Introduction

Hepatocellular carcinoma (HCC), as one of the most common and lethal tumors, ranks the third leading cause of cancer-related death around the world, with an overall 5-year survival rate less than 20% [1, 2]. Presently, the major treatments for HCC include surgical resection, microwave ablation, transcatheter arterial chemoembolization (TACE), targeted therapy with sorafenib or lenvatinib, immune checkpoint blockade (ICB), and liver transplantation [3, 4]. Despite the increasing improvements in early diagnosis and treatment in HCC, the survival of patients with HCC remains disappointed, due to the tumor heterogeneity ad specific tumor microenvironment (TME) [5,6,7,8]. Besides, the conventional clinicopathological parameters, such as histologic grade, TNM stage and vascular invasion, have limited predictive power for the clinical outcome of HCC patients, due to the complex molecular mechanisms of tumor regulation [9, 10]. Thus, it is necessary to explore the potential mechanisms underlying tumor development and progression for designing more individualized treatment options for HCC.

Hypoxia and desmoplasia are the key characteristic features of HCC, in which tumor desmoplasia accelerates hypoxia, and aggravated hypoxia activates stromal cells, such as cancer stem cells and hepatic stellate cells, resulting in severe desmoplasia [11,12,13,14,15]. Hypoxia-inducible factor-1α (HIF-1α), as a chief hypoxic regulator, plays an important role for the malignant phenotypes of HCC [16,17,18,19,20]. For instance, HIF-1α was found to promote the tumor progression, epithelial-mesenchymal transition, and vasculogenic mimicry formation by upregulate the expression of LOXL2 in HCC [19]. Besides, the hypoxic TME of HCC enhances the expression and function of HIF-1α targets (e.g. MDR1, GLUT-1 and VEGF) through stabilizing HIF-1α protein, thereby conferring sorafenib resistance in HCC [20]. Importantly, HIF-1α impairs anti-tumor immunity by increasing CD47 or PD-L1 expression, therefor leading to an immunosuppressive TME [21,22,23,24]. Consequently, HIF-1α may influence various cancer cell behaviors by regulator different genes expression. And several hypoxia-derived signatures have been reported to predict survival and therapeutic response of malignancy patients [25, 26], but it still remains further exploration whether hypoxia-induced HIF-1α-related genes could be an ensemble to divide HCC patients into different clusters with various clinical outcomes, genomic alteration, oncogenic pathway and therapeutic responses. Therefore, it is of great significance to construct a HIF-1α-specific prognostic model for assessing the molecular features of HCC with various hypoxia status.

In this study, using a series of bioinformatics and machine learning methods, HCC patients were classified into three subgroups according to HIF-1α related genes. And an HIF-1α score system based on the HIF-1α cluster was established for prognostic stratification in HCC. Then, the predictive powers of our model were respectively assessed through the C-indexes and ROC analysis. Furthermore, we also systematically investigated the oncogenic pathways, immune infiltration profile, and therapeutic response to TACE and ICB for HCC in different HIF-1α score groups.

Methods

Data processing

The mRNA expression profiles and corresponding clinical information of HCC patients were extracted from the Cancer Genome Atlas (TCGA, https://cancergenome.nih.gov/), the Gene Expression Omnibus (GEO, https://www.ncbi.nlm.nih.gov/geo/), and International Cancer Genome Consortium (ICGC, https://icgc.org/) databases. The raw counts of RNA-seq data from TCGA and ICGC HCC cohorts were normalized and log2 (x + 1) transformed. The dataset GSE14520 involves two independent microarray datasets (e.g. GSE14520–GPL571 and GSE14520–GPL3931), which were subsequently integrated into one cohort and proceeded for batch normalization using R packages “sav” and “limma”. Of the 371 HCC cases in the TCGA cohort, 343 were cases with overall survival (OS) longer than one month, and were enrolled in the subsequent survival analyses. And genes with zero expression levels in  > 50% of the tumor samples or genes with average expression  < 5 were excluded. The TCGA HCC cohort was used as the training set because it is an independent RNA-seq dataset with an appropriate sample size. And the RNA-seq data from GSE14520 cohort (validation cohort 1; N = 242) and the ICGC cohort (validation cohort 2; N = 229) were extracted for prognostic stratification validation. In addition, GSE104580 cohort (response: N = 81; non-response: N = 66) was used for the evaluation of therapeutic response to TACE. All datasets are freely available as public resources. Therefore, local ethics approval was not required.

