A Four-Chemokine Signature Is Associated with a T-cell–Inflamed Phenotype in Primary and Metastatic Pancreatic Cancer

Pancreatic ductal adenocarcinoma (PDAC) constitutes the majority of pancreatic cancers (1) and has a grim 5-year survival rate of 9% (2). With few recent major improvements in treatment or survival, novel therapeutic strategies are urgently needed (3). Immune-checkpoint blockade has shown sustained clinical success in other cancers, but to date has remained generally ineffective in PDAC (4–6). Reasons for this include our limited understanding of the molecular mechanisms governing the establishment of a functional antitumoral T-cell–inflamed phenotype and the resultant lack of treatment response biomarkers in PDAC (7, 8).

Currently used immune-checkpoint inhibitors targeting CTLA-4 and the PD-1/PD-L1 axis appear to require a preexisting, primed effector antitumor CD8⁺ T-cell infiltrate (8). Cancers with high tumor mutation burden (TMB) and genomic instability, such as microsatellite unstable colorectal cancers, display a high CD8⁺ infiltrate and good outcomes (9). On average, PDAC has low TMB (5, 10–13), being driven instead by chromosomal instability (14), which is thought to cause the lack of response to immune-checkpoint blockade (7). Yet, a small portion of PDAC patients presents with a substantial CD8⁺ T-cell infiltrate that is prognostically favorable (15–17), indicating the presence of occult drivers of antitumor immunity in this cancer type.

Chemokines are increasingly recognized for their chemotactic effects, driving primed effector T-cell recruitment into tumors (18). Strong associations of chemokines with CD8⁺ infiltration in different tumors have been demonstrated, including prostate, colorectal, sarcoma, melanoma, and lung (19–24). However, the role of chemokines in mounting an active T-cell–inflamed antitumor phenotype in PDAC has not been investigated. Here, we report on the expression of a distinct chemokine set (CCL4, CCL5, CXCL9, and CXCL10) that may be central to the establishment of a T-cell–inflamed phenotype in PDAC and could aid the identification of PDAC patients who might benefit from immunotherapies.

A schema of Materials and Methods is outlined in Supplementary Fig. S1A. Following informed consent and approval from Institutional Review or Institutional Research Ethics Boards, primary and metastatic human PDAC samples and clinical metadata were collected by the PanCuRx Translational Research Initiative at the Ontario Institute for Cancer Research (OICR). Samples underwent tumor enrichment by laser capture microdissection (LCM), and whole-genome and transcriptome sequencing, as described previously (25–27). Exclusion criteria included receipt of neoadjuvant therapy, mismatch-repair–deficient/homologous repair–deficient (MMRd/HRD) tumors and cellularity below 40%. Our discovery cohort comprised 113 primary PDAC resections and 107 liver metastases from the COMPASS trial (NCT02750657; ref. 26), with genomic data for 112 and 103 cases, respectively. Two validation cohorts were used, consisting of 182 PDAC patients from The Cancer Genome Atlas (TCGA; ref. 28) data portal (https://tcga-data.nci.nih.gov, 2018 release) and 92 PDAC patients from the Australian Pancreatic Cancer Initiative of the ICGC Au (29) data portal (https://dcc.icgc.org/projects/PACA-AU, downloaded October 2018). Research conducted using human subjects at the University Health Network, overseen by the UHN Research Ethics Board, adheres to the Tri-Council Policy Statement, The International Conference on Harmonization: Good Clinical Practice, including the Declaration of Helsinki and the Belmont Report.

