Abstract 171: A novel single-cell level approach integrating artificial intelligence (AI)-powered histomorphology labeling and spatial transcriptomics enables biomarker identification of treatment-resistance in salivary gland cancer (SGC)

Introduction: We integrated H&E-AI analysis and spatial transcriptomics (ST) to provide a synergistic analysis of tumor microenvironments (TME) beyond each technology alone. In this study, we focused on SGC, which encompasses several rare tumor types with unmet need. Methods: We developed an integrated analysis pipeline to combine cell- and tissue-level morphological information from H&E-AI (Lunit SCOPE IO, AI-powered H&E image analyzer) and gene expression profiles from ST (10x Genomics Xenium). We then analyzed 12 tumor samples provided by 10x Genomics for validation and, as a test set, 16 surgically resected SGC including 11 salivary duct carcinoma (SDC) treated with neoadjuvant nivolumab and cetuximab. Results: We first reciprocally validated H&E-AI and ST by confirming that ST captures well established cell-specific gene expression that corresponds to AI-predicted cell types (EPCAM, VWF, and TRAC mark carcinoma cells, endothelial cells, and lymphocytes with F1 = 0.88, 0.81, and 0.72, respectively). In the test SGC cohort (total 915,215 cells), we identified 20 cell type groups with distinct gene expression patterns. Due to high gene expression heterogeneity among putative tumor cell subtypes, ST based tumor cell detection using reference gene expression profiles was insufficient and required complementary morphological analysis from H&E-AI, which identified 10 groups as tumor cell subtypes harboring diverse morphologies (mean nuclear size per subtype: 27.14 - 40.92 µm2). Two patients with SDC had the same dominant tumor subtype (largest nucleus size, HER2/FOXA1/GATA3 positive), yet exhibited discordant responses to neoadjuvant therapy. Tumor cells from the SDC patient with recurrence (SDCw/Rec), compared to those from the SDC patient without recurrence (SDCw/oRec), exhibited upregulation of genes associated with immune evasion or epithelial-to-mesenchymal transition (EMT) such as MYC, TNC, and JUN (174%, 197%, 77% higher). Moreover, SDCw/Rec also had significantly fewer tumor-infiltrating lymphocytes (TIL) than SDCw/oRec (515 vs. 2,306 /mm2). Interestingly, the proportion of TIL with CXCL9 expression was far lower in SDCw/Rec [19.7% (1,636/8,312)] compared to SDCw/oRec [67.0% (25,658/38,267)]. Conclusion: An integrated ST and H&E-AI analysis allowed parallel transcriptomic and subcellular-level histomorphologic characterization of individual cells from tumor specimens at the micron scale, providing a highly visualizable solution for discovering novel resistance mechanisms. Based on this novel approach, we reported a SDC pair discrepant for recurrence after neoadjuvant therapy, where the recurred case exhibited 1) high intrinsic tumor signatures of EMT and immune evasion and 2) a low level of TILs along with low CXCL9 expression in each TIL. Citation Format: Jin Woo Oh, Sanghoon Song, Soohyun Hwang, Mingu Kang, Jinhee Lee, Sérgio Pereira, Hyunsu Kim, Sehhoon Park, Junhun Cho, Se-Hoon Lee, Han-Sin Jeong, Chang Ho Ahn, Siraj Ali, Chan-Young Ock, Myung-Ju Ahn. A novel single-cell level approach integrating artificial intelligence (AI)-powered histomorphology labeling and spatial transcriptomics enables biomarker identification of treatment-resistance in salivary gland cancer (SGC) [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2025; Part 1 (Regular Abstracts); 2025 Apr 25-30; Chicago, IL. Philadelphia (PA): AACR; Cancer Res 2025;85(8_Suppl_1):Abstract nr 171.