Consensus clustering analysis based on HIF-1α related genes in HCC

A total of 16 HIF-1α related genes were selected based on previous studies [27]. The prognostic values of these HIF-α related genes were investigated in the TCGA cohort, using Kaplan–Meier (KM) survival analysis. Subsequently, using R package ‘ConsensusClusterPlus’, Partitioning Around Medoids (PAM) algorithm was performed on the selected HIF-1α related genes to identify robust clusters of HCC, and the detail settings for clustering were shown as followed: the PAM clustering with up to 9 clusters; number of repetitions = 1000 bootstraps; pFeature = 1 (resampling 100% of any genes); pItem = 0.8 (resampling 80% of any tumor sample) and distance = euclidean. Then, the numbers of clustering was determined by three factors, including the consensus matrices, a consensus cumulative distribution function (CDF) curve and Calinski-Harabasz Index. Besides, the different clusters were also visualized in 3-demension by principal component analysis (PCA). Then, the OS differences among different HIF-1α clusters were assessed by KM survival curves, and log-rank test p value were calculated. The R package ‘pheatmap’ was used to visualize HIF-1α related genes expressions among different HIF-1α clusters. The association between HIF-1α clusters and clinicopathological characteristics (e.g. histologic grade, TNM stage, and vascular invasion) was assessed using the Fisher’s exact test or chi-square test. Under the threshold of |log2 fold change (FC)|> 1 and P < 0.05, differentially expressed genes between different HIF-1α clusters were determined with the help of the R package ‘limma’. Then, the overlapping differentially expressed genes (ODEGs) were extracted for further analysis.

Establishment and validation of the HIF-1α score system

Univariate Cox regression analysis was employed to assess the prognostic values of the ODEGs in the TCGA cohort. Then, LASSO regression analysis was conducted to determine the robust prognostic ODEGs, based on which an HIF-1α score system was subsequently constructed. The HIF-1α score = ∑the robust prognostic ODEGs expressions × the corresponding coefficients of multivariable Cox regression analysis. For external validation, the HIF-1α scores were also calculated for GSE14520 cohort and the ICGC cohort.

KM survival curve for the HIF-1α score system was also generated based on the optimal cutoff points from the X-tile (3.6.1) [28]. The predictive capabilities of the HIF-1α score system were assessed by ROC analysis and C-index. Besides, PCA was also conducted to assess the prognostic stratification ability. The correlations between the HIF-1α score system and clinicopathological parameters (e.g. TNM stage, vascular invasion, and histologic grade) were investigated using the chi-square test. Kruskal–wallis test or Wilcoxon test was utilized to assess the differences of HIF-1α score among different groups of these clinicopathological characteristics. In addition, with the help of R package “Survival ROC”, the ROC curves were generated for the HIF-1α score system and clinicopathological characteristics to compare their predictive discrimination for 1-, 3-, and 4-year OS.

Gene set variation analysis (GSVA)

The R package ‘gsva’ was applied to conduct GSVA to estimate the enrichment levels of 50 hallmark gene sets extracted from the Molecular Signature Database (https://www.gsea-msigdb.org/gsea/msigdb/collections.jsp#C5) in the TCGA cohort [29]. In addition, GSVA was also conducted to estimate the enrichment levels of 18 immunologic signatures obtained from previous studies [30, 31].

Association between the HIF-1α score and previously reported hypoxia scores

A total of three previously reported hypoxia scores based on the TCGA cohort were downloaded from the cBioportal database (http://www.cbioportal.org/), including Buffa, Winter, and Ragnum hypoxia scores [32,33,34]. The differences in these three hypoxia scores among these three HIF-1α score groups were then investigated. Pearson correlation analysis was used to evaluate the association between the HIF-1α score and these three hypoxia scores. KM survival analyses were conducted for these three hypoxia scores in the TCGA cohort. In addition, ROC analyses were conducted to compare their predictive discrimination for 1-, 3-, and 4-year OS.

