Abstract P6-04-15: AI-based prediction of tertiary lymphoid structures and lymphocyte immune infiltration in breast carcinomas

Abstract

Abstract Introduction Tertiary lymphoid structures (TLSs) and tumor-infiltrating lymphocytes (TILs) in breast carcinomas are prognostic for survival and predictive of certain therapy responses. The presence of TLSs and TILs are identified by manual pathological examination; however, this method often lacks reproducibility, limiting its use in routine clinical practice. Here, we demonstrate that morphological evaluation of whole slide images (WSIs) using an artificial intelligence (AI)-based analytic workflow comprised of convolutional neural network (CNN) deep learning models that accurately and reproducibly characterizes TILs, measured as the lymphocyte immune-infiltrated area (LIIA), and TLSs in the tumor microenvironment (TME) of breast carcinomas. Methods We collected a cohort of 445 TCGA breast cancer H&E WSIs, including clinical and sequencing data, and divided this cohort into luminal invasive lobular carcinoma (ILC) (n = 192), HER2-enriched (n = 110), and basal-like (n = 143) molecular subtypes. After 55 samples were excluded due to artifacts or incomplete clinical annotation, a total of 390 samples were analyzed. A combination of CNN-based deep learning models was used to detect and classify the tumor area, TLSs present in the TME, TLS density (number of TLS per mm2 of tumor), and lymphocyte-rich regions. The LIIA was calculated as the area of the stromal and TIL components of the TME. Validation was performed by manually annotating 10 random WSIs from the dataset. Spatial model predictions of the tumor and TLSs were combined to identify TLS locations. Each model’s predictions were verified by univariate (Kaplan-Meier) and multivariate (Cox regression) survival analyses, and the log-rank test was used to calculate overall survival. Additionally, the relationship between TLSs and LIIAs with CD274 expression (PD-L1) and a high tumor mutational burden (TMB > 10) was analyzed. Statistical analyses included Spearman’s rank correlation and Mann-Whitney tests. Results TLS were detected in 53% (n = 207) of the samples, with a mean density of 26.02 TLS/mm2 (Q3 = 5.53 TLS/mm2). TLS density was higher in basal-like subtype samples compared to luminal and HER2-enriched subtypes. While LIIA and TMB-high samples exhibited a significant relationship (p = 0.00001), no significant association was found between TME and TLS quantities or density. PD-L1 gene expression exhibited weak to moderate correlations with predicted LIIA in basal-like (r = 0.38, p = 0.00001) and HER2-enriched subtypes (r = 0.38, p = 0.0001). The luminal subtype had no significant correlation between PD-L1 expression and predicted LIIA. As a result, LIIA and TLS were characterized as positive prognostic factors for the basal-like subtype. After adjusting for age, stage, and grade, the LIIA and TLS density were found to be significant independent positive prognostic overall survival factors for the basal-like subtype (LIIA HR: 0.02, p = 0.003; TLS-high group HR: 0.09, p = 0.002). For the HER2-enriched subtype, TLS density was also a significant predictor (HR: 0.05, p = 0.035), while LIIA was not a statistically significant prognostic factor (HR: 0.0002, p = 0.08). Associations were not observed between the TLSs and LIIA between the ILC subtypes and survival outcomes. The same result was observed for univariate analyses. Conclusion The developed analytic pipeline accurately identified the presence of LIIA and TLS on H&E slides, demonstrating the potential of CNN for automated characterization of the breast cancer TME. AI-based TLS and LIIA quantification can be a robust tool for pathology processes, offering additional information to help in clinical decision-making. This approach can be used to detect features of immune morphology biomarkers in other cancer types. Citation Format: Vladimir Kushnarev, Daniil Dymov, Nadezhda Lukashevich, Lev Popyvanov, Anna Belozerova, Diana Shamsutdinova, Aida Akaeva, Yury Popov, Svetlana Khorkova, Ivan Valiev, Anastasia Zotova, Jessica H. Brown, Anna Love, Alexander Bagaev, Ekaterina Postovalova, Nathan Fowler. AI-based prediction of tertiary lymphoid structures and lymphocyte immune infiltration in breast carcinomas [abstract]. In: Proceedings of the 2022 San Antonio Breast Cancer Symposium; 2022 Dec 6-10; San Antonio, TX. Philadelphia (PA): AACR; Cancer Res 2023;83(5 Suppl):Abstract nr P6-04-15.