Purpose: Most high-grade serous ovarian cancer (HGSOC) patients develop recurrent disease after first-line treatment, frequently with fatal outcome. This work aims at studying the molecular biology of both primary and recurrent HGSOC.
Experimental Design: Gene expression profiles of matched primary and recurrent fresh-frozen tumor tissues from 66 HGSOC patients were obtained by RNA sequencing. Clustering analyses and pairwise comparison of the profiles between matched samples and subsequent functional alignment were used for the identification of molecular characteristics of HGSOC.
Results: Both primary and recurrent HGSOC samples presented predominant gene expression differences in their microenvironment, determined by a panel of genes covering all major pathways of immune activation together with a number of genes involved in the remodeling of extracellular matrix and adipose tissues. Stratifying tumor tissues into immune active and silent groups, we further discovered that although some recurrent tumors shared the same immune status as their primary counterparts, others switched the immune status, either from silent to active or active to silent. Interestingly, genes belonging to the B7-CD28 immune checkpoint family, known for their major role as negative regulators of the immune response, were overexpressed in the immune active tumors. Searching for potential tumor antigens, CEACAM21, a member of the carcinoembryonic antigen family, was found to be significantly overexpressed in immune active tissues in comparison with the silent ones.
Conclusions: The results illustrate the complexity of the tumor microenvironment in HGSOC and reveal the molecular relationship between primary and recurrent tumors, which have multiple therapeutic implications. Clin Cancer Res; 23(24); 7621–32. ©2017 AACR.
By comparing gene expression profiles of matched primary and recurrent fresh-frozen tissue samples from the same patients with HGSOC, we discovered that the predominant gene expression difference is their tumor microenvironment, presented by a panel of genes covering all major pathways of immune activation together with a number of genes involved in the remodeling of extracellular matrix and adipose tissues. The findings have important translational relevance, especially on the treatment of recurrent ovarian cancer. Not only the molecular nearness of recurrent HGSOC to their primary counterparts can guide the treatment by recurrence, but also the whole tumor microenvironment network can be taken into consideration by developing immune therapies or therapies targeting the tumor microenvironment. Finally, the immunosuppression mechanisms in HGSOC could be different as in other types of tumors. The functions of the B7-CD28 molecules should be carefully investigated before therapies can be developed.
High-grade serous ovarian cancers (HGSOC) account for about 75% of all ovarian carcinomas (1). Standard treatment includes debulking surgery and platinum-based chemotherapy. Although most patients show primary response, they frequently relapse and eventually develop fatal resistant disease. The 5-year survival rate of HGSOC is between 35% and 40% (2). Cancer recurrence and therapy resistance are considered the major causes of death.
Recent next-generation sequencing (NGS) data on primary ovarian cancer demonstrate that nearly all HGSOCs harbor a mutation in the TP53 gene (2). Other major molecular alterations are found in DNA damage and repair pathways, in which BRCA1 and BRCA2 genes are frequently mutated (3), and in the RB1/CCNE1 pathway (4). In addition, HGSOC is characterized with a high genetic instability and frequent promoter hypermethylation. Primary HGSOC is considered to be heterogeneous, comprising different tumor clones with distinct genetic characteristics (4–8).
Besides specific molecular characteristics of tumor cells, tumor microenvironment also plays an important role in the progression of HGSOC. Various molecules, pathways, or specific cell types involved in the immune activation, extracellular matrix (ECM) remodeling, and other stromal alterations were reported to affect the clinical outcome of the patients (9–15). Molecular subclassification of the primary HGSOC also suggested different microenvironments around the tumors (16).
We hypothesize that gene expression changes of both tumors and their microenvironment by recurrence might be determinants for uncontrolled proliferation and resistance of tumors to therapies. By defining the molecular characteristics of recurrent tumors, new therapeutic strategies might be developed against cancer progression. So far, recurrent HGSOC tumor samples have not been systematically collected and analyzed in depth (1) and hence remain poorly understood. To investigate the molecular characteristics of recurrent HGSOC and their microenvironment in comparison with their primary counterparts, we have collected matched primary and recurrent fresh-frozen tumor tissues from the same patients and analyzed expression profiles by RNA sequencing (RNA-seq) within the frame of an international project OCTIPS (Ovarian Cancer Therapy – Innovative Models Prolong Survival), funded by the European Commission's 7th Framework Programme.