Statistical analysis

The SPSS 25.0 software (SPSS Inc., Chicago, IL), R software version 4.0.5 and GraphPad Prism 8.0 software (GraphPad Software, Inc.) were used for all statistical analyses. The Kaplan–Meier method with log-rank test was performed to describe patients’ survival. Variables associated with survival were determined by univariate Cox regression analysis. Significant variables in univariate Cox regression analysis were further involved in LASSO regression analysis and multivariable Cox regression analysis. Kruskal–wallis or Wilcoxon tests was used to evaluated group differences, which were expressed as means ± standard deviation (SD). And correlation analyses were conducted using Pearson correlation coefficient. All statistical test were two-sided, and p value less than 0.05 was regarded as statistically significant.

Results

Three HIF-1α clusters of HCC based on consensus clustering analysis

Of 16 known HIF-1α related genes, 15 were notably correlated with worse OS of HCC patients in the TCGA cohort, including HIF1A, ALDOA, ENO1, ALDOC, PKM, HK1, GAPDH, PFKFB3, HK2, PDK1, SLC2A1, LDHA, PGK1, BNIP3, and SLC2A3 (Fig. 1A). Then PAM clustering were employed based on these 16 HIF-1α related genes to divide HCC samples into subgroups. A 3-cluster was selected as an appropriate solution for the consensus matrix with k = 3 or k = 4 deemed to be a cleanest separation among clusters, but the CDF curve and delta plot showed limited increase in area for k = 3 compared to k = 4. In addition, the Calinski-Harabasz Index for k = 3 was larger than k = 4 (Fig. 1B). According to the evidence above, the TCGA HCC samples were clustered into three groups by the selected HIF-1-related genes. 3D PCA further exhibited a significant spatial separation of tumor samples, which actually to some extend suggest that the selected HIF-1-related genes worked well in tumor sample clustering (Fig. 1C). The KM survival analysis showed significant differences in OS among these three clusters, in which HIF-1α cluster C exhibited the best survival and HIF-1α cluster B the worst (Fig. 1D). And a heatmap was performed to visualize the distribution of TNM stage, histologic grade, vascular invasion and the expression of these HIF-1α-related genes in different HIF-1α clusters (Fig. 1E). In addition the HIF-1α cluster was notably correlated with TNM stage and histologic grade, though not significantly associated with vascular invasion (Fig. 1F–H).

Fig. 1
figure 1

Consensus clustering based on HIF-1α related genes divided HCC samples into three clusters. A KM survival analyses for the selected HIF-1α-related genes were exhibited in forest plot. B Subgroups are identified based on the selected HIF-1α-related genes in TCGA HCC cohort by PAM consensus clustering upon their expression levels. Consensus clustering was based on 1000 resampled datasets and k was tested from 2 to 9. Consensus matrices, as well as the consensus cumulative distribution function (CDF) curve, delta area (change in CDF area) plot and Calinski-Harabasz Indexes are shown. C 3D PCA based on HIF-1α cluster exhibited a significant spatial separation of tumor samples. D The KM survival curve for OS showed a significant difference among these three HIF-1α clusters. E Heatmap for the 16 HIF-1α related genes expression and clinicopathological characteristics in different HIF-1α clusters. F–H HIF-1α cluster was significantly associated with TNM stage (F) and histologic grade (G), but not vascular invasion (H)