Tissue microarrays (TMA) were created as described previously (27) and stained for intratumoral CD3 (1:25; Ventana, 2GV6), CD8 (1:4; Dako, C8/144B), CD4 (prediluted; Dako, IR649), and FOXP3 (1:200; eBioscience, 14-4777-82; tumor cores were obtained from tumor centers). CD3 and CD8 TMAs were scanned using Leica Aperio AT2 under 20× objective. CD4 and FOXP3 TMAs were scanned using 3D HISTECH Panoramic SCAN under 20× objective. Of the 113 primary resected PDACs, a subset of 79 with available matching CD8⁺ tumor-infiltrating lymphocytes (TIL) counts and RNA-seq data were used to investigate the association between chemokine expression and CD8⁺ infiltration. CD8⁺ TILs were manually quantified by a pathologist and were defined as a CD8⁺ cell in direct contact by at least one dimension with a tumor cell. The number of CD8⁺ TILs was normalized by tumor area (total tissue area after exclusion of non-tumor regions) in μm² calculated using QuPath Image analysis software (30). CD8⁺ TIL counts were averaged across multiple tumor cores per patient, where available, and log₂ transformed. Total cell counts for additional cell populations in the 4-chemokine^hi and 4-chemokine^lo groups in primary resected PDACs (n = 58 of 113) were quantified using the Positive Cell Detection algorithm of QuPath Image analysis software and again normalized to tumor area. Inclusion criteria for automated quantification included (i) staining and image quality, (ii) sufficient tissue area for analysis, and (iii) samples with matched RNA-seq (Supplementary Fig. S1B). Tumors devoid of immune cell counts were included. Available matched tissues were for CD8⁺ (n = 39), CD3⁺ (n = 40), CD4⁺ (n = 21), and FOXP3⁺ cells (n = 32). Total cell counts for the 4-chemokine^hi and 4-chemokine^lo groups in liver metastases (n = 54 of 107) were calculated for CD8⁺ (n = 12) and CD3⁺ (n = 15) using the Immune Cell algorithm of HALO Image Analysis software (Indica Labs) and normalized by total tissue area. Slides were scanned using Leica Aperio AT2 under 40× objective. Samples where CD8⁺/CD3⁺ ratio was > 1 or with denominators containing zero were excluded. For log₂ transformation, samples with zero immune cells per area were substituted for 90% of the lowest immune count in the cohort. For primary resections, redundant cores were available and normalized per patient as indicated above.

Formalin-fixed paraffin-embedded blocks from 10 patients in the 113 primary resected cohort were retrieved and cut, following H&E review of blocks to select for sections with adequate viable tumor regions. Sections were stained for CCL4/MIP-1β (1:100; Abcam, catalog number EP521Y; stomach as positive control), CCL5/RANTES (1:200; Thermo Fisher Scientific, catalog number 701030; tonsil as positive control), CXCL9/MIG (1:20; RD Systems, catalog number MAB392; kidney as positive control), CXCL10/IP-10 (1:200; Santa Cruz, catalog number sc-101500; tonsil as positive control), CD11c (1:200; Abcam, catalog number EP1347Y), and CD8 (prediluted; Ventana/Roche, catalog number 5937248001) using a Ventana automated staining standard protocol (DDP Lab, UHN). Slides were scanned using Leica Aperio AT2 under 40× objective.

RNA-seq expression analysis was conducted using normalized transcripts per million (TPM). Genes were first filtered to include only protein-coding genes, including immunoglobulin and T-cell receptor genes. TPM was then computed, and values were gene-wise normalized by subtraction of median expression of log₂-transformed gene expression. Gene-wise normalization was performed separately on each cohort. To account for differences between samples due to differences in read mapping, normalized TPM values underwent sample-wise normalization using the median expression of 15 housekeeping genes (31). Subsequent gene-expression analyses were performed using normalized TPM values.

Differential gene-expression analysis was conducted using the edgeR package (v3.24.1; ref. 32). RNA-seq raw counts were filtered for protein-coding genes with a count value of ≥ 25 in a minimum of 75% of patients (33). Differentially expressed genes were defined at false discovery rate (FDR) < 0.05. The Immunology Database and Analysis Portal (ImmPort) database of InnateDB was used to identify genes related to immune function (34). Gene set analysis was conducted using the single-sample Gene Set Enrichment Analysis (ssGSEA) algorithm (35), implemented in the Gene Set Variation Analysis (GSVA) package (v1.30.0; ref. 36). Raw TPM values were filtered by omitting values less than 0.01. Genes expressed in less than 70% of samples were removed, followed by removal of genes with median TPM expression < 0.1. Fold changes to median expression were then calculated, and maximum fold changes were capped at 50-fold change. Fold-change TPM values were then log₂ transformed, and these values were used for ssGSEA. To account for the continuous nature of TPM values, the default Gaussian distribution for the kcdf argument of the gsva function was used (36), with a minimum and maximum gene set size of 2 and 500, respectively. ssGSEA enrichment scores were calculated using gene sets comprised of genes characteristic of immune cells (37) and pathways (38, 39), and T-cell–inflamed signatures (40–42), following removal of the 4-chemokine panel from gene sets. CYT (38) and Batf3-dendritic cell (Batf3-DC; ref. 43) scores were calculated by taking mean normalized gene-expression values of genes from gene sets.

