Establishing models of hepatic and pulmonary metastatic pancreatic ductal adenocarcinoma
To establish directly comparable lung and liver PDAC metastasis, we utilized a murine pancreatic adenocarcinoma cell line driven by Kras and Trp53 mutations (KPC) [12] injected into the two most common visceral sites of metastatic disease, liver, and lung. The use of this cell line was particularly important given its genetic similarity to human disease [13] and reflects the critical impact that mutations have on the immune response [14]. Similar to human disease, KPC cells are thought to be intrinsically limited in immunogenicity. A relatively low mutational burden in our KPC cell line was verified by whole exome sequencing and mutational analysis (Additional file 1: Figure S1A). The carcinogen-induced mouse pancreatic cancer cell line Panc02, which is known to harbor a high number of mutations including those that yield immunologically recognizable neoantigens [15], was used as a comparator. As a gross assessment of neoantigens related to the driver mutations in the KPC model, predicted MHC binding affinities for Kras(G12D) and Trp53(R172H) were also analyzed; no mutations in either of these genes were identified in the Panc02 cell line. As expected, neither driver mutation produced a putatively strong neoantigen as defined by a strong binding affinity of the mutant type peptide or by comparison of the mutant type to wild type binding affinity (Additional file 1: Figure S1B, C). To establish metastasis, KPC cells were injected intraportally by the hemispleen method [16] and intravenously via tail vein, leading to robust tumor burdens in the liver and lung by day 21 from the time of injection that could be validated by both gross and histological examination (Additional file 1: Figure S2). Using these models, both the immune and non-immune compartments of lung and liver microenvironments were analyzed with high-throughput molecular profiling to fully interrogate the differences between the two metastatic tissue sites.
Immune profiling with mass cytometry reveals enhanced immune activation in lung metastases
To characterize the microenvironment of the metastases, we profiled the immune compartment of normal liver, normal lung, KPC-bearing liver, and KPC-bearing lung with mass cytometry (Cytometry by Time-of-Flight; CyTOF). The CyTOF panel consisted of a total of 33 mass channels, including 16 for subtyping, 11 for functional analysis, 4 for barcoding, 1 for cell identification, and 1 for viability (Additional file 2: Table S1). By using four unique CD45 barcodes, we employed a batching strategy to enable multiplexed staining and data acquisition for robust analysis (Fig. 1a). The high-dimensional CyTOF data was then clustered using FlowSOM algorithm into a total of 20 metaclusters, which were then annotated into 13 final immune cell types (Fig. 1b and Additional file 3: Table S2). Annotations were based on known canonical expression profile of the different immune cell types, e.g. CD3+CD4+ clusters are annotated as helper T cells. When comparing between the two metastatic sites, a higher proportion of B cells were found in liver while a higher proportion of T cells, NK cells, monocytes, and CD11c+ dendritic cells were found in lung. Most of these organ-specific differences were also present in the absence of KPC metastasis, reflecting differences in the baseline immune infiltrate of these tissues. As expected, higher proportions of myeloid derived suppressor cells (MDSCs) and Tregs were noted in the KPC-bearing states for both liver and lung (Fig. 1c–e). When we assessed the functional profiles within each of the cell type clusters, lung generally exhibited higher mean metal intensities (MMI) of activation or co-stimulatory markers, CD69, ICOS, and CD27, and co-inhibitory markers, PD-L1, PD1, KLRG1, and BTLA (Additional file 1: Figure S3). A notable exception to this trend was higher MMI of LAG3 in T cell and NK cell clusters within the liver.
