Machine learning identification of hypoxia-related genes and prognostic risk-scoring model for effective survival stratification in hepatocellular carcinoma.

Abstract

e16261 Background: Hepatocellular carcinoma (HCC) is the fourth leading cause of cancer-related mortality globally, where treatment and prognostic assessment have important implications in clinical practice. Hypoxia, as a common feature within solid tumors, can directly change the tumor microenvironment, which affects the efficacy of cancer treatment and prognosis. In this study, we constructed and validated a hypoxia-based prognostic model using bioinformatics and machine learning. Methods: Two public datasets, GSE14520 and GSE41666, were collected from the Gene Expression Omnibus: (1) HCC tumor tissues compared to adjacent normal tissues (N = 214) and (2) HepG2 cells under normoxic and hypoxic conditions (N = 6). Differential expression analysis was performed to identify HCC characteristic genes and hypoxia-related genes, including their common genes (HCC-Hypoxia Overlap genes, HHOs). Using RNA-seq data of HCC patients (N = 367) from the TCGA Liver Cancer (LIHC) database, univariate Cox regression models were identified, and the Least Absolute Shrinkage and Selection Operator (LASSO) algorithm selected hypoxia-characteristic genes for the multivariate survival model. A hypoxia-related risk score was calculated based on the model of these characteristic genes and dichotomized cases into high-risk (HR) and low-risk (LR) groups. The model was validated using liver cancer cases (N = 232) from the International Cancer Genome Consortium database (ICGC-LIRI-JP). Results: Through differential expression analysis of the two datasets, we identified 52 HHOs. Univariate Cox analysis of these HHOs indicated that 21 genes were significantly associated with HCC patient survival. Through LASSO regression analysis, a total of 9 characteristic genes, including CENPA, KIF20A, DLGAP5, HMMR, UPB1, AFM, CABYR, PHLDA2, and N4BP2L1 were ultimately retained in the survival model. Based on these 9 genes, TCGA-LIHC samples were classified into HR and LR groups, and Kaplan-Meier (KM) analysis revealed significant differences in survival outcomes (p < 0.032). Risk scoring of the ICGC-LIRI-JP validation set classified samples into HR and LR. KM analysis showed that the survival times of patients in the HR group were significantly shorter than those in the LR group (p < 0.0001). Receiver Operating Characteristic Analysis analysis of the survival model showed area under the curve values of 0.815, 0.774, and 0.771 at 1, 2, and 3 years, respectively, demonstrating high performance in risk stratification. Conclusions: This study established a prognostic risk-scoring model based on 9 characteristic genes associated with hypoxia. This model can effectively stratify risks among HCC patients and demonstrate excellent performance in predicting survival. These findings may offer new biomarkers and therapeutic targets for the personalized treatment of HCC.