Confirmation of ProMisE: A simple, genomics-based clinical classifier for endometrial cancer
See editorial on pages 728-30, this issue.
We are grateful to the family and friends of Sarabjit Gill who have provided support for research in endometrial cancer through the British Columbia Cancer Agency.
Abstract
BACKGROUND
Classification of endometrial carcinomas (ECs) by morphologic features is irreproducible and imperfectly reflects tumor biology. The authors developed the Proactive Molecular Risk Classifier for Endometrial Cancer (ProMisE), a molecular classification system based on The Cancer Genome Atlas genomic subgroups, and sought to confirm both feasibility and prognostic ability in a new, large cohort of ECs.
METHODS
Immunohistochemistry (IHC) for the presence or absence of mismatch repair (MMR) proteins (to identify MMR deficiency [MMR-D]), sequencing for polymerase-ɛ (POLE) exonuclease domain mutations (POLE EDMs), and IHC for tumor protein 53 (p53) (wild type vs null/missense mutations; p53 wt and p53 abn, respectively) were performed on 319 new EC samples. Subgroups were characterized and assessed relative to outcomes. The prognostic ability of ProMisE was compared with that of current risk-stratification systems (European Society of Medical Oncology [ESMO]).
RESULTS
ProMisE decision-tree classification achieved categorization of all cases and identified 4 prognostic subgroups with distinct overall, disease-specific, and progression-free survival (P < .001). Tumors with POLE EDMs had the most favorable prognosis, and those with p53 abn the worst prognosis, and separation of the 2 middle survival curves (p53 wt and MMR-D) was observed. There were no significant differences in survival between the ESMO low-risk and intermediate-risk groups. ProMisE improved the ability to discriminate outcomes compared with ESMO risk stratification. There was substantial overlap (89%) between the p53 abn and high-risk ESMO subgroups; but, otherwise, there were no predictable associations between molecular and ESMO risk groups.
CONCLUSIONS
Molecular classification of ECs can be achieved using clinically applicable methods and provides independent prognostic information beyond established clinicopathologic risk factors available at diagnosis. Consistent, biologically relevant categorization enables stratification for clinical trials and/or targeted therapy, identification of women who are at increased risk of having Lynch syndrome, and may guide clinical management. Cancer 2017;123:802–13. © 2016 American Cancer Society.
INTRODUCTION
There are over 60,000 new cases of endometrial carcinoma (EC) diagnosed in North America each year,1, 2 with reported increases in both incidence and mortality rates globally. Still, EC is understudied; and, to date, no progress has been made to address its greatest challenge: irreproducible pathologic categorization, particularly of high-grade tumors,3, 4 which can lead to imprecise estimation of the risk of disease recurrence and death. This results in the over-treatment and under-treatment of thousands of women. Diverse EC subsets are lumped together in clinical trials, making interpretation of treatment efficacy impossible.
To address these shortfalls,5-7 we developed a novel molecular classifier called ProMisE (Proactive Molecular Risk Classifier for Endometrial Cancer) that assigns patients with endometrial cancer to 1 of 4 groups based on a combination of mutation and protein expression analyses8 (Fig. 1). This classification scheme was based on results from The Cancer Genome Atlas (TCGA) collaborative project that yielded 4 molecularly defined EC subgroups,9 but used cost-prohibitive methods for group assignment in routine clinical practice and depended on fresh-frozen tumor samples, requiring special sample handling. We and another research team8, 10 have tested more pragmatic methods that identify distinct subgroups with a prognostic signature consistent with the genomic subgroups identified through TCGA. We previously tested 16 models using key molecular parameters and settled on the best for further evaluation.8, 11
Steps in molecular classification with Proactive Molecular Risk Classifier for Endometrial Cancer (ProMisE) are illustrated. The first assessment is immunohistochemistry for the presence of mismatch repair (MMR) proteins to identify women, enabling rapid referral to the hereditary cancer program and possibly directing surgical or therapeutic decisions. Tumors are assessed next for polymerase-ɛ (POLE) exonuclease domain mutations (EDMs) and finally for protein 53 (p53) IHC, yielding 4 subgroups: MMR-D, POLE, p53 wild type (wt), and p53 null/missense mutations (abn).
Our objectives were, first, to test the ProMisE classifier in a new, larger confirmation cohort of cases, following Institute of Medicine guidelines11 for the development of omics-based tests, to confirm that this classification scheme yields prognostic information consistent with previous findings. Second, we wanted to compare the prognostic ability of ProMisE to traditional risk-stratification system commonly used to direct treatment (European Society for Medical Oncology [ESMO]).12
MATERIALS AND METHODS
Cohort Selection, Clinicopathologic Data, and Outcomes
After institutional review board approval was obtained, patients were identified retrospectively by searching for all EC patients from our institution who had clinical data and adequate specimen volume for molecular analysis. In a previous publication,8 we interrogated 16 molecular classification models, taking into account the order of testing and methodologies suitable for clinical application. Then, we selected and locked down the analytic methods of the ProMisE classification system, and we wished to evaluate its performance in a larger independent cohort of patients from our center (termed the confirmation cohort). Definitions used for the collected parameters and outcomes and molecular methods, including tissue microarray construction and immunohistochemistry, DNA extraction, sequencing, and ProMisE molecular subgroup assignment, were as described in previous publications8, 13 and further detailed in the Supporting Methods (see online supporting information).