Given the increased numbers of T cells in the lung metastatic microenvironment, we repeated the clustering analysis utilizing samples gated for CD3+ events for deeper profiling of T cells. All functional markers were incorporated into the clustering step, and a total of 20 resulting metaclusters were annotated into 18 T cell subtypes based on their expression profiles (Additional file 1: Figure S4 and Additional file 4: Table S3). As a percent of CD3+ cells, a much higher proportion of naïve helper T cells were seen in the lung compared to the liver (Fig. 2a), driving most of the difference seen in the T cell percentage of all CD45+ immune cells (Additional file 1: Figure S5). Compared to the normal tissue controls, KPC-bearing tissues had greater proportions of regulatory T cell clusters (“Treg_I” and “Treg_II”) and PD-1-expressing effector memory T cell clusters (“TcEM” and “ThEM_II”) along with decreased proportion of naïve cytotoxic T cell clusters (“TcN_I” and “TcN_II”) (Fig. 2a). Another important distinction was the tumor-associated presence (2% of CD3+ T cells) of Lag3+ Granzyme B+ cytotoxic T cells (“TcEFF_II”) within the liver but not in the lung microenvironment (Fig. 2a). Since metal intensity of Lag3 was relatively weak and since the Lag3-positive population was rare, we then performed fluorescent flow cytometry to validate these findings. Again, we confirmed the higher percentage of Lag3+ CD8 T cells and NK cells in the KPC-bearing liver compared to the KPC-bearing lung (Fig. 2b). In addition, to assess whether these changes in the proportions of T cell subtypes are truly tumor-responsive, we performed TCR repertoire analysis of the tissue-infiltrating T cells in both normal and KPC-bearing livers and lungs. Compared to normal tissue controls, KPC-bearing livers and lungs demonstrated a significantly higher sample clonality of T cells (Fig. 2c), suggesting that the differences in the T cell subpopulations are actually driven by the presence of the tumors within these metastatic sites. Together, these findings implied that compared with the liver, the lung metastatic microenvironment is generally composed of a higher level of tumor-associated immune cell infiltration and activation, especially T cells.
To further validate that the differences in the immune cell profiles are not just tissue-intrinsic differences but also altered in ways that are specific to KPC tumors, we performed immunohistochemistry (IHC) of four immune cell markers, CD8 (cytotoxic T cells), CD4 (helper T cells), CD68 (macrophage/myeloid cells), and B220 (B cells), to visually assess their infiltration within the two tissue sites (Additional file 1: Figure S6A-F). We specified our analysis to tumoral and adjacent normal regions separately and first quantified the density of cells. When comparing the number of cells per tissue area — in contrast to % of cells in CyTOF — there were higher density of CD4+ T, CD68+ myeloid, and B220+ B cells within the lung tumors compared to the liver tumors (Additional file 1: Figure S6G). There was a trend toward higher density of CD8+ T cells as well. To further characterize their relationships, e.g. the proximity of an antitumor effector CD8+ T cell to all other immune cells, we also measured the average distances between any given CD8+ T cell and CD4+ T cells, CD68+ myeloid cells, B220+ B cells, or another CD8+ T cell (Additional file 1: Figure S6H). The shortest distances among immune cells were observed within the tumors in the lungs, suggesting that the spatial coordination of immune cells is different between the lung and the liver TMEs. Importantly, the immune cell densities and distances were also higher and shorter, respectively, in the tumoral regions than their normal adjacent counterparts, suggesting that the immunologic profiles are not only different between the two sites but also shaped differentially by the tumor.
Transcriptomic analysis of the non-immune compartment implicate a role for liver parenchyma in establishing an immune-suppressive TME
To compare the features within the non-immune compartment, which includes the KPC cells, tissue-specific parenchymal cells, and stromal cells, KPC-bearing liver and lung samples were enzymatically dissociated into single cells and were subsequently processed by negative selection with a cocktail of magnetic beads targeting CD45 (pan-immune cells), CD31 (endothelial cells), and Ter119 (red blood cells) among others. The resulting samples were then analyzed by RNAseq along with KPC cells from 2D in vitro culture, normal lung samples, and normal liver samples as controls (Fig. 3a). PC analysis showed that samples from each of the groups clustered together (Fig. 3b). PC1 represented 71% of the variance and associated mostly with the contrast between KPC-bearing and normal samples, whereas PC2 represented 19% of the variance and associated with tissue type. Comparing the expression of KPC-associated genes, Pdx1, Muc1, Muc5ac, Sox9, Krt18, Krt19, and Cdh1 confirmed that the method successfully selected for KPC cells (Fig. 3c). Furthermore, the presence of parenchymal cell markers for liver (prothrombin and fibrinogen genes; F2, Fgb) and lung (surfactant protein genes; Sftpa1, Sftpb) was represented by the corresponding liver and lung samples, respectively, and not by the in vitro KPC cells. Also, substantial reduction of immune cells (Ptprc, CD3e, CD19 expression) in the KPC-bearing livers and lungs by the negative selection process was confirmed (Fig. 3c).
