Hypoxia Molecular Characterization in Hepatocellular Carcinoma Identifies One Risk Signature and Two Nomograms for Clinical Management

Abstract

Hypoxia is a universal feature in the tumor microenvironment (TME). Nonetheless, the heterogeneous hypoxia patterns of TME have still not been elucidated in hepatocellular carcinoma (HCC). Using consensus clustering algorithm and public datasets, we identified heterogeneous hypoxia subtypes. We also revealed the specific biological and clinical characteristics via bioinformatic methods. The principal component analysis algorithm was employed to develop a hypoxia-associated risk score (HARS). We identified the two hypoxia subtypes: low hypoxia pattern (C1) and high hypoxia pattern (C2). C1 was less sensitive to immunotherapy compared to C2, consistent with the lack of immune cells and immune checkpoints (ICPs) in C1, whereas C2 was the opposite. C2 displayed worse prognosis and higher sensitivity to obatoclax relative to C1, while C1 was more sensitive to sorafenib. The two subtypes also demonstrated subtype-specific genomic variations including mutation, copy number alteration, and methylation. Moreover, we developed and validated a risk signature: HARS, which had excellent performance for predicting prognosis and immunotherapy. We revealed two hypoxia subtypes with distinct biological and clinical characteristics in HCC, which enhanced the understanding of hypoxia pattern. The risk signature was a promising biomarker for predicting prognosis and immunotherapy.

Figure 1

Overall workflow diagram of our research.

Figure 2

Landscape of genomic variations in hypoxia-associated molecules in TCGA-LIHC cohort. (a) Principal component analysis for the expression profiles of 24 hypoxia-associated genes (HAGs) to distinguish tumors from normal samples in TCGA-LIHC cohort. Tumor group is marked with yellow, and normal group is marked with blue. (b) The expression of 24 HAGs between normal and tumor groups. Normal, blue; tumor, yellow. ns, P > 0.05; ^∗P < 0.05; ^∗∗P < 0.01; ^∗∗∗P < 0.001. (c) The mutation frequency of 24 HAGs. Each column represented individual patients. The upper bar plot showed TMB. The number on the right indicated the mutation frequency in each gene. The right bar plot showed the proportion of each variant type. (d) The location of copy number alteration (CNA) of HAGs on 23 chromosomes. (e) The CNA frequency of HAGs. The height of the column represented the alteration frequency. The loss frequency, blue dot; the gain frequency, red dot. (f) The correlation between methylation and expression of HAGs. (g) The interaction among 24 HAGs in HCC. The circle size represented the effect of each gene on the prognosis, and the range of values was calculated by the log-rank test was P < 1e − 8, P < 1e − 5, P < 0.001, and P < 0.05, respectively. Green dots in the circle, protective factors of prognosis; black dots in the circle, risk factors of prognosis. The lines linking regulators showed their interactions, and thickness showed the correlation strength between regulators. Negative correlation was marked with blue and positive correlation with red. The gene cluster A-D was marked with yellow, blue, red, and brown, respectively.

Figure 3

The prognostic significance and functional annotation of hypoxia subtypes in metacohort. (a) The CDF of consensus matrix for each k (indicated by colors). The lowest and flattest curve indicated the optimal k (k = 2). (b) Kaplan–Meier curves of OS for the two subtypes in metacohort. Log-rank test showed the P < 0.001. (c) Kaplan–Meier curves of RFS for the two subtypes in metacohort. Log-rank test showed the P < 0.001. (d) The activation of specific Hallmark pathways between the two subtypes. (e) The activation of specific KEGG pathways between the two subtypes.

Figure 4

The difference of ICP expression, immune cells infiltration, and immunotherapy response between C1 and C2. (a) The expression heatmap of ICPs between C1 and C2 in metacohort. High expression, red; low expression, blue. (b) The abundance of 23 immune cell subsets infiltration was compared between the C1 and C2 in metacohort. ns, P > 0.05; ^∗P < 0.05; ^∗∗P < 0.01; ^∗∗∗P < 0.001. (c) Correlations between immune checkpoints and HAGs in metacohort using Spearman analysis. Negative correlation was marked with blue, and positive correlation was marked with red. No asterisks represented no statistical significance; ^∗P < 0.05; ^∗∗P < 0.01. (d) Distribution of the immunotherapy response results predicted by TIDE algorithm between C1 and C2 in metacohort. Nonresponders, blue; responders, orange. (e) Submap analysis of the two subtypes and 47 previous melanoma patients with detailed immunotherapeutic information. NR represented nonresponders; R represented responders.

