A detailed description of the patient cohort was published previously (17). A total of 252 patient tumor samples with EC were collected retrospectively at the University Hospital Inselspital (Bern, Switzerland) and served as the testing cohort. All samples were reevaluated with the use of the 4th edition of the WHO classification and the 8th edition of the TNM classification (Table 1) (18, 19).

An additional 213 patients with early stage EC from the Hospital Universitario La Paz (Madrid, Spain) were included as the validation cohort (Table 1). The detailed description of the characteristics of the validation cohort was published previously (20).

For the testing cohort, a next-generation tissue microarray (ngTMA^®) was constructed using principles published previously (17, 21, 22). Punch biopsies with a diameter of 1.5 mm were collected in triplicate from the tumor center and invasive front (Figure 1 and Table 2).

For the validation cohort, tissue microarrays were made with selected central tumor areas based on hematoxylin and eosin evaluation. 1.2 mm tissue punches were arranged in duplicate using a TMA workstation (Beecher Instruments, Silve Spring, MD, USA) (20).

For the testing cohort, an Opal Kit with four fluorescence colors (Akoya Biosciences, MA, USA) was used to stain the slides (16, 23–30). Cell counts of each fluorescence channel were validated with conventional stained slides (Figure 2). The fluorescence staining was executed following routine protocols on a Leica Bond RX autostainer (Leica Biosystems, Nussloch, Germany). Four antigens were detected with different colors to display nuclei, CD3⁺ and CD8⁺ lymphocytes, and tumor cells (Table 3). For analysis, CD3 positive T-cells were defined irrespective of CD8 status, whereas the latter was defined as a subgroup. Tumor cells were defined as PanCK positive, stromal cells as negative for CD3, CD8 and PanCK.

The validation cohort was stained using two antibody panels and two staining cycles (Table 3). In both staining rounds, antibodies were amplified using Alexa fluorophores and tyramide signal amplification, counter-staining was done with 4,6-diamidino-2-phenylindole (DAPI). A detailed description of the staining process can be found here (16, 20).

All slides of the testing cohort were digitized in 8-bit with a Pannoramic 250 Flash II slide scanner (3D Histech, Budapest, Hungary) with fluorescence mode (Table 3). The scanned slides were imported in MRXS file format to the Case Viewer software (3D Histech, Budapest, Hungary) and exported as TILED TIFFs for further analysis.

For the validation cohort, TMAs were imaged as whole slides using a Zeiss Axio Scan.Z1 scanner (Zeiss Group, Oberkochen, Germany) with a Hamamatsu ORCA-Flash 4.0 V2 Digital CMOS Camera in 16-bit (Hamamatsu Corporation, Bridgewater, NJ, USA). The images were exported as Big TIFF with original raw channel data of four channels (DAPI, CD3, CD8 and PanCK) (20). After scanning, the images were merged into 4-channel RGB TIFFs and exported for further analysis.

Digital image analysis was executed with QuPath as open source tool (31, 32). Applying a script (Supplementary material), the TMAs were analyzed for cell count, percentage, and density of CD3⁺ and CD8⁺ lymphocytes in three different compartments (33). Tissue was detected with the ≪Pixel classifier≫ and defined as intra-tumoral, tumor neighborhood (close area of 50 microns), and tumor distant (Figures 1, 3 and Table 4) (34–36). The cut-off of 50 microns for a close to tumor region was taken from literature and appeared to be appropriate for TMA based analysis to allow appropriate tumor distant areas at core sizes of 1.2-1.5 mm (37). Cells were detected using the ≪Watershed cell detection≫ function. Appropriate thresholds for lymphocytes and tumor cells were calculated mathematically using the ≪Auto threshold≫ function of Fiji ImageJ (open source, GNU General Public License) with mean and maximum entropy models (Table 4) (16, 38). A minimum size of 0.01 mm² per ROI was set as threshold to avoid overestimation of cell densities due to area as denominator. Per case, the mean of cell densities of each ROI was calculated from multiple corresponding cores within the TMAs. Regarding the validation cohort, mean and maximum threshold levels had to be adapted to the different staining intensities and to the spectrum of 16-bit images (Table 4). The identical automated workflow as described above was used to detect and count CD3⁺ and CD8⁺ lymphocytes in each compartment.