Next, to assess the signaling pathways within the non-immune compartment that may be interacting with the immune compartment, genes significantly different between the liver and lung non-immune microenvironments (FDR-adjusted p < 0.05) were tested against the set of genes defining the PDL1 pathway from the Kyoto Encyclopedia of Genes and Genomes (KEGG) database. Since LAG3 was a notable feature in the liver immune TME, we also used GeneMania [17] to define a network of genes associated with Lag3 and Fgl1, a ligand of LAG3 predominantly found in the liver [18]. Our analysis revealed that both the PDL1 signaling pathway (Additional file 1: Figure S7) and FGL1-LAG3 network (Additional file 1: Figure S8) are significantly enriched in the liver. Significant enrichment of the FGL1-LAG3 network in the liver non-immune microenvironment was consistent with our cytometric analysis of the immune cells showing distinctive presence of LAG3+ T and NK cells in the liver. For an unbiased exploration, we also tested the differentially expressed genes against all signaling pathways in the KEGG database, revealing five other enriched pathways (PPAR, Wnt, p53, ErbB, Neurotrophin), three of which (PPAR, Wnt, Neurotrophin) could be attributed to the intrinsic parenchymal differences (Additional file 1: Figure S9).
In addition, to further interrogate the possible points of interplay between the immune and non-immune compartments of the metastatic microenvironment, we also focused our analysis on a select set of chemokines [19] and a set of immune markers [20] known to regulate immune function. Chemokines associated with pro-immune, immune-recruiting functions (Cxcl9, Cxcl10, Cxcl11, Cxcl14) were expressed at higher levels in the lung than in the liver (Fig. 4 and Additional file 1: Figure S10A). For Ccl5, Ccl22, Ccl27a, Ccl28, and Cxcl12, chemokines known to exert pro-tumor effects, substantially higher levels were noted in the liver. Among the immune regulatory molecules compared, Pdcd1lg2, Tgfb2, Tgfb3, Fasl, IL10, and Fgl1 were all expressed at higher levels in the liver than in the lung (Fig. 4 and Additional file 1: Figure S10B). These findings suggest that the non-immune compartment of the liver metastatic microenvironment creates a relatively immune-suppressed environment compared to that of the lung.
To characterize few key immune regulatory chemokines in more depth, namely CCL27/CCL28 (which attract Tregs via CCR10 [21, 22]) and CXCL12 (which attract immunosuppressive myeloid cells into the microenvironment via CXCR4 [23, 24]), and to validate the immune profiles from our KPC model in additional KPC models, we performed CyTOF again with a revised panel of antibodies spanning 40+ channels (Additional file 5: Table S4) that includes CCL27 and CXCL12 along with PDX1, pan-keratin, and podoplanin (for non-immune cells). In this run, sample barcodes were based on a combination of CD29 (integrin beta-1), CD98 (large neutral amino acid transporter), and CD45 to broaden the types of cells being examined. With a total of 33 annotated cell type clusters (Additional file 6: Table S5), we performed a side-by-side comparison of the metastatic tissue profiles from mice bearing our KPC, the 2838c3 KPC (“KPCc3”), or the 6419c5 KPC (“KPCc5”) tumors [25] along with normal controls. Again, across all models, we observed that T cells and dendritic cells were overall much more abundant in the lung, whereas B cells were more abundant in the liver (Additional file 1: Figure S11A). When comparing specific immune cell subtypes that are tumor-associated, i.e., significantly higher in proportion than in the normal controls, immunosuppressive myeloid cells and LAG3-high effector T cell populations were consistently more prevalent in the liver metastatic microenvironment (Additional file 1: Figure S11B). Expression of CCL27 and CXCL12 chemokines were highest in subpopulations of KPC cells (PDX1+, pan-keratin+; “KPC_II”, “KPC_III”; Additional file 1: Figure S11C) which were greater in proportion in the liver samples compared to the lung samples, corroborating the RNAseq data.
