Endometrioid Endometrial RNA Index Predicts Recurrence in Stage I Patients

Abstract

Purpose:

Risk prediction with genomic and transcriptomic data has the potential to improve patient outcomes by enabling clinicians to identify patients requiring adjuvant treatment approaches, while sparing low-risk patients from unnecessary interventions. Endometrioid endometrial carcinoma (EEC) is the most common cancer in women in developed countries, and rates of endometrial cancer are increasing.

Experimental Design:

We collected a 105-patient case-control cohort of stage I EEC comprising 45 patients who experienced recurrence less than 6 years after excision, and 60 Fédération Internationale de Gynécologie et d’Obstétrique grade-matched controls without recurrence. We first utilized two RNA-based, previously validated machine learning approaches, namely, EcoTyper and Complexity Index in Sarcoma (CINSARC). We developed Endometrioid Endometrial RNA Index (EERI), which uses RNA expression data from 46 genes to generate a personalized risk score for each patient. EERI was trained on our 105-patient cohort and tested on a publicly available cohort of 263 patients with stage I EEC.

Results:

EERI was able to predict recurrences with 94% accuracy in the training set and 81% accuracy in the test set. In the test set, patients assigned as EERI high-risk were significantly more likely to experience recurrence (30%) than the EERI low-risk group (1%) with a hazard ratio of 9.9 (95% CI, 4.1–23.8; P < 0.001).

Conclusions:

Tumors with high-risk genetic features may require additional treatment or closer monitoring and are not readily identified using traditional clinicopathologic and molecular features. EERI performs with high sensitivity and modest specificity, which may benefit from further optimization and validation in larger independent cohorts.

Introduction

Endometrial carcinoma is the most common gynecologic malignancy in developed countries and is the second most common in developing countries (1). Globally, incidence rates of endometrial carcinoma are rapidly increasing due to socioeconomic transitions and increasing rates of obesity (2, 3). Incidence of endometrial carcinoma is predicted to increase 55% from 2010 to 2030 in the United States (4). The most common histological type is endometrioid endometrial carcinoma (EEC), accounting for 75%–80% of all endometrial carcinomas. Most cases of EEC are low-grade, low-stage, and show a 5-year survival rate of 95% when treated with surgery alone (5–7). Traditional risk stratification systems to determine risk for recurrence use clinicopathologic information such as age, histologic grade, depth of myometrial invasion, and lymphovascular invasion.

With advances in molecular testing, additional prognostic subtypes have been identified in endometrial carcinoma. The Cancer Genome Atlas (TCGA) identified four molecular subtypes of endometrial cancer that correlated with survival outcomes: POLE ultramutated (POLEmut), microsatellite instability (MSI), p53 abnormal/TP53 mutated (p53abn), and those lacking any of these alterations (“Copy number-low”), hereafter referred to as nonspecific molecular profile (NSMP; ref. 8). POLEmut tumors are associated with high tumor mutational burdens and have highly favorable outcomes (96% 5-year survival). Within the NSMP subtype, CTNNB1 exon 3 mutation has variably been associated with decreased disease-free survival for EECs in some studies (9–12). While the prognostic value of the TCGA classification has been validated in several studies (13, 14), the feasibility of performing comprehensive genomic sequencing on all endometrial cancer patients has been questionable. As a surrogate for whole-exome sequencing, immunohistochemical analysis of p53 and mismatch repair protein (MLH1, PMS2, MSH2, and MSH6) expression and sequencing of the exonuclease domain of the POLE gene have been recommended to classify the tumors in the four molecular groups (15). The adoption of molecular classification into routine pathologic and clinical practice has become essential, and this is now included in staging recommendations (16). Additionally, molecular classification is now incorporated in national and international patient management guidelines (16–20), and the PORTEC4a trial is underway to prospectively evaluate the utility of molecular subtyping to inform patient management. Despite having largely favorable clinicopathologic features, some low-stage EECs recur after surgical excision and there is no reliable method for identifying such tumors at highest risks of recurrence (21). The new Fédération Internationale de Gynécologie et d’Obstétrique (FIGO) staging system for endometrial cancer attempts to account for this by “upstaging” EEC with abnormal p53 expression and “downstaging” tumors harboring pathogenic POLE mutations, regardless of histology (16). However, NSMP tumors are a heterogeneous group that comprises both aggressive and clinically low-risk endometrial carcinomas. Estrogen receptor (ER) expression has also been identified as a favorable prognostic factor (21); additional tools for NSMP risk-stratification are needed.

Here we performed a thorough analysis of the transcriptional landscape of EECs using EcoTyper (22, 23) in uterine-confined primary tumors to identify common cellular ecotypes enriched in EECs that recurred, which ultimately may inform disease classification and clinical decision making. Inspired by the methods used for risk-informing algorithms developed for endometrial cancer (21, 24, 25) and other cancers such as Complexity Index in Sarcoma (CINSARC; refs. 22, 23, 26, 27), we developed the Endometrioid Endometrial RNA Index (EERI). EERI is a score that is based on an optimized selection of 46 genes and gene weights. EERI was validated on an external dataset, stage I endometrioid patients from the TCGA Uterine Corpus Endometrial Carcinoma (TCGA-UCEC).

