Abstract 5131: Deep learning-based tumor microenvironment cell types mapping from H&E images of lung adenocarcinoma using spatial transcriptomic data

Abstract

Abstract Purpose Spatial characterization of cell types of the tumor microenvironment (TME) is a key to finding new targets as well as developing biomarkers for immuno-oncology treatment. Here, we develop and validate a deep learning model trained by integrating H&E images of lung adenocarcinoma and spatial transcriptomic data to decipher spatial mapping of multiple cell types in TME only using H&E images. Methods A total of 21 H&E images combined with spatial transcriptomic data of lung adenocarcinoma were used to develop a model. As each spot of spatial transcriptomic data consisted of a few cells, the cell types were inferred by CellDART, a domain adaptation-based method to estimating five cell types defined by single cell RNA-seq (scRNA-seq) data - B, T/NK, myeloid cells, fibroblasts, and epithelial cells. A convolutional neural network using H&E image patches as inputs was developed to predict the cell type scores for each corresponding spot. For the external validation, the model estimated the cell type scores from H&E-stained tissue image patches of lung adenocarcinoma of the Cancer Genome Atlas (TCGA-LUAD). Furthermore, the model inferred cell types from H&E images of lung adenocarcinoma tissues acquired by different patients and the results were compared with immunohistochemistry. Results The cell types inferred by the model using H&E image patches were significantly correlated with those derived by spatial transcriptomic data as an internal validation. The mean value of the deep learning-based cell types scores estimated by the TCGA-LUAD tissue images was significantly correlated with the cell type scores estimated by bulk RNA-seq data of corresponding TCGA data. When cell type inference maps of independent lung adenocarcinoma H&E images were spatially co-registered with immunohistochemistry slides, including CD3, CD20, and CD68, pixel-wise correlation analysis revealed positive significant correlations. Conclusions A deep learning model to infer spatial cell types of tumor microenvironment using H&E images was developed. The results of cell type scores were positively correlated with bulk RNA-seq-based immune cell enrichment scores and pixelwise analyses on immunohistochemistry images using typical cell type markers. This approach can provide objective and flexible deep learning-based models for characterizing tumor microenvironment. Citation Format: Hongyoon Choi, Kwon Joong Na, Jaemoon Koh, Young Tae Kim. Deep learning-based tumor microenvironment cell types mapping from H&E images of lung adenocarcinoma using spatial transcriptomic data [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2022; 2022 Apr 8-13. Philadelphia (PA): AACR; Cancer Res 2022;82(12_Suppl):Abstract nr 5131.