IRE1α regulates macrophage polarization, PD-L1 expression, and tumor survival

In the tumor microenvironment, local immune dysregulation is driven in part by macrophages and dendritic cells that are polarized to a mixed proinflammatory/immune-suppressive phenotype. The unfolded protein response (UPR) is emerging as the possible origin of these events. Here we report that the inositol-requiring enzyme 1 (IRE1α) branch of the UPR is directly involved in the polarization of macrophages in vitro and in vivo, including the up-regulation of interleukin 6 (IL-6), IL-23, Arginase1, as well as surface expression of CD86 and programmed death ligand 1 (PD-L1). Macrophages in which the IRE1α/X-box binding protein 1 (Xbp1) axis is blocked pharmacologically or deleted genetically have significantly reduced polarization and CD86 and PD-L1 expression, which was induced independent of IFNγ signaling, suggesting a novel mechanism in PD-L1 regulation in macrophages. Mice with IRE1α- but not Xbp1-deficient macrophages showed greater survival than controls when implanted with B16.F10 melanoma cells. Remarkably, we found a significant association between the IRE1α gene signature and CD274 gene expression in tumor-infiltrating macrophages in humans. RNA sequencing (RNASeq) analysis showed that bone marrow–derived macrophages with IRE1α deletion lose the integrity of the gene connectivity characteristic of regulated IRE1α-dependent decay (RIDD) and the ability to activate CD274 gene expression. Thus, the IRE1α/Xbp1 axis drives the polarization of macrophages in the tumor microenvironment initiating a complex immune dysregulation leading to failure of local immune surveillance.

both in the mouse [4,5] and in humans [6,7]. Because this phenomenon is considered at the 73 root of the dysregulation of local adaptive T cell immunity [8,9], much emphasis has been 74 placed on identifying common mechanisms driving the acquisition of tumor-promoting 75 properties by macrophages and dendritic cells in the TME [5, [10][11][12][13][14]. 76 The TME is home to environmental noxae such as hypoxia and nutrient deprivation 77 [15]. In addition, about 20% of tumors have a viral origin [16] and most (90%) solid tumors 78 carry chromosomal abnormalities [17]. These events, independently or collectively, can lead 79 to a dysregulation of protein synthesis, folding, and secretion [18,19], and the accumulation 80 of misfolded proteins within the endoplasmic reticulum (ER), triggering a stress response 81 termed the unfolded protein response (UPR) [20]. The UPR, an evolutionarily-conserved 82 adaptive mechanism [21], is mediated by three initiator/sensor ER transmembrane molecules: 83 inositol-requiring enzyme 1 (IRE1a), PKR-like ER kinase (PERK), and activating 84 transcription factor 6 (ATF6). In the unstressed state these three sensors are maintained 85 inactive through association with the 78-kDa glucose-regulated protein (GRP78) [22]. During 86 conditions, cross-priming of naïve CD8 + T cells by BMDC is greatly compromised [28]. In 115 line with this observation, Cubillos-Ruiz reported that the incubation of BMDC in ovarian 116 cancer conditioned media results in Xbp1 splicing, and that the conditional knock-out of 117 Xbp1 in dendritic cells improves antigen presentation and significantly reduces tumor growth 118 in vivo [36]. In line with these observation is a report showing that GRP78 in cancer cells 119 regulates macrophage recruitment to mammary tumors through metabolites secreted from 120 cancer epithelial cells [37]. Thus, UPR-driven cell-nonautonomous mechanisms play a 121 hitherto unappreciated role in orchestrating immune cells in the TME and driving their 122 dysregulation, so as setting the stage for failure of local immune surveillance. 123 We therefore decided to elucidate the mechanism(s) through which the UPR may 124 ultimately affect immune cells and perturb the TME to promote tumor growth. We focused 125 on macrophages as these cells represent the major population infiltrating most solid tumors in 126 humans, conspicuously more abundant than dendritic cells and other cells of myeloid origin 127 [38]. Relative to dendritic cells or myeloid derived suppressor cells (MDSCs) [39,40] little 128 is known about how the UPR affects macrophages during cancer development. Based on our 129 earlier report that BMDM can be polarized to a mixed IIS phenotype via a UPR-mediated 130 cell-nonautonomous mechanism [26] our initial goal was to verify whether this phenomenon 131 could be recapitulated in tumor-infiltrating macrophages in vivo in immunocompetent mice, 132 and what UPR pathway might contribute to their dysregulation. To this day, these questions 133 have remained largely unanswered. Here we show that the UPR and the IRE1a/XBP1 axis 134 are activated in macrophages during tumor growth, that the conditional knock-out of IRE1a 135 in macrophages regulates the acquisition of a mixed IIS phenotype and is also sufficient to 136 restrain tumor development in vivo. Importantly, we discovered that IRE1a signaling 137 regulates PD-L1 expression in murine and in tumor-infiltrating macrophages in humans. 138

