Abstract 5446: Immune phenotypes classified by deep learning-based H&E tissue analyzer demonstrate distinct immune landscape and transcriptomic features in ovarian cancer

Abstract

Abstract Introduction: Deep learning-based H&E slide analyzer, Lunit-SCOPE IO can classify the tumor microenvironment as three immune phenotypes (IPs): inflamed, immune-excluded and immune-desert. In our previous study, the IPs demonstrated distinct survival outcome and immunologic landscape in endometrial cancer. To further explore the application of IPs in other type of cancer, we applied Lunit-SCOPE in ovarian cancer (OV) and compared tumor microenvironments based on the IPs. Methods: H&E slide images, RNA-sequencing data, and clinical data were obtained from TCGA OV cohort. Lunit-SCOPE IO is a H&E slide analyzer using deep learning algorithm and was trained with 1,824 H&E slides of advanced NSCLC. We used the analyzer to classify H&E slide images from TCGA OV cohort to three IPs. Cibersort was used to assess immune cell infiltration across the IPs. The cytolytic activity score was obtained from geometric mean of GZMA and PRF1 RNA expression. We further assessed transcriptomic characteristics across IPs via differentially expressed genes analysis and gene set enrichment analysis (GSEA). Additionally, we compared CD274 gene expression across IPs. Finally, survival analysis was performed to evaluate survival outcome based on IPs. We used Mann-Whitney method to compare non-parametric variables between the IPs. Results: The proportion of inflamed, immune-excluded and immune-desert group in TCGA OV were 7 (10.2%), 13 (18.8%), and 49 (71%), respectively. Inflamed group showed significantly higher proportion of M1 macrophage compared with immune-desert group (6.6% vs. 2%, P=.012). We observed higher cytolytic activity score in inflamed group compared to immune-desert group (7.21 vs. 5.27, P=.016). In GSEA, inflamed group is significantly enriched in IL6 JAK STAT3 signaling, interferon gamma response, KRAS signaling up, and inflammatory response pathway (NES 2.19; adjusted P=.021, NES 1.70; adjusted P=.021, NES 1.62; adjusted P=.037, and NES 1.52; adjusted P=.037, respectively). Immune-desert is related to upregulated MYC target V2 pathway (NES -1.95; adjusted P=.037). Inflamed group did not show significant difference in CD274 expression, compared with others (Inflamed vs. immune-excluded, 4.97 vs. 4.19, P=.115 and Inflamed vs. immune-desert, 4.97 vs. 3.95, P=.058). Inflamed group did not show any significant difference in overall survival, compared with non-inflamed group (Inflamed vs. non-inflamed, HR 0.98; 95% CI 0.45 - 2.14; P=.958). Conclusion: Deep learning-based H&E slide analyzer can be used to classify tumor H&E slide images with OV into three IPs. In this study, the classification demonstrated distinct immunologic traits and transcriptomic characteristics in tumor microenvironment. The enrichment of interferon gamma response pathway in inflamed OV group may indicate opportunities for targeted immunotherapy. Citation Format: Horyun Choi, Youjin Oh, Chan-Young Ock, Kyunghyun Paeng, Young Kwang Chae. Immune phenotypes classified by deep learning-based H&E tissue analyzer demonstrate distinct immune landscape and transcriptomic features in ovarian cancer. [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2023; Part 1 (Regular and Invited Abstracts); 2023 Apr 14-19; Orlando, FL. Philadelphia (PA): AACR; Cancer Res 2023;83(7_Suppl):Abstract nr 5446.