Given the high prevalence and relapse rates of hepatocellular carcinoma (HCC), an increased capacity for early identification of patients most at risk for post-resection recurrence would help improve patient outcomes and prioritize health care resources. Here, we combined spatial multi-transcriptomics and proteomics approaches to characterize the tumor and immunological landscape of 61 samples. We observed a spatial and HCC-recurrence-associated distribution of natural killer (NK) cells in the invasive front and tumor center. Using artificial-intelligence alongside an extreme gradient-boosting algorithm, we developed the Tumor Immune MicroEnvironment Spatial (“TIMES”) score based on the expression of five NK-associated markers (SPON2, ZFP36L2, ZFP36, VIM, and HLA-DRB1) to predict HCC recurrence. We also demonstrated that TIMES score (HR = 29.6, P < 0.001) outperforms the current standard tools for patient risk stratification including the TNM (HR = 1.93, P = 0.113) and BCLC (HR = 1.55, P = 0.253) systems. In the clinic, we validated the model in 103 patients from three multi-centered cohorts achieve a real-world sensitivity of 90.00% and specificity of 90.24%. In the lab, following up on the individual marker with the highest prediction accuracy, in vivo models revealed that SPON2 increases IFN-γ secretion and enhances infiltration potential of NK cells at the invasive front. Additionally, we established the TIMES score on a publicly accessible website that can be easily achieved by different levels of pathology labs to facilitate global prediction of HCC recurrence risk and stratification of high-risk patients. With its ability to efficiently stratify high-risk patients, it exemplifying the utility of artificial intelligence to improve our understanding on TIME features that underlie tumor progression.

MotivationSpatially resolved transcriptomics (SRT) technologies have been developed to simultaneously profile gene expression while retaining physical information. To explore differentially expressed genes using SRT in the context of various conditions, statistical methods are needed to perform spatial differential expression analysis.ResultsWe propose that a new probabilistic framework, spatialDEG, can perform differential expression analysis by leveraging spatial information on gene expression with spatial information. SpatialDEG utilizes the average information algorithm and can be scalable to tens of thousands of genes. Comprehensive simulations demonstrated that spatialDEG can identify genes differentially expressed in tissues across different conditions with a controlled type-I error rate. We further applied spatialDEG to analyze datasets for human dorsolateral prefrontal cortex and mouse whole liver.AvailabilityThe R package spatialDEG can be downloaded from https://github.com/Shufeyangyi2015310117/spatialDEG.