2.1. Patient Cohorts and Sample Collection
We retrospectively collected tumor samples from patients diagnosed at stage I EOC from three independent Italian cohorts, for a total of 208 snap-frozen tumor biopsies. (i) 135 samples were collected at Gynecology Department of San Gerardo Hospital (Monza, Italy) and conserved in the tissue bank of the Department of Oncology of IRCCS Mario Negri Institute (Milano, Italy), (ii) 26 samples at the A. Nocivelli Institute for Molecular Medicine, Division of Obstetrics and Gynecology (ASST Spedali Civili of Brescia, University of Brescia, Brescia, Italy) and (iii) 47 samples at the Department of Gynecology-Oncology (University of Torino, Torino, Italy). Primary tumor tissues were collected at the time of diagnosis and first surgery prior to any chemotherapy treatment. Samples were submitted for a complete staging procedure according to the International Federation of Gynecological and Obstetrics (FIGO) criteria with a diagnosis of stage I EOC. Tumor histological types were determined following World Health Organization (WHO) standards.
Samples were randomly stratified into a training set and validation set to have sizable histotype groups inside each set; 76 samples of the Milano cohort were dedicated to the training set and 59 samples of Milano together with Brescia and Torino cohorts for the validation set. Histopathological and clinical annotations are described in Table 1
. The training set (n
= 76) included 16 clear cell (Cc), 19 endometroid (End), 17 mucinous (Muc) and 16 high grade serous (SerHigh) and eight low grade serous histotypes, while the validation sets (n
= 132) were 22 Cc, 55 End, 21 Muc, 26 SerHigh, and 8 SerLow histotypes. The training set underwent genome-wide characterisation, while the validation set was used for RT-qPCR experiments.
Training and validation sets have an average follow-up of nine and six years respectively, comparable mean age at the time of diagnosis, similar percentage of relapse, and comparable histo-pathological and clinical characteristics, such as grade and FIGO substage.
2.2. Stage I Classification Using Advanced Ovarian Cancer Expression Subtypes
Based on gene expression profiles, Tothill et al. [18
] identified six subtypes of invasive high grade EOC displaying distinct levels and patterns of immune cell infiltration with prognostic implications. The Cancer Genome Atlas (TCGA) subsequently revised these transcriptional subtypes and identified four classes [8
]: differentiated subtype (DIF) characterized by gene expression profile of well differentiated and low malignant potential cancers; mesenchymal subtype (MES) with high stromal content; immunoreactive subtype (IMR) which contains most of the immune transcripts associated with tumor inflammatory response and proliferative subtype (PRO) characterized by the expression of genes associated with cell division. Although molecularly different, these classes of patients were not found to be different in survival.
Here, we wonder whether these signatures can explain, at least in part, the transcriptome variability of stage I EOC samples and if specific histotypes are associated with one or more of these signatures. Figure 1
shows the results of the classification of our samples into the five mentioned subtypes along with their clinical information. All samples can be classified without ambiguity into one of the four classes, obtaining 26 DIF (34%) samples, 19 MES (25%), 21 IMR (28%), and 10 PRO (13%). Although transcriptional subtypes are distributed across all the histotypes, we found a significant association between Tothill’s classification and histotype (p
-value = 0.002—Supplementary File 1
). Specifically, end histotype samples are mainly PRO, MES subtype are mainly Cc and Muc and virtually lacks of Ser samples, which otherwise show a clear difference between high and low grades; we observed that SerHigh are preferentially IMR, and SerLow are almost exclusively DIF.
In agreement with this last result, we observed a significant association with grade (p
-value = 0.02—Supplementary File 1
). As expected, DIF subtypes are less commonly high grade samples compared to low grade as a consequence of the increasing nuclear atypia from low to high grades. We also observed an increasing presence of IMR samples along with the increase of tumor grade, suggesting that a high grade tumor can elicit an immune system response more than lower grade samples.
