Cluster analysis is an explorative analytical method, serving as a critical tool in psychology, psychiatry and related fields to map heterogeneous data into meaningful subgroups. Despite their extensive historical use, traditional clustering techniques suffer from a lack of stability, robustness, and generalisability. These issues stem from the inherent difficulties of the clustering optimization problem as well as the stochastic nature of algorithm optimizers. To address these challenges, we demonstrate the use of methods utilising ensemble learning techniques to combine clustering results from different algorithms, model specifications, and/or sampled sub-datasets to form a single, more reliable consensus of clustering solutions. We detail ensemble clustering principles, variations in base clustering generation models, and consensus methods. Detailed introductions in existing R libraries and practical examples using R code are provided to guide users in both implementing and optimising ensemble clustering models. We then include simulation studies of real-world data to demonstrate the substantial benefit of ensemble clustering compared with single-run clustering models. The resources presented here will enable researchers to apply advanced clustering techniques to decompose heterogeneous and complex psychological data into stable subgroups.

Supplementary Figures S1-S4

AIMS:Classification and risk stratification of endometrial carcinoma (EC) has transitioned from histopathological features to molecular classification, e.g. the ProMisE classifier, identifying four prognostic subtypes: POLE mutant (POLEmut) with almost no recurrence or disease-specific death events, mismatch repair deficient (MMRd) and no specific molecular profile (NSMP), with intermediate outcome and p53 abnormal (p53abn) with poor outcomes. However, the applicability of molecular classification is unclear in rare but aggressive histotypes of EC, e.g. de-differentiated and undifferentiated endometrial cancers (DD/UDEC). Here, we aim to assembled a cohort of DD/UDEC from a single institution and analysed the prognostic significance of ProMisE molecular subtypes and the expression of SWItch/sucrose non-fermentable (SWI/SNF) chromatin remodelling complex members, previously implicated in the pathogenesis of DD/UDEC. METHODS AND RESULTS:We accrued 88 DD/UDEC cases, assessed POLE status by Sanger sequencing and performed immunohistochemistry for p53, mismatch repair and SWI/SNF proteins on the tissue microarrays assembled. Assignment of molecular subtypes was possible in 80 tumours; POLE sequencing failed in the remaining eight cases. There were 12 (15%) POLEmut, 44 (55%) MMRd, 14 (17.5%) p53abn and 10 (12.5%) NSMP DD/UDEC. POLEmut DD/UDECs had excellent outcomes, but the other three molecular subtypes all had poor outcomes, with no significant differences among them. The loss of one or more SWI/SNF proteins [AT-rich interactive domain-containing protein 1A (ARID1A), ARID1B, SWI/SNF-related, matrix-associated, actin-dependent regulator of chromatin, subfamily A, member 4 (SMARCA4), SMARCA2], observed in 66% (55 of 83) cases, was not of prognostic significance. CONCLUSIONS:These results indicate that all molecular subtypes of DD/UDEC except POLEmut behave in an aggressive fashion. Further study is needed to determine whether these molecular alterations can be targeted with adjuvant therapy, in order to improve outcomes of patients with DD/UDEC.

Intratumoral heterogeneity (ITH) is spatial, phenotypic, or molecular differences within the same tumor that have important implications for accurate tumor classification and assessment of predictive biomarkers. The Canadian Ovarian Experimental Unified Resource (COEUR) has created a cohort of 437 FFPE tissue specimens from 108 tubo-ovarian high-grade serous carcinoma (HGSC) patients to quantify ITH across the anatomical sites and between primary and recurrence. We quantified the ITH of six clinically used immunohistochemical diagnostic and prognostic biomarkers (WT1, p53, p16, PR, CD8, and Ki67). Markers were stained on tissue microarrays and scored using a continuous or categorical interpretation of staining patterns. Two-way random effect and nested intraclass correlation were used to assess continuous markers, and Gwet’s AC1 was used for categorical markers. All biomarkers showed at least substantial agreement over several spatial comparisons, with WT1, p53 and p16 showing almost perfect agreement for most spatial comparisons. Similarly, categorical WT1, p53 and p16 showed almost perfect agreement for temporal comparisons, while the agreement for primary versus recurrence for PR, CD8 and Ki67 was only fair. We provide power calculations to achieve reliability of >0.60 and recommend testing emerging protein biomarkers to see whether they reach a clinically acceptable benchmark level of ITH.