For enrichment map, a ranked list of genes created using –log₁₀(P) × sign(log(fold change)) (44) from DGE edgeR output was used to generate enrichment scores using the GSEAPreranked module of GSEA (v3.0). Given the symmetrical nature of the ranked list, the weighted statistic was applied. Hallmark and C2:Canonical Pathway gene sets were downloaded from the MSigDB database (v6.2; ref. 39; http://software.broadinstitute.org/gsea/msigdb/genesets.jsp). Gene sets were capped at < 200 genes, with 1,000 permutations (44). The resulting differentially expressed genes between 4-chemokine^hi and 4-chemokine^lo patients, with FDR < 0.05, were visualized in an enrichment map using Cytoscape (v3.7.1), as described previously (44).

All bioinformatic and statistical analyses and plotting were performed using the R statistical programming language (v3.5.1). Plots were generated using the ggplot2 package (v3.1.0). Genomic data, including mutational burden and neoantigen counts, were obtained as described previously (27). P values adjusted for multiple tests were corrected using the Benjamini–Hochberg method.

Direct cell–cell contacts between CD8⁺ cytotoxic T cells and tumor cells are a prerequisite for effective antitumor killing (45). Thus, we first determined the number of TILs defined as CD8⁺ cells in direct physical contact with tumor cells by immunohistochemistry (IHC) of PDAC tissue microarrays. CD8⁺ TILs were absent in a subset of cases (15.19%; 12/79), whereas the majority of PDAC tumors contained CD8⁺ TILs in varying numbers (Fig. 1A). We leveraged matched RNA-seq data from these 79 patients to test if this tumor cell–adjacent CD8⁺ TIL infiltration was associated with specific chemokine expression. Importantly, the tissues had undergone tumor enrichment by LCM prior to RNA-seq profiling, so that this analysis of gene expression was also confined to the vicinity of tumor cells. Taking an unbiased approach, we found that among the 40 known human chemokines (46), only four showed a positive and significant correlation with CD8⁺ TIL counts (adjusted R² > 0.19; adjusted P < 0.001), namely, CCL4, CCL5, CXCL9, and CXCL10 (Fig. 1B). Further, these four chemokines were among the top 25 of all coding genes associated with CD8⁺ TIL counts (Supplementary Table S1).

Protein expression of these four chemokines in PDAC tissues (n = 10) was confirmed in our primary resected cohort by IHC (Fig. 1C). To further test whether expression of these chemokines was directly associated with tumor cell–adjacent CD8⁺ TILs, or instead resulted from underlying differences in mutational burden, we assessed potential associations with the number of single-nucleotide variants (SNV) and neoantigen counts. CCL4, CCL5, CXCL9, and CXCL10 expression was associated with neither SNV (Fig. 1D, top) nor neoantigen counts (Fig. 1D, bottom). Also, neither SNV nor neoantigen counts were associated with CD8⁺ TILs in PDAC tumors (Fig. 1E, left), except in MMRd/HRD (1 MMRd and 10 HRD) patients from a separate patient set (Fig. 1E, right). Overall, expression of four specific chemokines (CCL4, CCL5, CXCL9, and CXCL10) in PDAC was directly associated with the number of tumor cell–adjacent CD8⁺ TILs, a histologic metric of T-cell–inflamed tumors.

Next, we assessed whether CCL4, CCL5, CXCL9, and CXCL10 expression in PDAC was coregulated or marked independent patient subsets. Unsupervised hierarchical clustering of RNA-seq data from an expanded set of 113 treatment-naïve, non-MMRd/HRD, tumor-enriched, resectable PDAC revealed significant correlation among expression levels of all four cytokines (subsequently referred to as “4-chemokine signature”; Fig. 2A). Of note, CXCL11 expression also correlated with CCL4, CCL5, CXCL9, and CXCL10 expression (Fig. 2A) but was not included in the signature for lack of a strong positive correlation with CD8⁺ TIL counts (adjusted R² = 0.13; adjusted P = 0.004; Fig. 1B).