Materials and Methods

The Stanford-EEC cohort

FIGO stage 1 EEC with complete surgical resection and staging were retrospectively identified in our institution’s pathology database, as approved by our institutional review board (protocol 44615), where it was deemed appropriate to waive written consent. Pathologic samples were taken and studies were conducted in accordance with the Declaration of Helsinki. Clinical characteristics, including patient age, body mass index (BMI), race and ethnicity, and treatment course, were determined from the electronic health record. Cases included tumors with subsequent biopsy proven recurrence within 6 years of excision. The site of first recurrence was recorded as either vaginal cuff, locoregional, or distant metastasis. Sites of locoregional recurrence included vagina (other than cuff), pelvis, rectum, peritoneum, urethra, pelvic lymph nodes, and iliac lymph nodes. Assuming a population recurrence rate of 5%, estimated odds ratio of 7, and an alpha of 0.05, a sample size of n = 45 allows for a power of 0.90. The control group included stage 1 tumors matched by FIGO grade in patients with no evidence of recurrent or progressive disease with a minimum 6-year follow-up period from hysterectomy. The 45 case specimens were originally collected between 1998 and 2021 and the 60 control specimens were collected between 2002 and 2018. The outcome of interest for model development was binary case/control status.

Tissue microarrays

For each case, representative tumor areas were selected from hysterectomy specimens on formalin-fixed paraffin-embedded (FFPE) blocks based on the examination of hematoxylin and eosin (H&E)–stained slides. Each tumor was sampled at least in duplicate, using 1.0-mm diameter core punches.

Immunohistochemistry

Immunohistochemistry (IHC) for ER and progesterone receptor (PR) was determined by the review of pathology reports (biopsy or hysterectomy) of whole slide sections for 105 tumors. ER was not performed at the time of initial diagnosis in 13 tumors: of these, ER status was determined using tissue microarrays from hysterectomy material for four cases and three controls. In six patients, there was insufficient material to determine ER status. PR status was not performed at the time of diagnosis in 10 tumors and was not performed retrospectively due to lack of available material. Our institution used commercially available antibodies for ER (prediluted, clone SP1; Ventana Medical Systems, Roche, Indianapolis, IN) and PR (dilution 1:2, clone 1E2; Ventana Medical Systems). Per institutional protocol, intensity (1–3+) and percentage of overall nuclear positivity were recorded; we then simplified each tumor as either negative (<1%) or positive (≥1%).

DNA and RNA sequencing

Sequencing was performed using the Tempus xT Solid Tumor and Tempus xT-RNA assays on the FFPE block with the highest content of tumor from the hysterectomy specimen. FFPE blocks from nontumor sites from the hysterectomy specimen (e.g., cervix, myometrium) were selected for matched normal tissue. The Tempus xT Solid Tumor assay is a targeted 648-gene DNA next-generation sequencing (NGS) panel that uses matched tumor and normal samples to detect single-nucleotide variants, indels, and copy number variants, as well as chromosomal rearrangements in 22 genes with high sensitivity and specificity. It also includes the xT MSI assay of 239 sites to determine MSI status. A limit of detection of 5% was used for single-nucleotide variants. Detected variants were then prioritized and classified based on known hotspot regions within genes, internal and external somatic databases, primary literature, gene-disease relationships, and other features of somatic drivers. Detailed methods have been published previously (28). Tempus xT-RNA is a whole-transcriptome NGS assay that can also detect gene fusions and altered splicing events.

TCGA molecular classification

Using the molecular methods described above, the tumor mutational burden and pathogenic mutations were identified. All tumors were classified based on DNA sequencing analyses (29) into one of the four molecular groups established by TCGA: POLE ultramutated (POLEmut), MSI, p53 abnormal/TP53 mutated (p53abn), and NSMP (8). First, POLE mutation status was assessed, and if there was a pathogenic POLE mutation (30), tumors were classified as POLEmut. Next, the remaining tumors with MSI-high status were considered MSI, and then the microsatellite stable tumors with pathogenic TP53 mutations were labeled as p53 abnormal. Finally, those lacking any of the aforementioned molecular characteristics were classified as NSMP. Recurrence-free survival based on TCGA molecular classification is shown in Supplementary Fig. S1.

EcoTyper discovery

Gene-level transcripts per million (TPM) values were calculated from bulk RNA whole-transcriptome sequencing and analyzed by the EcoTyper machine learning platform. First, bulk tissue expression was analyzed by CIBERSORTx to identify cell type abundances in the tissue. The CIBERSORTx “Fractions” module used signature matrix-based deconvolution to estimate cell populations, which were then “purified” by the HiRes module to yield sample-specific cell-type-specific gene expression profiles. Next, the EcoTyper system used an iterative nonnegative matrix factorization (NMF) to recognize transcriptionally defined cell states. Once the cell states were identified, EcoTyper categorized patterns of co-occurrence of cell states across samples to discover cellular communities that recur across the population.

CINSARC

The CINSARC is an expression-based score that uses 67 genes implicated in mitosis and chromosome integrity (26). The CINSARC score was generated in R using a provided publicly available script (27, 31) to label patients in the C1 and C2 groups. The scores are generated using clustering of tumor gene expression, and using a leave-one-out-cross-validation method assigns patients to the C1 group if their expression profile is closer to the non-metastatic expression centroid, and the C2 group if they are closer to the metastatic expression centroid.