Statistical Analysis
The prognostic signature of ProMisE in the confirmation cohort was compared with that obtained in our previous publication as well as in the TCGA study by considering Kaplan-Meier survival plots. The Harell C-index was used to evaluate the ability of ProMisE to discriminate outcomes in new patients from the confirmation cohort using model parameters developed in the discovery cohort.8
The ProMisE molecular subgroups were characterized in the confirmation cohort using descriptive statistics; univariable associations of individual clinicopathologic parameters with ProMisE subgroups were compared using 1-way analysis of variance for continuous data (age at surgery, body mass index [BMI]) and the chi-square test for categorical data (stage, grade, histologic subtype at diagnosis, lymphovascular space invasion [LVSI], myometrial invasion, and lymph node status). Only patients who had complete clinicopathologic parameters were considered in the analyses. A missing value comparison was undertaken to ensure that missing values were not associated with subgroups.
Univariable and multivariable survival analyses of ProMisE were considered along with other clinicopathologic parameters of interest in association with overall survival (OS), disease-specific survival (DSS), and progression-free survival (PFS) as predictors. For multivariable survival analyses, a Cox proportional-hazards model was considered with ProMisE subgroups and prognostic factors available from the time of diagnosis, because our ultimate objective is to use ProMisE before surgical staging. Thus, we corrected for age, BMI, grade, and histotype in addition to molecular subgroup as assigned by ProMisE. We also performed multivariable analysis using additional parameters available from postsurgical staging, because several of these are known to be important prognostic factors (eg, stage, lymph node status, myometrial invasion, and the presence of LVSI). Adjuvant treatment status was included as a covariate in all multivariable models to account for the possible confounding effect of treatment.
In all Cox models, hazard ratios (HRs) with corresponding 95% confidence intervals (CI) and likelihood ratio test (LRT) P values were reported. The Firth penalized maximum-likelihood bias-reduction method was used to estimate the HR (indicated by a superscript F) when the number of censored cases exceeded 80%, and the corresponding CI were obtained using the profile likelihood. The proportional-hazard assumptions were graphically evaluated by means of smoothed Schoenfeld residuals.14
The discriminatory ability of several models combining ProMisE and other clinicopathologic parameters was compared using the Harell C-index and was internally validated in the full cohort (combined discovery8 + confirmation cohorts) using 1000 bootstrap samples to adjust for potential optimism in the performance.15, 16 Kaplan-Meier survival analysis was performed on the confirmation and full cohorts, plotting OS, DSS, and recurrence-free survival for each of the 4 ProMisE molecular subgroups as well as for ESMO clinical risk groups. The classification systems were then compared using cross-tabulation. All statistical hypothesis tests performed were 2-sided. Statistical significance was set at α = .05, and no attempts were made to correct for multiple comparisons.
RESULTS
Confirmation of Prognostic Signature in a New Cohort of Cases
The confirmation cohort consisted of 319 new cases drawn from 519 surgical specimens of EC, as outlined above (see Materials and Methods). Supporting Figure 1 (see online supporting information) outlines the reasons for excluding patients from the analyses; including having recieved neoadjuvant chemotherapy, missing molecular data, or follow-up < 2 years. The discovery cohort has previously been described8 (n = 141). Together, the confirmation and discovery cohorts encompassed 460 cases and are termed the full cohort.
The confirmation cohort included cases from a longer time period (1983-2013) than the original discovery series (2002-2009) (Supporting Fig. 2; see online supporting information). Overall follow-up in both cohorts was 5.2 years (reverse Kaplan-Meier analysis). Comparisons of clinicopathologic parameters, ESMO risk groups, and ProMisE molecular subgroups across the discovery, confirmation, and full cohorts are provided in Supporting Table 1 (see online supporting information).
We previously demonstrated8 the ability of the ProMisE molecular classifier to discriminate outcomes (OS, DSS, PFS) compared with ESMO risk groups. To evaluate ProMisE alone, ESMO alone, and ProMisE and ESMO combined, after accounting for treatment, we applied model parameters estimated from the discovery cohort (n = 141) to new cases (n = 319; the confirmation set), and we compared the Harrell C-index computed on the confirmation cohort with the previously derived index. We observed that the resulting C-index for ProMisE ranged from 0.66 to 0.67, depending on the endpoint selected (Supporting Table 2; see online supporting information). The discriminatory ability of ProMisE alone in the new cases was comparable to that of ESMO alone; and, when both were combined together, the performance was better than that of each alone. Kaplan-Meier survival curves for the confirmation cohort (Fig. 2) and the full cohort (Supporting Fig. 3; see online supporting information) were consistent with what was previously demonstrated in both our discovery cohort8 and TCGA data.9
Kaplan-Meier survival analyses are illustrated for the confirmation cohort (n = 319) according to Proactive Molecular Risk Classifier for Endometrial Cancer (ProMisE) molecular subgroups, including (A) overall survival (OS), (B) disease-specific survival (DSS), and (C) progression-free survival (PFS). CI indicates confidence interval; HR(F), hazard ratio using the Firth penalized maximum-likelihood bias-reduction method (with “F” indicating that the proportion of censored cases was >58% [corresponding confidence intervals were obtained using the profile likelihood]); MMR-D, mismatch repair-deficient; p53, tumor protein 53; p53 abn, null/missense p53 mutation; p53 wt, wild-type p53; POLE EDM, polymerase-ɛ exonuclease domain mutation.