To gain insights into the role of these four chemokines in PDAC immunobiology, we performed comparative gene-expression analysis. We divided the cohort into three groups (4-chemokine^hi, 4-chemokine^med, and 4-chemokine^lo; n = 29, 55, and 29, respectively) based on the quartiles of mean CCL4, CCL5, CXCL9, and CXCL10 coexpression (Fig. 2B). We investigated whether the 4-chemokine signature was associated with established transcriptional subtypes of PDAC (12, 47, 48) and observed no differences between groups (Supplementary Fig. S2). Differential gene-expression analysis revealed 245 upregulated and 35 downregulated genes in 4-chemokine^hi compared with 4-chemokine^lo tumors (Fig. 2C, left). Interestingly, the top upregulated gene was IDO1 (Fig. 2C, right), a marker of T-cell exhaustion (49). In total, 62 of the top 100 genes upregulated in the 4-chemokine^hi group (Supplementary Table S2) had well-known immune-related functions (Fig. 2C, right), indicating differences primarily in immune function between the two groups. This included several established markers of increased immune activity (IL2RB, IL2RA, ZAP70, ITK, CD8A, CD3E, and CD38; Fig. 2C, right), indicative of T-cell activation (50). To identify differentially expressed gene sets between 4-chemokine^hi and 4-chemokine^lo patients, a preranked list of differentially expressed genes was generated from DGE output of edgeR, followed by exclusion of CCL4, CCL5, CXCL9, and CXCL10. Gene set enrichment analysis, using the MSigDB Hallmark and C2:Canonical Pathways curated gene sets (51), revealed a pronounced enrichment in processes involved with immune regulation, T-cell signaling, cytokine signaling, and immune extravasation in the 4-chemokine^hi samples (Fig. 2D). Interestingly, we also observed upregulation of innate immunity pathways in the 4-chemokine^hi group, including pathways associated with immune activation, MHC class II presentation, and antigen presentation, and increased KRAS signaling (Fig. 2D). Taken together, expression levels of a coregulated 4-chemokine signature denoted a main difference in immune properties of human PDAC.

Prompted by the differences in immune-related gene and pathway expression between 4-chemokine^hi and 4-chemokine^lo PDAC tumors, we next determined immune cell profiles of the two patient groups. ssGSEA confirmed an expected main difference in expression of marker gene sets for cytotoxic cells (Fig. 3A), in addition to increased expression of marker gene sets for T-cell subsets [Th1 and Th2 helper cells, T-follicular helper cells (TFH) and T-effector memory cells (Tems)] in 4-chemokine^hi patients (Fig. 3A). Interestingly, marker scores for DCs (including activated and immature), neutrophils, and B cells were also elevated (Fig. 3A), indicating the presence of both effector and professional antigen-presenting immune populations in 4-chemokine^hi PDAC tumors. We furthermore quantified major T-cell populations by IHC in subsets of this cohort with available matched tissues. Again, and in concordance with the ssGSEA results, we observed increased amounts of CD8⁺ T cells per total tissue area in 4-chemokine^hi PDAC tumors, along with higher numbers of CD3⁺ T cells. Although we did not observe higher CD4⁺ helper T-cell counts, we did see increased FOXP3⁺ T-regulatory cells in this group, suggesting that an antitumor response has been initiated that consequently triggered the regulatory response (Fig. 3B). The ratios of CD8⁺/CD3⁺, CD8⁺/CD4⁺, and CD8⁺/FOXP3⁺ were similar in both groups, suggesting that the 4-chemokine signature not only identifies tumors with high CD8⁺ counts but broadly reflects increased lymphoid infiltrate required for, and in response to, active antitumor immunity.

We proceeded to further interrogate CD8⁺-specific metrics of functional antitumor immunity. 4-chemokine^hi PDAC tumors exhibited significantly higher cytolytic activity metric (CYT; Fig. 3B, bottom left), indicative of functionally active cytotoxic lymphocytes (38), and enrichment of gene sets involved with T-cell priming and activity, including T-cell costimulation and Type I IFN expression (Fig. 3C). T-cell exhaustion reflects a tumoral mechanism of immune evasion (49). We found increased expression of major markers of T-cell coinhibition and T-cell exhaustion (PDL1, PD1, TIM3, LAG3, TIGIT, CTLA4, and FASLG) in the 4-chemokine^hi group compared with the 4-chemokine^lo group (Fig. 3D). Importantly, we also observed enrichment of T-cell–inflamed gene-expression profiles (refs. 8, 42; Fig. 3E), including signatures recently shown to be predictive of response to immunotherapy across different cancer types (40, 41).

