A Digital Pathology Approach to Predict Spatial Subtype Signatures of Hepatocellular Carcinoma from Histologic Images

Summary: Background: Hepatocellular carcinoma (HCC) is a leading cause of cancer-related mortality. There are multiple classifications of HCC, including the Hoshida system, which stratifies HCC into 3 subtypes (S1, S2, and S3) with distinct pathomolecular features. Advances in precision medicine for HCC have been delayed by a scarcity of tissue biopsies due to historical concerns of safety risks. However, with the advancement of clinical trials for targeted therapies, tissue biopsies are being increasingly advocated for and even required. Thus, there is an emerging opportunity to pioneer the use of digital pathology approaches to advance precision medicine for HCC. Materials and Methods: We accessed publicly available spatial transcriptomics data, paired with corresponding hematoxylin and eosin (H&E) images, from 17 HCC slides. Analysis was performed to characterize the spatial distribution of S1, S2, and S3 subtype signatures. The H&E images were divided into tiles corresponding to the spatial transcriptomic spots. Using a deep learning foundation model for digital pathology, we extracted an embedding vector from each tile. Hoshida subtype signature scores were calculated for each spatial transcriptomic spot, and 2 spatial subtype clusters (HS1 and HS3) were derived using k-means clustering. A neural network was trained to predict these clusters from the H&E tile embeddings and was evaluated using leave-one-out cross-validation. External validation was performed by applying the trained model to H&E whole-slide images from HCC patients (n=352) in The Cancer Genome Atlas (TCGA). We assessed the correlation between the proportion of predicted HS1-positive tiles (HS1+) for each patient and bulk RNA-derived subtype signature scores, CTNNB1 mutation status, and overall survival. Results: Among the 17 HCC slides, our neural network model achieved a mean prediction accuracy of 0.74 for spatial transcriptomics-derived Hoshida subtypes across spots, using corresponding H&E tiles. When applied to the TCGA data, the model demonstrated a significant positive correlation between HS1+ and bulk RNA-derived S1 signature scores (p<0.0001). Conversely, we observed a negative correlation between HS1+ and the bulk RNA-derived S3 signature (p<0.0001). CTNNB1 mutation status was negatively associated with HS1+ (p=0.0019). In addition, patients with low HS1+ demonstrated improved overall survival compared to those with high HS1+ (p<0.0001). Conclusions: Our results suggest that deep learning can effectively predict the spatial distribution of gene expression-derived subtypes from H&E whole slide images. These predictions recapitulate established paradigms in HCC, while providing improved prognostic value compared to bulk RNA-based signatures alone. This highlights a promising role for digital pathology in HCC precision medicine. Satdarshan P. Monga and Yu-Chiao Chiu