Full size image

Establishment and validation of the HIF-1α score system

The DEGs between various HIF-1α clusters were shown by volcano plots (Additional file 1: Fig. S1A–C; Additional file 2: Table S1). And 597 ODEGs were identified among these three HIF-1α clusters (Additional file 1: Fig. S1D). And 317 ODEGs were notably correlated with HCC patients survival (P < 0.05) (Additional file 3: Table S2). Furthermore, nine robust prognostic ODEGs were identified by LASSO regression analysis (Fig. 2A, B), the KM survival analyses of which were shown in Fig. 2C. tenfold cross-validation was used to overcome over-fitting effect, and an optimal λ value of 0.059 was selected. Then, using multivariable Cox regression analysis, we established an HIF-1α score system based on these nine robust prognostic ODEGs in the TCGA cohort (training cohort), which was calculated as followed: HIF-1α score = (0.181207 × ADAMTS5 expression) + (0.004134 × EGLN3 expression) + (0.074128 × HOMER1 expression) + (0.087261 × LPCAT1 expression) + (0.143817 × NEIL3 expression) + (0.012743 × PFN2 expression) + (0.107185 × S100A9 expression) + (0.028619 × SLC1A5 expression) + (0.014439 × SPP1 expression). Then, patients were classified into low-, medium-, and high-HIF-1α score groups according to the most appropriate cutoff points [28]. The low-HIF-1α score group exhibited the best OS, and the high-HIF-1α score group the worst (Fig. 3A). Then, for external validation, we further calculated the HIF-1α scores respectively for the GSE14520 and ICGC cohorts. Of note, the results of KM survival analyses in the validation cohorts were consistent with that in the training cohort (Fig. 3B, C).

Fig. 2
figure 2

Identification of nine robust prognostic ODEGs for HCC. A, B LASSO regression analysis identified nine robust prognostic ODEGs for HCC. C KM survival analysis of the nine robust prognostic ODEGs

Full size image
Fig. 3
figure 3

Establishment and validation of an HIF-1α score system based on the HIF-1α cluster. A–C KM survival curves for the OS of HCC patients in different HIF-1α score groups in the TCGA (A), GSE14520 (B), and ICGC (C) cohorts. D–F ROC curve analysis of the HIF-1α score system for 1-, 3-, and 4-year OS prediction in the TCGA (D), GSE14520 cohort (E), and ICGC (F) cohorts. G–I PCA to confirm the cluster efficacy of the HIF-1α score system in the TCGA (G), GSE14520 (H), and ICGC (I) cohorts

Full size image

The HIF-1α score system showed a reliable accuracy of survival prediction, with a C-index of 0.736 (95%CI 0.691–0.781) in the TCGA cohort, 0.614 (95%CI 0.559–0.669) in the GSE14520 cohort and 0.668 (95%CI 0.582–0.754) in the ICGC cohort. Notably, the AUC values of the HIF-1α score for 1-, 3-, and 4-year OS prediction were 0.806, 0.741, and 0.739 in the TCGA cohort, 0.623, 0.639, and 0.638 in the GSE14520 cohort, and 0.741, 0.677, and 0.736 in the ICGC cohort, respectively (Fig. 3D–F). Besides, 3D PCA based on the HIF-1α score system further exhibited a significant spatial separation of tumor samples (Fig. 3G–I). These findings determined a significant predictive capability of the HIF-1α score system for HCC patients’ survival.

Association between HIF-1α score and clinicopathological features in HCC

The HIF-1α score system notably associated with advanced TNM stage, histologic grade and vascular invasion in HCC (Fig. 4A–C). Significant differences of HIF-1α scores were found among different groups of clinicopathological characteristics in HCC (Fig. 4D–F). Moreover, the HIF-1α score system was also idenfitied as an independent prognostic factor in HCC (Fig. 4G, H). Of note, the AUC values of the HIF-1α score system for 1-, 3-, and 4-year OS were higher than TNM stage, histologic grade, and vascular invasion (Fig. 4I–K).

Fig. 4
figure 4

The HIF-1α score system significantly associated with clinicopathological characteristics in HCC. A–C The HIF-1α score system significantly correlated with advanced TNM stage (A), histologic grade (B) and vascular invasion (C) in HCC. D–F Significant differences of HIF-1α score were found among HCC patients with different TNM stage (D), histologic grade (E), and vascular invasion (F). G Univariate Cox regression analysis showed that the HIF-1α score system, TNM stage, and vascular invasion were prognostic factors for HCC patients’ survival. H Multivariate Cox regression analysis showed that the HIF-1α score system and TNM stage were independent unfavorable prognostic factors for HCC patients’ survival. (I-K) ROC curves analysis for 1-, 3-, and 4-year OS prediction based on the HIF-1α score system, TNM stage, histologic grade, and vascular invasion