An effective T-cell response depends on presentation of antigens and priming of T cells by licensed professional antigen-presenting cells (APC; refs. 45, 50). We thus also interrogated the respective processes involved in APC maturation, and the innate arm of the immune system, which is implicated in the latter (52–54). In the above analysis, we had observed upregulation of gene sets specific for innate immune signaling pathways (Fig. 2D) and activated DCs (Fig. 3A). Furthermore, ADAMDEC1, a gene associated with an activated mature DC (55, 56) group (Fig. 2C), and GPB5 (Supplementary Table S2), a gene induced by IFN and promoting the NLRP3 inflammasome (57), were among the top upregulated genes in the 4-chemokine^hi. ssGSEA further revealed increased expression of gene sets involved in APC function, including MHC class I, APC coinhibition and costimulation in 4-chemokine^hi patients (Fig. 4A). Indeed, cells positive for the DC marker CD11c⁺ (41) colocalized with CD8⁺ T cells in 4-chemokine^hi patients (Fig. 4B). Furthermore, this group displayed increased expression of a Batf3-DC signature (ref. 43; Fig. 4C), indicating the presence of a key APC population for antigen cross-presentation and potentially effector T-cell recruitment (58). Innate immune-sensing pathways are important for APC stimulation and activation (8, 50). Interestingly, we observed both the stimulator of interferon genes (STING) and the NLRP3 inflammasome pathway (39) upregulated in the 4-chemokine^hi group (Fig. 4D). Together, these data suggest that 4-chemokine signature expression in PDAC is associated with innate immune-sensing pathways and APC stimulation that are crucial for T-cell activation.

The results described above show that increased expression of a 4-chemokine signature marks PDAC with a T-cell–inflamed phenotype and related innate immune-sensing pathways, independent of mutational burden. These findings were tested in two independent replication cohorts (92 tumors from the ICGC Au, ref. 59; and 182 PDAC from TCGA, ref. 28). Consistent with the observations from our cohort, the genes in the 4-chemokine signature were coregulated in both replication cohorts (Fig. 5A; Supplementary Fig. S3A). 4-chemokine^hi tumors (top quartile of 4-chemokine panel expression; Fig. 5B; Supplementary Fig. S3B) exhibited distinct expression patterns for immune cell–specific gene sets, including cytotoxic cells, other T-cell populations (pan T cells, Th1 cells, Tems, etc.), and activated DCs, B cells, and neutrophils (Fig. 5C; Supplementary Fig. S3C). Furthermore, we observed the same increase in CYT scores (Fig. 5D, left; Supplementary Fig. S3D, left) and enrichment in T-cell–inflamed gene-expression profiles predictive of immunotherapy efficacy (Fig. 5D, right; Supplementary Fig. S3D, right) in the 4-chemokine^hi PDAC tumors, along with upregulation of T-cell exhaustion genes (Fig. 5E; Supplementary Fig. S3E) and pathways involved with T-cell and APC priming and exhaustion (Fig. 5F and G; Supplementary Fig. S3F–S3G). Finally, the 4-chemokine^hi groups in the ICGC Au and the TCGA cohorts again showed increased expression of the Batf3-DC score, and the NLRP3 inflammasome (Fig. 5H; Supplementary Fig. S3H). STING signaling was upregulated in the ICGC Au (Supplementary Fig. S3H), whereas no differences were observed between cohorts in the TCGA data set (Fig. 5H, right).

The ICGC and the TGA cohorts comprise bulk samples, whereas our specimens were tumor cell–enriched by LCM. To control for a potential influence of tumor purity on the gene-expression data, we investigated CELLULOID-based cellularity estimates (14) for our samples. Mean cellularity of 4-chemokine^hi and 4-chemokine^lo samples was 68% and 75%, respectively (Supplementary Table S3A). To rule out the potential effects of these slight differences, we reanalyzed subsets with adjusted cellularity estimates for both groups. To do this, we created two cohorts from 4-chemokine^hi and 4-chemokine^lo subsets with similar cellularities by taking the average of the mean cellularity of each group, and sampling patients who were above and below 50% of this calculated average. This led to two comparable sample cohorts of 16 patients each, with adjusted mean cellularity estimates of 71% for both groups (Supplementary Fig. S4A; Supplementary Table S3B). These adjusted cellularity subgroups yielded comparable results (Supplementary Fig. S4B–S4D), rendering an influence of sample purity on our results highly unlikely. Overall, the 4-chemokine signature was consistently associated with a T-cell–inflamed phenotype and innate immune-sensing pathways across multiple cohorts, independent of tumor cellularity.