Auto-fluorescence from erythrocytes and tumor cells in FITC and CY5 channels was solved by applying basic anatomical principles to the script such as lymphocyte size and density of chromatin to define CD3⁺ and CD8⁺ lymphocytes. This allowed higher thresholds for lymphocyte signals compared to those of erythrocytes and tumor cells (Table 4).

Auto-fluorescent erythrocytes of the validation cohort were priorly removed using Ilastik-1.3.3post2 machine learning software (open source, GNU General Public License). The software was trained to differentiate between empty tissue, red blood cells, and “good” tissue with automatic removal of empty tissue and erythrocytes (20).

POLE mutational status of both cohorts was detected by hotspot Sanger sequencing of the catalytic region of POLE covered in exons 9, 12 (testing cohort), 13, and 14 as previously published (17, 20) and under consideration of the POLE risk score for variant classification (39). Only consensus pathogenic POLE mutations entered the analysis. Both cohorts were further classified into separate subgroups performing immunohistochemistries for MLH1, MSH2, MSH6, PMS2, and p53 proteins as described before (17, 20).

Statistical analysis of both cohorts comprised descriptive statistics, Spearman correlations and Chi-square tests. As cell counts did not appear as normally distributed in Kolmogorov-Smirnow and Shapiro-Wilks test, Kruskal-Wallis-test, Mann-Whitney-U-tests and non-parametric median comparison were applied. If multiple comparisons were analyzed with Kruskal-Wallis-test the significances were adapted with Bonferroni correction. ROC curve analysis was used to determine cut-offs for diagnostic accuracy for molecular subtypes. Cut-offs for survival analysis were initially based on medians of averaged CD3⁺ and CD8⁺ densities and furthermore determined with the final cut-off. Results were outlined with log-rank tests and Kaplan-Meier curves. The threshold for statistical significance was set at three different layers marked with asterixes *p ≤ 0.05, ^**p ≤ 0.01 and ^***p ≤ 0.001. All calculations were performed with SPSS, version 29.

CD3 and CD8 for conventional DAB-based brightfield IHCs were previously standardized during participation in the Immunoscore project in colorectal cancer at the Institute of Pathology Bern (12). In a first step, we ensured an accurate transfer to an immunofluorescence approach on the testing cohort. The Opal method showed high correlations of lymphocyte cell counts in comparison with conventional immunohistochemistry. A Spearman correlation analysis comparing the two different staining methods showed significant concordance with r_s = 0.90 for CD3⁺ (p < 0.001) and r_s = 0.93 for CD8⁺ (p < 0.001) (Figure 2).

Furthermore, 99.81% of CD8-positive cells were also CD3-positive (CD3⁺CD8⁺). The remaining 0.19% of CD3-negative CD8-positive cells (CD3^–CD8⁺) occured possibly to steric inhibition in multiplexing, but entered our analysis as CD8 positive. Hence, CD8⁺ represents almost completely a matured subgroup of cytotoxic T-cells of the general CD3⁺ population.

Next, we compared CD3⁺ and CD8⁺ cell counts per mm² with histological subtype, T-category, N-category, grading, lymphovascular invasion, hemangioinvasion, age and MELF pattern within the testing cohort. As mentioned above, cores were evaluated in total and the three detailed regions defined as intra-tumoral, close and distant.

In comparison of endometrioid versus non-endometrioid tumors, the testing cohort showed throughout higher lymphocytic counts in non-endometrioid tumors. However, differences in median cell densities for CD3⁺ did not reach significance. Regarding CD8⁺, these differences were evident as CD8⁺ in total areas of endometrioid subtype appeared with median values of 50/mm² versus 121/mm² in non-endometrioid (Mann-Whitney-U-test, p = 0.029). In tendency, this accounted for close areas (61/mm² versus 205/mm²; Mann-Whitney-U-test, p = 0.069) and distant areas with median values of 51/mm² versus 305/mm² (Mann-Whitney-U-test, p = 0.036), as well.

Within the T-category, CD3⁺ and CD8⁺ counts did not show relevant differences. Only in CD8⁺ median densities of distant areas a comparison between pT1a and pT3 cases showed an increase from 39/mm² to 92/mm² (Kruskal-Wallis-test all T-categories p < 0.001 and pairwise median test, p = 0.004).

