Identification of a novel subgroup of endometrial cancer patients with loss of thyroid hormone receptor beta expression and improved survival

Level 3 FPKM mRNA and miRNA sequencing data from endometrial carcinoma and adjacent normal tissue along with clinical metadata were downloaded from the Genomic Data Commons (GDC) web server in January 2018 using the GenomicDataCommons and TCGAbiolinks R packages (Fig. 1) [16]. All samples from the TCGA endometrial adenocarcinoma cohort were selected [9, 16]. mRNA and miRNA-sequencing data were normalized and processed according to standard GDC protocols [17]. The mRNA FPKM output mapped to 56,716 ensembl (ENSG) gene ids and was converted to transcripts per million (TPM) and subsequently log₂(TPM + 1) transformed to shrink the numeric range of the data. The miRNA FPKM data were similarly log₂(TPM + 1) transformed. Following an initial filtering step (see below), mRNA-sequencing data were checked for confounders using principal component analysis. To test whether NHR expression values were associated with the sequencing plate, a technical variable, a Fisher’s exact test was performed (Fig. 2).

Filtering of mRNA transcripts was performed to narrow down the possible space of candidate NHRs (see Figs. 1 and 3a). To remove genes that were rarely expressed, genes with zero expression in either tumor or adjacent normal endometrial tissue across > 80% of individuals were excluded. Next, for each gene, a boundary was determined to separate tumors into low and high expression groups. Specifically, a boundary two standard deviations below the mean of expression levels in adjacent normal tissue was utilized. Genes with a negative boundary were removed. Tumors were then classified according to their expression above and below the adjacent normal tissue boundary (either “high” or “low”, respectively). Genes with less than 20% or greater than 80% of samples in either the high or low expression tumor group were excluded to allow for sufficient samples for downstream statistical analysis.

Survival analysis was performed using R using the log-rank test as implemented in the ‘survival’ R package [18]. P-values were adjusted for multiple testing using the Benjamini-Hochberg method.

All Tier 1 oncogenes and tumor suppressors were downloaded from the COSMIC database in December 2017 [19]. Mutation data were downloaded from the GDC, and mutations were called using the MuTect2 pipeline [20]. DNA mutations designated as “high-impact,” meaning that they likely impacted the protein structure or the splicing of an mRNA, were selected. A binary matrix was created from the tumor mutation data, where a 1 corresponded to whether the patient had 1 or more high-impact mutations in that gene, and a 0 indicated the absence of a mutation in that gene for that patient (Fig. 6a). Association studies between mutation and expression status were performed using a 2 × 2 Fisher’s exact test (see Fig. 6a) and corrected for multiple testing using the Benjamini-Hochberg method [21].

All analyses and plots were performed in the statistical language R (version 3.6.0). An HTML document created using knitR and RMarkdown contains the code and workflow for all analysis performed in this study (https://​github.​com/​dpique/​endometrial-paper/​blob/​master/​2020_​endomet_​smry.​html). An R package receptLoss is available on Github (https://​github.​com/​dpique/​receptLoss) that facilitates the identification of tumors with expression loss for any dataset with gene expression data from tumor and adjacent normal tissue. This package is also available from Bioconductor (https://​bioconductor.​org/​packages/​release/​bioc/​html/​receptLoss.​html).

Our first objective was to develop an approach to identify nuclear hormone receptors (NHRs) whose expression is lost in a subset of endometrial tumors relative to adjacent normal endometrial tissue within a patient cohort. To accommodate low numbers of adjacent normal samples (N = 35) relative to tumor samples (N = 542), we assumed the adjacent normal data were generated from a single Gaussian distribution for each NHR. Then, using a lower bound defined by two standard deviations below the mean of this Gaussian distribution for the adjacent normal expression data (labeled as “B” in Fig. 3b), tumors were classified into two separate groups based on their expression values relative to this boundary. The advantage of this approach is that no assumptions are made as to whether the tumor data follow a particular distribution [22]. Therefore, patient subgroupings may be inferred in an unsupervised, distribution-independent manner from a low number of adjacent normal samples.