The majority of PDAC patients develop metastatic disease, most commonly in the liver (3, 60). In the final set of analyses, we thus determined whether association of the 4-chemokine signature with an antitumor phenotype was conserved in PDAC liver metastases. To this end, we analyzed 107 tumor-enriched liver biopsies from the prospective COMPASS trial (26). As in primary PDAC (Fig. 2A), the expression of CCL4, CCL5, CXCL9, and CXCL10 were coregulated (Fig. 6A) and 4-chemokine^hi metastases (Fig. 6B) displayed the changes in immune composition observed in primary tumors (Fig. 3A), including increased expression scores for gene sets indicative of T cells, cytotoxic cells, and activated DCs (Fig. 6C), whereas again no association with transcriptional subtypes (12, 47, 48) of PDAC was observed (Supplementary Fig. S2). Furthermore, these metastases had consistently higher CD8⁺ counts (Supplementary Fig. S5A) and displayed increased CYT scores and enrichment of T-cell–inflamed gene-expression profiles predictive of immunotherapy response (Fig. 6D). 4-chemokine^hi metastases had the same upregulation of all tested markers of T-cell exhaustion (Fig. 6E), gene sets indicative of T-cell activation and T-cell suppression (Supplementary Fig. S5B), DC stimulation, and MHC class I presentation (Supplementary Fig. S5C), and upregulation of the Batf3-DC score and increased STING and NLRP3 inflammasome signaling (Fig. 6F). Finally, as in primary tumors, these results were independent of mutational load and neoantigen counts (Supplementary Fig. S5D) and tumor cellularity post-LCM (Supplementary Fig. S6; Supplementary Table S4). Taken together, these data suggest a chemokine-regulated mechanism of T-cell–mediated antitumor immunity, conserved between primary tumors and metastatic sites.

In this study, we show the potential role of chemokines in establishing active T-cell–mediated immunity in PDAC. Our initial observation that, among all known human chemokines, only CCL4, CCL5, CXCL9, and CXCL10 are strongly associated with CD8⁺ TILs led us to discover that coexpression of this 4-chemokine signature broadly identifies both primary tumors and liver metastases with a T-cell–inflamed phenotype. Although our study, by design, can only demonstrate the capacity of chemokines to divide PDAC cases by their T-cell–inflamed phenotype, it is likely no coincidence that all four chemokines have major known functions in mounting effective antitumor immunity. CCL4 is essential for DC migration and subsequent DC-mediated effector T-cell activation and recruitment into the tumor (43) and CCL5 may serve as a central regulator of T-cell accumulation in solid tumors (61). In addition, CXCL9 and CXCL10 are known to almost exclusively recruit antitumor lymphocytes, including Th1 cells, NK cells, and CD8⁺ T cells (62, 63) and consistently associate with T-cell infiltration in numerous solid tumor types, including ovarian, esophageal, prostate, and lung cancers (19–24, 64, 65). Although expression of each of the chemokines individually correlated with CD8⁺ T-cell infiltration in PDAC, it is important to note that both CCL4 and CXCL10 were important in mediating T-cell recruitment in a mouse model of melanoma (43), whereas a recent report showed CCL5 and CXCL9 functionally cooperate in orchestrating CD8⁺ T-cell infiltration in an in vivo model of ovarian cancer (61), establishing a mechanism by which these chemokines functionally determine tumor immunoreactivity. Thus, our observation of strong and exclusive coregulation among these particular four chemokines across multiple data sets denotes that similar mechanisms operate in PDAC.