Tumor infiltrating CD11b + myeloid cells display the UPR/IIS signature in vivo 140
Previous in vitro studies indicated that BMDC and BMDM respond to a cell-nonautonomous 141 UPR developing a complex phenotype characterized by a UPR activation and a mixed pro-142 inflammatory/immune suppressive (IIS) phenotype [26,28]. Here as an initial step we 143 interrogated tumor-infiltrating myeloid cells (CD11b + ) to document these characteristics 144 during tumor growth in vivo. To this end, we implanted B16. CD11b + myeloid cells (Fig. S1). Of these ~50% expressed the F4/80 surface marker specific 151 of macrophages. We then compared the expression of the Venus protein in tumor-infiltrating 152 CD11b + cells to those in the spleen and bone marrow, both from tumor-distal and tumor-153 proximal femurs (Fig 1A). The Venus protein signal was significantly higher in tumor-154 infiltrating CD11b + cells relative to those in control tissues, suggesting a concurrent UPR 155 signaling with XBP1 splicing in the TME only. 156 Having established that XBP1 splicing occurs in tumor-infiltrating CD11b + cells, we 157 sought to detect other features of the IIS phenotype. To this end, we implanted B16.F10 cells 158 in wild-type C57BL/6 mice and isolated by positive selection CD11b + cells from tumor, 159 spleen and bone marrow 22 days post-implantation. Phenotypically, the isolated cells were 160 CD11b + and Gr1and showed the transcriptional upregulation of three key UPR genes: 161 Grp78, a downstream target of the ATF6 pathway, spliced Xbp1 (Xbp1-s) a downstream 162 product of the IRE1a pathway, and Chop, a downstream product of the PERK pathway (Fig.  163 1B). A transcriptional upregulation of all three genes suggested the activation of a classical 164 UPR. Contextually, CD11b + cells also showed the transcriptional upregulation of Il23p19, a 165 key pro-inflammatory cytokine gene, and Arginase-1 (Arg1), an immune suppressive enzyme 166 (Fig. 1C). 167 To see if the UPR/IIS signature also hallmarks CD11b + cells during spontaneous 168 tumor growth, we interrogated mice with mutations in the adenomatous polyposis coli (Apc) 169 gene ("Apc mice"), which develop small intestinal adenomas by 30 days of age [42]. We 170 pooled CD11b + cell infiltrates from adenomas from multiple Apc mice and probed the 171 expression of UPR genes, Il-23p19 and Arg1 relative to CD11b + cells isolated from either the 172 bone marrow or the spleen as controls. CD11b + cells from APC adenomas had increased 173 expression of UPR genes, Il-23p19 and Arg1 (Fig. 1D Fig. 2A). However, it significantly inhibited the transcriptional activation 188 of Il-6 and Il-23p19 (Fig. 2B) and trended towards inhibiting Arg1 (p=0.127) (Fig. 2C). 189 Previously, we showed that TERS CM promotes the expression of CD86 and PD-L1 in 190 BMDC [28]. Herein, we determined that ERAI BMDM treated with TERS CM also 191 upregulate CD86 and PD-L, and that such an upregulation that is markedly inhibited by 4µ8C 192 (Fig. 2D). 193 The involvement of the PERK pathway on the acquisition of the IIS phenotype by 194 BMDM was assessed using the small molecule GSK2656157, a preferential PERK inhibitor 195 [43]. GSK2656157 efficiently inhibited PERK phosphorylation ( Fig S3A) but had no effect 196 on the upregulation of Grp78, Il-6 and Arg1 induced in BMDM cultures by TERS CM (Fig.  197 S3B). Congruently, PERK inhibition had little to no effect on the surface expression of CD86 198 and PD-L1 ( Fig S3C). Collectively, these results suggest that BMDM polarization to the IIS 199 phenotype is IRE1a dependent. . IRE1a. Lactic acid induced Arg1 only, and 4HNE had no effect on any of the target genes 207 studied. Interestingly, 4µ8C reduced the induction of Arg1 by both LPS and lactic acid, 208 suggesting that the IRE1a may regulate the expression of this immune suppressive molecule 209 outside the context of the UPR (Fig. S4). 210