Materials and Methods
Patients and clinical materials
Sixty-six patients with high-grade serous epithelial ovarian cancer were included at the Department of Gynecology, European Competence Center for Ovarian Cancer; Campus Virchow Klinikum, Charité - Universitätsmedizin Berlin, Berlin, Germany; Division of Gynecological Oncology, Department of Obstetrics and Gynecology, Universitaire Ziekenhuizen Leuven, Katholieke Universiteit Leuven, Leuven, Belgium; Ovarian Cancer Action Research Centre and Imperial College, London, United Kingdom; Department of Gynecology and Gynecologic Oncology, University Medical Center Hamburg-Eppendorf, Hamburg, Germany; Department of Obstetrics and Gynecology, Comprehensive Cancer Center, Medical University of Vienna, Vienna, Austria; and Department of Gynecologic Oncology, Cancer Research Center Groningen, University of Groningen, University Medical Center Groningen, Groningen, the Netherlands. The studies were conducted in accordance with Declaration of Helsinki ethical guidelines. Informed consents were obtained from all patients prior to sample collection. All processes were approved by the local ethical committee (EK207/2003, ML2524, 05/Q0406/178, EK130113, EK366/2003, EK260/2003).
This retrospective study population consisted of matched primary and recurrent fresh-frozen tumors from HGSOC patients diagnosed between 1993 and 2013. All patients were operable and most of them were platinum sensitive.
RNA isolation and quality control
RNA extraction was performed using the “OCTIPS (#279113) Standard Sample Processing Protocol V2”, section “RNA extraction from tumor tissue” (17).
Quality control of RNA was performed using Agilent RNA 6000 Nano Kit (Agilent Technologies) and Agilent 2100 Bioanalyzer (Agilent Technologies) according to the manufacturer's instructions. RNAs with an RNA integrity number >4 were further processed.
RNA-seq and annotation
RNA libraries were created using the Illumina TruSeq RNA sample preparation kit V2 according to the manufacturer's instructions and sequenced on a HiSeq2000 (Illumina) using a V3 flowcell generating 1 × 50 bp reads. Raw sequencing reads were mapped to the transcriptome and the human reference genome (NCBI37/hg19) using TopHat 2.0 (18) and Bowtie 2.0 (19).
Data processing and prefiltering of genes
Genes with a mean read count ≤ 0.5 or with 10 or fewer observations different from zero, across all 132 samples, were excluded from the analysis, resulting in 28,235 genes eligible for analysis. Data was normalized to equal sequencing depth using the sequencing depth estimation method implemented in the “samr” package (20). For numerical reasons, 0.5 was added to all normalized expression read counts.
Unsupervised clustering analyses
The recurrent samples were clustered by Euclidean distance and the complete linkage clustering method. A total of 369 genes with interquartile ratios (ratio of third and first quartiles) greater than or equal to 10 and a third quartile read counts greater than or equal to 50 were considered for clustering. Clustering of primary samples was performed in the same way, taking 338 genes that passed the same criteria into account. To avoid disproportional influence of high gene expression values in the clustering, the normalized data were logarithmized. For graphical displays in heatmaps, values were truncated at the 99.9th percentile to increase the color intensity.
Fold change and the “area under the receiver operating characteristic curve” for the comparison of gene expressions between two clusters
The fold change (FC) with respect to two clusters of samples was calculated as the ratio of the group-wise medians of normalized read counts. Area under the receiver operating characteristic curves (AUROC) were computed to assess the concordance of gene expression between pairs of recurrent and primary samples, or to assess separation of the gene expression distributions of two groups of samples. For comparability across genes, AUROCs lower than 0.5 were flipped at 0.5 such that 0.5 and 1 were the lowest and highest achievable values, indicating no overlap (no concordance) or complete separation (full concordance) of gene expressions between the groups, respectively.
Functional annotation of differentially expressed genes
Genes, which had the highest expression difference in the two clusters identified by unsupervised clustering (FC > 5 or <0.2, AUROC > 0.75 and q3 > 50 read counts), were first assigned to their specific cell expression according to Angelova and colleagues (21) and examined for their specific functions in databases such as Wikipedia (https://en.wikipedia.org), WikiGenes (https://www.wikigenes.org/app/WikiGene), GeneCards (http://www.genecards.org), Ensembl (www.ensembl.org; ref. 22), NCBI Resources (https://www.ncbi.nlm.nih.gov/), and KEGG PATHWAY Database (www.genome.jp/kegg/pathway.html).