We assessed the T-cell–inflamed status of PDAC tumors by exploring three essential processes in the cancer-immunity cycle (45): (i) direct cell–cell contact between activated effector CTLs and their target cells, (ii) CTL-mediated killing of the target cell, leading to T-cell exhaustion, and (iii) professional APC-mediated cross-presentation of antigen to naïve CD8⁺ T cells. Strikingly, expression of the 4-chemokine signature marked the activity levels of all three of these processes in human PDAC, in primary tumors and metastases. First, among all currently known human chemokines, only expression of the four chemokines (CCL4, CCL5, CXCL9, and CXCL10) was directly associated with tumor-adjacent CD8⁺ TIL counts in PDAC. Tumor cell–adjacent CD8⁺ TIL counts served as a histologic proxy for active antitumor immunity. The importance of the physical interaction between TILs and tumor cells in 4-chemokine^hi patients stems from the mechanisms by which CD8⁺ effector T cells induce cell death (perforation of the cell membrane, enzymatic digestion through serine proteases, and extrinsic apoptosis through ligand binding), all of which require physical contact (50). In addition, RNA-seq data from tumor-enriched specimens obtained by LCM displayed a strong association between the 4-chemokine signature and presence of cytotoxic T-cell expression profiles in the immediate vicinity of the tumor cells. Second, the 4-chemokine signature associated with transcriptional metrics of effector CTL activation and function. Stratifying patients based on 4-chemokine signature expression segregated the cohort into the immunologically “cold” and “hot” phenotypes characterized by expression levels of genes involved with T-cell function (LCK, ZAP70, and IL2RB), T-cell exhaustion (PD1, PDL1, CTLA4, TIGIT, TIM3, and LAG3), and gene sets involved with T-cell stimulation and activity. Third, molecular metrics of professional APC activation were associated with the 4-chemokine signature, including pathways involved with DC activation, costimulation of APCs, and MHC class I presentation. Indeed, a recent study demonstrated CXCL9 expression colocalized with CD68⁺ and CD11c⁺ in epithelial ovarian cancer, suggesting TAMs and DCs produce this chemokine in a T-cell–inflammation context (61). The presence and activation of DCs represent a step crucial for the cross-presentation of antigen and recruitment of antitumor effector T-cell (8, 66), a process also dependent on chemokines (18, 62). Notably, CCL4 was recently shown to play a key role in recruiting Batf3-DCs, an important cross-presenting APC needed for T-cell–inflammation and response to immunotherapy in vivo for melanoma and PDAC (42, 43, 67). Supporting this, we found that expression of a Batf3-DC gene set associated with 4-chemokine signature expression across all four data sets analyzed, including liver metastases. This suggests DCs are involved in proper mounting of the cancer-immunity cycle, in both primary tumors and metastases in PDAC. Furthermore, activity of innate signaling pathways implicated in immune-mediated tumor control (52, 53, 68–71), including STING and NLRP3 inflammasome signaling, was increased in 4-chemokine^hi patients. Their potential functional roles in activating and therefore controlling DC/APC-mediated baseline antitumor immunity in PDAC requires further attention.

Our findings support recent evidence that chemokine expression, not only high tumor immunogenicity, drives T-cell infiltration and antitumor response (61, 72, 73). Although TMB works well in other cancers, here we also report that neither SNV counts nor neoantigens are associated with CD8⁺ T-cell infiltration or other metrics of antitumor immunity in the vast majority of PDAC tumors, apart from cases with discreet DNA-repair deficiencies, which comprise about 10% of PDAC (27, 74). We thus propose a model of two complementary mechanisms mediating antitumor immunity in PDAC (Fig. 6G), in which classic predictors such as TMB best applies to PDAC with DNA-repair deficiencies, whereas the 4-chemokine signature dictates a T-cell–inflamed phenotype in “typical” PDAC. Indeed, our data and work from other groups suggest that in PDAC and other cancers, there exists an alternative trigger aside from TMB and predicted neoantigen loads, which activates hallmarks of the cancer-immunity cycle and initiates a T-cell–inflamed phenotype. Although chemokine-mediated recruitment is necessary (18, 62), the mechanisms initiating T-cell–inflamed versus noninflamed tumors are poorly understood (8). Deregulation of oncogenic pathways as a basis for immunosuppression has been suggested (70). KRAS signaling was shown to be predictive of response to immunotherapy in colon adenocarcinoma and is known to drive an immunosuppressive environment in the development of PDAC (75–77). In an orthotopic model of PDAC, oncogenic KRAS^G12D signaling was shown to induce an immunosuppressive environment through GM-CSF–mediated recruitment of suppressive myeloid populations, impacting the ability of effector CD8⁺ T cells to function (76). Indeed, KRAS signaling was enriched in our 4-chemokine^hi patients, and it may be possible that aberrant signaling drives the immunosuppressive environment observed in this cohort. Other tumor-intrinsic mechanisms may affect chemokine expression. Wnt/β-catenin is associated with reduced CCL4 expression and subsequent Batf3-driven DC and T-cell recruitment in melanoma (42, 43, 70, 78). Further, recent data suggest that epigenetic silencing may lead to decreased CCL5 expression, which was associated with decreased CXCL9 expression, fewer TILs, and worse survival in an in vivo model of ovarian cancer (61). Indeed, several tumor-intrinsic pathways have been proposed to prevent successful antitumor immunity (70), and this warrants further investigation in PDAC. Thus, if tumor-intrinsic pathways drive antitumor immunity, it remains to be determined whether an antitumor phenotype is also conserved across sites. This chemokine-associated immunity was indeed observed in both primary and metastatic pancreas tumors, and metastatic pancreatic cancer cells have been shown to display mutational signatures akin to their parent tumors (79). It will be interesting to assess paired primary and metastatic tumors and whether daughter cells of primary tumors seed into distant metastatic sites and set up a similar niche.