Development of EERI

An initial set of 462 genes were selected, namely, 236 genes identified in the EcoTyper analysis, 53 genes that have previously been implicated in endometrial cancer pathogenesis, and 173 genes with the highest differential expression between the case and control groups from a volcano plot. TPM values from each patient for each of the 462 genes were used to optimize the weights and a constant using an evolutionary algorithm, maximizing for accuracy of prediction between a binary classification of “cases” and “controls” and a binary “EERI High-Risk” or “EERI Low-Risk.” Each patient received an EERI value that was a result of their TPM values and the optimized gene weights (W) and constant (b) using the following equation: $\mathrm{EERI}=\text{log}\left(|TPM\mathit{·}W|\right)\times \frac{TPM\mathit{·}W}{|TPM\mathit{·}W|}+b$. A value greater than zero was considered a positive or “EERI High-Risk,” while a negative value was considered “EERI Low-Risk.” The weights were optimized using evolutionary algorithms, inciting stochastic perturbations in the system to identify which members of the population enhance fitness (32). Evolutionary algorithms do not assume linearity or smoothness of a problem and optimize within given constraints (32, 33). A constant was subject to optimization to determine the “set-point” for the threshold between a positive and negative result. Additionally, sparse regression was utilized by incurring a penalty (a 0.01 deduction on the computed accuracy) for each additional gene that needed to be used, while priority was given to the combination that maximized computed accuracy. Genes are listed in Supplementary Table S1. The model was trained on the Stanford-EEC cohort and tested on the TCGA-UCEC cohort.

TCGA-UCEC Test Set

Bulk RNA sequencing obtained from the publicly available TCGA-UCEC (8) dataset was used as a test set for the EERI score. A total of 263 patients were found with all the following: endometrial cancer diagnosis, endometrioid subtype, stage I, and bulk RNA sequencing data. All samples in this dataset were collected from fresh frozen tissue. This is the largest dataset of publicly available EECs with bulk RNA sequencing available, generated by the TCGA Research Network: https://www.cancer.gov/tcga. The outcome of interest for model testing was time-to-recurrence. This outcome is right-censored for loss to follow-up.

Statistical analysis

For categorical variables such as the clinical characteristics, a χ² test was used to compute statistical significance of associations. Continuous variables were assessed first with an ANOVA test in the case of multiple groups. P-values were determined with a nonparametric Mann–Whitney U test for post hoc and pairwise groups. A 95% confidence interval was used for all tests and Bonferroni correction was used to adjust for multiple testing. Differences in recurrence-free survival were compared using log-rank tests of Kaplan–Meier curves. The multivariate Cox proportional-hazards model was used to calculate hazard ratios and 95% confidence intervals. The assumption of proportional hazards was verified by modeling Schoenfeld’s residuals (Supplementary Fig. S2).

Data availability

All the raw RNA sequencing data used in this report have been deposited in the National Center for Biotechnology Information Sequence Read Archive under the BioSample project PRJNA1064405 accession numbers SAMN39431340 to SAMN39431455. Other data generated in this study are available within the article and its supplementary data files. Raw data are available upon reasonable request to the authors.

Results

Clinicopathologic characteristics of Stanford-EEC cohort

A total of 105 FIGO stage IA (n = 67), IB (n = 34), and IC (n = 4) EEC were identified, including 45 cases with biopsy-proven recurrence (median time to recurrence 1.6 years, range 0.1–5.9 years) and 60 controls without recurrence. In the control group, follow-up ranged from 6.0 to 17.3 years (median 7.4 years). Twenty-four patients (24%) were treated with adjuvant vaginal brachy-therapy, which included five cases and 19 controls. Two patients (2%), one case and one control, were treated with adjuvant pelvic external beam radiation (EBRT). Three patients (3%), all within the case cohort, received both adjuvant vaginal brachytherapy and EBRT. Seventy-five patients (71%) did not receive adjuvant treatment, which included 40 controls and 35 cases. The administration of adjuvant radiation did not have a significant correlation with recurrence in this cohort (P = 0.46). However, the type of radiation used differed, with brachytherapy alone being used more frequently in the control group than in the cases (32% vs. 11%, P = 0.03) and EBRT alone used less in the controls than in the cases (0% vs. 7%, P = 0.03).

Anatomic location of first recurrence included vaginal cuff (n = 18), locoregional (n = 14), and distant metastasis (n = 13). The most common sites for distant metastasis at first recurrence were abdominal wall (n = 3), omentum (n = 2), and lung (n = 2). The majority (87%) of tumors were low histologic grade (FIGO grade 1–2). Stage and lymphovascular space invasion (LVSI) were the only variables with significantly different representation between the two groups. There were more patients with stage IA disease in the recurrent group than in the control group (84% vs. 48%, P < 0.01). In the control group, more patients had positive LVSI than in the recurrent group (29% vs. 15%, P = 0.03).