Full size image

Association between the HIF-1α score system and genomic mutation

Consistent with published studies, we also found that mutations in TP53, TTN, CTNNB1, MUC4, MUC16, ALB, APOB, RYR2, PCLO, and ABCA13 are the top 10 genetic alterations in HCC (Fig. 5A) [35]. Missense mutation was the most common type of genetic alteration (Fig. 5A). And the C > T transversion was the most common nucleotide change (Fig. 5A). In addition, TP53 and CTNNB1 alterations notably associated with the HIF-1α score system (Fig. 5A). TP53 alteration was significantly correlated with higher HIF-1α score, while CTNNB1 alteration lower HIF-1α score (Fig. 5B–E).

Fig. 5
figure 5

Association between HIF-1α score system and somatic mutation in HCC. A The top 10 most commonly mutated genes in HCC patients. B The HIF-1α score system significantly correlated with TP53 mutation. C TP53 altered HCC exhibited Higher HIF-1α score. D HIF-1α score system significantly correlated with CTNNB1 mutation. E CTNNB1 altered HCC exhibited Lower HIF-1α score

Full size image

Analysis of functional annotations among different HIF-1α score groups

The differentially enriched oncologic pathways among different HIF-1α score groups were shown in Additional file 4: Table S3. And 27 differential oncologic pathways were found overlapped among these three HIF-1α score groups (Fig. 6A). Of note, the high HIF-1α score group exhibited the highest enrichment levels for hypoxia-related terms (e.g. hypoxia and glycolysis), immune-related processes (e.g. inflammatory response, IL2-STAT5 signaling, TNF-α signaling via NF-кB, IL6-JAK-STAT3 signaling, allograft rejection, and complement), crucial oncogenic pathways (e.g. mTORC1, MYC, PI3K-AKT-mTOR, and Hedgehog signalings), and proliferation processes (e.g. G2M checkpoint, E2F targets, mitotic spindle), while the low HIF-1α score group the lowest. However, metabolic processes (e.g. oxidative phospholation, xenobiotic metabolism, bile acid metabolism, and peroxisome), on the contrary, were notably suppressed in tumors with HIF-1α scores, but markedly enriched in tumors with low HIF-1α scores (Fig. 6A).

Fig. 6
figure 6

Association between HIF-1α score and oncologic biological processes and hypoxia status in HCC. A Heatmap to exhibit 27 differential crucial oncologic biological pathways overlapped among the three HIF-1α score groups. B–D Stepwise increasing Buffa (B), Ragnum (C), and Winter (D) hypoxia scores were found from low to high HIF-1α score group. (E–G) HIF-1α score significantly associated with Buffa (E), Ragnum (F), and Winter (G) hypoxia score systems. H–K KM survival 6curves for OS of HCC patients according to the HIF-1α (H), Buffa (I), Ragnum (J), and Winter (K) score systems in the TCGA cohort. L–N ROC curves for 1- (L), 3- (M), and 4- (N) year OS prediction based on the HIF-1α, Buffa, Ragnum, and Winter score systems in the TCGA cohort

Full size image

Comparison of our HIF-1α score system with other previously reported hypoxia score systems

Of note, stepwise increasing Buffa, Winter and Ragnum hypoxia scores were observed from low to high HIF-1α score groups (Fig. 6B–D). These findings verified our HIF-1α score system a reflection of the hypoxia status in HCC. Besides, the HIF-1α score system notably associated with these previous reported hypoxia score systems (Buffa: Cor = 0.55, P < 0.0001; Ragnum: Cor = 0.61, P < 0.0001; Winter: Cor = 0.59, P < 0.0001) (Fig. 6E–G). Compared with these previously reported hypoxia score system, KM survival analyses indicated that our HIF-1α score system exhibited the best stratification ability of prognosis (Fig. 6H–K). In addition, the ROC analyses also suggested that the HIF-1α score system possessed the optimal predictive capability for 1-, 3-, and 4-year survival in HCC (Fig. 6L–N). Taken together, these findings suggested the superior predictive value of the HIF-1α score system for patient prognoses.