Characterizing ProMisE Subgroups
Statistically significant univariable associations between molecular subgroups and all clinical and pathologic parameters measured are provided in Table 1. Supporting Table 3 (see online supporting information) presents details on follow-up and the number of events according to ProMisE subtypes. Supporting Table 4 (see online supporting information) provides further details on the molecular features in the full cohort. Missing data analysis revealed an association with BMI but no other associations with clinicopathologic parameters or outcomes (Supporting Table 5; see online supporting information).
| No. of Patients (%) | ||||||
|---|---|---|---|---|---|---|
| Variable | Total | MMR-D | POLE EDM | p53 wt | p53 abn | P Value |
| No. of patients | 319 | 64 | 30 | 139 | 86 | |
| Age, y | ||||||
| Mean ± SE | 66.9 ± 0.7 | 66.3 ± 1.4 | 62.2 ± 2.2 | 65.2 ± 1.1 | 71.7 ± 1 | .000 |
| Median | 68.1 | 67.2 | 61.5 | 65.6 | 71.7 | |
| Missing | 3 | 1 | 0 | 2 | 0 | |
| BMI, kg/m2 | ||||||
| Mean ± SE | 31.3 ± 1.2 | 33.9 ± 5.3 | 25.9 ± 0.8 | 32.2 ± 1.2 | 29.7 ± 0.9 | .000 |
| Median | 27.9 | 28.5 | 26.7 | 27.9 | 28.3 | |
| Missing | 42 | 7 | 6 | 25 | 4 | |
| Stage | ||||||
| I | 221 (70.2) | 47 (74.6) | 27 (93.1) | 109 (79.6) | 38 (44.2) | .000 |
| II-IV | 94 (29.8) | 16 (25.4) | 2 (6.9) | 28 (20.4) | 48 (55.8) | |
| Missing | 4 | 1 | 1 | 2 | 0 | |
| Grade | ||||||
| 1-2 | 123 (38.6) | 12 (18.8) | 9 (30) | 99 (71.2) | 3 (3.5) | .000 |
| 3 | 196 (61.4) | 52 (81.2) | 21 (70) | 40 (28.8) | 83 (96.5) | |
| Histology | ||||||
| Endometrioid | 215 (67.4) | 48 (75) | 24 (80) | 126 (90.6) | 17 (19.8) | .000 |
| Nonendometrioid | 104 (32.6) | 16 (25) | 6 (20) | 13 (9.4) | 69 (80.2) | |
| LVSI | ||||||
| No | 189 (62.6) | 34 (54.8) | 17 (60.7) | 107 (81.1) | 31 (38.8) | .000 |
| Yes | 113 (37.4) | 28 (45.2) | 11 (39.3) | 25 (18.9) | 49 (61.3) | |
| Missing | 17 | 2 | 2 | 7 | 6 | |
| Myometrial invasion | ||||||
| None | 49 (15.7) | 3 (4.7) | 3 (10.3) | 32 (23.4) | 11 (13.4) | .001 |
| <50% | 145 (46.5) | 34 (53.1) | 16 (55.2) | 67 (48.9) | 28 (34.1) | |
| >50% | 118 (37.8) | 27 (42.2) | 10 (34.5) | 38 (27.7) | 43 (52.4) | |
| Missing | 7 | 0 | 1 | 2 | 4 | |
| Lymph node status | ||||||
| Negative | 150 (47.6) | 41 (64.1) | 19 (63.3) | 34 (24.6) | 56 (67.5) | .000 |
| Not tested | 146 (46.3) | 22 (34.4) | 11 (36.7) | 98 (71) | 15 (18.1) | |
| Positive | 19 (6) | 1 (1.6) | 0 (0) | 6 (4.3) | 12 (14.5) | |
| Missing | 4 | 0 | 0 | 1 | 3 | |
| Treatment | ||||||
| None | 163 (52.6) | 35 (55.6) | 12 (40) | 94 (68.6) | 22 (27.5) | .000 |
| Any | 147 (47.4) | 28 (44.4) | 18 (60) | 43 (31.4) | 58 (72.5) | |
| Missing | 9 | 1 | 0 | 2 | 6 | |
| ESMO | ||||||
| Low risk | 95 (30) | 10 (15.6) | 7 (24.1) | 77 (55.8) | 1 (1.2) | .000 |
| Intermediate risk | 49 (15.5) | 13 (20.3) | 10 (34.5) | 18 (13) | 8 (9.3) | |
| High risk | 173 (54.6) | 41 (64.1) | 12 (41.4) | 43 (31.2) | 77 (89.5) | |
| Missing | 2 | 0 | 1 | 1 | 0 | |
- Abbreviations: BMI, body mass index; ESMO, European Society of Medical Oncology; LVSI, lymphovascular space invasion; MMR-D, mismatch repair-deficient; p53, tumor protein 53; p53 abn, null/missense p53 mutation; p53 wt, wild-type p53; POLE EDM, polymerase-ε exonuclease domain mutation; SE, standard error.
- a Testing was done using a 1-way analysis of variance with no assumption on the homogeneity of variances.
Several observations about molecular subgroups and phenotypes, which were previously described in the discovery cohort,8 were noted again in the confirmation cohort, including younger age and lower BMI among women with polymerase-ɛ (POLE) exonuclease domain mutation (EDM) tumors; older age; and lower BMI associated with tumors that have null/missense tumor protein 53 (p53) mutations (p53 abn tumors), low stage (93.1% stage I), high grade (70% grade 3), and predominantly endometrioid (80%) tumors in the POLE EDM subgroup. The p53 abn tumors had aggressive pathologic features (96.5% grade 3, 80.2% nonendometrioid histology, 61.3% LVSI, and the highest proportion of lymph node positivity [14.5%]) and advanced stage, but we also noted aggressive pathologic features (81.2% grade 3, 25% nonendometrioid histotype, 45.2% with LVSI, 25.4% advanced stage II-IV) associated with mismatch repair (MMR)-deficiency (MMR-D ECs) (Table 1) compared with the wild type p53 (p53 wt) subgroup.
