Multimodal single-cell analysis reveals the prognostic role of myeloid subpopulation–driven PI3K/AKT signaling in hepatocellular carcinoma

Background: Hepatocellular carcinoma (HCC) represents a major global health burden, characterized by rising incidence, marked heterogeneity, and poor prognosis. Reliable biomarkers for prognostic stratification and therapeutic targeting remain urgently needed. Methods: We conducted an integrative multi-omics analysis by combining TCGA-LIHC bulk RNA-seq data with single-cell RNA-seq datasets. Autoencoder-based feature extraction and machine learning models were applied for prognostic biomarker identification, followed by survival analysis and single-cell atlas construction. Key pathway proteins were validated using Western blot, and cellular functions were assessed with Transwell assays. Results: Using multi-gene representations derived from an autoencoder, we first identified the PI3K/AKT pathway as a critical survival-related pathway in hepatocellular carcinoma. Next, integrated single-cell analysis uncovered 12 distinct cellular clusters, among which the myeloid subpopulation showed pronounced transcriptional dysregulation, with its functional state driven by PI3K/AKT signaling. To further refine these findings, we intersected 2,567 upregulated genes from myeloid cells with 105 PI3K/AKT pathway–related genes obtained from GSEA, identifying 24 overlapping genes. Building on this, we compared 12 machine learning algorithms and selected LASSO for feature optimization, resulting in a robust predictive model with an ROC-AUC of 0.949. In parallel, we screened survival-related candidate genes and validated them using the external dataset GSE76427; through three independent approaches, RAC1 and SLC2A1 were consistently identified as core genes. Incorporating analyses of immune infiltration, immunotherapy response, and cell–cell communication, Western blot experiments further confirmed that the regulatory roles of RAC1 and SLC2A1 are closely linked to modulation of the PI3K/AKT pathway. Finally, functional assays demonstrated that targeting RAC1 and SLC2A1 effectively inhibited cell migration and invasion. Conclusion: This study establishes a comprehensive framework that integrates multi-omics profiling, single-cell transcriptomics, and machine learning to dissect the molecular heterogeneity of HCC. We identified the PI3K/AKT signaling axis as a key prognostic driver, particularly within tumor-associated myeloid subpopulations, and uncovered candidate biomarkers such as RAC1 and SLC2A1 that are strongly associated with adverse clinical outcomes. These findings not only highlight the pivotal role of immune–tumor interactions in shaping HCC progression but also provide mechanistic insights that may inform the development of targeted and combination therapies.