Supervised clustering analyses
We used the 126 immune-related genes found by unsupervised clustering and the subsequent functional analysis to define the immune status of primary and recurrent samples by clustering similarly as described above. This clustering is referred as “supervised” clustering.
Heatmap presentation of expression of selected genes with known functions
To illustrate the within-gene differences in the expression of selected genes, logarithmized normalized read counts were centered and scaled to unit variance, separately performed for the primary and the recurrent samples. The transformed values were truncated at two SDs to increase the color intensity of the heatmaps. Samples were grouped by immune status and were ordered alphabetically within these groups. In particular, no clustering was performed on either samples or genes. Because of its equivalence to the AUROC, the Mann–Whitney U test was used to compare gene expressions between active and silent samples. P values were corrected for multiple testing within the two groups of genes using the Bonferroni–Holm method (23). Adjusted P < 0.05 was considered to indicate statistical significance.
Pairwise FC and selection of differentially expressed genes in paired samples
For each gene and each pair of samples, the pairwise FC was calculated as the ratio of normalized read counts. Genes were considered to be differentially expressed in a pair of samples, if the FC > 5 or <0.2 and at least one of the two samples had read count > 50. In this way, we generated two gene lists for each pair of samples. One included genes overexpressed in primary samples and the other included genes overexpressed in recurrent samples. For each subgroup, overexpressed genes presented in more than 50% of the samples were extracted and presented for primary and recurrent samples.
Comparison of clinical outcome of patients with different immune status
Cumulative probabilities of survival were estimated using the product–limit method, censoring for end of follow-up. Crude and adjusted HRs with 95% confidence intervals (CI) were estimated by Cox regression. For the OCTIPS study cohort, operable recurrent tumor was the inclusion criterion, and therefore, cumulative survival probabilities could not be unbiasedly estimated if the date of primary diagnosis served as baseline (24). Therefore, we compared the mortality hazard after cancer diagnosis between patients with active and silent tumors using a delayed-entry (entry at recurrence) Cox regression model (25). Time from primary diagnosis to recurrence was compared between patients with active and silent primary tumors using the Mann–Whitney U test. Two-sided P values <0.05 or 95% CIs excluding parity were considered as statistically significant.
Analyses of TCGA gene expression data
A total of 488 patients with ovarian cancer were included in The Cancer Genome Atlas (TCGA) data. We included all HGSOC and excluded cases missing information on residual tumor and on survival, obtaining a total of 421 patients. TCGA_489_UE.txt was used to access the gene expression data and TCGA Supplementary Table S1 for the clinical information. We aligned the 126 genes to the TCGA annotation and could find 59 overlapping genes. Supervised clustering was performed using these 59 genes in the same way as described above to define the immune active and silent cluster. Overall and relapse-free survival were compared between the two groups.
Analyses of RNA-seq data of Patch and colleagues (4)
In a recent study (4), RNA-seq data from 80 primary HGSOC using TruSeq RNA Sample Preparation v2 Kit and HiSeq2000 instrument (Illumina) were provided. RNA-seq data were downloaded from the European Genome-phenome Archive (EGA; http://www.ebi.ac.uk/ega/), which is hosted at the EBI, under accession number EGAS00001000397: RNA-seq data (exp_seq.OV-AU.tsv). The corresponding clinical data were filed donor.OV-AU.tsv and specimen.OV-AU.tsv. The access to datasets was approved by the specified Data Access Committee. We defined the immune activation status of these 80 primary tumors by supervised clustering using the 126 immune relevant genes in the same way as for our primary and recurrent samples. Overall and relapse-free survival were compared between the two groups similarly as described above. The information on residual tumor after debulking surgery was not available.
Transcriptome sequencing data have been deposited in the EGA, which is hosted by the EBI and CRG, under accession number EGAS00001002660.
Formalin-fixed paraffin-embedded (FFPE) tissues were sectioned at 3 μm. The IHC staining of EpCAM, CD45, CD8, and CD20 was performed with the Dako LSAB+ System-HRP Kit (Dako) including the second antibodies following the manufacturer's instructions. Primary antibodies were diluted with Dako REAL Antibody diluent (Agilent Technologies) and incubated overnight at 4°C. FLEX Negative Control Mouse Cocktail (Agilent Technologies) and Negative Control Rabbit IgG (Biocare Medical) were used as isotype controls. For the staining of the immunoglobulins, DAB- substrate was applied directly after primary antibody incubation. Nuclei were stained with hematoxylin solution modified according to Gill III (Merck Millipore) before mounting the slide with Kaisers Glyceringelatine (Merck Millipore).
