Machine-learning Approach to Reveal a Novel Hepatic Stellate Cell-based Classification for Prognostic Prediction in Patients with Hepatocellular Carcinoma

Abstract

Abstract Background: Hepatocellular carcinoma (HCC) is one of the major concerns regarding public health globally. Cancer-associated fibroblasts (CAFs) play a vital role in HCC progression. The identification of CAF-associated HCC subtypes and the development of CAF-related HCC precise treatment strategies are unmet needs. Methods: A total of 288 CAF signatures were obtained from previous studies. Consensus clustering analysis was employed to identify the CAF-related subtypes in HCC. Enrichment analysis, CIBERSORT, and ESTIMATE were applied to comprehensively evaluate heterogeneity across the HCC subtypes. Four machine-learning methods, including Least Absolute Shrinkage and Selector Operation regression, Elastic Net, survival Support Vector Machine, and Neural Network, were used to construct the prognostic model (HC score). The immunotherapy cohort was enrolled to explore the potential of the HC score in predicting immunotherapy responsiveness. Results: Based on the CAF signatures, we identified two HCC subtypes: HCf-inactive and HCf-active subtypes. The two HCC subtypes had significantly different immune features, fibrosis features, and prognoses. Furthermore, we constructed a CAF-related gene prognostic model HC score based on the integration of four machine-learning methods. The HC score predicted the outcomes in patients with HCC compared with traditional clinicopathological features. Moreover, the HC score could also effectively predict the microenvironment characteristics of HCC. The immunotherapy cohort analysis indicated that the HC score had great potential to help identify the immunotherapy candidates. Conclusions: We identified two CAF-related HCC subtypes: HCf-inactive and HCf-active subtypes. A gene prognostic model was developed that could well predict HCC prognosis, microenvironment status, and immunotherapy responsiveness.