Abstract 2095: Automated mapping and analysis of stromal cells in tumor microenvironment in pancreatic adenocarcinoma andcholangiocarcinoma using deep learning

Abstract

Abstract Objective: A characteristic histologic feature of pancreatic adenocarcinoma (PAD) and cholangiocarcinoma (COL) is extensive desmoplasia (DP) alongside leukocytes and stromal cells (SCs). Advances in mRNA-sequencing have enhanced our understanding of cancer biology in relation to selective changes in expressivity amongst SCs. DP changes secondary to aberrant expression in SCs creates a barrier to absorption and penetration of therapeutic drugs, but few models exist to analyze the spatial and architectural elements composing the complex tumor microenvironment (TME) in association with mRNA levels. Methods: The histopathology images (H&E stain) and mRNA-seq of 178 PAD and 36 COL patients (pts) were obtained from the Cancer Genome Atlas (TCGA) and analyzed with the deep learning (DL) algorithm, which characterizes histological features in comparison to the corresponding mRNA-seq, allowing for rapid automated analysis of large quantities of data. Ninety genes enriched in leukocytes (CD8+ T cells, B cells, CD4+ regulatory T cells, macrophages, neutrophils, NK cells, and plasmacytoid dendritic cells), 7 genes for cytolytic activities (GZMA, PRF1, GZMH, GZMK, NKG7, CD3E, and CD247), and 5 genes involved with fibroblastic and DP changes (PDFGRA, ACTA2, COL1A1, COL1A2, and PDPN) were analyzed. For each pt, mRNA levels of select genes were analyzed against histologic features, including degree of DP reaction, number of leukocytes, and degree of leukocyte clustering and isolation from tumor cells. Results: DL analysis demonstrates that the number of fibroblasts and degree of DP correlates with and predicts the mRNA expression of genes associated with fibroblastic and DP changes. The mRNA level of CXCL12 correlates with the degree of leukocyte clustering and spatial isolation in PAD and COL. The 5 genes associated with DP and fibrosis do not have a linear relationship with CXCL12 mRNA levels (R2 <0.1) in COL and (R2=0.2196 to 0.6279) in PAD. Cytolytic activity, measured by the mRNA levels from 7 genes, does not correlate with CXCL12 expression (R2 <0.1) in COL, and (R2=0.3530 to 0.6060) in PAD. Conclusion: A DL model enables automated analysis and mapping of DP changes within stromal and malignant cells, revealing the spatial and architectural relationship in the TME with varying gene expression. This demonstrates that the degree of leukocyte clustering and isolation from tumor cells correlates with CXCL12 mRNA levels in PAD and COL. CXCL12 expressivity appears to be a contributing factor, limiting access of leukocytes to tumor cells and diminishing an important mechanism combating tumor progression. Varying degrees of DP and cytolytic activities of immune cells within the TME were also observed in association with CXCL12 expression in PAD and COL. Further biomarker-driven prospective studies in the context of immunotherapy and anti-fibrosis are warranted. Citation Format: Sunyoung S. Lee, Jin Cheon Kim, Seongwon Lee, Jilliam Dolan, Andrew Baird. Automated mapping and analysis of stromal cells in tumor microenvironment in pancreatic adenocarcinoma andcholangiocarcinoma using deep learning [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2018; 2018 Apr 14-18; Chicago, IL. Philadelphia (PA): AACR; Cancer Res 2018;78(13 Suppl):Abstract nr 2095.