Figure 5

The clinical characteristics and prognosis of hypoxia subtypes in TCGA cohort. (a) The expression heatmap of 24 HAGs in TCGA-LIHC cohort. Survival status, age, gender, BMI, vascular invasion, histology grade, AJCC stage, and hypoxia subtypes were displayed as patient annotations. (b, c) Kaplan–Meier curves of OS (b) and RFS (c) between C1 and C2 in TCGA-LIHC cohort. (d–f) Composition percentage of clinical characteristics such as age (d), gender (e), and BMI (f) between C1 and C2. (g–i) Composition percentage of AJCC stage (g), grade (h), and vascular invasion (i) between C1 and C2. (j, k) Distribution of the estimated IC50 of sorafenib (j) and obatoclax (k) between C1 and C2 in TCGA-LIHC cohort.

Figure 6

The genomic alterations of hypoxia subtypes in TCGA cohort. (a) Significantly mutated genes (SMGs) in two hypoxia subtypes. Each column represented individual patient. The upper bar plot showed TMB, and the percentage on the left showed the proportion of samples with mutations. The right bar plot indicated the MutSigCV q-value. The mutation alternations types were indicated by different colors. (b) Mutation signatures extracted from the two hypoxia subtypes. The three mutation signatures with the highest cosine similarity to COSMIC signatures exhibited in C1 and C2, respectively. The etiology of each signature and the cosine similarity between the original and the reconstructed mutation signatures were indicated. (c) Gain (red) or loss (blue) frequencies of copy number alterations (CNAs) in the chromosomes. (d–g) The burden of copy number gain (d) and loss (e) in arm level were compared between C1 and C2. The burden of copy number gain (f) and loss (g) in focal level were compared between C1 and C2. (h) The distribution of fraction genome altered (FGA), fraction genome gained (FGG), and fraction genome lost (FGL) in C1 and C2. ^∗P < 0.05; ^∗∗∗∗P < 0.0001.

Figure 7

The DNA methylation modification of two hypoxia subtypes in TCGA-LIHC cohort. (a) The distribution of global methylation level (GML) in two hypoxia subtypes. (b) The correlation between GML and proliferation score. (c) Spearman correlation analysis between GML and 23 immune cells. The circle size represented the strength of correlation. (d, e) The methylation-driven genes (MDGs) in C1 (d) and C2 (e). Each column represented individual patients. The percentage on the left showed the proportion of samples in the whole that this gene was identified as an MDG. The right bar plot indicated the total number of samples identified as an MDG in each gene.

Figure 8

The development and validation of HARS in TCGA-LIHC and NCI cohorts. (a) Volcano plot displayed the differentially expressed genes (DEGs) between the two subtypes. Red dots represented upregulated genes, blue dots represented downregulated genes, and grey dots represented genes with no significance. (b, c) GO (b) and KEGG (c) enrichment analysis of 299 DEGs. The significantly biological functions were extracted with adjusted P value < 0.05. (d) Using MCODE plug-in of Cytoscape software, we extracted a key module including 10 genes. (e) The expression difference of 10 key genes between C1 and C2. ^∗∗∗P < 0.001. (f) Univariate Cox regression analysis revealed the prognosis significance of 10 key genes, all of them were risk factors. (g, h) Kaplan–Meier curves for OS (g) and RFS (h) between low HARS and high HARS groups in TCGA-LIHC cohort. (i) Estimation of the prognosis prediction by receiver operating characteristic curve (ROC) in TCGA-LIHC cohort. (j, k) Kaplan–Meier curves for OS (j) and RFS (k) between low HARS and high HARS groups in NCI cohort. (l) Estimation of the prognosis prediction by ROC in NCI cohort.

Figure 9

The construction of two nomograms predicting OS and RFS, respectively. (a–c) Nomogram for predicting the 1-, 3-, and 5-year OS of HCC patients (a). Calibration (b) and ROC curve (c) for evaluating the performance of nomogram predicting the 1-, 3-, and 5-year OS for HCC patients. (d–f) Nomogram for predicting the 1-, 3-, and 5-year RFS of HCC patients (d). Calibration (e) and ROC curve (f) for evaluating the performance of nomogram predicting the 1-, 3-, and 5-year RFS for HCC patients. Blue line, 1-year; red line, 3-year; yellow line, 5-year.

Figure 10

The performance of HARS for predicting the immunotherapy response. (a–c) The Kaplan–Meier analysis (a), immunotherapy response ratio (b), and ROC curve (c) of HARS in GSE100797 cohort. (d–f) The Kaplan–Meier analysis (d), immunotherapy response ratio (e), and ROC curve (f) of HARS in GSE91061 cohort. (g-i) The Kaplan–Meier analysis (g), immunotherapy response ratio (h), and ROC curve (i) of HARS in VanAllen cohort. (j–l) The Kaplan–Meier analysis (j), immunotherapy response ratio (k), and ROC curve (l) of HARS in IMvigor210 cohort.