Supplementary Figures S1-S13

Supplementary Tables S1-S11

OBJECTIVES:Optimal management of patients with stage IA p53abn endometrial cancer without myoinvasion, classified as intermediate risk in the 2020 European Society of Gynaecological Oncology, European Society for Radiotherapy and Oncology, and European Society of Pathology (ESGO-ESTRO-ESP) guidelines, and the 2022 European Society of Medical Oncology (ESMO) guidelines, is currently unclear. Practice varies from surgery alone to adjuvant radiation±chemotherapy. Our aim was to assess the risk of disease recurrence in patients with stage IA p53abn endometrial cancer without myoinvasion compared with stage IA with myoinvasion (<50%). METHODS:Stage IA p53abn endometrial cancers were identified from retrospective cohorts. Cases were segregated into stage IA with no myoinvasion, including (1) tumor restricted to a polyp, (2) residual endometrial tumor, and (3) no residual tumor in hysterectomy specimen, versus stage IA p53abn with myoinvasion (<50%), with treatment and outcomes assessed. RESULTS:There were 65 stage IA p53abn endometrial cancers with no myoinvasion (22 polyp confined, 38 residual endometrial tumor, 2 no residual in hysterectomy specimen, 3 not specified) and 97 with myoinvasion. There was no difference in survival outcomes in patients with stage IA without myoinvasion (16% of patients recurred, 19% if there was residual endometrial disease) compared with stage IA with myoinvasion (17%). The risk of recurrence was lowest in patients with stage IA p53abn endometrial cancer without myoinvasion treated with chemotherapy±radiation (8%). Most recurrences in patients with stage IA without myoinvasion were distant (89%), with no isolated vaginal vault recurrences, and all except one distant recurrence occurred in patients who had not received adjuvant chemotherapy. CONCLUSION:The recurrence rate in patients with stage IA p53abn endometrial cancer without myoinvasion was 16%, highest in the setting of residual endometrial disease (19%), and exceeding the threshold where adjuvant therapy is often considered. The high frequency of distant recurrences observed may support chemotherapy as part of the treatment regimen.

Objective Previous research suggests serum CA125 reflects extra-uterine disease in patients with endometrial carcinoma (EC). Our objective was to determine if CA125 can identify patients with extra-uterine and/or nodal metastases, the association of this biomarker with EC molecular subtype, and to explore an optimal cutoff in this context. Methods We assessed the association of CA125 levels with clinicopathologic and outcomes data on a cohort of 1107 molecularly classified EC. Results Abnormal CA125 (>35kU/L) was associated with higher stage and lymph node metastases (LNM) in all EC and in each molecular subtype on univariate (p < 0.01) and multivariate (p < 0.05) analyses. POLEmut had the lowest median CA125 level and proportion of CA125 abnormal patients, and p53abn the highest proportion (p < 0.001). CA125 > 35 kU/L had a sensitivity of 0.82, specificity 0.53, positive-predictive-value 0.92, and negative-predictive-value 0.31 for LNM, with similar values for stage>I. CA125 > 35 kU/L was associated with worse overall (OS), disease-specific (DSS), and progression-free survival (PFS) in all EC, p53abn (OS, DSS, PFS), NSMP (OS, DSS), and MMRd (OS, DSS) subtypes. CA125 > 35 kU/L demonstrated a relative risk (RR) of 2.50 with presence of stage III/IV disease (p < 0.001) and RR 18.4 for the presence of synchronous endometrial and ovarian carcinomas (SEOC)/co-existing adnexal malignancies (CAM) (p < 0.001). An exploratory cut point, optimized for correlation with DSS (CA125 > 24 kU/L) show similar association with clinical parameters and survival outcome. Conclusions CA125 levels are associated with molecular subtype, stage>I disease, and SEOC/CAM. CA125 remains a useful clinical tool in the triage of EC in the era of molecular classification.

