Machine learning-driven comprehensive profiling of tumor heterogeneity and sialylation in hepatocellular carcinoma.

Hepatocellular carcinoma (HCC) exhibits profound cellular heterogeneity, the understanding of which is critical for improving prognosis and therapy. Using single-cell RNA sequencing of 32,247 cells from human HCC samples, we characterized the tumor ecosystem and identified five malignant hepatocyte subpopulations with distinct molecular profiles and stage-specific enrichment. Among these, the S100A6⁺ C1 and S100A9⁺ C4 subpopulations were predominantly associated with advanced tumors and actively remodeled the tumor microenvironment through enhanced signaling pathways such as MDK and MIF. We further identified PGAM2 as a key transcriptional regulator in early-stage tumors, whose activity correlated with sialylation-a process linked to immune evasion. Based on these findings, we developed a prognostic model integrating PGAM2 and sialylation-related genes, which robustly stratified patients into high- and low-risk groups with significantly different survival outcomes, immune contextures, and predicted therapeutic responses. Functional experiments validated AGRN, a component of the signature, as a functional driver of HCC proliferation and invasion. Collectively, our results decode the cellular and molecular heterogeneity of HCC, provide a clinically relevant prognostic tool, and highlight potential targets for further investigation.