All 4 cases with POLE EDMs and recurrence or death were reviewed. Although 1 of these was an advanced stage (IIIC1) mixed serous-endometrioid carcinoma, the remaining were endometrioid and early stage (IA/B). Three of 4 patients received adjuvant chemotherapy and radiation yet still experienced recurrence and died of their disease. The fourth patient received no adjuvant therapy and died of intercurrent illness (at age 92 years) without documented recurrence.
The order of molecular testing for ProMisE, as illustrated in Figure 1, means that some patients may have more than 1 molecular feature but get categorized based on the first feature identified in the diagnostic tree (eg, MMR-D). In our full cohort of 460 cases, 16 (3.4%) had more than 1 molecular feature (Table 2).
| Patient | Age, y | Stage | Histology and Grade | ESMO | Treatmentb | Recurrence Status | PFS, y | OS, y | MMR IHCc | POLE EDMd | TP53 IHCd | TP53 Mutd | Other Mutationsd | Comments |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| 1 | 56 | IA | CC3 | 3 | EBRT, VB | No | 19.96 | MSH6 loss | Wild type | 2+ | R175H | None | MSH2 unknown | |
| 2 | 73 | IB | EM3 | 3 | None | No | 11.12 | MSH6 loss | Wild type | 2+ | R175H | None | MSH2 unknown on 4-panel testing | |
| 3 | 75 | IB | EM3 | 3 | None | Yes | 0.75 | 0.75 | PMS2 loss | Wild type | 2+ | N239D | None | MLH1 loss on 4-panel testing |
| 4 | 48 | IA | Undiff3 | 3 | None | Yes | 0.29 | 0.29 | MSH6 loss | Wild type | 0 | Frameshift | None | Both MSH6 and MSH2 loss on 4-panel testing |
| 5 | 71 | IV | S3 | 3 | CP | Yes | 3.00 | 10.10 | MSH6 loss | Wild type | 2+e | Wild type | None | p53 IHC resulted in mixed assignment, fluidigm revealed no mutation; MSH6, MSH2 present on 4-panel testing |
| 6 | 64 | IV | S3 | 3 | CP | Yes | 0.60 | 1.11 | PMS2 loss | Wild type | 2+ | D281H | None | MLH1 loss on 4-panel testing |
| 7 | 42 | IIIC1 | S/CC3 | 3 | CP, EBRT, PA | No | 5.12 | PMS2 loss | Wild type | 2+e | Wild type | FBXW7: R505C | p53 IHC mixed on older panels; MLH1 loss on 4-panel testing | |
| 8 | 55 | IA | Undiff3 | 3 | None | No | 6.98 | PMS2 loss | Wild type | 2+e | Wild type | None | MLH1 loss on 4-panel testing | |
| 9 | 59 | IIIC1 | EM3 | 3 | CP, EBRT, PA | No | 2.50 | PMS2 loss | Wild type | 2+ | R273H, R273C | None | PMS2 loss, MLH1 unclear | |
| 10 | 65 | IV | EM3 | 3 | CP, EBRT, PA | No | 5.21 | PMS2 loss | Wild type | 2+e | Wild type | PP2R1A: R183W | MLH1 loss on 4-panel testing | |
| 11 | 65 | IA | S3 | 3 | CP. EBRT | No | 2.73 | MSH6 loss | Wild type | 2+ | R248Q | PP2R1A: S95L, E216K; FBXW7: R224X | ||
| 12 | 59 | IIIC1 | EM3 | 3 | EBRT, VB | No | 9.67 | Intact | V411L (59%) | 2+ | E286K (43%) | PP2R1A: E297K(44%); FBXW7: R658X(45%) | ||
| 13 | 61 | 1A | EM3 | 2 | CP | No | 5.24 | Intact | P286S (35%) | 0 | N200T (25%) | FBXW7: R658X(42%) | Increased immune markers, PTEN + (strong), ER++ | |
| 14 | 1A | SC3 | 3 | CP | No | 5.70 | Intact | V411 | 2+ | Wild type | PP2R1A:R183W, E471D; FBXW7:R479Q, S438Y | |||
| 15 | 57 | 1B | EM3 | 3 | CP, EBRT, VB | No | 5.32 | Intact | A456P | 2 | R169X, Y31C | PP2R1A:R221W; FBXW7: R689W | See discordance fluidigm p53 and IHC possible core heterogeneity | |
| 16 | 55 | 1A | S3 | 3 | None | No | 7.03 | PMS2 loss | F367L | 1+ | D41N | None |
- Abbreviations: C, carboplatin; CC, clear cell; EBRT, external-beam radiotherapy; EDM, exonuclease domain mutation; EM, endometrioid; ESMO, European Society of Medical Oncology; ER, estrogen receptor; FBXW7, F-box and WD repeat domain containing 7; IHC, immunohistochemistry; MLH1, mutL homolog 1; MMR, mismatch repair; MSH2, mutS homolog 2; MSH6, mutS homolog 6; Mut, mutation; P, paclitaxel; p53, tumor protein 53; PA, para-aortic boost; PFS, progression-free survival; PMS2, PMS1 homolog 2; POLE, polymerase-e; PP2R1A, protein phosphatase 2, structural/regulatory subunit A, a; PTEN, phosphatase and tensin homolog; S, serous; Undiff, Undifferentiated; VB, vaginal brachytherapy; y, years.
- a Eleven cases had MMR-D and p53 abn (classified as MMR-D) identified, 4 had POLE EDM and p53 abn (classified as POLE EDM) identified, and a single case had MMR-D and POLE mutations (classified as MMR-D) identified. Molecular parameters and clinicopathologic features are outlined. Results from additional genes tested as part of the select sequencing panel (TP53, PP2R1A, and FBXW7) were available and were interrogated for these “double-feature” cases.