Abstract Background: Mismatch repair deficient (MMRd) endometrial cancer (EC) represents one third of all ECs. Identifying patients with MMRd EC enables testing for Lynch Syndrome (LS) and access to FDA-approved immune checkpoint blockade (ICB) therapy. Recent data have highlighted the diversity within MMRd tumors, suggesting worse outcomes and lower response to ICB in patients with MLH1 loss. Our aim was to characterize a cohort of MMRd ECs to elucidate if clinically meaningful substratification of this molecular subtype (MLH1 loss vs. other MMRd) could be achieved. Methods: MMRd ECs were identified from retrospective institutional and population-based cohorts (1994-2016) with clinicopathologic data, immunohistochemistry (IHC) assessment of estrogen receptor (ER) and L1CAM and CTNNB1 mutation status recorded. Multiplex IHC for immune markers (CD3, CD8, CD79a, CD138, PD-1, PD-L1, FoxP3, IDO-1) was assessed and compared in patients with MLH1 loss or PMS2/MLH1 loss (when 2 antibody MMRd testing performed) vs. isolated MSH2, MSH6, PMS2 or MSH2/MSH6 loss (remainder of MMRd). Results: 655 MMRd ECs were identified, 52% of cases assessed with 4 antibody IHC (MLH1, PMS2, MSH2, MSH6) and 48% with 2 (PMS2 and MSH6). This included 32 (5%) patients with MLH1 loss and 488 (75%) with PMS2/MLH1 loss (together 80% of MMRd cohort), 1% (n= 9) MSH2, 11% (n=75) MSH6, 4% (n=24) PMS2, and 4% (n=24) with loss of MSH2/MSH6). 76 cases were confirmed to have MLH1 hypermethylation but 67% of cases with loss of MLH1 or PMS2/MLH1 did not have methylation testing performed. The majority MMRd ECs were FIGO stage I (75%) and endometrioid histotype (90%). Patients with loss of MLH1 or PMS2/MLH1 were older (p<0.001), had higher BMI (p<0.001) and had tumors with more LVI (p=0.031), deep myoinvasion (p<0.001) and grade 3 (p=0.046) compared to remainder of MMRd EC. 9% of the MMRd cohort were ER negative, 8% with loss of MLH1 or PMS2/MLH1 and 16% in remainder of MMRd. Inferior outcomes were observed for progression-free survival and overall survival in patients with MLH1 or PMS2/MLH1 loss compared with the remainder of MMRd. ER, L1CAM and CTNNB1 status were not associated with clinical outcomes across the total MMRd cohort, nor within MLH1 or PMS2/MLH1 loss. There was diversity in the immune landscape within MMRd EC with significantly lower CD8 levels in MLH1 or PMS2/MLH1 loss compared to the remainder of MMRd EC. Furthermore, 25% of the MLH1 or PMS2/MLH1 loss group were tumor infiltrating lymphocyte (TIL) ‘low’/immune cold by cluster analysis compared to 11% in the remainder of MMRd EC. Only 16 patients (2.4%) were identified as having LS but 82% of these MMRd patients had not been tested (31% non-LS testing attributable to identification of hypermethylation of MLH1). Conclusion: Substratification within MMRd ECs can provide prognostic and predictive information, with loss of MLH1 and/or its dimer partner PMS2 identifying a subset of MMRd EC with inferior outcomes that may be attributed to lower TIL. ER, L1CAM, and CTNNB1 mutation status do not add prognostic refinement within MMRd EC. Citation Format: Amy Jamieson, Jennifer Pors, Samuel Leung, Derek Chiu, Stefan Kommoss, Aline Talhouk, David G. Huntsman, Naveena Singh, Blake Gilks, Jessica N. McAlpine. Substratification of mismatch repair deficient (MMRd) endometrial cancers can provide prognostic and predictive refinement [abstract]. In: Proceedings of the AACR Special Conference on Endometrial Cancer: Transforming Care through Science; 2023 Nov 16-18; Boston, Massachusetts. Philadelphia (PA): AACR; Clin Cancer Res 2024;30(5_Suppl):Abstract nr B026.