PD-L1 was stained with Leica BOND staining device (Leica Biosystems), PD-1 and CD1a with Ventana Benchmark staining device (Roche Diagnostics), both following the company's instructions.
Antibodies and dilutions
The following antibodies were used: anti-EpCAM (IgG, rabbit, clone E144; Abcam), 1:300; anti-CD45 (IgG, rabbit, clone E19-G; DB Biotech), 1:800; anti-CD8 (IgG1, mouse, clone 144B; Abcam), 1:25; anti-CD20 (IgG2a, mouse, clone L26; Thermo Fisher Scientific), 1:200; anti-CD1a (IgG1, mouse, clone 010; Dako), 1:50; anti–PD-1 ready to use (IgG1, mouse, clone NAT105; Ventana Medical Systems, Inc.); anti-PD-L1 (rabbit, clone E1L3N; Cell Signaling Technology), 1:200; anti-Ig [goat anti-human IgG, IgM, IgA (H + L), HRP-conjugated, Thermo Fisher Scientific], 1:100.
The study cohort consisted of 66 patients (aged from 21 to 74 years, median 56) with recurrent HGSOC. Matched fresh-frozen tumor samples from both primary and recurrent tumors were collected. All patients underwent cytoreductive surgery at first diagnosis as well as at recurrence and received platinum-based first-line chemotherapy. Median interval between primary operation and diagnosis of recurrence was 20 (interquartile range, 14–28) months, which is representative of a normal HGSOC population (26). Patients with previous malignant disease were only included if they had at least 5-year disease-free interval at the first diagnosis of HGSOC (Supplementary Table S1).
Identification of immune active and immune silent HGSOC samples
As recurrent HGSOCs present very different clinical manifestations that most possibly reflect different molecular subtypes, we first examined whether there would be strong expression differences within the 66 recurrent tumors. Unsupervised clustering analysis of gene expression profiles revealed the existence of two distinct clusters of tumor tissues C1 and C2, containing 36 and 30 samples, respectively (Fig. 1A). These two clusters were virtually unchanged, when we performed the analyses with varied numbers of genes that were subjected to clustering, indicating the stability of the clusters.
To define the major biological differences between these two clusters, we assigned the 142 most differentially expressed genes to their biological functions. Although only 2 genes (PCP4 and PHOX2A) were overexpressed in cluster C2, the remaining 140 had overexpression in C1 (Supplementary Table S2). Of these 140 genes (Fig. 1B), 126 (89%) were directly related to immune response, corresponding to genes specifically expressed in B (8%) or T cells (23%), monocytes (1%), and genes coding for proteins involved in immunoglobulin construction (42%) or for proteins with other direct functions in the immune activation (15%). These results clearly indicated that recurrent HGSOC samples presented significant gene expression differences of tumor microenvironment.
In the same way, unsupervised clustering of the 66 primary tumors generated two clusters, which also presented outweighing expression differences in immune-related genes (Supplementary Table S3). Of the 113 most differentially expressed genes, 90 were specifically expressed in various types of B and T cells, or encoded immunoglobulins and proteins directly related to immune activation.
Our data thus showed that HGSOC could be stratified into two subgroups according to the expression of immune-specific genes. Hence, we reclustered the recurrent and primary tumors separately using the 126 immune related genes identified in the recurrent samples. By doing so, primary and recurrent tumors were classified as immune active (higher expression of the immune genes) and immune silent (lower expression) (Supplementary Fig. S1A and S1B).
We therefore defined 4 subgroups of matched samples according to the immune status (Supplementary Table S4) and named them active–active (16 pairs), silent–active (21 pairs), active–silent (9 pairs), and silent–silent (20 pairs). Supplementary Figure S1C shows examples of expression difference of two immune-related genes (IGKV4-1 and CD79A) in these subgroups.
