A novel signature derived from immunoregulatory and hypoxia genes predicts prognosis in liver and five other cancers

HCC data sets (n = 839) used in this study were GSE14520 (n = 242) from the Liver Cancer Institute, LIRI-JP (n = 226) from the International Cancer Genome Consortium (ICGC) database and TCGA-LIHC (n = 371) from the Cancer Genome Atlas (TCGA) Research Network (Additional file 1) [11]. Full clinical characteristics of HCC patients were listed in Additional file 2. Among these patients, underlying etiology differed between the cohorts with GSE14520 comprising of patients with HBV, while LIRI-JP and TCGA-LIHC consisting of patients with HBV, HCV and/or non-alcoholic steatohepatitis.

Gene expression data for 24 other cancer types (n = 20,241) generated by TCGA Research Network [11] were downloaded from Broad Institute GDAC Firehose. Illumina HiSeq rnaseqv2 Level 3 RSEM normalized gene expression profiles were converted to log₂(x + 1) scale. Expression values of genes associated with tumor-infiltrating T cells were corrected for levels of infiltration by dividing expression values of each gene with CD3D values.

To determine differentially expressed genes between tumor and adjacent normal liver tissues in the GSE14520 cohort, the Bayes method and linear model were employed using the R package limma. P-values were adjusted using the false discovery rate controlling procedure of Benjamini–Hochberg. Genes with log₂ fold change of > 1 and adjusted P-values < 0.05 were considered significant.

Expression scores for hypoxia, tumor-infiltrating T cells, 45-gene signature and 8-gene signature were calculated for each patient in all the cohorts by taking the average log₂ expression values. Nonparametric Mann–Whitney–Wilcoxon test was used to compare the distribution of hypoxia scores in tumor and adjacent non-tumor samples. Risk scores for each patient were determined by taking the sum of Cox regression coefficient for each of the signature genes multiplied with its corresponding expression value. Nonparametric Spearman’s rank correlation was employed to assess the relationship of gene scores and risk scores with tumor hypoxia (hypoxia score).

We employed the Cox proportional hazards model to investigate the association between patient survival duration and one or more risk factors, e.g. 45-gene or 8-gene score, tumor stage and other clinical variables. Univariate analysis was performed to determine the influence of individual risk factors on overall survival. As multiple covariates can potentially influence patient prognosis, multivariate analyses were performed by including risk factors that were significantly associated with overall survival identified in univariate analysis (P < 0.05). Hazard ratios (HR) were determined from Cox models with HR greater than one indicating that a covariate was positively associated with event probability (increased hazard) and negatively associated with survival length. Cox regression analyses were performed using the R survival and survminer packages. Proportional hazards assumption was supported by a non-significant relationship between scaled Schoenfeld residuals and time using the R survival package. In addition, Kaplan–Meier and log-rank tests were used in univariate analyses of the gene signatures in relation to patient survival and were performed using the survival and survminer packages. Patients were median-dichotomized into low and high-risk groups based on mean expression scores of signature genes and Kaplan–Meier estimates of survival probability over time were generated. Difference between high and low-risk groups were tested using the log-rank test.

Time-dependent receiver operating characteristic (ROC) curve analysis was used to assess the predictive performance of both 45-gene and 8-gene signatures in comparison with standard tumor staging parameters. The R survcomp package was employed to compute time-dependent ROC curves. ROC curves depicted true positive rates (sensitivity) versus false positive rates (1-specificity). The area under the curve (AUC) is a measure of how well the gene signatures can predict patient survival where AUC ranges from 0.5 (for an uninformative marker) to 1 (for a perfect predictive marker).

To determine whether the gene signatures (45-gene or 8-gene signature) can distinguish tumor and non-tumor samples, multidimensional scaling (MDS) analyses were performed using the R vegan package to visualize samples’ distance in the reduced 2-dimensional space. Euclidean genetic distances between each sample were investigated by MDS ordination. Permutational multivariate analysis of variance (PERMANOVA) was employed to test for differences between tumor and non-tumor samples.

Analysis of biological pathway enrichment on the 45-gene set was conducted using GeneCodis [12] against the Kyoto Encyclopedia of Genes and Genomes (KEGG) and Gene Ontology (GO) databases. The Enrichr [13] tool was used to identify transcription factors from the ChEA database that are potential regulators of the 45 genes. Functional protein association network of the 45 genes was determined using the STRING [14] database.