- b Treatment refers to primary adjuvant treatment and does not include treatment that may be received at recurrence.
- c MMR IHC results are shown for 2-panel testing (MSH6 and PMS2).
- d Mutation frequencies are indicated in percentages where known: all mutations shown have been validated.
- e The intensity of p53 staining was mixed in older studies (1 + and 2+). Shading is per ProMisE category; Orange for mismatch repair deficiency (MMR-D), blue for cases with confirmed POLE EDMs, green for p53 wt and red for p53 abn. p53 status is described by IHC but also by mutation status. Yellow highlights other mutations identified through sequencing (not part of ProMisE molecular classification) but shared in characterization of these double ‘feature’ cases. Shading is per ProMisE category; Orange for mismatch repair deficiency (MMR-D), blue for cases with confirmed POLE EDMs, green for p53 wt and red for p53 abn. p53 status is described by IHC but also by mutation status. Yellow highlights other mutations identified through sequencing (not part of ProMisE molecular classification) but shared in characterization of these ‘double feature’ cases.
Establishing ProMisE as an Independent Prognostic Marker in EC: Confirmation Cohort
Univariable survival analysis revealed that all demographic, clinical, and pathologic parameters, and known or suspected prognostic factors, were significantly associated with outcomes except BMI (Table 3). Kaplan-Meier survival analysis in the discovery cohort8 had demonstrated near overlap of the p53 wt and MMR-D curves. With the larger number of cases in the confirmation cohort, these 2 intermediate outcome survival curves clearly separate. Figure 2 and Supporting Figure 3 (see online supporting information) illustrate the Kaplan-Meier analyses for OS, DSS, and PFS in the confirmation and full cohorts, respectively (log-rank P ≤ .001 for all), and the numbers and types of events in the confirmation cohort are outlined in Supporting Table 3 (see online supporting information).
| OS | DSS | PFS | |||||||
|---|---|---|---|---|---|---|---|---|---|
| Variable [Ref] | No./Total No. of Events | HR (95% CI)a | LRT P | No./Total No. of Events | HR (95% CI)a | LRT P | No./Total No. of Events | HR (95% CI)a | LRT P |
| ProMisE subgroup [p53 wt] | 92/319 | — | 67/308 | — | 62/254 | — | |||
| MMR-D | 2.21 (1.22-3.92)F | .0000 | 2.81 (1.38-5.74)F | 0.0000 | 3.30 (1.52-7.22)F | .0000 | |||
| POLE EDM | 0.78 (0.25-1.93)F | 0.74 (0.15-2.38)F | 0.51 (0.06-2.12)F | ||||||
| p53 abn | 3.54 (2.18-5.84)F | 5.09 (2.85-9.54)F | 7.84 (4.22-15.59)F | ||||||
| ESMO [low risk] | 92/317 | — | 67/306 | - | 62/253 | — | |||
| Intermediate risk | 1.03 (0.40-2.43)F | .0000 | 2.43 (0.69-9.03)F | .0000 | 11.29 (0.92-1560)F | .0000 | |||
| High risk | 3.44 (2.01-6.30)F | 9.28 (3.97-28.35)F | 105.63 (NA-13300)F | ||||||
| Age at surgery | 90/316 | 1.05 (1.03-1.07) | .0000 | 65/305 | 1.03 (1.01-1.05) | .0037 | 62/253 | 1.02 (1.00-1.05) | .0287 |
| BMI | 76/277 | 0.98 (0.95-1.00) | .0716 | 54/266 | 0.98 (0.95-1.01) | .0953 | 57/224 | 1.00 (1.00-1.01) | .4077 |
| Stage [I] | 90/315 | — | 65/304 | — | 62/253 | — | |||
| II-IV | 3.94 (2.60-5.99)F | .0000 | 5.61 (3.43-9.39)F | .0000 | 7.53 (4.51-12.95)F | .0000 | |||
| Grade | |||||||||
| 1-2 | 92/319 | — | 67/308 | — | 62/254 | — | |||
| 3 | 2.78 (1.73-4.66)F | .0000 | 6.32 (3.15-14.73)F | .0000 | 37.66 (10.28-332.10)F | .0000 | |||
| Histology | |||||||||
| Endometrioid [1-2] | 92/319 | — | 67/308 | — | 62/254 | — | |||
| Nonendometrioid | 2.56 (1.70-3.86) | .0000 | 3.59 (2.22-5.87)F | .0000 | 5.55 (3.33-9.54)F | .0000 | |||
| LVSI [no] | 84/302 | — | 61/292 | — | 57/242 | — | |||
| Yes | 3.37 (2.19-5.27)F | .0000 | 4.45 (2.65-7.72)F | .0000 | 6.22 (3.58-11.35)F | .0000 | |||
| Lymph nodes tested [Negative] | 90/315 | — | 65/304 | — | 62/252 | — | |||
| Not tested | 0.60 (0.38-0.94) | .0181 | 0.63 (0.37-1.06)F | .0134 | 0.10 (0.04-0.23)F | .0000 | |||
| Positive | 1.51 (0.74-3.08) | 2.17 (0.99-4.30)F | 2.43 (1.29-4.31)F | ||||||
| Treatment | |||||||||
| [None] | 89/310 | — | 64/299 | — | 60/248 | — | |||
| Any | 1.65 (1.08-2.52) | .0184 | 2.56 (1.54-4.42)F | .0002 | 3.07 (1.80-5.42)F | .0000 | |||
| Myometrial invasion | |||||||||
| [None] | 89/312 | — | 64/301 | — | 62/252 | — | |||
| <50% | 1.66 (0.76-4.27)F | .0000 | 1.23 (0.46-3.99)F | .0000 | 1.23 (0.51-3.54)F | .0000 | |||
| ≥50% | 4.74 (2.26-11.84)F | 5.67 (2.39-17.45)F | 4.95 (2.21-13.64)F | ||||||
- Abbreviations: BMI, body mass index; DDS, disease-specific survival; HR, hazard ratio; LRT, likelihood ratio test; LVSI, lymphovascular space invasion; MMR-D, mismatch repair-deficient; PFS, progression-free survival; NA, not available; p53, tumor protein 53; p53 abn, null/missense p53 mutation; p53 wt, wild-type p53; ProMisE, Proactive Molecular Risk Classifier for Endometrial Cancer; POLE EDM, polymerase-ε exonuclease domain mutation.