Matched primary and recurrent samples presented various immune activation status
To obtain an overview of important immune-related molecular processes in paired samples, we compared the expression of selected genes covering T- and B-cell response, natural immunity, as well as immune checkpoint modulation, irrespectively if they were included in the 126 genes. As shown in Fig. 2, not only the genes specific for the immune cells (CD8, CD4, FOXP3, CD79A, CD20, all P < 0.001; and CD1a, P = 0.08 and P = 0.06, in primary and recurrent samples, respectively), but also some genes encoding their effectors such as IFNγ (P < 0.01), granzyme B (GZMB, P < 0.01), and perforin (PRF1, P < 0.01) were significantly overexpressed in the immune active tissues, thus strengthening the notion of an overall activation of the immune response. Interestingly, lymphotoxin alpha (LTA) and lymphotoxin beta (LTB) presented significant overexpression in immune active recurrent tumor samples (P < 0.01) but did not differ in their expression in primary tumor samples significantly. Of note, macrophage migration inhibitory factor (MIF), which is considered to regulate innate immunity, did not show significant difference in expression between immune active and silent samples.
Remarkably, the major ligands and receptors (PD-L1, PD-L2 with their receptor PD-1; CD80, CD86 with their receptors CTLA4 and CD28) of the CD28/B7 costimulatory pathway, known to negatively regulate the immune checkpoint, were all significantly overexpressed in the immune active samples (all P < 0.01). Interestingly, although the ligand ICOSLG did not show any difference in active versus silent, its receptor ICOS showed significantly overexpression in active samples (P < 0.001). CD276 (B7-H3) and VTCN1 (B7-H4) were not differentially expressed in immune active versus silent samples (Fig. 2).
Further, we performed IHC staining of CD45, CD8, CD20, CD1A, PD-L1, PD-1, and immunoglobulins together with EpCAM in 6 active and 6 silent tissue samples. Figure 3 showed examples of the IHC staining results in immune active tissues. Abundant CD8+ T cells as well as CD20+ B cells were found in immune active tumor tissues (Fig. 3B, C, F, and G). This was accompanied by a strong immunoglobulin staining (Fig. 3D and H), which was colocalized with the CD20+ cells. The antigen-presenting CD1a protein was expressed in dendritic cells, most of which were found interspersed within tumor cells (Fig. 3I). PD-L1 staining was mainly found in tumor cells (Fig. 3J), whereas the receptor PD-1 was clearly expressed in lymphocytes (Fig. 3K). This staining pattern was found in other 4 active tissues but in none of the 6 silent tissues (Supplementary Fig. S2). The results were in concordance with those generated from RNA-seq, confirming that the strong expression of the immune genes in HGSOC tumor tissues was accompanied by strong lymphocytic infiltrations in tumors. To examine whether the difference could be caused by sampling, we evaluated CD8, CD20, and immunoglobulin protein expression in multiple FFPE blocks from 5 patients. As a result, all sections from the same patient showed similar staining pattern.
Matched primary and recurrent HGSOCs presented varied patterns of microenvironment changes
The pairwise comparison of gene expression profiles of tumor pairs in each of the four subgroups presented diverse pattern of gene expression differences (Fig. 4). In the active–active and the silent–silent subgroups, no differentially expressed genes could be identified, with the sole exception of CCL19, which was overexpressed in 9 of 16 recurrent samples of the active–active subgroup (Fig. 4A and D). This indicated that primary and recurrent tumors showed little if any expression difference in these two subgroups. In the active–silent as well as in the silent–active subgroups, the active samples showed overexpression of immune-related genes as expected, and there were no notable overexpressed genes in the silent samples (Fig. 4B and C). Furthermore, in the divergent subgroups, immune active samples presented elevated expression of a number of genes involved in remodeling of ECM. These did not only include genes responsible for the degradation of ECM, such as ADAMTS2 and FAP, but also genes such as COL5A2, COL3A1, or HAS1 (hyaluronan synthase), which code for the components of ECM. In addition, numerous genes involved in adipose tissue remodeling were overexpressed in most of the immune active primary tumor samples (FABP4, GPD1, PLIN, ADIPOQ, and TUSC5) in the active–silent subgroup (Fig. 4C). We examined these adipose-related genes in the silent–active sample pairs to verify whether we may have missed them by using stringent selection criteria. Indeed, a number of active recurrent samples showed overexpression of the genes too (e.g., 10/21 had FC >5 higher expression of FABP4).
We also compared expression of genes known to play important roles in epithelial ovarian cancer formation and progression, such as TP53, CDKN1A (p21), BRCA1, BRCA2, PARP1, CCNE1, ERBB2 (HER-2), PAX8, VEGFA, and MUC1 (1, 3), but did not observe any notable expression difference between active and silent tumor tissues in both primary and recurrent samples (Fig. 5A). In addition, we examined the expression of ABCB1 (MDR1), which has been associated with acquired resistance of HGSOC (4). Interestingly, we observed a significant increase of MDR1 expression in immune active tumors (Fig. 5C; P < 0.001 in primary tissues; P = 0.04 in recurrent tissues).
