Document resource
Background Hepatocellular carcinoma (HCC) is the most prevalent type of liver cancer. Incidence of HCC has increased over the past several decades, and despite incredible advances in treatment, including immunotherapies, mortality remains high and the five-year survival rate is just 18%. Although immunotherapy has already proved promising for both early and late stage disease, a significant fraction of patients undergoing treatment fail to respond. Understanding the cellular and molecular dynamics at play during therapy could lead to fewer failed outcomes, more personalized treatments, and mechanistic insight into disease processes.Methods We first generated a dataset comprising spatial proteomic (Akoya Phenocycler) and spatial transcriptomic (10x Visium) profiling of human HCC tissue. Samples in this study derive from a recent Phase II clinical trial and include pretreatment biopsies and posttreatment resections from matched patients. Next, we devised a deep learning framework to integrate the multi-omics spatial data and predict therapeutic response from local cell microenvironments ( figure 1A). Our spatially-resolved computational approach allowed us to group microenvironments by their predictiveness or cellular compositions. We interpret these features from our models, and further associate each microenvironment with a corresponding gene expression signature from co-registered ST data.Results Using a graph neural network (GNN) trained on pre-treatment spatial proteomics data, we were able to predict post-treatment response with ROC-AUC > 0.9 ( Figure 1B). Models trained using Visium spatial transcriptomics data also performed well, and predictions on the spot level were spatially concordant with GNN prediction. Using spatial model predictions, we identified a niche that envelops lymphoid aggregates (LAs) and physically separates them from tumor bulk. LAs classified as ‘interfaced’ based on the presence of this niche were almost found in nonresponders (figure 2). We further quantified the cell interactions and expression patterns unique to this interface niche. Our analysis prioritized several genes with previously reported links to immunotherapy resistance, including SPP1, as well as several novel candidates. Extending these findings, we found that the expression of the interface signature was enriched in single-cell transcriptomic data of breast cancer nonresponders to immunotherapy.Conclusions Integrative spatial omics analysis is a powerful tool to interrogate the expression pathways and cell interactions that underlie successful immunotherapy responses. We report a novel pipeline to register and systematically interrogate these multimodal data. We further report a novel lymphoid aggregate interface microenvironment that points towards new therapeutic or clinical stratification strategies.Abstract 478 Figure 1(A) Schematic of interpretable deep learning to predict immunotherapy response (B) Model prediction performance on HCC patients. Includes baseline spatial transcriptomics (ST) and multiplexed immunofluorescence (MIF) predictorsAbstract 478 Figure 2Characterization of an interface niche in immunotherapy nonresponders. Images show the classification of LAs as exposed or interfaced. LA: lymphoid aggregate; R./N.R.; responder/nonresponder; Pre/Post: pretreatment/posttreatment