- a A superscript “F” indicates that Firth CIs were computed using the profile likelihood and may not match the LRT P values.
Multivariable survival analysis using parameters that were available at diagnosis and molecular subgroup as assigned by ProMisE demonstrated that only ProMisE subgroup was associated with all measured outcomes (OS, P = .021; DSS, P = .016; and PFS, P = .001). Age was associated with OS (P < .001), only but with an HR just above 1 (Table 4). Accounting for parameters that were available postsurgical staging (stage, lymph node status, myometrial invasion, and LVSI), in addition to that were parameters available at diagnosis, we were insufficiently powered to demonstrate an independent association of ProMisE with outcomes (P values were not significant for OS, DSS, or PFS).
| OS, 76 of 272 Events | DSS, 54 of 261 Events | PFS, 55 of 219 Events | ||||
|---|---|---|---|---|---|---|
| Parameter [Ref] | HR (95% CI)a | LRT P | HR (95% CI)a | LRT P | HR (95% CI)a | LRT P |
| ProMisE molecular subgroup [p53 wt] | ||||||
| MMR-D | 1.90 (0.88-4.04)F | .0211b | 1.32 (0.51-3.35)F | .0156b | 0.64 (0.25-1.60)F | .0011b |
| POLE EDM | 1.01 (0.26-2.99)F | 0.42 (0.04-1.88)F | 0.19 (0.02-0.81)F | |||
| p53 abn | 2.61 (1.27-5.72)F | 2.28 (1.02-5.58)F | 1.75 (0.84-3.96)F | |||
| Age | 1.04 (1.02-1.07)F | .0007b | 1.02 (0.99-1.05)F | .2669 | 0.99 (0.96-1.01)F | .3774 |
| BMI | 1.00 (0.96-NA)F | .2439 | 1.00 (0.95-NA)F | .1874 | 1.01 (1.00-1.01)F | 0.2394 |
| Grade [grade 1-2] | ||||||
| Grade 3 | 1.04 (0.50-2.21)F | .9592 | 2.13 (0.79-6.19)F | .1523 | 28.05 (6.19-268.05)F | .0000b |
| Histology [endometrioid] | ||||||
| Nonendometrioid | 1.31 (0.68-2.49)F | .4958 | 1.51 (0.69-3.16)F | .3651 | 1.35 (0.69-2.70)F | .3716 |
| [No Treatment] | ||||||
| Any treatment | 1.14 (0.67-1.93)F | .5993 | 1.16 (0.61-2.27)F | .6148 | 0.84 (0.46-1.63)F | .6529 |
- Abbreviations: BMI, body mass index; DDS, disease-specific survival; HR, hazard ratio; LRT, likelihood ratio test; MMR-D, mismatch repair-deficient; PFS, progression-free survival; p53, tumor protein 53; p53 abn, null/missense p53 mutation; p53 wt, wild-type p53; ProMisE, Proactive Molecular Risk Classifier for Endometrial Cancer; POLE EDM, polymerase-ε exonuclease domain mutation; Ref, reference category.
- a A superscript “F” indicates that Firth CIs were computed using the profile likelihood and may not match the LRT P values.
- b These P values indicate statistically significant difference.
Comparison of ProMisE Subgroups and ESMO Risk Groups
Kaplan-Meier survival plots of ESMO risk groups against outcomes revealed essentially indistinguishable outcomes between the low-risk and intermediate-risk groups (Fig. 3). Molecular subgroups were cross-tabulated with ESMO risk groups in the full cohort to assess associations between the molecular versus clinical classification systems. The majority of cases (89.5%) in the full cohort that were classified as p53 abn were classified as high-risk by ESMO criteria, however, the diversity among the other molecular subgroups was profound (Table 1 and Supporting Fig. 4; see online supporting information).
Kaplan-Meier survival analyses are illustrated for the confirmation cohort (n = 319) according to European Society for Medical Oncology (ESMO) endometrial cancer risk group, including (A) overall survival (OS), (B) disease-specific survival (DSS), and (C) progression-free survival (PFS). CI, confidence interval; HR(F), hazard ratio using the Firth penalized maximum-likelihood bias-reduction method (with “F” indicating that the proportion of censored cases was >58% [corresponding confidence intervals were obtained using the profile likelihood]).
Internal validation of prognostic models for all cases that have been evaluated to date (full cohort: n = 460) is illustrated. The ability to discern outcome was assessed for European Society for Medical Oncology (ESMO) risk group assignment alone, Proactive Molecular Risk Classifier for Endometrial Cancer (ProMisE) alone, ProMisE with ESMO, ProMisE with select parameters available at the time of diagnosis, and ProMisE with select parameters available at postsurgical staging (highest C-index). DSS indicates disease-specific survival; OS, overall survival; PFS, progression-free survival.