Within cases with nodal metastasis CD3⁺ densities showed no differences. Elevated medians in CD8⁺ densities were found in tendency in intra-tumoral (p = 0.087, Mann-Whitney-U-test) and significantly in close areas (p = 0.008, Mann-Whitney-U-test) with median values of 62/mm² and 94/mm² in comparison to the corresponding median values of 46/mm² and 51/mm² of the pN0 group, respectively.

In the testing cohort, grading was paralleled with stepwise higher CD3⁺ cell densities, with median values of 137/mm² in G1, 153/mm² in G2 and 196/mm² in G3 tumors (Kruskal-Wallis-Test, p = 0.036), particularly pronounced in distant areas with medians of 140/mm² in G1, 178/mm² in G2 and 259/mm² in G3 tumors (Kruskal-Wallis-Test, p = 0.011). For CD8⁺, advanced grading presented with total cell densities of 39/mm² in G1, 96/mm² in G2 and 209/mm² in G3 tumors (Kruskal-Wallis-Test, p < 0.001), which could be split into differences from 46/mm² to 75/mm² to 100/mm² in close areas (Kruskal-Wallis-Test, p < 0.001), as well as differences from 38/mm² to 60/mm² to 96/mm² in distant areas (Kruskal-Wallis-Test, p < 0.001), respectively.

Lymphovascular invasion showed significantly higher CD8⁺ cell densities in close and distant areas, with medians of 92/mm² and 89/mm² for lymphovascular invasion versus 53/mm² and 46/mm² 143/mm² for L0 cases (Mann-Whitney-U-Test p = 0.031 and p < 0.001, respectively). In hemangioinvasion similar differences were found in cell densities in distant areas with CD3⁺ values of 260/mm² (V1) to 166² (V0) and CD8⁺ values of 94/mm² (V1) to 45/mm² (V0) (Mann-Whitney-U-Test p = 0.004 and p < 0.001, respectively).

No significant dependencies were found between the CD3⁺ and CD8⁺ immune infiltrate and age (Spearman correlation) or MELF pattern (Kruskal-Wallis-Test).

As expected, the differences of CD3⁺ and CD8⁺ cell densities and TCGA groups were most pronounced (Kruskal-Wallis-Test, p < 0.001). In the testing cohort, the highest medians for CD3⁺ and CD8⁺ cell densities in total areas were found in POLE mutated tumors, followed by those with MMRdef. The lowest values were found in NSMP tumors, with p53 mutated tumors in between (Figure 4A).

Comparison of the three compartments showed the highest medians of CD3⁺ cell densities in the vicinity of the tumor, followed by areas distant from the tumor, and the lowest values were found for intra-tumoral areas. These findings were consistent for three TCGA groups, except for p53abn tumors (Kruskal-Wallis-Test, p < 0.001) (Figure 4C).

CD8⁺ cells behaved similarly (Kruskal-Wallis-Test, p < 0.001). Only with regards to the highest cell densities, the intra-tumoral region in POLEmut cases was more pronounced. Tumors with p53 mutation had the lowest values of intra-tumoral CD8⁺ in comparison to the vicinity of the tumor and distant CD8⁺ cells (Figure 4E).

ROC curve analysis of CD3⁺ and CD8⁺ densities across the three regions of intra-tumoral, close, and distant compared with different TCGA subgroups revealed the best discriminative power for intra-tumoral CD8⁺ densities for the distinction of the POLE status (AUC 0.800, p = 0.001). With respect to the opposite data in the NSMP group (AUC 0.325, p < 0.001), intra-tumoral CD8⁺ densities show discriminatory power between POLEmut and POLEwt cases.

To achieve a good predictor for POLEwt cases, the optimal cut-off was an intra-tumoral CD8⁺ density as low as 56/mm² in the testing cohort. For pragmatic applicability, a cut-off of 50 truly intra-tumoral CD8⁺ cells per mm² was further on used. Applying this threshold in our model, it excludes 51.2% (129/252) POLE wild-type cases from sequencing with a sensitivity of 100.0% and a specificity of 53.3%. By contrast, it doubles the positive predictive value in secondary sequencing from native 4.0 to 8.1% in a pre-selected series.

