Abstract 5444: Integrated analysis of gene expression signatures and AI-based detection of tertiary lymphoid structures and stromal tumor-infiltrating lymphocytes in breast cancer H&E samples

Abstract Background: Transcriptomic and morphological features of the tumor microenvironment (TME) are associated with favorable prognosis and can serve as surrogate markers for immune checkpoint inhibitor response in certain breast cancer subtypes. Tertiary lymphoid structures (TLS) and stromal tumor-infiltrating lymphocytes (sTILs) are integral components of the TME. New techniques for TLS characterization could improve reproducibility and automate the current workflow. Here, we correlated AI-based imaging platform predictions of TLS and sTILs on breast cancer H&E whole slide images (WSIs) with expression data. Design: WSIs from 390 TCGA breast cancer samples were used to detect normalized tumoral TLS and tumor stroma infiltration areas by a convolutional neural network (CNN). High and low groups were defined by IQR1 and IQR3. Gene counts were compared using differential expression analysis followed by gene set enrichment analysis (GSEA) with hallmark gene sets. Kassandra's cell type percentage predictions, functional gene signature (Fges) single-sample GSEA, and PROGENy pathway activity median-scaled scores were compared with the Mann-Whitney U test. The Fges consisted of 4 TLS and 29 TME signatures and were clustered into 4 groups: fibrotic, myeloid, T cell, and TLS. Group medians were analyzed. Somatic deep amplifications, deletions, and missense mutations of cancer pathways were compared with Fisher’s exact test. The FDR method was used to adjust p-values. Results: FDCSP, CXCL13, CD79A, CCL19, and IGLL5 were upregulated in the TLS-high group (logFC > 1.5, p < 0.003), and CXCL13, CXCL9, CCL19, CD3E, and CD79A were upregulated in the sTIL-high group (logFC > 2, p < 4e-9). GSEA identified interferon γ and α response and allograft rejection signatures in TLS- and sTIL-high groups (p = 0.006, NES > |2.5|). sTIL-high samples expressed high levels of E2F targets, while the sTIL-low group was enriched with myogenesis and epithelial-mesenchymal transition gene sets. The TLS-high group was enriched in the TLS Fges cluster (p = 1e-5) and contained high B, CD4+ T, and CD8+ T cell content (p < 2e-4). sTIL samples expressed high levels of B, CD4+ T, CD8+ T, and myeloid cells (p < 0.001) and TLS and myeloid clusters (p < 1e-5). TGFβ and VEGF pathways were more active in sTIL-low samples (p < 0.001), while JAK-STAT was most active in the sTIL-high group (p = 0.0006). In the basal subtype (p < 0.001), the TLS-high group was associated with NOTCH2 amplification; the sTIL-low group had increased NRG1 deletions; and the sTIL-high group contained BRCA2 mutations. Conclusions: RNA-based methods used were concordant with AI-based methods of TLS and sTIL detection. Our CNN-based method of identifying immune structures in the TME with morphological features on H&E WSIs may automate traditional pathology workflows with additional validation. Citation Format: Nadezhda Lukashevich, Vladimir Kushnarev, Daniil Dymov, Anastasia Zotova, Anna Belozerova, Ivan Valiev, Lev Popyvanov, Anna M. Love, Nathan Fowler, Alexander Bagaev, Ekaterina Postovalova. Integrated analysis of gene expression signatures and AI-based detection of tertiary lymphoid structures and stromal tumor-infiltrating lymphocytes in breast cancer H&E samples. [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 5444.