Comparison and Internal Validation of Prognostic Models
We computed the Harrell C-index on all cases evaluated to date (full cohort; n = 460), comparing the discriminatory ability of ProMisE molecular classification, ESMO risk-group stratification, a combination of both ProMisE and ESMO, as well as ProMisE with the addition of select clinicopathologic parameters available from the time of diagnosis or postsurgical staging to discern outcomes (OS, DSS, and PFS). ProMisE alone appeared to perform as well as ESMO (Fig. 4), and the ability to discern outcomes was improved with ProMisE and the addition of select parameters.
DISCUSSION
There is a need for biologically informative molecular tools to improve both 1) reproducibility of EC categorization, and 2) risk stratification to guide optimal surgery, adjuvant therapy, and cancer surveillance regimens for women with EC. Neither is adequately achieved by the current systems. Lack of a consensus among expert pathologists in morphologic assessment of histotype and grade is a long-standing problem3, 4 despite attempts to refine diagnostic criteria,17-19 suggesting that reproducibility in assignment of these subjective morphologic variables cannot be significantly improved. The lack of diagnostic reproducibility in our current classification exerts a pervasive, negative effect; in clinical research, it is impossible to compare cohorts diagnosed in different centers, which, in turn, makes it impossible to specifically study subgroups or move toward subgroup-specific treatment. The lack of diagnostic reproducibility also means that the same patient will be offered different treatments at different centers. Although there is clearly an intent to offer the best treatment by all practitioners, the lack of uniformity in treatment reflects uncertainty about what is best, and this state of uncertainty is a result, at least in part, of the challenges posed by the lack of diagnostic reproducibility,5-7, 20, 21 as noted in previous publications.
The only way to break this impasse is to move toward a more robust molecular classifier, similar to what was suggested by the genomic analyses in TCGA.9 Our group and the Leiden/Translational Research in Postoperative Radiation Therapy for Endometrial Carcinoma (TransPORTEC) team8, 10, 22 were inspired by this but wished to achieve genomic categorization using more pragmatic approaches. Stelloo et al22 recently published their experience evaluating early stage endometrioid endometrial adenocarcinomas using microsatellite instability (MSI) analysis and a 14-gene panel. Those authors categorized tumors in to 4 molecular subgroups and, as in our previous publication and this current report, they evaluated whether genomic classification combined with clinicopathologic features would provide more prognostic information than the latter alone. They concluded that integrating clinicopathologic and molecular factors improved the risk assessment of patients with early stage EC, stating that assessing this integrated risk profile in daily practice is feasible and that it holds promise to reduce over-treatment and under-treatment.
The TCGA, Leiden, and ProMisE molecular classification systems have enabled us to characterize these 4 prognostic subgroups, and the biologic relevance of these molecular features is apparent. The observation that the p53 abn subgroup encompasses the highest proportion of high-grade, advanced stage, nonendometrioid histotypes and arises in older, thinner women comes as no surprise to clinicians or pathologists, but the emerging phenotype of women with ECs that harbor POLE EDMs is of interest; these are younger, thinner women who have tumors with surprisingly aggressive pathologic features (70% have grade 3 tumors, 35% have deep myometrial invasion, and approximately 40% have LVSI) despite favorable outcomes. The MMR-D subgroup had uterine factors (clinicopathologic features in the uterus itself)23 that were very similar to those observed in the POLE subgroup (ie, comparable proportions of high-grade tumors and deep myometrial invasion with slightly higher [45%] LVSI), yet they had the worst observed outcomes of any group except the p53 abn subgroup and actually received less adjuvant therapy than the POLE cohort. The incorporation of molecular features that reflect tumor biology/behavior becomes essential as clinicians increasingly withhold lymph node dissection for patients with apparent early stage disease (approximately 35% of both the MMR-D and POLE subgroups were incompletely staged).
Successful translation to the clinical setting requires methods that are easy to perform and interpret, relatively low in cost, and thus can be achieved at any cancer center. Arguably, our methods, using MMR IHC rather than MSI assay (comparable results,24 no normal sample required, and common practice in any pathology laboratory) and detecting mutations in a focused area of a single gene rather than gene panels, are even more feasible for translation.
ProMisE molecular classification is a decision tree of binary parameters: the presence or absence of a protein or of a mutation. We did observe a very small subset of patients (3.4%) who had more than 1 discerning molecular parameter identified (see Table 2). It will be important to know how to classify such uncommon cases in practice, but a multicenter study will likely be needed to gather enough cases to fully understand their natural history. Some comments are possible, however, based on this case series. Four women had both POLE EDM mutations and p53 abnormalities and were classified by ProMisE as part of the POLE EDM group. The clinical course and molecular landscape of these tumors (evidence of increased mutations across multiple genes) was more consistent with the POLE-ultramutated phenotype; this is in agreement with the findings of Hussein et al,25 who reported that such POLE EDM/p53 abn tumors are more appropriately classified as POLE EDM; thus, in the ProMisE algorithm pulling out POLE-mutated cases as an early step and classifying these tumors as POLE EDM would seem appropriate. It has been reported that POLE EDMs and MMR-D are mutually exclusive.13, 26, 27 We identified 1 patient who had a validated POLE mutation and confirmed PMS1 homolog 2 (PMS2) loss (known family history). Her clinical course was favorable, and the best categorization of her tumor is uncertain.