Human rapid autopsy specimens recapitulate immune signaling differences observed in mouse models
To correlate the findings from the KPC mouse model with human disease, five matched liver and lung metastatic samples from deceased PDAC patients were identified in the Johns Hopkins Rapid Autopsy Biobank. Sample sources and their general treatment histories are tabulated separately (Additional file 5: Table S4). Immunohistochemical staining of canonical immune markers, CD4, CD8, CD20, and CD68, were first performed and compared between the two organ sites. Similar to what was observed in the mouse models, the intratumoral density of CD4+ and CD8+ T cells was significantly greater at the pulmonary metastatic site (Additional file 1: Figure S12A-C). The intratumoral density of CD20+ B cells was also higher in the lung, consistent with mouse IHC analysis. Notably, when comparing the markers on a per-tissue block basis, there were not only inter-patient variability and inter-site variability, but also intra-site variability (Additional file 1: Figure S12D). Regarding checkpoint markers, PD1 and LAG3 expression were assessed by positive expression of each marker per T cell density (Additional file 1: Figure S12A, E). High PD1 expression was detected in both liver and lung samples without a clear pattern. LAG3 expression was also variable, but there was an observable trend, in which the highest expressing samples were from the liver, which was also observed in the mouse CyTOF data. Overall, the findings in the matched human samples were consistent with what were observed in the mouse model.
Site-specific microenvironment features are detectable in primary tumors of lung and liver
Finally, to test the hypothesis that the site-specific features associated with metastatic pancreatic lesions stem from the intrinsic tissue microenvironment, tumor-associated and TME-associated characteristics were compared between lung and liver primary cancers using the Cancer Genome Atlas (TCGA) database. To first determine whether the molecular profiles in the lung and liver metastatic sites based on the KPC mouse model were preserved in the lung and liver TCGA datasets, we employed our transfer learning algorithm projectR [26] to project the principal component dimensions derived from the gene expression data for the cancer cells isolated in the KPC mouse metastatic model (Fig. 3b) onto TCGA datasets. Upon projecting onto cancer and adjacent normal samples from primary liver, lung, and pancreas TCGA datasets, we again observed that while PC1 explained the differences between the normal and tumor samples, PC2 captured the tissue sites differentially (Additional file 1: Figure S13A). This indicated that tumor-associated and tissue-associated signals were separately and consistently relatable across different species and primary tumor types.
To delineate whether this site-specific signal was at least in part tumor-driven and not solely parenchymal/stromal, we explored the MetMap dataset [11], a publicly available dataset based on RNAseq profiling of uniquely barcoded breast cancer cell lines injected in vivo for spread onto five different metastatic sites. Because these are barcoded cell lines, the dataset specifically represents cancer cell-intrinsic differences. Based on PC analysis, we found that PC1 is most highly represented by the cancer cells within the liver (Additional file 1: Figure S13B). Then, upon projecting PC1 from MetMap onto our KPC dataset, we found that PC1 was much more strongly recapitulated by the KPC samples from liver (Additional file 1: Figure S13C), suggesting that the lung and liver metastatic sites also differentially shapes tumor-intrinsic features.
Next, to discern what tissue-specific features were being preserved, we also compared the TME-related features in the lung and liver TCGA datasets. Based on MIXTURE deconvolution [27] of the RNAseq datasets, lung cancers demonstrated significantly higher levels of immune cells, including CD8+ T cells, CD4+ T cells, Tregs, dendritic cells, and macrophages (Fig. 5). The differences were also present in the adjacent normal tissues, suggesting that the features are driven by the tissue microenvironment. Importantly, these findings were consistent with the pancreatic cancer mouse model and rapid autopsy analyses. Similarly, differential expression of select genes involved in immune regulation of cancers as analyzed in the mouse models and the human datasets for pancreatic metastatic disease were compared between the lung and liver primary cancers. Again, many of the immune-modulatory genes differentially expressed in the mouse model were site-specifically recapitulated, e.g., higher CCL2, CXCL5, CXCL9, CXCL11, CXCL14, CXCL17, and IDO1 in lung cancers and higher CXCL12, FGL1, and LAG3 in liver cancers (Fig. 6a). Gene set analysis showed that the chemokine, PDL1 pathway, and immune regulatory gene lists were highly enriched (FDR p values < 0.1, Fig. 6b). Several genes were also, however, inconsistent with the mouse model data including CCL5, CCL20, CCL22, CCL28, CXCL10, and PDL2, suggesting that these features may not be as robustly tissue-specific across different primary cancers. Taken together, these results strongly support the concept that many of the TME signatures in metastatic sites can be explained by the features intrinsic to the tissue. Furthermore, consistent with our hypothesis, the presence of baseline differences in the non-tumor-bearing tissues is particularly apparent. Our analysis suggests that for a given tissue site, particular pathways are involved similarly in modulating — suppressing or stimulating — the immune response to different tumor types. In all of the datasets analyzed, liver exhibited a relatively greater presence of features that characterize an immune-suppressive microenvironment compared to lung.