In the Stanford-EEC cohort, all TCGA molecular groups were represented, the majority of which were NSMP (62%), followed by MSI (28%), p53 abnormal (9%), and POLEmut (2%). There were more patients with p53 abnormality in the recurrent group than in the controls (13% vs. 5%); however, the distribution among the four TCGA molecular subtypes was not significantly different (P = 0.43). There were no significant differences between cases and controls with respect to patient age, race, ethnicity, BMI, TCGA molecular subtype, CTNNB1 status, histologic grade, ER status, PR status, or age of the sample. The clinical characteristics of cases and control are summarized in Table 1.

Table 1.

Clinical characteristics of cases and controls in the Stanford-EEC cohort.

	Total	Case	Control	P value
Number of patients	105 (100%)	45 (43%)	60 (57%)
Median time to recurrence/follow-up (years)		1.6 (0.1–5.9)	7.4 (6–17.3)
Recurrence location
Vaginal cuff	18 (17%)	18 (40%)	—
Locoregional	14 (13%)	14 (31%)	—
Distant	13 (12%)	13 (29%)	—
No recurrence	60 (57%)	—	60 (100%)
Stage				P < 0.01
Stage IA	67 (64%)	38 (84%)	29 (48%)
Stage IB	34 (32%)	6 (13%)	28 (47%)
Stage IC	4 (4%)	1 (2%)	3 (5%)
TCGA molecular subtype				P = 0.43
MSI	29 (28%)	13 (29%)	16 (27%)
p53 abnormal	9 (9%)	6 (13%)	3 (5%)
POLE	2 (2%)	0 (0%)	2 (3%)
NSMP	65 (62%)	26 (58%)	39 (65%)
NSMP beta-catenin status				P = 0.14
CTNNB1 normal	37 (57%)	11 (42%)	26 (67%)
CTNNB1 abnormal	28 (43%)	15 (58%)	13 (33%)
Grade				P = 1
Low (G1–G2)	91 (87%)	39 (87%)	52 (87%)
High (G3)	14 (13%)	6 (13%)	8 (13%)
LVSI status				P = 0.03
Negative	75 (79%)	31 (79%)	44 (79%)
Suspicious	5 (5%)	5 (13%)	0 (0%)
Positive	22 (23%)	6 (15%)	16 (29%)
ER status				P = 0.92
ER+	95 (98%)	39 (98%)	56 (98%)
ER−	2 (2%)	1 (1%)	1 (1%)
PR status				P = 0.09
PR+	89 (94%)	34 (36%)	55 (58%)
PR−	6 (6%)	5 (5%)	1 (1%)
MMR status				P = 0.71
MMR intact	67 (64%)	27 (60%)	40 (67%)
MMR loss	22 (21%)	9 (20%)	13 (22%)
MMR unknown	16 (15%)	9 (20%)	7 (12%)
Race				P = 0.48
White	56 (53%)	27 (60%)	29 (48%)
Asian	17 (16%)	6 (13%)	11 (18%)
Black or African American	3 (3%)	0 (0%)	3 (5%)
Native Hawaiian or other Pacific Islander	2 (2%)	1 (2%)	1 (2%)
Other race	21 (20%)	7 (16%)	14 (23%)
Unknown	6 (6%)	4 (9%)	2 (3%)
Ethnicity				P = 0.32
Hispanic or Latino	18 (17%)	6 (13%)	12 (20%)
Not Hispanic or Latino	82 (78%)	35 (78%)	47 (78%)
Unknown ethnicity	5 (5%)	4 (9%)	1 (2%)
Menopausal status				P = 0.62
Premenopausal	17 (16%)	9 (20%)	8 (13%)
Postmenopausal	87 (84%)	35 (80%)	52 (87%)
Age				P = 0.76
<50	8 (8%)	4 (9%)	4 (7%)
50–60	32 (30%)	16 (36%)	16 (27%)
60–70	40 (38%)	17 (38%)	23 (38%)
70–80	16 (15%)	6 (13%)	10 (17%)
>80	9 (9%)	2 (4%)	7 (12%)
BMI				P = 0.96
Under 20	5 (5%)	2 (4%)	3 (5%)
20–25	21 (20%)	10 (22%)	11 (18%)
25–30	25 (24%)	9 (20%)	16 (27%)
30–35	20 (19%)	8 (18%)	12 (20%)
Over 35	34 (32%)	16 (36%)	18 (30%)
Adjuvant treatment				P = 0.46
None	75 (71%)	35 (78%)	40 (67%)
Brachytherapy and/or EBRT	30 (29%)	10 (22%)	20 (33%)
Type of adjuvant treatment				P = 0.03
Brachytherapy	24 (23%)	5 (11%)	19 (32%)
EBRT	2 (2%)	1 (2%)	1 (2%)
EBRT and brachytherapy	3 (3%)	3 (7%)	0 (0%)
Block age				P = 0.08
<5 years	9 (9%)	7 (16%)	2 (3%)
5–10 years	66 (63%)	29 (64%)	37 (62%)
10–15 years	22 (21%)	5 (11%)	17 (28%)
>15 years	8 (8%)	4 (9%)	4 (7%)

Clinicopathologic characteristics of the TCGA-UCEC cohort

Within the TCGA-UCEC cohort (34), FIGO stage IA (n = 133), IB (n = 109), and IC (n = 21) EEC were identified. All patients were included regardless of follow-up time to best represent the native patient population. This cohort contained 22 patients with documented recurrence (median 1.6 years, range 0.2–4.4 years) and 241 patients without evidence of recurrence. In the nonrecurrent patients, follow-up ranged from 0.1 to 11.4 years (median 2.8 years). Twenty-three patients (9%) were treated with adjuvant radiation, which included three with recurrence and 20 without recurrence. Type of radiation was not specified in this dataset. A total of 222 patients (85%) did not receive adjuvant treatment, which included 19 with recurrence and 203 without recurrence. Eighteen patients did not have a record of adjuvant radiation. The administration of adjuvant radiation did not have a significant correlation with recurrence in this cohort (P = 0.72).