To investigate what could be the trigger of the strong immune activation in HGSOC, we searched for potential tumor antigens that could be overexpressed in immune active samples. Alpha-fetoprotein (AFP), carcinoembryonic antigens CEACAM1, CEACAM3 (ENSG00000170956), CEACAM4, and CEACAM19, CA-125, trophoblast glycoprotein (TPBG), tyrosinase (TYR), and melanoma-associated antigens MAGEA1, MAGEB1, and MAGEB2 showed either no or very low expression levels (median ≤ 3, the third quartile ≤ 13 read counts) or did not present any expression difference in immune active versus silent samples (Fig. 5B). Notably, CEACAM21 (ENSG00000007129) showed an overexpression in the immune active samples (Fig. 5D; primary samples: P = 0.056; recurrent samples: P < 0.001).
Immune activation status and patient outcome
No difference could be found for the interval from primary diagnosis to recurrence between patients with active versus silent primary tumors (P = 0.517), for mortality after recurrence between patients with active and silent recurrent tumors (P = 0.454), and for mortality after primary diagnosis between patients with active and silent primary samples (P = 0.089). Adjusting the presence of residual tumors in the analyses of mortality gave similar, nonsignificant results.
When performing supervised clustering of the TCGA samples, we could align 59 of 126 genes and define a cluster of 45 samples as immune active and 376 samples as immune silent (Supplementary Fig. S3). Patients with immune active status showed a statistically significant better disease-free survival as the silent ones, whereas no difference in overall survival could be found in the two groups (Fig. 6A and B).
In addition, we used the 126 immune-related genes to perform the supervised clustering of the 80 primary HGSOCs in the cohort of Patch and colleagues (4). Two clusters were defined consisting of 54 immune active and 26 silent samples (Supplementary Fig. S4). Even though Kaplan–Meier curves showed a slightly better outcome of the active group over the silent one, there was no statistically significant difference in overall survival or relapse-free survival between the two groups of patients (Fig. 6C and D).
This study was designed to characterize recurrent HGSOC and their microenvironment in comparison with their primary counterparts at the transcriptomic level. To reach this aim, matched primary and recurrent HGSOC samples were collected from 66 patients that underwent debulking surgery at both primary diagnosis and upon recurrence. The unique design of the study allowed pairwise analyses based on a real “case–control” comparison, reducing the perturbation of individual gene expression difference to a minimum. The relatively large sample size for this type of collection further enabled the acquisition of new knowledge on HGSOC.
We first discovered that both primary and recurrent tumors presented predominant molecular differences in their microenvironment, determined by a panel of genes covering all major pathways of immune activation accompanied by numerous genes involved in ECM and adipose tissue remodeling. By defining tumors as immune active and silent ones, we further discovered that although some paired primary and recurrent tumors shared the same immune status, being either both active or both silent, other pairs switched the status, either from silent to active or active to silent. Using the comparison analyses, we obtained results, which were based on expression difference of numerous functional related genes involved in a certain biological process and thus represented genuine and reliable molecular difference between individual samples. Our data strongly suggest that the appearance of recurrent tumors was not driven by large-scale changes of gene expression regulation in tumor cells, but could rather be determined by the host immune reactions.
Immune status of the tumor samples was defined by 126 genes, which were identified by unsupervised clustering. The active samples had an overexpression of the major genes responsible for immune activation, some of which were immune cell specific, suggesting a lymphocyte infiltration. This substantiated that the overexpression of the 126 genes signed for a strong immune reaction of the host against the tumor. Along with T- and B cell–specific markers, elevated levels of IFNγ, granzyme B (GZMB), and perforin (PRF1) were also measured in immune active tissues, indicating a bona fide immune response. The presence of infiltrating immune cells in ovarian cancer was first reported nearly 40 years ago (15). Numerous studies elicited the favorable prognostic significance of increased numbers of various types of immune cells or overexpression of individual immune cell markers (9–11, 27–29). Our works revealed that the immune activation in HGSOC was an overall reaction of the host rather than a response of a single cell type or the expression alteration of a single gene.