Of note, data from the validation cohort were based on different TMA core size, staining and scanning protocols. A side-to-side comparison to DAB-stained brightfield immunohistochemistry was not accessible. However, the digital pathology script was exerted identically. Thus, the total cell counts per mm², and the three sub-ordinated regions, were defined identically.

With exception of distant areas, the median cell densities for CD3⁺ and CD8⁺ were each significantly higher than in the testing cohort depicted in Figures 4B, D, F (Mann-Whitney-U-test, each p < 0.001 for total areas, intra-tumoral and close). As shown later, this finding affects mainly the specificity of the proposed cut-off.

Regarding the above-mentioned associations of CD3⁺ and CD8⁺ with clinical-pathological parameters, only the TCGA subgroups remained a stable parameter.

The correlation with histological subtype presented with inverse values in comparison to the testing cohort, as endometrioid cases in comparison to non-endometrioid cases showed higher values. This accounts for CD3⁺ with medians of 541/mm² and 115/mm² in comparison to 352/mm² and 7/mm² in close and distant areas, respectively (Mann-Whitney-U-Test p = 0.016 and p < 0.001). For endometrioid versus non-endometrioid subtype, CD8⁺ revealed median values of 272/mm² versus 140/mm² in total, 358/mm² versus 155/mm² in close subregions, as well as 79/mm² versus 2/mm² in distant regions (Mann-Whitney-U-test, p = 0.024, p < 0.001, p < 0.001, respectively).

In the validation cohort, the analysis of grading revealed opposite data again, as low-grade cases showed significantly higher values of CD3⁺ and CD8⁺ counts than high-grade cases. For CD3⁺ this was shown in comparison between G2 and G3 tumors revealing median differences of 552/mm² to 172/mm² in close regions and 794/mm² to 125/mm² in distant regions (Kruskal-Wallis test p = 0.01 and p < 0.001 followed by pairwise median test p = 0.020 and p = 0.010). Regarding CD8⁺, the G1 versus G3 cases showed median values of 355/mm² to 242/mm² in total areas (Kruskal-Wallis test p = 0.048 followed by pairwise median test p = 0.042), 577/mm² to 290/mm² in close regions (Kruskal-Wallis test p < 0.001 followed by pairwise median test p < 0.001), and 400/mm² to 43/mm² in distant regions (Kruskal-Wallis test p < 0.001 followed by pairwise median test p < 0.001), respectively.

Due to the composition of the validation cohort the T-category could only compare between pT1a versus pT1b/pT2 cases, which in intra-tumoral regions showed decreased differences for CD3⁺ with values from 343/mm² to 265/mm² (Mann-Whitney-U test, p = 0.031) and CD8⁺ with values from 253/mm² to 155/mm² (Mann-Whitney-U test, p = 0.031).

The pN-category could not be tested as the validation cohort consisted of nodal negative cases only. Lymphovascular invasion was only in CD8⁺ close areas associated with significantly decreased medians from 356/mm² in L0 cases to 242/mm² in L1 cases (Mann-Whitney-U test, p = 0.048). Hemangioinvasion did not differ in CD3⁺ and CD8⁺ counts. The presence of MELF pattern showed significant decreases from 283/mm² to 176/mm² medians of CD8⁺ in total areas (Mann-Whitney-U-test, p = 0.035), which was even more pronounced in close subregions with a decrease from 385/mm² to 247/mm² (Mann-Whitney-U test, p = 0.016).

In summary, none of the conventional histopathological parameters was confirmed in its correlation with CD3⁺ and CD8⁺ densities between the testing and validation cohort, indicating that molecular subtype is the main influencing factor on lymphocyte infiltrates.

Regarding TCGA, the validation cohort corroborates the previous findings. As depicted in Figures 4A, B, the distribution patterns appear very similar.

Out of the 12 POLE mutated cases in the validation cohort, 11 had an intra-tumoral CD8⁺ mean density of 98/mm² or higher. Hence, the proposed cut-off of 50/mm² CD8⁺ lymphocytes showed a sensitivity of 91.7% in the validation cohort, but a decreased specificity of 13.4% as CD8⁺ counts were generally higher in the validation cohort. Of note, the single POLEmut case as outlier showed exceedingly low values of only 6/mm² in total areas. Combining both cohorts, the cut-off of 50/mm² intra-tumoral CD8⁺ lymphocytes can predict POLEwt cases with a sensitivity of 95.5% and a specificity of 32.3% (Figure 5).

