Abstract 7398: Unveiling the tumor microenvironment of hepatocellular carcinoma using AI trained by spatial transcriptomics: A preliminary study to predict response to immunotherapy

Abstract

Abstract Background. Immune checkpoint inhibitors (ICIs), particularly targeting the PD-1 pathway, are promising in treating hepatocellular carcinoma (HCC). However, their variable effectiveness among individuals calls for a better understanding of the tumor microenvironment (TME) and reliable predictive biomarkers. Here, we employ spatial transcriptomics (ST) to develop a deep-learning model that aims to investigate the TME in HCC using Hematoxylin and eosin (H&E) staining images. Our focus is on identifying cell types within the TME that are linked to the response to ICI in a cohort of patients treated with nivolumab. Methods. Our study leveraged 21 Visium ST datasets from HCC to develop a model aimed at inferring the cellular composition of TME from H&E images that matched with ST data. The comprehensive TME cell types enrichment score, encompassing Tumor Endothelial Cells (TECs), Tumor-Associated Macrophages (TAMs), B cells, T cells, and Cancer-Associated Fibroblasts (CAFs) in each spot, was calculated by the CellDART algorithm, a method that maps cell type by integrating with a reference scRNA-seq data. The model was trained in a patch-wise manner, utilizing a pre-trained ResNetRS50 as the backbone model. For the internal validation, Spearman correlation coefficients were evaluated for predicted cell types using H&E, and the trained model was correlated with cell type enrichment scores estimated by ST datasets of 4 samples independent of the training set. The model was applied to H&E images of 16 patients with HCC who underwent nivolumab treatment. The model predicted the 5 cell types enrichment score in the tumor and tumor-adjacent normal liver tissues. Results. The internal validation set results demonstrate robust Spearman correlations between our model predictions and actual cellular compositions in the TME. Specifically, we observe strong correlations for T cells (rho = 0.66), TAMs (rho=0.46), CAFs (rho=0.32), B cells (rho=0.30), and TECs (rho=0.24). When we applied the model to the H&E images of the cohort of patients who underwent nivolumab, immunotherapy non-responders showed a significantly higher predicted enrichment score of CAFs in tumor-adjacent regions (Mann-Whitney test, p<.01). Upon stratifying patients based on stromal CAF prediction, those with low stromal CAF levels exhibited better overall survival (Log-rank test, p<.05). Conclusions. We present a deep learning model to analyze TME in HCC solely on H&E images trained by ST data. Our findings showed a potential relationship between CAFs in tumor-adjacent regions and non-responders to nivolumab treatment in HCC. These insights underscore the potential to predict nivolumab treatment responders using H&E images combined with the deep learning model. This approach could provide a significant advancement in personalized treatment strategies for HCC patients. Citation Format: Dongjoo Lee, Haenara Shin, Seungho Cook, Daeseung Lee, Hongyoon Choi, Won-Mook Choi, Changhoon Yoo, Kwon Joong Na. Unveiling the tumor microenvironment of hepatocellular carcinoma using AI trained by spatial transcriptomics: A preliminary study to predict response to immunotherapy [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2024; Part 1 (Regular Abstracts); 2024 Apr 5-10; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2024;84(6_Suppl):Abstract nr 7398.