In addition, our work revealed the complex network of tumor environment in HGSOC, presented by the cooverexpression of genes involved in the remodeling of ECM and adipose tissue along with the immune genes. HGSOC tumor tissues present intricate admixtures of tumors, infiltrating immune cells, and stromal cells, making it difficult to isolate specific cell contingents by means of microdissection (Fig. 3; Supplementary Fig. S2). RNA-seq of the total tumor enabled us to identify differentially expressed genes, most of which are cell specific. Tumor cells are usually embedded within the ECM and separated in space from immune cells. To execute their immune functions, some immune cells like CD8+ T cells must come into contact with epithelial tumor cells. In lung cancer, active T-cell motility was observed in loose fibronectin and collagen regions. Furthermore, reducing ECM constructive components can increase the number of T cells in contact with tumor cells (30). As a consequence, the destruction of the ECM induced by immune cells could trigger a process resembling wound healing (31, 32), whereas ECM components could be produced by other stromal cells or tumor cells. Therefore, we found an overexpression of genes not only responsible for the degradation (32, 33) but also for the production of ECM components in immune active samples, demonstrating that the remodeling of the ECM could be a dynamic process, in which different molecules and cells interact in the microenvironment in response to tumors. Besides, we also found numerous genes responsible for the remodeling of adipose tissues, which were predominantly overexpressed in most of the immune active tumor samples. Advanced ovarian cancer is often embedded in fat tissue (34, 35). In the same sense that active immune cells have to traverse the ECM, they also have to travel across the adipose tissue to come into contact with tumor cells. For the development of immune therapies against HGSOC, it is essential to understand the network of tumor microenvironment. If we could destroy the adipose tissue and the ECM around tumors, the accessibility of immune cells and other therapeutics might be increased and thus the therapy efficacy might be improved, an issue that has already drawn the attention of scientists (36).
In particular, overexpression of the B7-CD28 gene family members (PD-1/PD-L1, CD80/CTLA4 among others), which are known to negatively regulate T-cell activation (37–40), was also found in immune active samples. It is in line with a recent report, showing that PD-1+ T infiltrating lymphocytes and PD-L1+ tumor cells were both favorable prognostic factors in ovarian cancer (41). The inverse function of the immune checkpoint molecules in controlling the T-cell activation in other types of cancers was also suggested by several other independent studies (42–44). As a matter of fact, some of the B7-CD28 members, such as CTLA-4, have been reported to brake T regulatory cell proliferation driven by CD28 (45) or to inhibit CD28 costimulation by depleting the ligands (46). Similar findings suggest that the immune system can also facilitate tumor progression (47). Our data support the complex regulation of the immune system by the B7-CD28 family genes.
The identification of tumor-associated antigens (TAA) is of great importance because they represent potential targets of tumor vaccines (48). We found a significant overexpression of CEACAM21 in immune active samples. CEACAM21 is a member of the carcinoembryonic antigen family, which is usually expressed in gastrointestinal tissues during embryonal development and is silenced before birth. Very little is known concerning this gene, making it a very interesting target for further investigations. However, this observation must be validated on an independent HGSOC cohort and the real function of CEACAM21 needs to be unraveled before being conclusive on its true status as a tumor antigen. We have established a series of HGSOC cell lines (49), which could serve as models to study the impact of CEACAM21 on immunogenicity.
We found overexpression of ABCB1 (MDR1), which was correlated with platinum resistance in HGSOC in a recent study (4). It is of note that MDR1 was also reported to be expressed in CD34+ stem cells, dendritic cells, NK cells, and cytotoxic T lymphocytes, suggesting that it may protect immune cells against stress-induced or bystander lysis (50). Thus, the upregulation of MDR1 in the immune active samples might also be part of the immune response in these tumors.
The findings reported here lead to some considerations for ovarian cancer research and therapy development. First, the B7-CD86 family members showed a concordant overexpression with the host immune activation. This suggests that immune regulation in HGSOC might be different than in melanomas. We could show a disease-free survival difference between immune active and silent tumors in the TCGA data, implying that these genes might have other functions in immune modulation. However, this difference could neither be confirmed by our own data nor by the data from Patch and colleagues, indicating that further studies are needed to verify the results. Studies with a larger cohort of samples and a thorough investigation on the functions of the gene family must be performed to design innovative immune therapy strategies. Second, the coexpression of immune genes with that of ECM and adipose tissue remodeling genes is intriguing, because enzymes for both ECM construction and destruction have been found. It could be of important therapeutic consequences, as loosening the structure of the ECM or adipose tissue would not only facilitate contact of immune cells with tumor cells, but might also increase exposure of cancer cells to drugs. Hence, it will be important to fine tune immune therapy strategies in HGSOC. Finally, some recurrent tumors are quite similar as their primary counterparts while others are not. If recurrent tumors could be analyzed at molecular levels and compared with their primary counterparts, the scheme of treatment could be adjusted according to the primary response of the patients.