For prognostic analysis both cohorts were combined and tested for conventional histological parameter, TCGA subtypes, medians of CD3⁺ and CD8⁺ densities per region and the above-mentioned cut-off. Information was available for recurrence free survival and overall survival with a follow-up period up to 10 years. As expected, the conventional parameter showed a highly significant stratification of survival data with log-rank test, in detail histological subtype (RFS p < 0.001; OS p = 0.002), T-category, N-category, grading, lymphovascular and hemangioinvasion (each RFS p < 0.001; OS p < 0.001). Only MELF pattern did not contribute to prognostics. These findings are consistent with the previous published data (5, 7, 8, 40).

In a first step, CD3⁺ and CD8⁺ counts were analyzed with the median of each subcategory as cut-off for survival analysis. Best stratification was reached for CD3⁺ intra-tumoral (RFS p = 0.001; OS p < 0.001) and close subregions (RFS p = 0.004; OS p < 0.001), which also accounted for CD8⁺ in intra-tumoral (RFS p = 0.001; OS p < 0.001) and close regions (RFS p = 0.001; OS p < 0.001). Distant regions contributed less to prognostics.

Next the above-mentioned cut-off of 50 CD8⁺ cells/mm² was applied to the survival analysis and compared to the performance of TCGA molecular subtypes (Figure 6). Of note, none of the CD3⁺ or CD8⁺ derived parameters exceeded the excellent prognostics POLEmut status can provide.

The cut-off of 50 intra-tumoral CD8⁺ cells per mm² can be integrated and depicted as a pretest within a clinical decision flowchart prior to POLE mutation analysis (Figure 7) (15).

The clinical strategies for reduction of molecular testing further encompass ESGO low-risk cases (e.g., with low grading and no LVSI) and should encounter the existence of multiple classifier carcinomas with combinations of POLE, MMRdef, and p53abn in up to 3% of cases (41).

In total, the immune infiltrate pretest proposes a POLEwt prior to sequencing in 72 (33.0%) from the original 218 cases. If resources and access to molecular testing are limited, the pretest might bring down healthcare costs by a third and nearly doubles the likelihood for positive sequencing results. Together with clinical-pathological assessment, POLE sequencing can be further minimized to only one-third of all cases in our cohorts.

The recent WHO classification in EC strengthened the role of molecular characterization in ECs, which represents great progress in the understanding of tumor biological and prognostic differences (1). However, in clinical decision making, preselection of cases to submit for further molecular testing is mandatory in terms of healthcare system cost-effectiveness and timely start of radio- and chemotherapy if necessary. The recently introduced ESGO risk classification shows the value added by molecular data, particularly for FIGO I and II staged ECs (3). First algorithms with immunohistochemistry for MMR proteins and p53, in combination with histological subtype and conventional histopathological parameters, can safely reduce the number of molecular tests by nearly half (3, 15). This is, for instance, because very low-risk cases will not need POLE testing to underline a further decreased risk (15).

POLE mutations are rare, with rates of 6–8% in EC, which urges for even more restrictive pretest developments (2). Tumor-associated immune responses are well described for POLE as well as MMRdef cases (6, 7, 20, 42). Inevitably, the use of immune infiltrate hardly distinguishes MMRdef from POLEmut cases. However, low immune infiltrates remain a predictor of POLE wild-type. Our quantitative and spatial analysis of CD3⁺ and CD8⁺ immune cells as a search tool revealed intra-tumoral CD8⁺ cell densities as the strongest discriminator. For usefulness in daily practice, we consider a cut-off of up to 50/mm² intra-tumoral CD8⁺ lymphocytes as the best applicable value to assume a POLE wild-type situation. Below this threshold, only one POLE mutation was detected. Of note, this patient showed no particularities like immunosuppressive therapies, co-morbidities and had so far no recurrence. In combination with the abovementioned clinical-pathological algorithms, the rate of molecular testing could be further reduced to approximately a third of cases (15).