Level 3 mutation datasets were downloaded from GDAC. Kaplan–Meier analysis and log-rank tests were employed to determine the association of somatic mutations in combination with the 45-gene or 8-gene signatures, on overall survival.

All graphs were generated using the ggplot2 package in R. Heatmap was generated using the R pheatmap package.

Since both hypoxia and high density of infiltrating Tregs are linked to poor clinical outcomes, we hypothesized that these features could be employed to predict prognosis in patients with hepatocellular carcinoma (HCC). Hypoxia inducible factors (HIFs) are transcription factors (TFs) that play key roles in regulating cellular responses to hypoxia [15]. The alpha subunits (HIF-1α and -2α) heterodimerize with the beta subunit (HIF-1β) to orchestrate physiological responses in hypoxia [16]. To develop our gene signature uniting hypoxia and tumor-infiltrating T-cells, we utilized three gene sets: (1) pan-cancer hypoxia genes (52 genes) [6]; (2) HCC-infiltrating Tregs (401 genes) and (3) HCC-infiltrating exhausted CD8⁺ T cells (82 genes) [8] (Fig. 1). We utilized a HepG2 hepatoma HIF chromatin immunoprecipitation sequencing (ChIP-seq) dataset, GSE120885 [17], to determine which of these genes were bound by HIFs (Fig. 1).

To determine the extent of hypoxia in cancers, we interrogated in vivo mRNA expression patterns of the 52 hypoxia genes [6] in 25 cancer types including HCC (n = 20,662) retrieved from The Cancer Genome Atlas (TCGA) [11] (Additional file 1). Hypoxia scores for each tumor and non-tumor samples were determined by obtaining the mean expression values (log₂) of the 52 hypoxia genes [6]. We observed that tumor samples were significantly more hypoxic than non-tumor samples in 20 out of 25 cancers, which included HCC (Additional file 3A). Multidimensional scaling analyses revealed that the 52 genes can distinguish tumor from non-tumor samples in these cancers, hence hypoxic transcriptional states can be used as a proxy for identifying cancerous cells (Additional file 3B).

We predict that patients with more hypoxic tumors would have higher expression of tumor-infiltrating T cell genes since hypoxia could promote the maintenance of immunological escape via the suppressive function of Tregs [10]. The HCC-infiltrating T cell gene set uniting both Tregs and exhausted CD8⁺ T cells consisted of 438 genes collectively, with 45 genes found to be in common between the two cell types (Fig. 1). Infiltrating T cell expression scores for each patient were calculated as mean expression values (log₂) of the 438 genes (Fig. 1). Patient hypoxia scores significantly correlated with tumor-infiltrating T cell scores in three independent HCC cohorts (n = 839; Additional file 2): GSE14520 (ρ = 0.564, P < 0.0001); TCGA-LIHC (ρ = 0.390, P < 0.0001) and LIRI-JP (ρ = 0.633, P < 0.0001) (Additional file 3C), suggesting that highly hypoxic tumors have transcriptional profiles associated with immunosuppressive functions.

Since HIFs are the key TFs that regulate numerous important aspects of oncogenesis [18], we hypothesize that identifying genes that are directly bound by HIFs may have more profound implications on prognosis. Using the HIF ChIP-seq dataset, we observed that 79 of the 438 tumor-infiltrating T cell genes were direct HIF targets (Fig. 1). Notably, Spearman’s correlation coefficients between the 79 genes and hypoxia score were significantly higher: GSE14520 (ρ = 0.669, P < 0.0001), TCGA-LIHC (ρ = 0.469, P < 0.0001) and LIRI-JP (ρ = 0.694, P < 0.0001) (Additional file 3D), suggesting a role for HIFs in regulating tumor lymphocytic infiltration in HCC.

Levels of T cell infiltration were corrected by dividing the expression of each of the 79 genes with CD3D expression. Of the 79 genes, 26 were upregulated more than 1.5-fold in the training cohort, GSE14520 (Additional file 3E). On the other hand, ChIP-seq analysis revealed that 23 of the 52 hypoxia genes mentioned earlier were direct HIF targets (Fig. 1). The final gene set was comprised of 45 genes that encompass 26 tumor-infiltrating lymphocyte genes and 23 hypoxia genes, with 4 genes being in common (Fig. 1; Additional file 4).