Anatomic location of first recurrence included vaginal cuff (n = 5), locoregional (n = 4), distant metastasis (n = 7), and unknown locations (n = 6) for which no location was recorded. The majority (51%) of tumors were of low histologic grade (FIGO grade 1–2).

In the TCGA-UCEC cohort, all molecular subtypes were represented, including NSMP (40%), MSI (39%), p53 abnormal (11%), and POLEmut (10%). There were more p53 abnormal tumors in the recurrent group (27% vs. 10%); however, the distribution between the four molecular subtypes was not significantly different (P = 0.06). There were no significant differences between recurrent and nonrecurrent tumors with respect to patient age, race, ethnicity, BMI, TCGA molecular subtype, or histologic grade. The clinical characteristics of cases and control are summarized in Table 2.

Table 2.

Clinical characteristics of recurrent and nonrecurrent tumors in the TCGA-UCEC cohort.

	Total	Recurrent	Nonrecurrent	P value
Number of patients	263 (100%)	22 (8%)	241 (92%)
Median time to recurrence/follow-up (years)		1.6 (0.2–4.4)	2.8 (0.1–11.4)
Recurrence location
Vaginal cuff	5 (2%)	5 (23%)	—
Locoregional	4 (2%)	4 (18%)	—
Distant	7 (3%)	7 (32%)	—
Recurrence location unknown	6 (2%)	6 (27%)	—
No recurrence	241 (92%)	—	241 (100%)
Stage				P = 0.97
Stage IA	133 (51%)	10 (45%)	123 (51%)
Stage IB	109 (41%)	10 (45%)	99 (41%)
Stage IC	21 (8%)	2 (9%)	19 (8%)
TCGA molecular subtype				P = 0.06
MSI	102 (39%)	10 (45%)	92 (38%)
p53 abnormal	30 (11%)	6 (27%)	24 (10%)
POLE	25 (10%)	0 (0%)	25 (10%)
NSMP	106 (40%)	6 (27%)	100 (41%)
Grade				P = 0.39
Low (G1–G2)	155 (59%)	10 (45%)	145 (60%)
High (G3)	107 (41%)	12 (55%)	95 (39%)
Race				P = 0.89
White	194 (74%)	16 (73%)	178 (74%)
Asian	12 (5%)	0 (0%)	12 (5%)
Black or African American	40 (15%)	5 (23%)	35 (15%)
Native Hawaiian or other Pacific Islander	3 (1%)	0 (0%)	3 (1%)
American Indian or Alaska Native	2 (1%)	0 (0%)	2 (1%)
Unknown	12 (5%)	1 (5%)	11 (5%)
Ethnicity				P = 0.13
Hispanic or Latino	5 (2%)	1 (5%)	4 (2%)
Not Hispanic or Latino	178 (68%)	10 (45%)	168 (70%)
Unknown ethnicity	80 (30%)	11 (50%)	69 (29%)
Menopausal status				P = 0.29
Premenopausal	20 (8%)	0 (0%)	20 (9%)
Postmenopausal	223 (92%)	22 (100%)	201 (91%)
Age				P = 0.61
<50	27 (10%)	0 (0%)	27 (11%)
50–60	68 (26%)	5 (23%)	63 (26%)
60–70	95 (36%)	10 (45%)	85 (35%)
70–80	57 (22%)	6 (27%)	51 (21%)
>80	16 (6%)	1 (5%)	15 (6%)
BMI				P = 0.62
Under 20	6 (2%)	1 (5%)	5 (2%)
20–25	35 (13%)	5 (23%)	30 (12%)
25–30	55 (21%)	6 (27%)	48 (20%)
30–35	49 (19%)	3 (14%)	45 (19%)
Over 35	121 (46%)	7 (32%)	103 (43%)
Adjuvant treatment				P = 0.72
None	230 (87%)	19 (86%)	203 (84%)
Brachytherapy and/or EBRT	23 (9%)	3 (14%)	20 (8%)