A recent meta-analysis demonstrated a sensitivity of 85.0% and specificity of 66.0% in detecting POLEmut cases by TILs, if the MMR status was known (43). This is in line with our findings that the TIL reaction grows with increasing tumor mutational burden in terms of hyper-mutational type in MMRdef cases and ultra-mutational type in POLEmut cases (5–8, 44–47). Of note, adding more immune cell markers, as recently performed in extended multi-plexing, might even decipher more discriminators (20).

Elevated intra-tumoral cytotoxic CD8⁺ T-cells could indicate a pro-apoptotic reaction that might contribute to the excellent prognosis of POLEmut cases, which is not as pronounced in MMRdef cases (5). Our data imply very basically a biological activation of CD3⁺ cells near a tumor to a recruitment of cytotoxic CD8⁺ cells to the intra-tumoral area, which is most pronounced in POLEmut cases. However, the questions of tumor heterogeneity and other co-founders affecting the immune reaction were not part of this study. Additionally, POLEmut immune induction fits very well to the concept of immune-ablative cancer treatment, but represents a natural course with excellent prognosis even without checkpoint inhibition (47). POLEmut cases with little immune reaction might represent the rare candidates for advanced stages and worse outcome (48). The tumor biological functional roles of TILs in POLEmut cases with respect to checkpoint blockade must be further elucidated. Some studies found higher stromal TILs than intraepithelial, and one study found higher TILs in MMRdef tumors compared to POLE (7, 8, 49), which can be explained by the similarities of ultra- and hyper-mutated cancers with possibly similar tumor mutational burden.

Relating to conventional prognosticators, no consistencies were found in comparison between histological subtype, grading, and TNM-classification, including LVI status. In tendency, a positive correlation between grading and TILs has been described but might be influenced by the existence of serous or clear cell carcinomas (40, 49–51).

The prognostic effect of TILs has been widely discussed as pro (49, 51, 52), and contra (7, 8, 40). Since there is a positive correlation between POLE mutated tumors and a higher CD8⁺ intra-tumoral immune infiltrate (5–8, 14, 44–46), it would be logical that their prognostic significance might be dependent on case numbers of POLE mutated ECs. This has recently been shown by a meta-analysis of almost two thousand EC cases (53). Additionally, our data elucidate a stronger role of CD3⁺ and CD8⁺ cell densities in intra-tumoral and close subregions on prognosis. Hence, more precise definitions and distinctions of intra-tumoral tumor infiltrating lymphocytes (iTILs) from stromal tumor infiltrating lymphocytes (sTILs) are warranted.

POLE sequencing remains the gold standard for appropriate molecular subtyping as outliers based on immune cell densities like in our study occur (47). Together with the validation cohort, we found less than five per cent POLEmut cases, which represents a low rate of POLEmut according to literature (6–8, 39, 49). If real-life rates of POLEmut remain as low, increased pretest probabilities might justify POLE sequencing costs and expand its use in healthcare systems with more limited resources. Additionally, risk factors justifying molecular subtyping might only be evident after surgery and timely indication for radio- or chemotherapy might be compromised, if molecular analysis is delayed. In such situations conventional pretests might point toward therapeutic options until sequencing information is available.

Another limitation is the complexity of multiplexed immunofluorescence, which is not yet used as a standard staining technique for daily diagnostic usage.

The differences between both cohorts in cell counts and results for histological subtype and grading could only partially be explained by the clinical-pathological differences, and are also attributable to the different staining protocols and scanning conditions. This is reflected in the available literature as well. Different studies show great ranges in intra-tumoral and stromal CD8⁺ densities between 14/mm² and 650/mm² (Supplementary material) (5, 6, 8, 9, 52, 54). To properly apply the immune infiltrate pretest, it would be necessary to harmonize staining conditions.

A strength is the high correlation of the testing cohort immune infiltrate counts between the fluorescence method and the gold standard; conventional DAB staining validated for the Immunoscore. As a consequence, we maintained the cut-off from the testing cohort as a reference point. However, the variability in immunohistochemistry between laboratories might need round-robin tests or even laboratory specific reference values from other biomarkers (20). Additionally, our proposed cut-off needs validation in clinical trials before clinical usage.