Taken together, our results suggest a different involvement of the immune system and of the cell signaling pathways (especially those dedicated to cell proliferation, cell dedifferentiation, and growth) in different stage I EOC grades and histotypes. To better explore these hypotheses, we decided to (i) study the immune system involvement of our stage I samples using dedicated computational approaches and (ii) provide a characterization of cell signaling pathways across the different histotypes.
2.3. Immuno-Phenotype of Stage I EOC Patients
The levels of infiltrating immune cells in tumors are often associated with growth, cancer progression, and poor patient outcome [19
]. To understand the immune system contribution to the transcriptome of stage I EOC samples, we used two different approaches: (i) a deconvolution method proposed by Chen et al. [21
], and (ii) a score method proposed by Charoentong et al. [9
The deconvolution model by Chen et al. [21
] uses gene expression data to estimate and quantify the activation of signatures representative of 22 types of different immune cells. Samples have been divided by histotypes and grades and analyzed separately. Figure 2
A shows the estimated proportion of immune cells composition obtained in our samples.
Globally, infiltrated immune cells seem to be a very low fraction of the total amount of profiled cells (an average of 0.5%). Few weak but significant differences in the amount of cell types can be appreciated across histotypes. In particular, NK cells resting are significantly more abundant in Muc (p
= 0.06), macrophages M2 are significantly more abundant in Cc (p
= 0.006), while Treg are highly abundant in SerLow (p
= 0.04) (see Supplementary File 2
). Moreover, a higher amount of T cells gamma delta appears high compared to low grades (p
= 0.01; Supplementary File 2
) and a higher amount of NK cells resting are present in low with respect to high grade samples (p
= 0.02). Finally, abundances of immune cell types do not show significant association with survival (OS or PFS) using multivariate models adjusted for histotypes and grades (Supplementary File 2
The immunoscore proposed by Charoentong et al. [9
] evaluates the expression of immune biomarkers belonging to four categories: (i) antigen processing molecules (MHC), (ii) checkpoints immunomodulators (CP), (iii) immune effector cells (EC) and (iv) suppressor cells (SC). The expression of the biomarkers within these categories are summarized using a z-score called immunophenoscores (IPS) and graphically represented by immunophenogram (IPG). We tested if immunogenicity differences across tumor grades and histotypes can be appreciated using IPG and IPS and if the patient immune system at the diagnosis can have any effect on the disease progression. For detailed description of IPG and IPS, see Methods Section 3.4
In agreement with the the cell deconvolution approach, all samples, independently by histotype and grade, showed low fractions of infiltrated immune cells. Many genes of the CP class often showed expression values under the detection threshold (gray slices) in both activating and suppressing molecules (Figure 2
Considering the global immune-landscape, we found two main clusters of patients characterized by different levels of expression of MHC, EC, SC, and CP expressed biomarkers (Figure 2
C). The cluster on the left (hereafter called cluster 2) is moderately enriched in Ser samples (both high and low grade) (63% vs. 37% p
= 0.07) and interestingly patients in cluster 2 have a significantly longer survival than the other patients when adjusted for Grade (p
= 0.03, HR = 0.28, CI 95% = 0.08–0.92) (Figure 2
E). Going deeper in the comparison across histotypes, we found that the immune checkpoints regulators are highly abundant in Muc and SerLow samples (p
= 0.01), while effector cells are more abundant in SerHigh histotype and low abundant in Muc (p
= 0.03) (Supplementary File 3
). Moreover, although no significantly different MHC levels can be appreciated across subtypes (Supplementary File 3
), we found that cluster 2 has a significantly higher expression of MHC genes with respect to cluster 1 (p
MHC genes are highly upregulated compared with the other group of immuno-related genes (Figure 2
B,D). High expression of MHC genes in ovarian cancer has been already observed [22
] by Gooden and colleagues in a cohort of 270 ovarian cancers, including early and advanced stages, along with tumor infiltrating CD8+ T lymphocytes (CTLs). According to our results, Gooden and colleagues found that HLA-E is highly expressed in virtually all the EOC tumors along with others human leukocyte antigen (HLA) molecules of both class I and II, suggesting that these tumors are characterized by an intact antigen processing apparatus and by abundant CTL infiltration.