We next assessed the ability of the 45-gene signature (HCC45) to predict overall survival (OS) in three independent HCC cohorts (n = 839). The signature can successfully discriminate between tumor and non-tumor samples in these cohorts (Additional file 5A). Tumor samples were notably less tightly clustered implying significant intertumoral heterogeneity (Additional file 5A). We next determined HCC45 scores, calculated for each patient as the mean expression of the 45 genes. Patients were ranked according to their HCC45 scores and divided into high- or low-risk groups using the median cutoff. The OS rates were significantly reduced in high-risk patient groups with log-rank P values of 0.012, 0.0042 and 0.0043 in the GSE14520, TCGA-LIHC and LIRI-JP cohorts respectively (Fig. 2a). Disease-free survival rates were also significantly lower in high-risk patients (P = 0.026) (Additional file 5B).

To test the independence of HCC45 over current staging systems, we performed subgroup analysis by applying the signature to patients within each tumor, node, metastasis (TNM) stages. The signature successfully identified high-risk patients, particularly among early-stage (1 and/or 2) patients (Fig. 2b). To evaluate the predictive performance of HCC45 on 5-year OS, we performed the receiver operating characteristic (ROC) analysis in comparison with tumor staging parameters. The area under the curves (AUC) for HCC45 were 0.759 (GSE14520), 0.804 (TCGA-LIHC) and 0.883 (LIRI-JP) (Fig. 2c). All HCC45 AUCs were higher than that of TNM staging (Fig. 2c). Moreover, combining HCC45 with TNM staging considerably improved its predictive ability: AUC = 0.800 (GSE14520), 0.830 (TCGA-LIHC) and 0.890 (LIRI-JP) (Fig. 2c).

We assessed two common mutations in HCC, TP53 and CTNNB1. Mutation in the tumor suppressor TP53 is associated with poor prognosis [19, 20]. Results on β‐catenin CTNNB1 mutations have been mixed; some demonstrating poor outcomes [21, 22], while others demonstrating favorable prognosis [23]. Both TP53 (P < 0.001) and CTNNB1 (P = 0.025) mutations were associated with poor prognosis (Table 1). Mortality in patients with high HCC45 scores and TP53 mutation were ~ 45% higher than in patients with low HCC45 scores harboring wild-type TP53 at 2 years (Fig. 2d). Likewise, joint relation of HCC45 and CTNNB1 predicted a ~ 35% increase in mortality at 4 years between the lowest and highest risk groups (Fig. 2d).

We next examined the relation of HCC45 to other clinical parameters and observed that it remained highly prognostic after adjustment for tumor size, cirrhosis, fibrosis, TNM stages, Barcelona Clinic Liver Cancer stages, vascular invasion, TP53/CTNNB1 mutation status and/or alpha-fetoprotein levels in a multivariate Cox regression analysis: GSE14520 hazard ratio [HR] (HR = 1.89, P = 0.0052), TCGA-LIHC (HR = 1.61, P = 0.033) and LIRI-JP (HR = 2.20, P = 0.024) (Table 1). Both cirrhosis and HCC45 harbor complementary prognostic information contributing to elevated HR to 4.59 (95% CI 1.11–19.03, P = 0.036) (Table 1).

Gene Ontology analysis of HCC45 revealed enrichment of biological pathways linked to hypoxia, metabolism, inflammation and cancer (Additional file 6). Moreover, HCC45 was enriched for targets of several well-known cancer-related transcription factors such as MYC and RUNX1 (Additional file 6) and were functionally connected as a group (protein–protein interaction network enrichment: P < 1e−16) (Additional file 7).

Having identified that HCC45 predicts outcome in HCC, we evaluated the performance of each gene as a prognostic factor (Fig. 1). Of the 45 genes, univariate Cox regression analysis revealed that 14, 19 and 28 genes were significantly associated with unfavorable prognosis in GSE14520, TCGA-LIHC and LIRI-JP respectively (Figs. 1; 3a). Eight genes (HCC8) were found to confer prognostic information in all HCC cohorts: CA9, CCL20, CORO1C, CTSC, LDHA, NDRG1, PTP4A3 and TUBA1B (Fig. 3a; Additional file 8). Expression of HCC8 not only distinguished tumor from non-tumor sample (Additional file 9) but also increased according to tumor progression (Additional file 10). As expected, the signature successfully identified high-risk patients in full HCC cohorts, GSE14520 (P = 0.0001), TCGA-LIHC (P = 0.00016) and LIRI-JP (P = 0.0003) cohorts, and in patients stratified by tumor stage (Fig. 3b). The 5-year disease-free survival rates were also lower in high-risk patients as determined by the 8-gene signature (Additional file 11). Furthermore, as measured by Wald Chi square statistics and log-rank tests, HCC8 is superior in predicting death than HCC45 (Figs. 2a; 3b; Table 1).