Loss of IRE1a-Xbp1 in macrophages attenuates the IIS phenotype, PD-L1 expression 211 and tumor growth in vivo 212
Earlier reports showed that XBP1 is required for the development and survival of bone 213 marrow derived DC [44], and that the deletion of XBP1 in lymphoid DC [40,45]   To ascertain the physiological relevance of these findings, we next assessed the 250 survival of Ern1 and Xbp1 CKO mice implanted with B16.F10 melanoma cells. We reasoned 251 that survival would constitute an optimal initial read-out for the complex interactions between 252 cancer cells and immune cells in the TME with focus on the IRE1a-XBP1 axis in myeloid 253 cells. Survival in Ern1 CKO mice was significantly greater (p=0.03) than in control Ern1 fl/fl 254 mice (Fig. 4A). By contrast, Xbp1 CKO mice survived longer than control Xbp1 fl/fl mice but 255 the difference was non-significant (Fig. 4A). Based on survival data we isolated F4/80 256 tumor-infiltrating macrophages of tumor-bearing Ern1 CKO mice to assess the UPR/IIS and 257 Cd274 gene expression status. Xbp1s, Il-23p19, Arg1 and Cd274 genes were all markedly 258 reduced in Ern1 CKO macrophages compared to their Ern1 fl/fl counterpart (Fig. 4B). 259 Together, these results point to macrophage IRE1a as a key negative regulator of TME 260 immunodynamics and tumor growth in vivo. 261

Loss of RIDD regulation in Ern1 CKO macrophages 262
Because the IRE1a-XBP1 axis also regulates PD-L1 expression and both Ern1 and Xbp1 263 CKO BMDM showed significantly-reduced surface PD-L1 protein expression compared to 264 fl/fl BMDM (Fig. 3E), we decided to distinguish the relative contribution of Xbp1 splicing 265 and RIDD to this phenomenon. To this end, we performed RT-qPCR on Ern1-and Xbp1 266 CKO BMDM treated or not with TERS CM relative to fl/fl controls. We found that Cd274 The genotype of each mouse used in this experiment is shown in Fig. S5. Upon TERS CM 277 treatment Ern1 expression in Ern1 CKO macrophages was 1.79-fold over that of untreated 278 cells compared to 3.26 fold in fl/fl macrophages (Fig. 5B). We found that consistent with the 279 flow cytometry data, Cd274 (PD-L1) expression was markedly increased in macrophages 280 (44.45-fold) but only moderately increased in Ern1 CKO macrophages (4.11-fold, Fig 5C). 281 Thus, both genetic and chemical inhibition of IRE1a signaling yielded concordant results. 282 Next, we performed a comprehensive analysis of RIDD activity using a set of 33 283 putative RIDD target genes previously defined [49]. We found that only half (sixteen) of 284 these genes behaved as bona fide RIDD targets in TERS CM-treated BMDM (i.e., decreased 285 expression after TERS CM treatment in fl/fl macrophages) (Fig. 5D, upper panel). We found 286 that in Ern1 CKO macrophages, there was a clear loss of a "RIDD signature" compared to 287 fl/fl macrophages, both basally and after TERS CM treatment (Fig. 5D, lower panel). When 288 considered together through an analysis of the mean z-score for the 16 genes, it became 289 apparent that TERS CM induction of RIDD activity was much more effective in fl/fl than in 290 Ern1 CKO macrophages (Fig. 5E). Collectively, these results show that macrophages 291 lacking Ern1 lose RIDD regulation, suggesting that RIDD may be implicated in the 292 regulation of PD-L1 expression. 293 In the same analysis we found that Tabpb (

