A novel molecular classification system based on the molecular feature score identifies patients sensitive to immune therapy and target therapy.

Background: Hepatocellular carcinoma (HCC) is heterogeneous and refractory with multidimensional features. This study aims to investigate its molecular classifications based on multidimensional molecular features scores (FSs) and support classification-guided precision medicine. Methods: Data of bulk RNA sequencing, single nucleotide variation, and single-cell RNA sequencing were collected. Feature scores (FSs) from hallmark pathways, regulatory cell death pathways, metabolism pathways, stemness index, immune scores, estimate scores, etc. were evaluated and screened. Then, the unsupervised clustering on the core FSs was performed and the characteristics of the resulting clusters were identified. Subsequently, machine learning algorithms were used to predict the classifications and prognoses. Additionally, the sensitivity to immune therapy and biological roles of classification-related prognostic genes were also evaluated. Results: We identified four clusters with distinct characteristics. C1 is characterized by high TP53 mutations, immune suppression, and metabolic downregulation, with notable responsiveness to anti-PD1 therapy. C2 exhibited high tumor purity and metabolic activity, moderate TP53 mutations, and cold immunity. C3 represented an early phase with the most favorable prognosis, lower stemness and tumor mutations, upregulated stroma, and hypermetabolism. C4 represented a late phase with the poorest prognosis, highest stemness, higher TP53 mutations, cold immunity, and metabolic downregulation. We further developed practical software for prediction with good performance in the external validation. Additionally, FTCD was identified as a classification-specific prognostic gene with tumor-suppressing role and potential as a therapeutic target, particularly for C1 and C4 patients. Conclusions: The four-layer classification scheme enhances the understanding of HCC heterogeneity, and we also provide robust predictive software for predicting classifications and prognoses. Notably, C1 is more sensitive to anti-PD1 therapies and FTCD is a promising therapeutic target, particularly for C1 and C4. These findings provide new insights into classification-guided precision medicine.