Tumors with high HIF-1α scores were associated with high non-response rate to TACE

GSE104580 cohort was utilized to investigate the predictive capability of the HIF-1α score system for therapeutic response to TACE. Then, stepwise increasing portions of TACE-non-response HCC patients were found from low to high HIF-1α score groups (Fig. 7A). Higher HIF-1α scores were found in TACE-non-responsive tumors when compared with TACE-responsive tumors (Fig. 7B). Besides, we evaluated the predictive capability of the HIF-1α score system for therapeutic response to TACE, the AUC of which was 0.793 (95%CI 0.719–0.868) (P < 0.0001) (Fig. 7C). These findings suggested that the HIF-1α score system performed well in the prediction of therapeutic response to TACE.

Fig. 7
figure 7

Tumors with high HIF-1α score exhibited high non-response rate to TACE. A Stepwise increasing portions of TACE-non-response HCC patients were found from low to high HIF-1α score groups. B Higher HIF-1α scores were found in TACE-non-responsive tumors when compared with TACE-responsive tumors. C ROC curve analysis to assess the diagnostic capability of the HIF-1α score system for the prediction of therapeutic response to TACE

Full size image

Tumors with high HIF-1α scores correlated with immunosuppressive phenotype and high non-response rate to immune checkpoint blockade

The 18 immune-related signatures were shown in Fig. 8A, including macrophages, neutrophils, myeloid-derived suppressor cells (MDSC), gamma delta T cells, regulatory T cells (Tregs), T follicular helper cells, plasmacytoid dendritic cells, activated dendritic cells, immature dendritic cells, immature B cell, activated B cell, natural killer cells, cytolytic activity, activated CD8+ T cells, type 17 T-helper cells (Th17 cells), type 1 T-helper cells (Th1 cells), type 2 T-helper cells (Th2 cells), and Th2/Th1. Stepwise increasing infiltrations of pro-tumor immune cells were observed from low to high HIF-1α score groups, including MDSC, Tregs, Th2 cells, T follicular helper cells, and dendritic cells (Plasmacytoid, activated and immature) (Fig. 8B). In addition, tumors with high HIF-1α score possessed increased enrichment level of Th2/Th1 (Fig. 8B).

Fig. 8
figure 8

Tumors with high HIF-1α score notably associated with immunosuppressive phenotype and high non-response rate to immune checkpoint blockade. A Heatmap to exhibit the 18 immune-related terms in the three HIF-1α score groups. B Enrichment levels of MDSC, Tregs, Th2 cells, T follicular helper cells, and dendritic cells (Plasmacytoid, immature, and activated) in the three HIF-1 score groups. C Stepwise increasing portions of ICB-non-response HCC patients were found from low to high HIF-1α score groups. D Higher HIF-1α scores were found in ICB-non-responsive tumors when compared with ICB-responsive tumor. E The IPS-CTLA-4 and IPS-PD-1/PD-L1/PD-L2 were significantly lowest in patients in the high HIF-1α score group. F–K Higher HIF-1α scores significantly associated with higher PD-1 (F), PD-L1 (G), CTLA-4 (H), B7-H3 (I), CD47 (J), and CD80 (K) expressions in HCC. L–Q Tumors in high HIF-1α score group exhibited highest expressions of PD-1 (L), PD-L1 (M), CTLA-4 (N), B7-H3 (O), CD47 (P), and CD80 (Q), while those in the low-HIF-1α score group exhibited the lowest