Ecotype E3 and E4 abundance is associated with recurrence and CTNNB1 status

Four ecotypes (CE1, CE2, CE3, CE4) were identified by the EcoTyper Discovery pipeline with unique cell types and cell states. CE1 contains high fractions of fibroblasts, dendritic cells, and neutrophils. CE2 consists of predominately fibroblasts and neutrophils. Both CE3 and CE4 are composed of low fractions of immune cells and predominately endothelial and epithelial cells existing in polarized cell states. The epithelial cells in the CE3 ecosystem were characterized by high expression of genes implicated in oncogenesis (WT1, SET, WNT7B), immune-mediating genes (C4A, C4B, CCL21, CCR7, CD200, SIGLEC10), and metabolism and cellular stress (UCN, VNN1, ALOX15B). In addition, endothelial cells in the CE3 ecotype had high expression of genes associated with tumor invasion (VASP, CNN2, COTL1, SLIT1), inflammation (IFI16, KLRK1, SOD2, COX6B1, CA12, TNIP1), and chemotherapy resistance (SLC31A2, EIF3G). Conversely, the CE4 ecotype has epithelial cells rich in expression of tumor suppressors (TP53, CDKN2A, EAF2), MUC16 (also known as CA125), and hormone receptors (GHRH, SSTR1, AKR1C2). The endothelial cells in the CE4 ecotype had high levels of genes associated with apoptosis (CASP12, PDCD1, APOPT1, BNIP3), DNA repair mechanisms (NHEJ1, PARPBP, TDP2), and immune regulation (TLR1, CD38, CD48, CD69, IL7R).

Of the four carcinoma ecotypes, only CE4 demonstrated significantly different expression between the case and control groups. Tumors that did not recur, on average, had higher abundance of CE4 (P < 0.01), Fig. 1A. Tumors that were classified as NSMP and wild-type CTNNB1 status had the highest abundance of CE4, while NSMP tumors with CTNNB1 exon 3 mutations had the highest amount of CE3 (Fig. 1B).

Figure 1.

Ecotypes E3 and E4 abundance is associated with recurrence and CTNNB1 status. A, Abundance of ecotypes by recurrence status. ANOVA f-value = 0.03. B, Abundance of ecotypes by CTNNB1 status. ANOVA f-value = 0.02 and post hoc testing with Mann–Whitney U test significant.

CINSARC classifies EECs into low- and high-risk groups

The CINSARC classification was computed for the patients in our cohort. The CINSARC class C1 is defined as the non-metastatic centroid of gene expression, while the C2 class is the metastatic centroid. In the Stanford cohort, CINSARC stratified patients into high- and low-risk groups. The median recurrence-free survival for patients designated as C1 (low-risk) was greater than in the C2 (high-risk) group (6.6 vs. 5.2 years, P = 0.04; Fig. 2A and B). However, in the TCGA-UCEC cohort, which has a lower percentage of patients with recurrence, CINSARC classification did not accurately predict recurrence-free survival (Fig. 2C and D). Median recurrence-free survival was not significantly different between the C1 and C2 groups (3.1 vs. 3.4 years, P = 0.76), with a hazard ratio of 1.4 (95% CI, 0.6–3.5). Risk stratification with CINSARC utilizes a leave-one-out cross validation method to train and predict on the same cohort; however, with more samples a reference cohort could be created, allowing for a more robust prediction with this gene panel in the future. While the genes used to characterize patients into CINSARC and EcoTyper Discovery groups were different, the C1 group was comprised of 65% of patients with high CE4 abundance, and the C2 group was comprised of 48% of patients with high CE3 abundance, suggestive of some overlap between these two risk prediction methods.

Figure 2.

Survival estimates by CINSARC designation. A, Average recurrence-free survival by CINSARC groups and (B) prediction table in the Stanford-EEC cohort. C, Average recurrence-free survival by CINSARC groups and (D) prediction table in the TCGA-UCEC cohort. P-values by Mann–Whitney U test. Kaplan–Meier survival curves in the (E) TCGA-UCEC cohort by CINSARC designation. P-value from log-rank test.

EERI accurately stratifies EEC patients into high- and low-risk groups

Forty-six genes and weights were identified using evolutionary algorithms, maximizing for accuracy of predicting recurrence with the Stanford-EEC cohort as the training set and the TCGA-UCEC cohort as the test set. EERI uses the expression of 46 genes to stratify patients into high- and low-risk groups (Fig. 3A and B). On a receiver operator curve (ROC), EERI was able to predict with area under the curve (AUC) of 0.94 in the Stanford-EEC cohort (training set) and 0.81 in the TCGA-UCEC cohort (test set) shown in Fig. 3C. In the Stanford-EEC cohort, EERI was able to stratify patients into high- and low-risk groups with an accuracy of 94% (sensitivity of 96%, specificity of 93%). Among the TCGA-UCEC cohort, EERI had an accuracy of 81% (sensitivity of 91%, specificity of 80%) as shown in Fig. 3D and E. Kaplan–Meier curves stratified by EERI designation are shown in Fig. 3F, based on 6-year recurrence with administrative censoring at that time. Among all patients in both the Stanford-EEC and TCGA-UCEC cohorts, 59.3% patients designated as EERI High-Risk experienced a recurrence while only 2.6% of the EERI Low-Risk recurred. False negatives and false positives were not significantly associated with any clinical or pathologic variable, including use of adjuvant radiation, when compared to cases as listed in Supplementary Tables S2 and S3. Within the NSMP and MSI molecular subtypes, the EERI High-Risk classification was associated with recurrence in both the Stanford-EEC and TCGA-UCEC cohorts (Supplementary Fig. S3; P < 0.01). Survival curves for patients with stages II–IV disease in the TCGA cohort are provided in Supplementary Fig. S4.

Figure 3.

Performance of EERI. Heatmap of TPM by gene per group in the (A) Stanford-EEC cohort and (B) TCGA-UCEC cohort. C, EERI ROC and AUC in the Stanford-EEC and TCGA-UCEC cohorts. Prediction table based on EERI designation in the (D) Stanford-EEC cohort and (E) TCGA-UCEC cohort. F, Kaplan–Meier survival curves and risk tables for the TCGA-UCEC cohort.