Abstract While endometrial cancer (EC) has an overall favorable prognosis, some patients do poorly and may benefit from refinements of current classification systems. The TCGA-inspired pragmatic molecular classification tool Proactive Molecular Risk Classifier for Endometrial Cancer (ProMisE) has been integrated into international guidelines for EC risk stratification and management. ProMisE stratifies ECs into four prognostic groups: POLEmut, NSMP (no specific molecular profile), MMRd (mismatch repair deficient), and p53abn, where POLEmut has the best prognosis and p53abn has the worst prognosis. Our objective was to determine if proteomic profiling could provide additional prognostic or predictive information for EC patients, across or within ProMisE molecular subtypes. Global proteome profiling of FFPE samples, that had clinicopathologic and outcome data, was performed on 184 ECs encompassing all four ProMisE subtypes, including replicate samples of the same tumor, and both biopsy and final hysterectomy specimens. To ensure representation of each subtype, we aimed for an approximately equal distribution; 40 (27 %) MMRd, 33 (22 %) POLEmut, 42 (28 %) NSMP and 33 (22 %) p53abn, rather than the population-based distributions. There was high reproducibility in the proteomic profiles of intra-tumor replicate samples, and between matched biopsy and hysterectomy tumor samples (Pearson’s correlation >0.9). Consensus clustering generated four clusters, named ‘Adhesion’, ‘Immune’, ‘Proliferation’, and ‘Metabolic’ based on proteins enriched in each cluster. The Proliferation Cluster had the worst outcomes and the highest proportion of stage III/IV, serous, and p53abn tumors than the other clusters. The Immune Cluster had the most favorable outcomes, despite having a relatively substantial proportion of stage III/IV, serous, and p53abn tumors. We correlated protein expression with common mutations, including ARID1A mutations that cause loss of ARID1A protein expression, found in up to 60% ECs. ARID1A positive tumors also express proteins in the retinoic acid signaling pathway, while ARID1A-deficient tumors express proteins indicative of neutrophil infiltration. Comparing molecular subtypes, we found p53abn ECs were enriched in proteins associated with poor outcomes in many tumor types, such as GRB7. We validated elevated GRB7 expression in p53abn tumors using immunohistochemistry and found an association with worse disease specific survival (DSS) across the whole cohort (HR=2.39). Nucleolin expression was associated with worse prognosis in the NSMP subtype (DSS HR=9.88), while BABAM1 expression was associated with better prognosis within p53abn tumors (DSS HR=2.43). Knockout of BABAM1 (part of the BRCA complex) in cell lines resulted in increased sensitivity to PARP inhibition. Proteomic analysis of EC identifies candidate prognostic markers that may further refine current molecular classification and help guide treatment decisions. New therapeutic interventions could be developed to target proteins and pathways identified by EC proteomic profiling. Citation Format: Dawn R. Cochrane, Gian Luca Negri, Jutta Huvila, Juliana Sobral de Barros, Forouh Kalantari, Nissreen Mohammad, David Farnell, Emily Thompson, Amy Lum, Sandra E. Spencer, Amy Jamieson, Samuel Leung, Derek Chiu, Martin Koebel, Stefan Kommoss, Friedrich Kommoss, Blake Gilks, Lien Hoang, David Huntsman, Gregg B. Morin, Jessica N. McAlpine. Proteomic profiling of endometrial carcinomas [abstract]. In: Proceedings of the AACR Special Conference on Endometrial Cancer: Transforming Care through Science; 2023 Nov 16-18; Boston, Massachusetts. Philadelphia (PA): AACR; Clin Cancer Res 2024;30(5_Suppl):Abstract nr PR002.

OBJECTIVES:We have previously shown that DNA based, single test molecular classification by next generation sequencing (NGS) (Proactive Molecular risk classifier for Endometrial cancer (ProMisE) NGS) is highly concordant with the original ProMisE classifier and maintains prognostic value in endometrial cancer. Our aim was to validate ProMisE NGS in an independent cohort and assess the performance of ProMisE NGS in real world clinical practice to address if there were any practical challenges or learning points for implementation. METHODS:We evaluated DNA extracted from an external research cohort of 211 endometrial cancer cases diagnosed in 2016 from Germany, Switzerland, and Austria, across seven European centers, comparing standard molecular classification (NGS for POLE status, immunohistochemistry for mismatch repair and p53) with ProMisE NGS (NGS for POLE and TP53, microsatellite instability assay) for concordance metrics and Kaplan-Meier survival statistics across molecular subtypes. In parallel, we assessed all patients who had undergone a new NGS based molecular classification test (n=334) comparing molecular subtype assignment with the original ProMisE classifier. RESULTS:A total of 545 endometrial cancers were compared. Prognostic differences in progression free, disease specific, and overall survival between the four molecular subtypes were observed for the NGS classifier, recapitulating the survival curves of original ProMisE. In 28 of 545 (5%) discordant cases (8/211 (4%) in the validation set, 20/334 (6%) in the real world cohort), molecular subtype was able to be definitively assigned in all, based on review of the histopathological features and/or additional immunohistochemistry. DNA based molecular classification identified twice as many 'multiple classifier' endometrial cancers; 37 of 545 (7%) compared with 20 of 545 (4%) with original ProMisE. CONCLUSION:External validation confirmed that single test, DNA based molecular classification was highly concordant (95%) with original ProMisE classification, with prognostic value maintained, representing an acceptable alternative for clinical practice. Careful consideration of reasons for discordance and knowledge of how to correctly assign multiple classifier endometrial cancers is imperative for implementation.