Full size image

Then, ImmuneCellAI (http://bioinfo.life.hust.edu.cn/ImmuCellAI/#!/) was utilized for the prediction of therapeutic response to immune checkpoint blockade (ICB) [36]. Stepwise increasing portions of ICB-non-response HCC patients were found from low to high HIF-1α score groups (Fig. 8C). Higher HIF-1α scores were found in ICB-non-responsive tumors when compared with ICB-responsive tumors (Fig. 8D). To further verify the predictive ability of HIF-1α score system in patients’ response to ICB, we investigated the differences in immunophenoscore (IPS) among various HIF-1α score groups, which was obtained from the Cancer Immunome Atlas (https://tcia.at/) and used to predict the efficacy of ICB [37]. Two subtypes of IPS, including IPS-CTLA-4 and IPS-PD-1/PD-L1/PD-L2, were selected for further analysis as the surrogates of HCC patients’ response to anti-CTLA-4 and anti-PD-1/PD-L1/PD-L2 treatment. The IPS-CTLA-4 and IPS-PD-1/PD-L1/PD-L2 were significantly lowest in patients in the high HIF-1α score group (Fig. 8E). Besides, HIF-1α scores positively correlated with PD-1, PD-L1, CTLA-4, B7-H3, CD47, and CD80 expressions, which are crucial for the immunosuppressive phenotype (Fig. 8F–K) [38,39,40,41,42,43]. And tumors with high HIF-1α score possessed increased expressions of PD-1, PD-L1, CTLA-4, B7-H3, CD47, and CD80 (Fig. 8L–Q). Taken together, our findings suggest an immunosuppressive status and high non-response rate to ICB in tumors with high HIF-1α score, identifying the HIF-1α score system as a potential indicator of the immunogenic profiles in HCC.

Discussion

The current study classified HCC patients in the TCGA cohort into three HIF-1α clusters based on the 16 reported HIF-1α related genes. Then, we also established an HIF-1α score system based on nine prognostic ODEGs among various HIF-1α clusters, which exhibited satisfactory predictive capability for survival. Furthermore, compared with other previous hypoxia score systems, our model possessed the optimal predictive efficacy for OS in HCC. Putting these together, we constructed a novel HIF-1α score system for HCC, which works well in prognostic prediction and hypoxic status evaluation of HCC patients.

Some of the nine genes involved in the HIF-1α score system have been reported to play crucial roles in liver carcinogenesis. Previous study by Zhu et al. demonstrated that ADAMTS5 is an unfavorable prognostic biomarker with a potential role in extracellular matrix remodeling and metabolic regulation in HCC, and could be a promising therapeutic target for patients with HCC [44]. HOMER1 was identified as an unfavorable prognostic biomarker associated with glycolysis in HCC, but its potential oncologic role in HCC still remains further works [45]. LPCAT1 overexpression promotes the malignant phenotypes of HCC through directly inhibiting STAT1 [46]. NEIL3 overexpression enables the repair of oxidative DNA damage at telomeres in mitosis, which is critical to prevent senescence of HCC cells [47]. The oncologic role of PFN2 has been reported in various malignances, but little is known in HCC, and for the first time, we identified PFN2 as a novel unfavorable prognostic biomarker in HCC [48,49,50,51]. Upregulation of S100A9 by HBx-mediated NF-κB activation promotes tumor progression in HCC [52]. Notably, it has been reported that the EGLN3 is recruited by HIFAL to crosstalk with PKM2, thereby enhancing HIF-1α transactivation [53]. SPP1 overexpression promotes HCC progression and metastasis by activating CCR1 expression via the PI3K-AKT-HIF-1α signaling pathway [54]. SLC1A5 is a critical glutamine transporter in malignancies, and activation of which captures huge amounts of glutamine [55, 56]. It should be noted, based on data mining, that all the nine genes utilized in this model should be HIF-1α related genes, though most of which have not been found to be modulated by HIF-1α. Therefore, further works should be conducted to explore how HIF-1α modulates or interacts with these genes.

