The association between the tumor immune microenvironments and clinical outcome in low‐grade, early‐stage endometrial cancer patients

Abstract

Endometrial tumors show substantial heterogeneity in their immune microenvironment. This heterogeneity could be used to improve the accuracy of current outcome prediction tools. We assessed the immune microenvironment of 235 patients diagnosed with low-grade, early-stage endometrial cancer. Multiplex quantitative immunofluorescence was carried out to measure CD8, CD68, FOXP3, PD-1, and PD-L1 markers, as well as cytokeratin (CK), on tissue microarrays. Clustering results revealed five robust immune response patterns, each associated with specific immune populations, cell phenotypes, and cell spatial clustering. Most samples (69%) belonged to the immune-desert subtype, characterized by low immune cell densities. Tumor-infiltrating lymphocyte (TIL)-rich samples (4%) displayed high CD8 + T-cell infiltration, as well as a high percentage of CD8/PD-1 + cells. Immune-exclusion samples (19%) displayed the lowest CD8 + infiltration combined with high PD-L1 expression levels in CK + tumor cells. In addition, they demonstrated high tumor cell spatial clustering as well as increased spatial proximity of CD8 + /PD-1 + and CK/PD-L1 + cells. FOXP3 and macrophage-rich phenotypes (3% and 4% of total samples) displayed relatively high levels of FOXP3 + regulatory T-cells and CD68 + macrophages, respectively. These phenotypes correlated with clinical outcomes, with immune-exclusion tumors showing an association with tumor relapse. When compared with prediction models built using routine pathological variables, models optimized with immune variables showed increased outcome prediction capacity (AUC = 0.89 versus 0.78) and stratification potential. The improved prediction capacity was independent of mismatch repair protein status and adjuvant radiotherapy treatment. Further, immunofluorescence results could be partially recapitulated using single-marker immunohistochemistry (IHC) performed on whole tissue sections. TIL-rich tumors demonstrated increased CD8 + T-cells by IHC, while immune-exclusion tumors displayed a lack of CD8 + T-cells and frequent expression of PD-L1 in tumor cells. Our results demonstrate the capability of the immune microenvironment to improve standard prediction tools in low-grade, early-stage endometrial carcinomas.

Weakly supervised deep learning to predict recurrence in low-grade endometrial cancer from multiplexed immunofluorescence images

“…This suggests that the primary tumors were well represented in the TMAs. In a subset of these patients, we have previously shown a very good correlation between the immunofluorescence results of TMAs and whole-slide tumor sections 22,27 .…”

“…Despite the prognostic value of immune infiltration in endometrial cancer, relatively little is known about whether these immune features could accurately be associated with the risk of recurrence in patients with the low-grade disease. Recently, we have shown that the immune infiltration features in a primary tumor resection studied with multiplexed IF (PD-L1, PD-1, CD8, CD68, FOXP3, and CK) could reliably predict recurrence in patients with low-grade, early-stage endometrial cancer, outperforming the molecular subtypes 22 . Here, we have confirmed this.…”

“…Although several studies have shown the prognostic value of immune cell subtypes in high-grade endometrial cancer, particularly T cells and macrophages, there is still no evidence that a single marker alone could be used to improve risk stratification in patients with lowgrade, early-stage disease 20,21 . Using a comprehensive multiplexed IF panel, we have recently shown that the tumor-immune microenvironment milieu is the most important factor for predicting tumor recurrence in low-grade, early-stage endometrial cancer, outperforming currently used prognostic variables, including molecular subtypes 22 .…”

“…1 Here, we have expanded our analysis from 235 patients to 250 patients with long-term follow-up and 32 tumor recurrences 22 . Our previous model was developed based on algorithms trained by experienced pathologists.…”

“…The novelty here comes from the artificial intelligence nature of the approach. The WSDL model was developed using (i) our previously validated panel of seven markers (PD-L1, PD-1, CD8, CD68, FOXP3, CK, and DAPI) 22 , (ii) NaroNet, designed to extract the complex tumor-immune interrelations at multilevel features (local phenotypes, cellular neighborhoods, and tissue areas) using unsorted multiplexed IF images (in the current approach, no images were excluded from the analyses, while previously only images with a minimum tumor content of 20% were analyzed), and (iii) a tenfold cross-validation approach (in our previous approach, all samples were used to train the model). Lastly, we evaluated the robustness of the model and the power of NaroNet to overcome technical variabilities.…”