A signature of hypoxia-related factors reveals functional dysregulation and robustly predicts clinical outcomes in stage I/II colorectal cancer patients

The gene expression profiles of CRC tissue samples obtained from six public cohorts, including 309 CRC patients from the CIT/GSE39582 gene microarray dataset that served as the discovery cohort, were retrospectively analyzed. The two largest individual data sets—CIT/GSE39582 and The Cancer Genome Atlas (TCGA)—were used for training and independent validation. The meta-validation cohort consisted of the remaining four microarray data sets—GSE14333, GSE17536, GSE37892 and GSE33113—which were obtained from the Gene Expression Omnibus (GEO) database. All datasets are from the GPL570 platform ([HG-U133_Plus_2] Affymetrix Human Genome U133 Plus 2.0 Array). TCGA cohort data was downloaded from Broad GDAC Firehose, and the other data sets were obtained directly in their processed format from the GEO database through Bioconductor package ‘GEOquery’. Transcripts per million (TPM) of level 3 RNA-Seq data in log2 scale was applied to calibrate the gene expression levels in TCGA cohort. The ‘combat’ algorithm of the R package ‘sva’ and the z-scores were used to correct the batch effects, in order to standardize microarray data across multiple experiments and compare them independent of the original hybridization intensities. The data of 1877 CRC patients enrolled from Sep 27 to Dec 26, 2018 was also included.

To construct a prognostic HGS, annotated functional database MSigDB (version 6.2) [13‐15] was used to identify a list of hypoxia-related genes with the keyword “hypoxia”, and the HGSs measured by all platforms were selected. The log-rank test was used with 1000 randomizations (80% of samples each time) to evaluate the association between each HGS and clinical outcome in the training dataset. Genes that were repeatedly significant were selected as the candidates of the hypoxia signature. To minimize the risk of over-fitting, Cox proportional hazards regression model was applied with the least absolute shrinkage and selection operator (LASSO) (glmnet, version 2.0-16). The penalty parameter was estimated by tenfold cross-validation in the training data set at 1 SE beyond the minimum partial likelihood deviance.

A time-dependent receiver operating characteristic (ROC) curve (survival ROC, version 1.0.3) at 5 years was plotted using Kaplan–Meier estimation, and used to determine the optimal HGS cutoff to separate patients in the training data set into the low-risk and high-risk groups. The HGS corresponding to the shortest distance between the ROC curve and the point representing 100% true positive rate and 0% false-positive rate was used as the cutoff value. Univariate analysis was used to evaluate the prognostic value of the HGS in stage I/II CRC patients, and in patients at all stages in the training and independent validation cohorts. In the multivariate analyses, HGS was combined with other clinical and pathological variables.

To investigate the biological characteristics of the HGS, enrichment analysis was conducted for differentially expressed genes (DEGs) between the risk groups in TCGA CRC data set using R package ‘gProfileR’. Gene Set Enrichment Analysis (GSEA) was further performed using Bioconductor package ‘HTSanalyzeR’ to predict the significant pathways [16].

All statistical analysis was performed in R software (version 3.5.1; http://​www.​Rproject.​org). Descriptive statistics were computed for all variables, and expressed as mean ± standard deviation (SD) or medians and interquartile ranges (IQR) for continuous factors, and as frequencies for categorical factors. Continuous values were compared using Student-t tests between different groups. Log-rank test was used to evaluate results of the univariate analysis of HGS and other clinico-pathological factors with disease free survival (DFS). Multivariate analysis was performed with the Cox proportional hazards regression model. The C-index was calculated by ‘survcomp’ (version 1.32.0). P values less than 0.5 were considered statistically significant.