In this study, HCC with high HIF-1α scores are considered much more aggressive, due to the notable associations with shorter OS, more advanced TNM stage, histologic grade and vascular invasion. By investigating the functional annotations among HIF-1α high, medium, and low score tumors, MYC signaling, mTORC1 pathway, and Hedgehog signaling were found significantly enriched in tumor with high HIF-1α score. MYC has been reported to be one of the major drivers of HCC initiation and progression [57, 58]. MYC-upregulated tumor cells were found to have enhanced glucose and glutamine utilization and overexpression of critical glycolytic and glutaminolytic enzymes [59,60,61]. And mTORC1 pathway has also been identified as a critical oncogenic signaling pathway for tumor cell growth, survival, and metabolism [62,63,64]. Notably, mTORC1 pathway has been found to enhance the transcription and translation of HIF-1α, and could serves as a central activator of the hypoxia response (e.g. Warburg effect) by inducing multiple signaling pathways [65,66,67]. Sun et al. also demonstrated that dual inhibition of both mTOR and glycolysis synergistically impairs the proliferation and tumorigenesis of mTOR hyperactive cells [68]. Hedgehog signaling facilitates the metastasis and drug resistance of HCC in combination with other oncogenic pathways, such as MYC pathway and TGF-β signaling [69,70,71]. Taken together, these findings demonstrate that MYC, mTORC1 pathways, and Hedgehog signaling are crucial oncogenic drivers in high HIF-1α score tumors, which might provide a preliminary theoretical basis for combination treatment using MYC inhibitor and HIF-1α inhibitor, or using mTORC1 inhibitor and HIF-1α inhibitor, or using Hedgehog signaling inhibitor and HIF-1α inhibitor in these highly aggressive and hypoxic HCCs.

The current study indicated that the HIF-1α score system performed well in the prediction of therapeutic response to TACE, in which tumors with high HIF-1α score represented high non-response rate to TACE. Consistently, Wei et al. revealed that activation of HIF-1α/pAKT loop leads to HCC non-response to TACE [72]. Moreover, previous study by Martin et al. also showed that hypoxia signaling and glycolysis-related pathways are upregulated in HCCs insensitive to TACE, and targeting PKM2, an HIF-1α-related gene, could improve the response rate to TACE [73]. Thus, for tumors with high HIF-1α, strategies to inhibit HIF-1α signaling or glycolysis-related pathway may be helpful for recovering or improving the efficacy of TACE.

In addition to activation of multiple critical oncogenic pathways, we revealed that HCCs with high HIF-1α scores were more immunosuppressive and higher non-response rate to ICB (e.g. PD-1/PD-L1 and CTLA-4 inhibition). In particular, tumors with high HIF-1α scores possessed high pro-tumor immune cells infiltrations, especially for MDSC and Tregs. These tumors had high expression of important immune checkpoint proteins, including PD-1, PD-L1, CTLA-4, B7-H3, CD47, and CD80. Previous studies revealed that MDSCs exhibit immunosuppressive activities in TME, facilitating immune escape and non-response to ICB of tumor cells [74,75,76,77]. MDSCs were also able to facilitate the pro-tumor activities of Treg cells, thereby dampening anti-tumor immunity of effective T cells in HCC [78, 79]. Chiu et al. demonstrated that HIF-1α increases the transcription of CCL26 in tumor cells to recruit MDSCs to the hypoxic regions of tumor [80]. Notably, tumors with high HIF-1α scores also showed significantly upregulations of TNF-α signaling via NF-кB and IL6-JAK-STAT3 signaling, which were critical for the accumulation and immunosuppressive capacity of MDSCs [81,82,83,84,85,86,87]. Targeting MDSC trafficking has been reported to enhance anti-PD-1 efficacy, representing a promising therapeutic strategy against HCCs [74, 80, 82]. It is also reported that hypoxia-induced HIF-1α upregulates PD-L1 expression in several solid tumors, which was consistent with our findings that tumors with higher HIF-1α score exhibited higher expression PD-L1 [88]. Taken together, HCCs with high HIF-1α scores exhibited a highly immunosuppressive TME. And low/medium HIF-1α score tumors seem to be more suitable for immunotherapy. For highly immunosuppressive tumors, treatment strategies to restrain pro-tumor immunity by targeting HIF-1α or suppressing desmoplasia or hypoxia may be promising for enhancing the effect of immunotherapy.

Conclusions

In conclusion, our study developed a novel HIF-1α score system to discriminate HCC with different hypoxia status and immunogenic cold TMEs, which exhibited reliable predictive efficacy for patient prognoses, TACE therapeutic response, and immunotherapy response. Besides, our study present the distinctive functional and immunogenic profiles for tumors with high HIF-1α scores and low/medium HIF-1α scores, which might provide preliminary rationales for individual treatment options of these HCC patients.