The immune system can be manipulated by cancer cells to avoid destruction. Two key proteins involved in this process are PD-1 (programmed cell death protein 1) and PD-L1 (programmed cell death ligand 1). PD-1 sits on the surface of immune cells, particularly T cells, and acts as a brake on immune activity. When cancer cells express PD-L1, they can bind to PD-1 and essentially switch off the immune attack against them.
Drugs called immune checkpoint inhibitors (ICIs) - such as pembrolizumab - work by blocking this PD-1/PD-L1 interaction, allowing the immune system to recognize and destroy cancer cells again. These drugs have transformed treatment for many cancers, including endometrial cancer (EC).
However, a key challenge remains: there is no universally agreed-upon method for measuring PD-1 and PD-L1 expression in tumor tissue. Different labs use different scoring systems, which can lead to inconsistent treatment decisions. This study sought to compare the traditional approach (pathologists visually examining stained tissue slides) with a newer machine learning approach to assess PD-1/PD-L1 expression in endometrial cancer.
The researchers studied 60 patients with early-stage endometrial cancer (specifically, endometrioid carcinoma) who underwent surgery at Gyeongsang National University Hospital in South Korea between 2002 and 2009. The patients had a mean age of 51 years, and the majority had T1-stage disease, meaning the cancer was largely confined to the uterus.
Tissue samples were arranged into a tissue microarray (TMA) - a technique where small cores from multiple tumor samples are embedded in a single block for simultaneous staining. Each core was stained with antibodies targeting PD-1 and PD-L1, allowing researchers to see where and how strongly these proteins were expressed.
Two pathologists visually scored each core, separately assessing expression in tumor cells (TC) and immune cells (IC). The same stained slides were then digitally scanned and analyzed using Genie, a machine learning morphometric analysis tool developed by Leica Biosystems. Genie measured the proportion of positive cells across the entire core and counted positive cells per square millimeter. Statistical comparisons were made using Fisher's exact and chi-square tests.
PD-1 expression was found almost exclusively in immune cells rather than tumor cells: only 1 out of 60 cores showed PD-1 positivity in tumor cells, while 23 cores (38%) had high PD-1 expression in immune cells. PD-L1 showed more expression in both cell types, with 10 cores positive in tumor cells and 26 cores (43.3%) showing high expression in immune cells.
The most important finding was that both pathologists and Genie independently identified a statistically significant relationship between PD-1 expression and tumor size. Pathologist-scored PD-1 in immune cells correlated with tumor size (p=0.026), and Genie-measured PD-1 proportion correlated similarly (p=0.033). This agreement between human and machine methods is clinically significant.
Genie also found an association between PD-1 expression and histological grade (p=0.028), a measure of how abnormal cancer cells appear under the microscope. Pathologists found that PDL1 expression in immune cells correlated with histological grade (p=0.003). There were some differences between the two methods in other measurements, but the core finding - PD-1 expression tracking with tumor size - was confirmed by both approaches.
The finding that pathologists and Genie reached similar conclusions about PD-1 expression is important because machine learning analysis offers several practical advantages. It can reduce the workload of pathologists, decrease inconsistencies between different readers and institutions, and potentially provide more objective, reproducible measurements across different laboratories.
Current clinical guidelines recommend PD-1/PD-L1 testing to guide immunotherapy decisions in endometrial cancer. However, there is no single standardized method - three different scoring systems (TPS, CPS, and IC) are used for different cancer types, and no unified approach covers all tumors. This study suggests that machine learning could help standardize these assessments.
One noted limitation of the Genie method is the difficulty in separately training it to distinguish PD-L1 or PD-1 staining patterns across different cancer cell types. The authors suggest that with more training data and algorithm refinement, Genie could be extended to predict prognosis from small biopsy samples taken before surgery, which would be particularly useful for patients with inoperable tumors.
Pembrolizumab is currently recommended for endometrial cancer patients whose tumors have high microsatellite instability (MSI-H) or high tumor mutational burden (TMB-H). The ability to reliably measure PD-1 and PD-L1 expression adds another layer of information that could help identify patients who will benefit most from immunotherapy.
The fact that PD-1 expression tracked with tumor size suggests it may reflect the immune response to tumor burden - larger tumors may attract more immune cells, which upregulate PD-1. This relationship could have prognostic value, as PD-1 expression may indicate the degree of immune engagement with the tumor.
The authors propose that the Genie-based approach could eventually be applied to small curettage biopsies - tissue samples taken without surgery - to assess immune status before deciding on treatment. This would be particularly valuable for patients who cannot undergo surgery, allowing immunotherapy decisions to be made from minimal tissue samples.
This study represents the first validation of machine learning morphometric analysis for assessing PD-1 expression in endometrial cancer tumor microenvironments. Both pathologists and the Genie algorithm identified the same significant association between PD-1 expression in immune cells and tumor size, demonstrating that automated analysis can replicate expert human evaluation.
The study enrolled a relatively small cohort of 60 patients at a single institution, which is an acknowledged limitation. The results require validation in larger, multi-center studies with more contemporary patient cohorts. Additionally, the Genie method cannot yet separately analyze tumor and immune cell populations for PD-L1, which limits its full applicability.
Despite these limitations, this work establishes a proof of concept that machine learning can reliably assist in evaluating immune checkpoint biomarkers. As more training data becomes available and algorithms improve, AI-assisted pathology could play an increasingly important role in guiding personalized cancer immunotherapy in endometrial and other gynecological cancers.