EERI high-risk confers greater risk of recurrence than other classifiers

To compare EERI to clinical and pathologic variables, hazard ratios (HR) for risk of recurrence-associated EERI were computed for the TCGA-UCEC cohort using a Cox multivariate proportional-hazards regression. Clinical variables accounted for were age, BMI, menopausal status, and use of adjuvant radiation. Pathologic features that were considered were stage at diagnosis (IA–IC), tumor grade (I–III), and p53 status. CINSARC and EcoTyper Designations were also included as covariates. When the concurrent clinical and pathologic variables were accounted for, EERI High-Risk designation was found to be significantly associated with recurrence in the TCGA-UCEC cohort (HR 9.86; P < 0.001) as seen in Fig. 4A. Estimated survival functions for those with EERI High- and Low-Risk scores with all other covariates fixed at their average values are shown in Fig. 4B. No other classifier had a significant HR in the Cox multivariate proportional-hazards model.

Figure 4.

Six-year recurrence-free survival Cox proportional-hazards model with covariates. A, HR of recurrence and associated variables in the TCGA-UCEC cohorts by Cox multivariate testing. Error bars represent 95% confidence intervals. B, Survival based on Cox regression adjusted for average covariates in (A).

Discussion

While most patients with early-stage EEC have an excellent prognosis, patients who showed recurrence have poor outcomes with limited treatment options (6, 35, 36). Even EEC with ultralow-risk features, including grade 1 morphology and lack of myometrial invasion, experience 3% recurrence, usually within the first 36 months of diagnosis (37). Prior studies of low-grade low-stage EEC have demonstrated that patients with recurrent disease were older (35, 37) and were most commonly non-White; however, in our study these clinical characteristics did not differ between the case versus control groups (35, 37). Risk stratification for early-stage EEC based on clinicopathologic features alone has shown poor accuracy in multiple studies (38), and the underlying mechanisms for progression of disease in this group is largely unknown.

The incorporation of molecular classification in addition to clinicopathologic features shows improved risk stratification of ECs, particularly in high-risk cancers (13, 15, 39). POLEmut tumors, regardless of tumor grade and stage, have demonstrated excellent prognosis in numerous studies (8, 40, 41). Conversely, p53 abnormal tumors have demonstrated significantly worse prognosis (15, 39, 40). All TCGA molecular subgroups are represented in low-stage EEC, which is corroborated in our study. There was no significant difference in TCGA molecular subgroup distribution between our recurrent and nonrecurrent cohorts, which both showed similar molecular classification distributions to those seen on meta-analyses of EEC (42). Notably there were only two patients with tumors harboring POLE mutations, and both were in the control group. Efforts have been made to better predict outcome in very low-risk EEC. Several previous studies have demonstrated CTNNB1 exon three mutations to be associated with higher recurrence rates (10, 12, 43, 44); however, CTNNB1 mutations also occur in up to 28% of EECs with favorable outcome (44). PIK3CA (45) and chromosome 1q gains (35) have also been correlated with disease recurrence in low-grade, low-stage EEC. However, larger studies assessing a combination of somatic genetic alterations, RNA expression, and tumor microenvironment are lacking.

This study aimed to develop novel tools for predicting risk of recurrence in low-risk patients with early-stage EEC. Using EcoTyper, we identified four distinct cellular ecotypes that were composed of different epithelial, endothelial, and immune cell populations. The abundance of two ecotypes (CE3 and CE4) was correlated with outcomes (Fig. 1). Both ecotypes were composed of primarily epithelial and endothelial cells with no significant immune cell populations. High CE3 expression was associated with higher rates of recurrence. The CE3 ecotype was characterized by high expression of genes involved in oncogenesis (WT1, SET, WNT7B), inflammation (IFI16, KLRK1, SOD2, COX6B1, CA12, TNIP1, C4A, C4B, CCL21, CCR7, CD200, SIGLEC10), tumor invasion (VASP, CNN2, COTL1, SLIT1), and chemotherapy resistance (SLC31A2, EIF3G). The CE4 ecotype was associated with low rates of recurrence, characterized by high levels of genes associated with tumor suppression (TP53, CDKN2A, EAF2), apoptosis (CASP12, PDCD1, APOPT1, BNIP3), DNA repair (NHEJ1, PARPBP, TDP2), and immune regulation (TLR1, CD38, CD48, CD69, IL7R), and hormone receptors (GHRH, SSTR1, TPT). These findings suggest that there may be a specific microenvironment of endothelial and epithelial cells within the primary tumor that increases the risk for recurrence. CE3 and CE4 abundance was not associated with any specific molecular subtype. However, within the NSMP subset, CE3 was associated with mutant CTNNB1, while CE4 correlated with wild-type status, consistent with findings demonstrated in prior studies (10, 12, 35, 44).