Abstract Introduction: Recent publications have identified ‘low risk’ (LR-NSMP) and ‘high risk’ no specific molecular profile (HR-NSMP) endometrial cancers (ECs). Although LR-NSMP (low grade (G1/2) estrogen receptor (ER) positive tumors have extremely low rates (<2%) of death from disease and are considered candidates for de-escalation of adjuvant therapy, little is known about the newly defined entity of HR-NSMP (G3 and/or ER-negative), including optimal management, with 27% of HR-NSMP patients dying from their disease We aimed to perform in-depth characterization of a cohort of HR-NSMP ECs to identify additional prognostic or predictive features that may inform management. Methods: Clinicopathologic data collection, immunohistochemistry (IHC), and next generation sequencing was performed on a cohort of 148 HR-NSMP ECs testing for associations with outcomes. Three subgroups of HR-NSMP were assessed: i) low grade without any ER expression (G1/2ER-, n=40), ii) high grade with ER (G3ER+, n=67), and iii) high grade without ER expression (G3ER-, n=41). Results: In a univariate analysis, advanced stage (III/IV), positive lymph node (LN) status, no LN assessment, and PIK3CA mutations were associated with inferior progression-free survival (PFS), disease specific survival (DSS), and overall survival (OS). HR-NSMP patients with positive LNs had a hazard ratio (HR) of death of 7.55 (95% CI: 3.41−16.73) compared to node negative. Additionally, patients with no LN assessment had an increased risk of death with a HR of 2.33 (95% CI: 1.12−4.85) compared to node negative. Within the three HR-NSMP subgroups, the worst OS (HR 7.06, CI 4.12-12.09, p< 0.001), DSS (HR 20.43, CI 10.14-41.19, p< 0.001), and PFS (HR 3.69, CI 1.59-8.63, p<0.001) were observed in patients with both adverse features (G3ER-). PIK3CA mutations were found in 32.7% of all HR-NSMP tumors tested (G1/2ER- 45.5%, G3ER+ 42.3%, G3ER- 6.7%) and associated with inferior OS (HR 4.84, CI 1.21-19.41, p=0.014), DSS (HR 4.84, CI 1.21-19.41, p=0.014), and PFS (HR 3.22, CI=1.08-9.66, p=0.028). There was a trend towards decreased OS and DSS in HR-NSMP patients with overexpression of HER2 IHC. L1CAM IHC overexpression and CTNNB1 mutation status were not associated with outcomes when assessed across all HR-NSMP or within subgroups. Conclusion: Among HR-NSMP ECs, the worst clinical outcomes were observed in patients harboring both high grade and ER negative tumors. In contrast to LR-NSMP ECs where patients who had no lymph node assessment had excellent outcomes mirroring node negative status, HR-NSMP patients who had no nodal assessment had a high rate of recurrence and death from disease suggesting the importance of comprehensive staging in this cohort. PIK3CA appears to be a prognostic stratification feature within HR-NSMP and represents a potential therapeutic pathway. More work is needed to understand the association of HER2 expression and ER status and whether HER2 -targeted therapies represent actionable opportunities for these patients. Citation Format: Andrea Neilson, Amy Jamieson, Derek Chiu, Samuel Leung, Amy Lum, Jennifer Pors, Stefan Kommoss, Aline Talhouk, David G. Huntsman, Naveena Singh, C. Blake Gilks, Jessica N. McAlpine. High risk no specific molecular profile (HR-NSMP) endometrial cancer can be stratified into three subgroups based on tumor grade and estrogen receptor status with differing clinicopathologic characteristics and outcomes [abstract]. In: Proceedings of the AACR Special Conference on Endometrial Cancer: Transforming Care through Science; 2023 Nov 16-18; Boston, Massachusetts. Philadelphia (PA): AACR; Clin Cancer Res 2024;30(5_Suppl):Abstract nr B031.

Supplementary Fig. S1. CONSORT diagram of patient selection and cohort construction. Supplementary Fig. S2. Plasma cell infiltrates before and after NACT. The data was generated and analyzed as in Figure 1. Supplementary Fig. S3. Example of intraepithelial PD-1+ FoxP3+ TIL in a quadruple-color IHC of PD-L1 CD8 PD-1 FoxP3. Supplementary Fig. S4. Unsupervised clustering of TIL based on post-NACT data.