Multivariate models of HCC8 revealed that this reduced gene set sufficiently served as an independent prognostic risk factor: GSE14520 (HR = 1.72, P = 0.02), TCGA-LIHC (HR = 2.23, P = 0.0009) and LIRI-JP (HR = 2.58, P = 0.012) (Table 1). Predictive value for 5-year OS increased when the HCC8 signature was used in combination with TNM staging: GSE14520 AUC = 0.67 versus 0.73 and TCGA-LIHC AUC = 0.73 versus 0.76 (Fig. 3c). When TP53 or CTNNB1 mutation status and HCC8 were jointly used, patients with high HCC8 levels and mutant TP53 (P < 0.0001) or CTNNB1 (P < 0.0001) had the lowest chances of survival (Fig. 3d).

We calculated risk scores for each patient by taking the sum of Cox regression coefficient for each of the 8 genes multiplied with its corresponding expression value [24]. Risk scores significantly correlated with hypoxia scores in all 3 cohorts, indicating that patients with more hypoxic tumors have higher risk for death as predicted by HCC8 (Additional file 12), overall suggesting that it was sufficient and effective in predicting death.

Hypoxic tumors recruit CD4⁺ Tregs to suppress effector T cell function and promote tumor tolerance [10]. We hypothesized that HCC8 could predict outcome in other solid malignancies in addition to HCC. HCC8 could indeed discriminate between tumor and non-tumor samples in five other cancer types (Additional file 13). Remarkably, our prognostic 8-gene signature derived from HCC was a significant adverse prognostic factor for OS in these five cancer types: head and neck (HR = 1.64, P = 0.004), renal papillary cell (HR = 2.31, P = 0.047), lung (HR = 1.45, P = 0.03), pancreas (HR = 1.96, P = 0.006) and endometrium (HR = 2.33, P = 0.003) (Fig. 4a; Additional file 14). Multivariate analysis adjusting for tumor stage revealed that the HCC8 signature retained independent prognostic relevance in these cancers: head and neck (HR = 1.66, P = 0.003), lung (HR = 1.45, P = 0.029), pancreas (HR = 1.91, P = 0.009) and endometrium (HR = 1.99, P = 0.016) (Additional file 14). Importantly, while TNM staging was not a reliable predictor of OS in pancreatic cancer on its own, our 8-gene signature successfully predicted high-risk patients when accounting for tumor stage (P = 0.009) (Additional file 14).

The predictive value of the 8-gene signature outperformed the TNM staging system: head and neck (AUC = 0.610 versus 0.606), renal papillary cell (AUC = 0.701 versus 0.640) and pancreas (AUC = 0.694 versus 0.593) (Fig. 4b). When used in conjunction with staging information, additive effects on the predictive performance of HCC8 were observed: head and neck (AUC = 0.661), renal papillary cell (AUC = 0.768), lung (AUC = 0.671) and pancreas (AUC = 0.701) (Fig. 4b). Importantly, risk scores derived from HCC8 showed strong positive trends with tumor hypoxia, suggesting that high risk patients had more aggressive tumors: head and neck (ρ = 0.729, P < 0.0001), renal papillary cell (ρ = 0.766, P < 0.0001), lung (ρ = 0.856, P < 0.0001), pancreas (ρ = 0.844, P < 0.0001) and endometrium (ρ = 0.374, P < 0.0001) (Additional file 15).

TP53, KRAS and CDKN2A mutations are commonly observed in lung cancer [25]. CDKN2A (P = 0.03), but not TP53 (P = 0.2) or KRAS (P = 0.4), was significantly associated with poor prognosis (Additional file 14). When used in combination with our HCC8 signature, patients with high HCC8 score and mutant CDKN2A had ~ 30% higher chances for death at 2 years than patients with low HCC8 and wild-type CDKN2A (Additional file 16).