We applied the receptLoss approach to detect tumor-specific expression loss for the expression profiles of the 47 previously-identified NHRs that were captured by the filtered gene expression dataset [1]. Any NHRs that lost expression in a subset of endometrial tumors relative to normal tissue were selected after initial filtering (Fig. 3a, step 2b). An example of the distribution of one NHR in both tumor (blue histogram) and adjacent normal samples is shown in Fig. 3b. There are two subgroups of tumors, separated by a boundary (labeled as “B” in Fig. 3b) drawn by the threshold defined by two standard deviations below the mean of the adjacent normal tissue expression data. To quantify the degree of separation between the two tumor subgroups defined by the NHR expression in adjacent normal tissue, a ∆μ statistic was developed. The ∆μ statistic measures the difference between the means of the two tumor groups for each gene (Fig. 3c). The distribution of all ∆μ values for each of the 21 NHRs that met initial filters (see Methods and Fig. 3c, blue histogram) versus the 6536 genes in the genome is shown in Fig. 3c.

Next, NHRs with ∆μ values that fell 1 standard deviation above all genes were selected. Six NHRs with relatively large ∆μ values exhibit a loss of expression in a subset of tumors relative to normal tissue (Fig. 4). Three of these NHRs – progesterone receptor (PGR), estrogen receptor 1 (ESR1), and androgen receptor (AR) – have been previously reported to lose expression in subsets of endometrial carcinomas [10, 23], which confirms that our approach can capture NHRs known to lose expression in endometrial carcinoma. The proportion of tumors that lose expression for each gene varies widely, along with the shapes of the gene expression distribution. These properties further illustrate how this approach is more flexible than standard analyses that assume identical distributions in both tumor and normal groups. The degree of expression loss is most prominent with PGR, which has the largest ∆μ value and whose tumor gene expression values are bimodally distributed. These results suggest that our approach adequately captures a distinct subset of NHRs that exhibit expression loss in a subset of tumors relative to normal tissue.

It has been observed that the loss of expression of estrogen and progesterone, two of the best characterized NHRs, is associated with worse outcomes in endometrial cancer [15]. The identification of new NHRs whose expression is associated with clinical outcomes could aid in the development of new prognostic biomarkers or therapeutic targets. Three out of the six NHRs (PGR, ESR1, and THRB) showed significant differences in five-year survival analysis between tumors that lost expression versus those that did not (log-rank test, q-value < 0.001, see Methods) (Fig. 5). Of note, the loss of THRB expression was associated with improved 5-year survival. This contrasts with the pattern observed for PGR and ESR1, wherein the loss of these receptors is associated with worse 5-year survival (Fig. 5). In the TCGA cohort, there were no significant differences between AR loss of expression and 5-year survival, though other studies have found conflicting results regarding AR expression and clinical outcomes in endometrial cancer [23, 24]. These results show a previously unreported relationship between THRB expression and endometrial cancer with regards to survival outcome.

In order to identify additional clinical correlations with THRB expression loss, the relationship between THRB expression (loss versus no loss) and several clinical prognostic factors (stage, grade, histology, and TCGA molecular subtypes) was examined [9]. The clinical stage (χ² test, q = 0.063), grade (χ² test, q = 0.036), and histology (χ² test, q = 0.031) of endometrial tumors were not significantly associated with THRB expression at the q < 0.01 level. However, there was a strong association (χ² test, q = 4.94 × 10^− 4) between the loss of THRB expression and the integrative molecular subtype defined by TCGA [9]. Among the 4 integrative molecular subtypes defined by TCGA, the microsatellite instability (MSI) subtype was significantly enriched among tumors that lose THRB expression (28/122 = 23.0%) compared with tumors that do not lose THRB expression (37/420 = 8.8%) (Fisher’s exact test, odds ratio = 3.07, q = 9.76 × 10^− 5). Next, the relationship between THRB expression and the degree of MSI was examined using the MSI-specific classification approach reported by TCGA (χ² test, q = 4.44 × 10^− 6) [9]. Endometrial tumors that lost THRB expression had a greater proportion of tumors classified as high-grade MSI (MSI-high: 52/122 = 42.6%) relative to endometrial tumors that did not lose THRB expression (MSI-high: 75/420 = 17.9%) (Fisher’s exact test, odds ratio = 3.40, q = 1.12 × 10^− 7). The previously unreported association between loss of THRB expression and microsatellite instability in endometrial carcinoma forms the basis for future mechanistic studies and links THRB expression loss to a molecular subtype established by TCGA.