Moreover, although not significant, we observed a moderate increase of IPS levels for Muc and Ser samples (both high and low grades) compared to Cc and End (Figure 2
Finally, as expected, different grades have slightly different values of IPS (Supplementary File 3
). This seems to be mainly due to differences in expression levels of specific immune cells e.g., EC and SC classes dedicated to respectively to effector and suppressor cells show significant differences between high and low grade (activated CD8+ T cells and CD4+ T cells, Tem CD8+ and Tem CD4+ cells: G3 vs. G2 p
= 0.02; Tregs and Myeloid-derived suppressor cells (MDSCs); G3 vs. G2 p
= 0.08 Supplementary File 3
). Specifically, samples showing the most severe nuclear atypia and the worst prognosis have higher levels of EC and lower level of SC, suggesting that the activity of both these tumor infiltrated T cells might cause a differential tumor visibility to the immune system [23
]. Finally, we do not find any association between IPS and survival (Supplementary File 3
In conclusion, both approaches indicate that Ser samples seem to be more immuno-reactive with respect to other subtypes having activated effector cells, small abundances of NK cell resting, and altered checkpoints.
2.4. Transcriptional Alterations of Ovarian Cancer Stage I
To highlight subtype-specific transcriptional alterations, we identified differentially expressed genes across histotypes and performed a Gene Ontology (GO) enrichment analyses (Supplementary File 6
). However, given the small number of differentially expressed genes (using FDR
< 0.05 as threshold), not all the subtype comparisons show significant results; in particular, the analysis using Molecular Function categories gives poor results (Supplementary File 6
). On the other hand, using Biological Processes, we found that SerHigh seems to be the most different subtype with respect to the others. Specifically, high and low grade Serous differ for expression in genes involved in cell cycle regulation and DNA repair, while SerHigh and End for genes involved in development, cell–cell communication, migration and Wnt signaling. Finally, SerHigh and Cc show expression differences for genes involved in extracellular structure organization metabolic and catabolic processes (Supplementary File 6
). As expected, our results indicate SerHigh as the subtype with a more aggressive phenotype in terms of proliferation, deficient DNA repair, and migration.
The GO functional analyses reported above give a general description of the biological processes characterizing subtypes without taking into account (i) the links among these processes and (ii) the impact that the non-coding part of the transcriptome has on expression alterations. To fill these gaps, we compared integrated transcriptomes (mRNA and miRNA) using a combined network-based approach as described in Materials and Methods. Briefly, each histotype is compared with all the others, and the results are combined into a unified network in which our previously identified prognostic circuit [13
] was integrated.
The outcome of our analysis is a network composed of 166 genes and 70 miRNAs broadly covering around 13 pathways (Supplementary Material 4
). The network has been summarized in Figure 3
A, which represents a detailed and comprehensive picture of all the biological processes differentiating EOC histotypes.
The first result that emerges from the structure of the network is that transcriptional differences across histotypes are found at the coding gene level, while at the pathway level all histotypes showed a high degree of overlap. This indicates that every histotype shows alterations on specific coding transcripts and gene-family members, but that the same biological processes are involved. Moreover, we found that histotype-specificity seems to be principally guided by small non-coding RNAs as most of them modulate the cross-talk between different areas of the network with a histotype-specific expression.
Secondly, our results indicate a central role of cAMP-PKA-CREB1 signaling axis as the main pathway associated with histotype specificity. This pathway promotes the expression of cAMP-dependent target genes and requires the activation of CREB1 by cAMP-dependent protein kinase A (PKA). This signal has been already linked to the platinum resistance in advanced ovarian cancer by Dimitrova et al. [24
] showing that patients with high cAMP have low progression-free survival.
cAMP-PKA-CREB1 signaling is linked to a large range of other biological signals such as metabolic pathways, BAD oncogene, and, exclusively in Cc, the regulation of intracellular availability of Calcium. It is clear from the network topology that miRNAs elements play central role in histotype-specificities modulating the interplay between different areas of the network such as RAC1, MAPK signaling pathway, cell cycle, and growing factor receptors.