Even given these challenges, molecular classification appears to offer prognostic information superior to current standards. Deficiencies across all risk-stratification systems for EC have been highlighted in recent publications.5-7 By using the ESMO 2013 classifier,12 which is purportedly the strongest of the available traditional systems,6 our data revealed that essentially only 2 outcome groups were discernible, with the low-risk and intermediate-risk group survival curves overlapping. In contrast, molecular classification yielded 4 distinct subgroups with significantly different survival curves. Cross-tabulation of ESMO and ProMisE revealed significant overlap of cases in the high-risk and p53 abn subgroups but otherwise tremendous diversity of risk-group assignment within the molecular subgroups. It is clear that there may be both over-treatment and under-treatment of women based solely on application of the ESMO risk-assessment tool. Specifically, the POLE subgroup, which we know is associated with excellent outcomes, has an almost equal distribution across risk groups, with a resulting wide variation in therapy. The optimal treatment within each molecular subgroup will need to be investigated through collaborative trials; however, with a robust clinically applicable classifier in hand, such trials will become possible.
On multivariable analysis including the key clinicopathologic parameters available at diagnosis, the ProMisE molecular classification remained significantly associated with OS, DSS, and PFS. We have approximately 600 additional cases under investigation from an independent cancer center (validation cohort11 that, when combined with this series, provides over 1000 cases for statistical modeling, enabling us to determine which key features (at diagnosis or after staging) may further strengthen ProMisE. Priority will be given to the parameters available at diagnosis (eg, not stage, which can only be determined after comprehensive surgical procedure), because recently demonstrated that ProMisE subgroups can be successfully determined in diagnostic endometrial samples (ie, obtained before surgery) and thus have the potential to inform management from the earliest time point.28 This is particularly relevant for young women with EC who are interested in future fertility and wish to delay definitive hysterectomy or for surgical candidates who are deemed vulnerable (eg, elderly), for whom less aggressive surgery may be favored if the risk of metastatic disease is determined to be either low or sufficiently high that the patient can receive adjuvant treatment without needing the information generated by extended staging surgery to guide this decision. Future treatment algorithms could, for example, consider progesterone therapy and delayed hysterectomy in POLE EDM or p53 wt tumors. Completion of the validation phase in the next independent 600 case series will also enable us to cross the “bright line,” as described in the Institute of Medicine guidelines,11 positioning ProMisE for evaluation of clinical utility and use, confirming the appropriateness of testing on diagnostic samples, and translating ProMisE into clinical practice.
In the short term, ProMisE offers benefits that can be immediately realized, including: 1) implementation of MMR IHC for all patients to help identify women who may have Lynch syndrome and should be referred for counseling and testing to distinguish somatic events from germline mutations and for testing of family members, as appropriate; 2) stratification of clinical trials to enable both the interpretation of completed trials and the enrollment of future trials according to ProMisE subgroup; and 3) the identification of opportunities for targeted therapy. Given the success demonstrated with immune blockade in cancers with MMR-D29-32 (including ECs), clinicians may consider directing therapy based on molecular subgroup with high neoantigen load and associated immune infiltrates (MMR-D or even POLE EDM in the rare cases of recurrence). However, further investigations are needed to define how well ProMisE or components of ProMisE will serve as biomarkers to predict response to therapies.
Possible weaknesses of this series include the nature of the patient cohort, which was obtained from a single tertiary referral center over a wide time interval and was not representative of the overall distribution of ECs in the general population. The quality and quantity of DNA extracted from older FFPE tissue blocks can be a concern; however, our success rate in sequencing and validation was very high (only 8 of 519 cases [1.5%] had missing POLE mutation status). As in other series, cases with very low tumor volume may have been less represented in our cohorts. Surgery and receipt of adjuvant treatment varied considerably in these cohorts, consistent with the changing algorithms over the past 3 decades.
In summary, ProMisE addresses the greatest obstacles faced by clinicians and patients in the management of endometrial cancer; namely, the inability to consistently classify EC tumors and deficient risk-stratification systems for directing care. ProMisE provides biologically relevant information to patients and physicians, using molecular data to group patients with EC based on their risk of recurrence and death. We have developed methods that are simple enough to be performed in any cancer center on formalin-fixed, paraffin-embedded material and at relatively low cost, enabling easy translation into the clinic. We are encouraged by the success of ProMisE in diagnostic samples providing earlier information that is particularly relevant for young women struggling with difficult decisions about their reproductive health.28 Molecular classification will enable us to advance both clinical management and research for EC, as has been achieved with subtype-specific approaches to other malignancies over the last several years.
FUNDING SUPPORT
Jessica N. McAlpine is supported in part by a Canadian Institute of Health Research New Investigator Award and the BC Cancer Foundation Clinical Investigator Award.
CONFLICT OF INTEREST DISCLOSURES
The authors have no individual conflicts of interest to declare. The British Columbia Cancer Agency has filed a US patent (no. 62192230) for the ProMisE molecular classifier (pending).
AUTHOR CONTRIBUTIONS
Aline Talhouk: Conceptualization, methodology, software, validation, formal analysis, data curation, writing–original draft, writing–review and editing, and visualization. Melissa K. McConechy: Methodology, investigation, and writing–original draft. Samuel Leung: Software, validation, formal analysis, data curation, and visualization. Winnie Yang: Investigation. Amy Lum: Investigation. Janine Senz: Investigation. Niki Boyd: Conceptualization. Judith Pike: Investigation. Michael Anglesio: Resources. Janice S. Kwon: Investigation and writing–review and editing. Anthony N. Karnezis: Investigation and writing–review and editing. David G. Huntsman: Conceptualization, methodology, resources, and writing–review and editing. C. Blake Gilks: Conceptualization, methodology, investigation, resources, writing–original draft, and writing–review and editing. Jessica N. McAlpine: Conceptualization, methodology, investigation, writing–original draft, writing–review and editing, visualization, supervision, project administration, and funding acquisition.
