of deceased patients. Clinical features of patients are specified in Supplementary Table S4. IHC analysis was performed in this cohort of primary naïve breast cancer tumors and local recurrences obtained from patients with breast cancer at the CUN. Tumors were fixed in 10% buffered formalin after surgical removal and paraffin-embedded using standard protocols. Tumors were classified according to the World Health Organization 2004 classification. For survival analysis, RFS and overall survival were calculated from the date of surgery to the date of recurrence or death according to the clinical information. For the evaluation, an H-score was calculated after IHC analysis by two expert pathologists in blind records. Patients were stratified into two groups according to the median of Enpp1 H-score.

The TMAs included 114 samples derived from 56 paired samples (primary and LRF) obtained from 28 patients (HER2⁺, n = 18; TNBC, n = 10) who failed locoregionally, and 48 single samples of HER2⁺ (n = 24) or TNBC (n = 24) patients who remained disease free for at least three years of follow-up after primary treatment. It also included 10 unpaired samples of TNBC or HER2⁺ patients who relapsed locally or elsewhere (TNBC, n = 8; HER2⁺, n = 2). In total, four TMAs were constructed using a 1-mm diameter needle in a manual tissue arrayer (MTA-1, Beecher Instruments). For better representation of the tumors, three cores obtained from different areas were included in the TMA. Consecutive 3-µm sections were obtained. TMAs were constructed by the Morphology Core Facility at the CIMA (Pamplona, Spain).

Statistical analysis was performed using GraphPad Prism 8.0. Survival curves were compared with log-rank test. Differences among groups were analyzed either by one-way ANOVA followed by Tukey post hoc test or by Brown–Forsythe and Tamhane T2 test. For nonparametric statistics, data were analyzed by Kruskal–Wallis test followed by Mann–Whitney multiple comparisons test with Bonferroni adjustment. Student t test was used where indicated.

Reagent tables are included in the Supplementary Section.

All RNA-seq and microarray data have been deposited in the Gene Expression Omnibus (accession number GSE193692).

N. Perurena reports other support from the Government of Spain during the conduct of the study. T. Walle reports grants from the European Commission, German Cancer Aid, and the German Cancer Research Center during the conduct of the study; grants from CanVirex outside the submitted work; and stock ownership in Roche, BionTech, Bayer AG, and AstraZeneca. V. Strnad reports grants, personal fees, and nonfinancial support from Nucletron Operations B.V., an Elekta Company, and personal fees from Siemens outside the submitted work. I. Melero reports grants and personal fees from Bristol Myers Squibb, Roche, Alligator, AstraZeneca, and Genmab, and personal fees from F-Star, Numab, Gossamer, Hotspot, Highlight Therapeutics, Biolinerx, and Amunix outside the submitted work. No disclosures were reported by the other authors.

B. Ruiz-Fernández de Córdoba: Formal analysis, validation, investigation, visualization, methodology, writing–original draft, writing–review and editing. H. Moreno: Data curation, formal analysis, funding acquisition, validation, investigation, visualization, methodology. K. Valencia: Data curation, validation, investigation, methodology, writing–review and editing. N. Perurena: Investigation, writing–review and editing. P. Ruedas: Formal analysis, validation, investigation. T. Walle: Data curation, formal analysis, validation, investigation. A. Pezonaga-Torres: Validation, investigation, visualization, methodology. J. Hinojosa: Validation, investigation, methodology. E. Guruceaga: Data curation, software, formal analysis. A. Pineda-Lucena: Investigation, methodology. M. Abengózar-Muela: Resources, data curation, investigation, visualization. D. Cochonneau: Investigation, visualization, methodology. C. Zandueta: Investigation, visualization, methodology. S. Martínez-Canarias: Validation, investigation, visualization. A. Teijeira: Validation, investigation, visualization, methodology, writing–review and editing. D. Ajona: Investigation, methodology, writing–original draft, project administration. S. Ortiz-Espinosa: Validation, investigation, visualization, methodology. X. Morales: Data curation, validation, investigation, visualization, methodology. C. Ortiz de Solórzano: Validation, writing–review and editing. M. Santisteban: Resources, writing–original draft. L.I. Ramos-García: Data curation, software, formal analysis. L. Guembe: Validation, investigation, methodology. V. Strnad: Investigation, visualization, methodology. D. Heymann: Formal analysis, validation, investigation, visualization, methodology. S. Hervás-Stubbs: Formal analysis, validation, investigation, visualization, methodology. R. Pío: Validation, methodology. M.E. Rodríguez-Ruiz: Data curation, formal analysis, investigation, methodology. C.E. de Andrea: Resources, data curation, supervision, validation, investigation, visualization, methodology. S. Vicent: Resources, validation, investigation, visualization, methodology, writing–review and editing. I. Melero: Formal analysis, investigation, methodology, writing–review and editing. F. Lecanda: Conceptualization, resources, formal analysis, supervision, funding acquisition, writing–original draft, project administration, writing–review and editing. R. Martínez-Monge: Conceptualization, resources, formal analysis, supervision, funding acquisition, investigation, writing–original draft, project administration, writing–review and editing.

This work was supported by FIMA, the Fondo de Investigación Sanitaria-Fondo Europeo de Desarrollo Regional “Una manera de hacer Europa” to R. Martínez-Monge (PI16/01847, PI19/01884, PI17/00411, and PI20/00419), Foundation AECC (PRYES211377MART), and the Gobierno de Navarra (ref. 34/2021). F. Lecanda was funded by Cancer Research Thematic Network of the Instituto de Salud Carlos III (RTICC RD12/0036/0066), SAF2015-71606R, RTI2018-094507-B-100 financed by MCIN/AEI/10.13039/501100011033 and by FEDER “Una manera de hacer Europa.” F. Lecanda was also funded by “la Caixa” Foundation, Caja Navarra Foundation, and the Foundation AECC (PROYE20083LECA). This study was also supported by the Foundation for Applied Medical Research (FIMA) and CIBERONC (CB16/12/00443). S. Vicent was supported by the Spanish Ministry of Economy and Competitiveness (MINECO, SAF2013-46423-R and SAF2017-89944-R), the European Commission (618312 KRASmiR FP7-PEOPLE-2013-CIG), Worldwide Cancer Research (16-0224), Red Temática de Investigación Cooperativa en Cáncer (RD12/0036/0040), La Caixa-FIMA agreement, ANOC, and Maria Eugenia Burgos de la Iglesia's family. C.O. de Solórzano was supported by the Spanish Ministry of Science, Innovation and Universities (RTI2018-094494-B-C22 and RTC-2017-6218-1; MCIU/AEI/FEDER, UE). The authors wish to thank Mrs. Ruth Breeze for editorial assistance. We are indebted to the members of the Morphology and Image Core facility, especially Cristina Ederra and Mikel Ariz and all members of the Animal Core Facility at CIMA. Graphics were designed with the free version of BioRender.

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