We analyzed the CIT gene microarray dataset (GSE39582) and created the discovery subset with 309 eligible CRC patients (Fig. 1). After exclusion of genes with MAD > 0.5 and less median expression, 1636 genes were retained for further analysis. Following selection of 80% of the repeatable genes via 1000 random Cox univariate regressions, we identified 106 genes that were associated with DFS, of which 14 hypoxia-related genes were selected to construct the HGS using LASSO Cox regression for stage I/II CRC patients (Fig. 2). The risk scores were calculate using the formula derived from the Cox model as follows: Risk score = − 0.013 × exp(mRNA expression level of MDM2) + 0.0733 × exp(mRNA expression level of VEGFA) + 0.112 × exp(mRNA expression level of ORAI3) + 0.043 × exp(mRNA expression level of MVD) − 0.060 × exp(mRNA expression level of TRAF3) − 0.003 × exp(mRNA expression level of CYB5R3) − 0.003 × exp(mRNA expression level of ZBTB44) − 0.045 × exp(mRNA expression level of CASP6) + 0.082 × exp(mRNA expression level of FBP1) − 0.026 × exp(mRNA expression level of CCNG1) − 0.032 × exp(mRNA expression level of FAM117B) − 0.025 × exp(mRNA expression level of PRELID2) − 0.129 × exp(mRNA expression level of RRP1B) + 0.014 × exp(mRNA expression level of GAS6). Based on time-dependent ROC curve analysis, the optimal cutoff of HGS for stratifying patients in the training set into the high and low risk groups was determined to be a satisfactory RFS cutoff at 5 years (Fig. 3b, e and h). The incidence of tumor recurrence was higher among the patients in the high-risk group compared to the low-risk group when the entire CIT dataset (n = 566) was used as a training cohort (Fig. 4a, P < 0.001).

The prognostic significance of HGS was assessed with additional CRC transcription data sets that included clinical and prognostic data. Clinicopathological characteristics of three cohorts are listed in Table 1. The validation datasets consisted of TCGA datasets (n = 624) and the meta-validation cohort (n = 687), including GSE17536, GSE33113, GSE37892 and GSE14333. No significant difference was seen between the clinico-pathological features of the training and validation cohorts (Table 2). The DFS was significantly higher in the low-risk compared to the high-risk HGS group in all three cohorts (training cohort: HR = 4.35, 95% CI 2.30–8.23, P < 0.001; validation: HR = 2.14, 95% CI 1.09–4.21, P = 0.024 and meta-validation cohort: HR = 1.91, 95% CI 1.08–3.39, P = 0.024) (Fig. 3c, f and i). We compared HGS with Oncotype DX to further evaluate its prognostic value and robustness (Table 3), and found that HGS had a more optimized C-index in both training and TCGA cohorts (training cohort: 0.80 vs 0.65, TCGA cohort: 0.70 vs 0.61, Table 3). Further data mining indicated a prognostic value of HGS in all CRC cohorts (GSE39582 cohort: HR = 2.01, 95% CI 1.46–2.77, P < 0.001; TCGA cohort: HR = 1.72, 95% CI 1.14–2.62, P = 0.010; meta-validation cohort: HR = 1.94, 95% CI 1.34–2.8, P < 0.001) (Fig. 4c, f and i). Similar results were obtained in the AUC analysis (Fig. 4b, e and h).

Univariate and multivariate analyses were used to determine whether patient age, gender, tumor stage, tumor location, pathological gene status and HGS were associated with prognosis in stage I/II CRC patients. The univariate analysis showed that HGS was significantly associated with a poor outcome in the three cohorts, (GSE39582 cohort: HR = 8.66, 95% CI 4.37–17.17, P < 0.001; TCGA cohort: HR = 2.59, 95% CI 1.08–6.25, P = 0.04; and meta-validation cohort: HR = 8.25, 95% CI 3.09–22.03, P < 0.001, Table 2). After adjusting for other factors in the multivariate analysis, it remained an independent prognostic factor (GSE39582 cohort, HR = 7.54, 95% CI 3.78–15.06, P < 0.001; TCGA cohort, HR = 2.59, 95% CI 1.08–6.25, P = 0.04; and meta-validation cohort, HR = 7.25, 95% CI 2.72–19.29, P < 0.001, Table 2).

Gene ontology and KEGG pathway enrichment analysis of the DEGs and the GSEA showed a significant enrichment of metabolic pathways such as mTROC1 (P = 0.0001), G2-M (P = 0.0001), mitosis (P = 0.0001), oxidative phosphorylation (P = 0.0001), MYC (P = 0.0001), and PI3K–AKT–mTOR (P = 0.0039) (Fig. 5).