A link between IRE1a and PD-L1 expression in human tumor-infiltrating macrophages 302
The data reported herein suggest that Cd274 gene expression in murine macrophages is 303 positively regulated by IRE1a. Recently, Xu et al. [51] reported that PD-L1 protein 304 expression in murine MYC tg :KRAS G12D tumor cells is decreased by a small molecule that 305 enables the cell to resume translation while the eIF2a downstream from PERK remains 306 phosphorylated. Therefore, we decided to study the relationship between CD274 (PD-L1) 307 gene expression and the two major UPR pathways, IRE1a and PERK, across multiple human 308 cancers. We began by interrogating the relative contribution of ERN1 (IRE1a) and EIF2AK3 309 (PERK) to CD274 gene expression. In this analysis, we queried The Cancer Genome Atlas 310 (TCGA) collection of RNA-sequencing expression data for bulk samples from thirty-one 311 tumor types. Across these data, we observed that ERN1 correlates strongly with EIF2AK3 312 (Pearson correlation coefficient = 0.55; p < 1e-200) (Fig. S8A), and that both ERN1 (p ≤ 313 1.46e-51) and EIF2AK3 (p ≤ 1.62e-44) correlate positively with CD274, suggesting that the 314 UPR plays a role in CD274 gene expression. These correlations prompted us to further 315 interrogate the relationship between CD274, ERN1 and EIF2AK3, with respect to levels of 316 infiltrating macrophages in bulk tumor samples approximated by a macrophage score derived 317 from the geometric mean of three genes expressed by macrophages (CD11b, CD68, and 318 CD163). We found a positive correlation between ERN1 and CD274 within the high 319 macrophage infiltration group ( > 70th percentile) (Spearman correlation coefficient 0.18; p 320 < 1.3e-21) (Fig. S8B). By contrast, the low macrophage infiltration group (< 30 th percentile) 321 had a much weaker correlation (Spearman correlation coefficient 0.06; p < 0.001) (Fig.  322 S8B). On the other hand, EIF2AK3 and CD274 within the high macrophage infiltration group 323 ( > 70th percentile) had a lower correlation (Spearman correlation coefficient 0.09; p < 1.9e-324 7) than in the corresponding Ern1 group (Fig. S8C). Finally, EIF2AK3 and CD274 within the 325 low macrophage infiltration group (< 30th percentile) had a surprisingly higher correlation 326 (Spearman correlation coefficient 0.15; p < 8.32-15) than in the respective high macrophage 327 infiltration group (Fig. S8C). Collectively, this analysis suggests that when macrophage 328 infiltration is high, Ern1 is a better predictor of CD274 gene expression than EIF2AK3. 329 We also integrated the macrophage score with ERN1 and EIF2AK3 to predict CD274 330 expression in an ordinary least squares (OLS) linear regression model, including the tumor type as a covariate (Table 1). We found that this model assigns significant, positive 332 coefficients for the interaction terms of macrophages with ERN1 (ERN1*Macrophages, beta 333 coefficient = 0.0012, p < 0.023) but not EIF2AK3 (EIF2AK3*Macrophages, beta coefficient 334 = 0.0007, p < 0.155), suggesting that ERN1 but not EIF2AK3 is predictive of CD274 gene 335 expression within tumor-infiltrating macrophages in individual tumor types (Table 1). To 336 validate these results we analyzed RNA-Seq data generated from macrophages isolated from 337 thirteen patients with either endometrial or breast cancer [52]. We found a strong Pearson 338 correlation coefficient between ERN1 and EIF2AK3 in these data (correlation coefficient 339 0.738; p < 0.003), suggesting UPR activation. Since IRE1a activity is a multistep and 340 complex process [53] and may not be completely captured by ERN1 expression levels, we 341 derived a systemic representation of pathway activity controlled by IRE1a and by 342 comparison PERK. We collected sets of downstream genes in the IRE1a and PERK 343 pathways [54], and derived aggregate scores for each pathway from the mean expression 344 signal of all detectable genes after z-score transformation. Since the transformed pathway 345 scores could potentially amplify noise from genes with low expression, we applied filters to 346 include only genes in each pathway with levels beyond a specific threshold (Fig. S9). We 347 varied this filter threshold from zero to one thousand raw counts and then included the 348 pathway activity scores in multiple OLS linear models to predict CD274 across tumor-349 infiltrating macrophage samples (Fig. 6A). We found that a filter threshold of 100 counts 350 effectively reduced noise while preserving signal from 84% of detectable genes in both the 351 IRE1a and PERK pathways. In this model, the IRE1a score predicted CD274 expression 352 with a significant positive beta coefficient (beta coefficient = 21.043, p-value = 0.040), while 353 the PERK score was non-significant (beta coefficient = 36.842, p-value = 0.103). This 354 pattern of significant IRE1a coefficient and nonsignificant PERK coefficient was consistent 355 across all filter thresholds (Fig. 6B). Comparing models wherein CD274 expression was 356 explained by IRE1a activity alone or by both IRE1a and PERK activity using the Aikake 357 information criterion analysis shows that a model containing both is as 0.54 times as probable 358 as the IRE1a alone to minimize the information loss (DAIC = 1.23). Taken together, these 359 analyses suggest that the activation of CD274 gene expression in tumor-infiltrating 360 macrophages depends primarily on the IRE1a pathway.  Importantly, 4µ8C also inhibited the TERS CM-induced upregulation of Arg1, and that of 414 pro-angiogenic vascular endothelial growth factor (VEGF) (Fig. S10). Since IL-6 and IL-23 415 are known to bias T cell differentiation towards inflammatory (Th17) or regulatory T cells 416 Nested OLS models with ERN1 only and ERN1 + PERK were compared using the Aikake 640 information criterion (AIC). For each model, the AIC was calculated as AIC = 2k -2ln(L), 641 where k represents the number of estimated parameters, and L represents the likelihood 642 function for the model. Models were compared using the formula exp((AICmin − AICi)/2), 643 which represents the relative likelihood of model i with respect to the best available model. 644

Statistical analysis 645
To determine if differences between groups were statistically significant for PCR 646 experiments, groups were compared using unpaired student's t-tests with Welch's correction.