It is very important to examine whether the immune activation status would have an effect on the patients' outcome. Our cohort was rather small for survival analyses, when stratifying the patients by residual tumor status, one of the most important factors affecting both disease-free and overall survival. In the cohort from a public dataset with 80 primary tumors, the residual tumor status was missing, and the results of survival advantage of patients with immune activation were not significant. Studies with larger sample cohort with available information on rest tumor will be needed to clarify the impact of immune activation on clinical outcome of the patients.
The study produced a rather large dataset. On the basis of the variance in tumor cell contents of the samples, we decided to use stringent conditions to analyze differentially expressed genes in all analytic procedures to obtain reliable results. This might lead to the nonobservance of some genes, which had less expression changes or/and lower expression. Alternative analyses using lower stringency could be performed to investigate such genes or their related biological processes. However, results from such analyses must be validated using bigger sample cohorts.
In conclusion, we observed a difference of a complex microenvironment network including immune activation and remodeling of tissues around tumors across all tumor samples and defined tumors as immune active and immune silent. By subgrouping paired tumor lesions into active–active, active–silent, silent–active, and silent–silent ones, we found that there was no notable difference between primary and recurrent tumors in the active–active and silent–silent subgroups, suggesting that these recurrent tumors resemble their primary counterparts. Other tumors presented major molecular difference by switching their immune status, either from silent to active or active to silent, accompanied by the corresponding changes in the microenvironment. Interestingly, major genes of the B7-CD28 immune checkpoint family, which are considered to negatively regulate the immune activation, as well as CEACAM21, which is a member of the carcinoembryonic antigen family, and MDR1, were also overexpressed in immune active tissues in comparison with the silent ones.
Disclosure of Potential Conflicts of Interest
H. Gabra has ownership interests (including patents) in Papyrus Therapeutics, is a consultant/advisory board member for Aprea and Pharmacokinesis, and reports receiving commercial research grants from Aprea, Morphotek, and Takeda. No potential conflicts of interest were disclosed by the other authors.
Conception and design: C. Kreuzinger, I. Vergote, H. Gabra, G. Heinze, S. Polterauer, E. Berns, D.C. Castillo-Tong
Development of methodology: C. Kreuzinger, I. Vergote, G. Heinze, E. Berns, D.C. Castillo-Tong
Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): C. Kreuzinger, D. Smeets, E.I. Braicu, A. Wolf, S. Darb-Esfahani, I. Vergote, A. Vanderstichele, B. Boeckx, D. Lambrechts, H. Gabra, G. B.A. Wisman, F. Trillsch, R. Horvat, S. Polterauer, E. Berns, D.C. Castillo-Tong
Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): C. Kreuzinger, A. Geroldinger, J. Sehouli, J. Koller, K. Joehrens, I. Vergote, B. Boeckx, D. Lambrechts, F. Trillsch, G. Heinze, R. Horvat, S. Polterauer, D.C. Castillo-Tong
Writing, review, and/or revision of the manuscript: C. Kreuzinger, A. Geroldinger, D. Smeets, E.I. Braicu, J. Sehouli, S. Darb-Esfahani, I. Vergote, A. Vanderstichele, B. Boeckx, H. Gabra, G. B.A. Wisman, F. Trillsch, G. Heinze, S. Polterauer, E. Berns, C. Theillet, D.C. Castillo-Tong
Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): C. Kreuzinger, D. Smeets, E.I. Braicu, J. Koller, A. Wolf, S. Darb-Esfahani, B. Boeckx, D. Lambrechts, R. Horvat, E. Berns, D.C. Castillo-Tong
Study supervision: E.I. Braicu, G. Heinze, E. Berns
This work was supported by European Commission Seventh Framework Programme under grant agreement No. 279113-2 (to OCTIPS Consortium; www.octips.eu).
The documentation of clinical and patient data was managed with "AlcedisTRIAL the web based documentation system" of Alcedis GmbH, Winchesterstr. 3, 35394 Giessen, Germany. We thank Elisabeth Maritschnegg and Jasmine Karacs for their technical support.
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.