We confirm the role of Cc miRNA biomarkers miR-30a-3p and miR-30a-5p [16
] in cell cycle regulation, as well as of Muc miRNA biomarkers, miR-192 and miR-194 [17
], in regulating multiple steps of cell cycle, Activin and Inhibins pathways, and Integrins.
Apart from their role in physiological events, integrins are also involved in many pathological conditions such as inflammation and tumor progression. Here, different histotypes are characterized by the deregulations of specific integrins: ITGA1, ITGA11, ITGB1 are downregulated in Ser samples while ITGA10, ITGB4 in Cc, ITGB8 in Muc and ITGB3 in End. The involvement of ITGA1 at least in serous histotype suggests an MYC dependent regulation. Very recently, a new competing enhancer regulatory mechanism between PVT1 and MYC has been demonstrated [25
]. PVT1 promoter partially insulates downstream enhancer elements from regulating MYC expression, and thus functions as an intra-TAD boundary element. In this way, PVT1 expression that we found associated with poor prognosis [15
], impairs MYC expression that at the same time is a key regulator of ITGA1 [26
] downregulated in high and low grade Ser samples. ITGA1 is found to be a key player of cell adhesion-mediated drug resistance in ovarian cancer [27
], thus its strong downregulation in Ser histotype suggests a more therapy-sensitive phenotype of serous ovarian tumors compared to other histotypes.
The only link to the immune system is reported through the presence of SOCS2 that is a target of miR-192 and is downregulated in Ser samples. SOCS2 is a negative regulator of the IFN-gamma signaling pathway, which, in turn, is a key regulator of balance between immune activation and immune suppression [28
]. Specifically, loss of SOCS2 enables robust tumor-immune rejection by expanding DC-based T cell priming and antigen-specific adaptive immunity. In agreement with our previous observations, our network suggests a higher immune-reactive behavior of Ser (with low levels of SOCS2 for both high and low grade serous) and Muc (with high levels of miR-194-5p that targets SOCS2) than in other histotypes.
2.5. Histotype-Specific Transcripts Validation
To validate our network results, a list of 53 (30 miRNAs and 23 genes) transcripts were selected and their expression values were assessed using qRT-PCR in the training set. The selection was made according to literature evidence, expression differences across histotypes, and topological position within the network.
In the training set, a total of 20 molecular elements (38%) were confirmed as histotype-specific (7 over 23 genes and 13 over 30 miRNAs) (Supplementary Material 5
). Apart from the already published mucinous (miR-192 and miR-194) and clear cell (miR-30a-3p and miR-30a-5p) miRNA biomarkers [16
], we confirm some known coding and non coding elements known to be involved in tumor progression e.g., MDM2, ESR1, E2F3, Cyclin Dependent Kinases (CDKN1A, CDKN2A, CDK4 and CDK6), hsa-miR-214-3p and hsa-miR-29a/b/c, and hsa-let-7a.
These 20 elements were further tested in an independent validation set (Table 2
and Supplementary Material 5
). Ten out of the 13 miRNAs (77%) and 5 of the 8 genes (63%) were validated as histotype specific (Figure 3
B and Supplementary Material 5
). As novel biomarkers, we found miR-214-3p and let7a-5p showing significantly high levels in Muc and SerLow histotypes, miR-96 showing low levels in Muc and SerLow, and mir-29b that is highly expressed only in SerLow subtypes. On the other hand, looking at the coding genes, we found that Muc histotype has a significantly lower and higher expression levels of respectively MDM2 and CDKN2A with respect to other subtypes, CDK6 is lowly expressed in Muc and low grade subtypes, while CDKN1A is upregulated in low grade serous samples. Instead, the Cc subtype is characterized by low levels of E2F3.