Taken together, expression of a 4-chemokine signature consisting of CCL4, CCL5, CXCL9, and CXCL10 in PDAC was associated with a variety of histologic and molecular hallmarks of a T-cell–inflamed phenotype and the APC-mediated processes required for mounting functional effector CD8⁺ T-cell antitumor immunity. Indeed, a recent report by Dangaj and colleagues demonstrated CCL5 and CXCL9 expression strongly associated with CD8A expression across several cancer types, including kidney, breast, lung, uterus, colon, and ovarian (61). Although CCL4 and CXCL10 did not associate across all cancer types investigated, CCL4 was associated with CD8A expression in both lung and kidney cancer, and CXCL10 in kidney cancer (61), suggesting that, together, the four chemokines in our panel could predict T-cell inflammation in PDAC tumors as well.

Together, this study shows that specific chemokines are directly associated with transcriptional and cellular metrics of a T-cell–inflamed phenotype and APC-mediated processes in PDAC, both of which are required for antitumor immunity and could potentially serve as predictors of response to immunotherapy, especially in cases with low mutational burden. We hope this correlation in patients will instigate mechanistic studies to shed light on the potential functional role of these chemokines in establishing antitumor immunity in PDAC. Indeed, all four chemokines in our panel have been associated with T-cell inflammation and response to immune-checkpoint blockade, including anti–PD-1 and anti–PD-L1 monoclonal antibodies, in other cancers (40–42). Importantly, this association was seen in both primary PDAC and liver metastases. This has therapeutic implications as the majority of PDAC are diagnosed at advanced stages (3), most with metastatic disease in the liver. Understanding the underlying immunobiology driving effector T-cell infiltration in PDAC may uncover new methods to increase CD8⁺ T-cell infiltration into these tumors and help identify patients for immunotherapeutic strategies.

No potential conflicts of interest were disclosed.

Conception and design: J.M. Romero, B. Grünwald, S. Gallinger

Development of methodology: J.M. Romero, B. Grünwald, P.P. Bavi, S.E. Fischer, S. Gallinger

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): J.M. Romero, B. Grünwald, G.H. Jang, P.P. Bavi, A. Jhaveri, M. Masoomian, S.E. Fisher, A. Zhang, R.E. Denroche, I.M. Lungu, A. De Luca, J.M.S. Bartlett, J. Xu, N. Li, S. Dhaliwal, S.B. Liang, D. Chadwick, F. Vyas, P. Bronsert, R. Khokha, F. Notta, G.M. O'Kane, J.M. Wilson, J.J. Knox, A. Connor, Y. Wang, G. Zogopoulos, S. Gallinger

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): J.M. Romero, B. Grünwald, S. Gallinger

Writing, review, and/or revision of the manuscript: J.M. Romero, B. Grünwald, T.L. McGaha, F. Notta, P.S. Ohashi, S.J. Done, G.M. O'Kane, J.M. Wilson, A. Connor, S. Gallinger

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): J.M. Wilson

Study supervision: S. Gallinger

This study was conducted with the support of the Ontario Institute for Cancer Research (PanCuRx Translational Research Initiative) through funding provided by the Government of Ontario, the Wallace McCain Centre for Pancreatic Cancer supported by the Princess Margaret Cancer Foundation, the Terry Fox Research Institute, and the Canadian Cancer Society Research Institute. The study was also supported by charitable donations from the Canadian Friends of the Hebrew University (Alex U. Soyka) and the Pancreatic Cancer Canada Foundation. S. Gallinge is the recipient of an Investigator Award from OICR and the Lebovic Chair in Hepatobiliary/pancreatic Surgical Oncology. J.M. Romero was supported by a Canadian Institutes of Health Research CGSM studentship. B. Grünwald was supported by the Princess Margaret Cancer Foundation, the European Molecular Biology Organization (ALTF 116-2018), and the Alexander von Humboldt Foundation (DEU 1199182 FLF-P). G. Zogopoulos is a clinical research scholar of the Fonds de recherche du Québec, Santé. We acknowledge the contributions of team members from the OICR Genomics and Bioinformatics platform (genomics.oicr.on.ca). We also thank the following individuals who provided patient samples: Jim Biagi and Sulaiman Nanji, Kingston General Hospital, Kingston, ON, Canada; Gloria M. Petersen, Mayo Clinic College of Medicine, Rochester, MN, USA.

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.