To understand the relationship between high-impact mutations in DNA sequence and loss of NHR expression, an analysis of all samples with available somatic mutation data (N = 399) was performed using a 2 × 2 Fisher’s exact test (Fig. 6a). The strongest relationship was found between mutations in PTEN, a tumor suppressor, and PGR expression (OR = 3.0, q = 9.1 × 10^− 5). Furthermore, a significant relationship between the presence of RNF43 mutations and the loss of THRB expression was identified (odds ratio = 0.295, q = 0.002) (Fig. 6b). RNF43 encodes a ubiquitin ligase that downregulates Wnt signaling activity in pancreatic adenocarcinoma cells [25, 26]. Though the role of RNF43 in endometrial cancer has not been previously described, endometrial cancers with increased Wnt signaling activity are associated with poor outcomes [27]. In addition, patient tumors harboring mutations in NSD1, a histone methyltransferase [28], are inversely associated with the loss of THRB expression (odds ratio = 0.269, q = 0.002) (Fig. 6b). Inactivating mutations in NSD1 are associated with genomic hypomethylation and improved survival in head and neck squamous cell carcinoma [29], though no such link has been reported between NSD1 and survival in endometrial cancer. These findings highlight novel interactions between known drivers of carcinogenesis and THRB expression that may form the basis for future experimentation. For instance, it would be of interest to further understand the transcriptional regulatory relationships between THRB and Wnt signaling, as both directly promote pro-growth transcriptional changes [30].

Progesterone and estrogen receptor co-expression is a feature of many endometrial cancers and is associated with a favorable prognosis [10]. We wondered whether the novel NHR-based subgroups existed independently of known ESR1 and PGR-based subgroups, since this could help define novel molecular subgroups for prognostication purposes. To address this question, we first tested whether any co-expression existed between each of the 15 possible pairings between the 6 NHRs using a two-sided Fisher’s exact test (Fig. 7a). We identified the expected strong association of co-expression between PGR and ESR1 (q-value < 2.42 × 10^− 53 and odds ratio = 182), which is consistent with previous findings from endometrial cancer cohorts [31]. No association was present between THRB and any other nuclear hormone receptor, including PGR or ESR1 (q-value > 0.01). These data demonstrate that the loss of THRB expression occurs independently of other well-characterized NHRs and form the basis for a novel molecular subgroup.

One potential application of this finding is that THRB expression could be used to refine survival prognostication models that utilize both ESR1 and PGR [10]. Consistent with the established literature, endometrial cancers that lose both ESR1 and PGR expression (DN, for double negative) have poor 5-year survival compared with cancers that do not lose both ESR1 and PGR (DP, for double positive) (q-value = 1.23 × 10^− 6, log-rank test) (Fig. 7b, left panel). However, when DN tumors are further subdivided by THRB expression status, DN tumors that express THRB have a poor prognosis (5-year probability of survival = 55.8%, Kaplan-Meier estimate), while DN tumors that do not express THRB have an excellent prognosis (5-year probability of survival = 93.7%, Kaplan-Meier estimate) (q-value = 0.056, log-rank test). Since THRB is expressed in a majority of DN tumors (103/121, or 85%), THRB could be investigated as a potential therapeutic target within a large subset of DN endometrial tumors that otherwise lack targeted treatment options and have a poor prognosis. Indeed, modulating thyroid hormone receptor beta signaling has been investigated successfully as a therapeutic strategy in murine models of hepatocellular carcinoma [32].

miRNAs are small RNA molecules that govern the expression of target mRNAs with complementary sequences. In cancer, the expression of miRNAs is often dysregulated, and miRNAs have been proposed as both therapeutics [33] and therapeutic targets [34] in cancer. Previous studies have established relationships between increased miRNA expression and decreased THRB mRNA expression in papillary thyroid carcinoma and in renal cell carcinoma [35, 36]. To determine whether there was a relationship between miRNA expression and THRB expression in this EC cohort, a differential miRNA expression analysis was performed between tumors with high versus low THRB expression. We observed 3 differentially expressed miRNAs whose expression was increased in tumors that lost THRB expression (q < 0.001, log₂(|fold change|) > 1, Fig. 8a). The miRNA with the most significant q-value, miRNA-146a, was shown in a previous study to directly interact with THRB mRNA in papillary thyroid carcinoma [35]. miRNA-146a remained significantly differentially expressed when performing the same analysis only on endometrial cancers with endometrioid histology (Additional file 1). However, there were no significantly differentially expressed miRNAs among endometrial cancers with serous histology. Furthermore, miRNA-146a is expressed at relatively low levels in adjacent normal endometrial tissue (Fig. 8b). A trend is observed across both tumor and adjacent normal tissue in which decreasing levels of THRB are associated with increasing levels of miRNA-146a (Fig. 8b). The elevated expression of miRNA-146a in tumors that lose THRB expression suggests a potential biological mechanism through which THRB expression is lost in endometrial cancer. Future studies could also examine whether miRNA-146a itself could serve as a potential therapeutic for the high THRB subgroup of patients with poor 5-year survival.