Artificial intelligence (AI) –powered H&E whole-slide image (WSI) analysis of tertiary lymphoid structure (TLS) correlates with immune phenotype and related molecular signatures in non–small-cell lung cancer.

Abstract

e20520 Background: Recent data suggested immune infiltrates in TLS around tumor bed may be a favorable predictive factor for immunotherapy in various cancers. TLS have been suggested to be correlated with inflamed immune phenotype and relevant inflammatory gene signatures of adaptive immunity. To analyze this relationship, here, we developed an AI model to assess TLS objectively in H&E WSI, and assessed its correlation with immune phenotype and immunologic signatures. Methods: H&E images, relevant gene expression profiles and clinical data from The Cancer Genome Atlas (TCGA) lung cancer dataset (LUAD and LUSC, N = 913) were used for the analysis. Lunit SCOPE TLS, an AI-powered H&E WSI analyzer, was developed with 3.59 x 109 μm2 annotated area and 1,439 H&E stained WSI of 18 cancer types to segment TLS within tumor microenvironment (TME). Based on spatial tumor-infiltrating lymphocyte (TIL) density, IPs were classified into inflamed IP (IIP) as high intratumoral TIL (iTIL), immune-excluded IP (IEP) as low iTIL and high sTIL, and immune-desert IP (IDP) as low TIL overall. The infiltration of immune cells, the activity of related pathways and biological processes, and the signature scores of interest gene sets were analyzed by using the CIBERSORT, DESeq2 and GSEA tools. Results: Of 913 samples, 69.7% (636/913) contain TLS. The median value of the proportion of TLS area within TME was 0.12%, which was applied for the cut point of TLS-high vs -low group. The proportion of TLS-high was not significantly different according to EGFR mutation (mutation vs wild type: 45.5% vs 50.2%), KRAS mutation (49.3% vs 50%), but numerically decreased in the merging set of other driver mutations including ALK, ROS1, RET, MET, NTRK1-3 (36% vs 50.3%, p = 0.223). Interestingly, TLS-high proportion was significantly different according to immune phenotype, as TLS were present in 57.6% of IIP (186/323), 49.8% of IEP (241/484), and 28.3% of IDP (30/106, p < 0.001). TLS-high group was positively correlated with memory B cells (fold change [fc] 1.93, p < 0.001), CD8+ T cells (fc 1.20, p < 0.001), and M1 Macrophages (fc 1.17, p < 0.001), and negatively correlated with neutrophils (fc 0.59, p < 0.001) and M2 macrophages (fc 0.91, p = 0.006). This result was additionally supported by GSEA analysis of GO:BP or Hallmark gene sets which showed TLS-high was associated with B cell receptor signaling pathway (normalized enrichment score [NES] 3.16, p = 0.001), immunoglobulin production (NES 3.07, p = 0.002), and interferon-gamma response (NES 2.20, p = 0.002). Epithelial-mesenchymal transition was negatively associated with TLS-high (NES -1.54, p < 0.001). Conclusions: TLS is associated with inflamed immune phenotype and infiltration of memory B cells as well as CD8+ T cells in non-small cell lung cancer.