In CINSARC, a validated prognostic model in sarcomas, we used an RNA expression signature based on 67 genes implicated in mitosis and chromosome integrity to identify patients at risk of recurrence. It has also been validated in many other epithelial cancer types, including colorectal carcinoma (26, 31). In the Stanford-EEC cohort, CINSARC score was predictive of recurrence-free survival and supports the hypothesis that there are disparate genetic signatures of low-stage EECs that progress. The low rate of recurrence for early-stage endometrial cancer may make the application of CINSARC for risk stratifying less feasible than other methods presented in this study.

To better determine potential molecular and genetic risk factors for future progression, we developed the EERI, a risk prediction score that utilizes the RNA expression of 46 genes to stratify patients into low- and high-risk groups for predicting the likelihood of recurrence in early-stage EEC. EERI was able to prognostically stratify patients with greater accuracy than TCGA molecular classification, and outperformed two preexisting algorithms, EcoTyper and CINSARC. EERI was trained on the Stanford-EEC cohort, in which it was able to stratify patients into low- and high-risk groups with high sensitivity (96%) and high specificity (93%). In the test set, the TCGA-UCEC cohort, sensitivity remains high (91%) while specificity suffers somewhat (80%). This may be due to overfitting of the small training set, and/or inherent biologic ambiguity between recurrent and nonrecurrent endometrial cancers. The strength of EERI is its high sensitivity with a negative predictive value of 97% (Stanford-EEC) and 99% (TCGA-UCEC). Notably, the positive predictive value is 91% in the Stanford-EEC cohort, for which there is a higher frequency of recurrence, and much lower at 30% in the TCGA-UCEC cohort. If the TCGA cohort is more representative of EECs in general, EERI’s overestimation of high-risk individuals may lead to overtreatment if utilized alone. Therefore, EERI still warrants validation and further optimization in larger datasets to increase robustness before being utilized clinically.

EERI was optimized through evolutionary algorithms that identified 46 genes that had the greatest impact on recurrence out of 462 genes. In the 46 genes utilized by EERI, several of the poor prognostic genes were associated with cell-cycle regulation (PPARG, RXRB, TRIM46, RET, DHDH), metabolism (KHK, PLD5), beta-catenin (CCDC85A), and innate immunity (DEFB4A, ZCCHC3, CTS11). Several of the protective genes were associated with immune activity (CD8A, IL1RN, KLHDC7B, MYEOV), tissue repair (TFF1, MMP10), cellular differentiation (LEFTY1, CST6), and Wnt inhibition (DKK1). Other notable genes included in EERI have roles in the immune system (CD68, CD300E) cell-cycle regulation (CCNJ), transcription (SET), and differentiation (EN2).

Our findings are consistent with previous studies that have demonstrated the potential of genomic data to improve risk prediction in endometrial cancer. Many RNA-based scores have been developed for breast cancer (Ocotype Dx, MammaPrint, Prosigna, Breast Cancer Index, and EndoPredict; ref. 46), which have paved the way for more precise data-driven approaches to prognostication of early-stage cancers. Another example utilized in this paper is CINSARC. Several studies have validated the prognostic value of CINSARC in soft tissue sarcomas (26, 31). However, our study demonstrates that EERI outperforms CINSARC in predicting the risk of recurrence in low-risk patients with early-stage EEC.

Notably, the feasibility of performing sequencing assays on all endometrial cancer patients has not been established, and the TCGA molecular classification system has not been widely adopted into clinical practice at all institutions and laboratories. The EERI scoring system hopes to overcome the limitations of previously developed prognostic tools. However, EERI has been validated in only two cohorts, the Stanford-EEC (105 patients) and the TCGA-UCEC cohort (263 patients) and robustness outside of these cohorts has yet to be determined. Further study of additional cohorts from multiple institutions is needed. As such, future work will focus on testing EERI with additional datasets and refining the gene panel as more data become available.

In conclusion, our study demonstrates the potential of RNA-based risk prediction algorithms to improve risk stratification and clinical decision making in low-risk endometrial cancer. By identifying patients who are unlikely to experience recurrence, EERI may help clinicians make informed treatment decisions, sparing ultralow-risk patients from the potential harms of unnecessary adjuvant therapy. Future studies are needed to validate the clinical utility of EERI in larger, multicenter cohorts and to investigate the potential of EERI to guide the development of more personalized and targeted therapies for endometrial cancer.

Supplementary Material

Note

Supplementary data for this article are available at Clinical Cancer Research Online (http://clincancerres.aacrjournals.org/).

Translational Relevance.

Here we present the Endometrioid Endometrial RNA Index (EERI), which has a sensitivity of 91% and specificity of 80% in predicting 6-year recurrence for patients with stage I endometrioid endometrial carcinoma. Tested in an independent cohort, the The Cancer Genome Atlas–Uterine Corpus Endometrial Carcinoma cohort, an EERI high-risk designation is associated with a 30% increase in recurrence and a hazard ratio of 9.9 [95% confidence interval (CI), 4.1–23.8; P < 0.001]. This study demonstrates the potential of RNA-based risk prediction algorithms to improve risk stratification and clinical decision making in low-risk endometrial cancer. By identifying patients who are unlikely to experience recurrence, EERI may help clinicians make informed treatment decisions, sparing ultralow-risk patients from the potential harms of unnecessary adjuvant therapy. Future studies are needed to validate the clinical utility of EERI in larger, multicenter cohorts and to investigate the potential of EERI to guide the development of more personalized and targeted therapies for endometrial cancer.