HIF-1 transcription activity: HIF1A driven response in normoxia and in hypoxia

The human SKNBE2 (ATCC #CRL-2271) cell line was grown in Dulbecco’s modified Eagle’s medium supplemented with 10% heat inactivated fetal bovine serum (Sigma), 1 mM L-glutamine, penicillin (100 U/ml) and streptomycin (100 μ g/ml) (Invitrogen), at 37 °C, under 5% CO2 in a humidified atmosphere. The cells exposed to hypoxia were grown at 0.5% oxygen for 2 h. The cells used for all the experiments were re-authenticated and tested as mycoplasma-free. Early-passage cells were used and cumulative culture length was less than 3 months after resuscitation.

To knock-down HIF1A expression, the pGIPZ lentiviral shRNAmir that targets human HIF1A were purchased from Open Biosystems (Thermo Fisher Scientific, Inc.). We used two different shRNAs against HIF1A: V2LHS_132152 (RHS4430–98513964) (shHIF1A#A) and V2LHS_236x718 (RHS443098513880) (shHIF1A#B). A non-silencing pGIPZ lentiviral shRNAmir was used as the control (RHS4346). The production of lentivirus particles and cells infection was performed as previously described [12]. To obtain 100% GFP-positive cells, puromycin was added into the medium for an additional 10 days.

Cell pellets were resuspended in a hypo-tonic buffer (10 mM HEPES-K +, pH 7.5, 10 mM KCl, 1.5 mM MgCl 2, 0.5 M dithiothreitol) in the presence of a protease inhibitors cocktail (Roche). The cells were lysed by addition of ice-cold 0.5% NP-40 for 10 min. The nuclei were pelleted at 1000 x g for 2 min at 4 °C and nuclear protein extraction and concentrations was determined as previously described [12]. Protein membranes were probed with anti-HIF-1α (610,959; BD Biosciences) and horseradish-peroxidase-conjugated anti-mouse secondary antibody (1:4000 dilution; ImmunoReagent). Positive bands were visualized using the ECL kit SuperSignal West Pico Chemiluminescent Substrate (Pierce). A rabbit anti-H3 antibody (ab1791 Abcam) was used as the control for equal loading.

Total RNA was isolated from NB cell line using TRIzol LS Reagent (Invitrogen) according to manufacturer’s instructions; samples quality and library construction is described in Additional file 1. cDNA Sequencing was accomplished using an Illumina HiSeq™ 2000 platform according to the manufacturer’s protocols (Analysis performed at BIOGEM facility). Illumina paired end sequencing protocol yielded about 20 millions of 2x101nt reads with high quality bases (mean quality of 34) and mean % GC of 46.

Sequencing data were analyzed with the set of open source programs of the Tuxedo suite: TopHat v2.0.14 (for sequence alignment) and Cufflinks v2.1.0 (for differential expression analysis), following the pipeline published in Nature Protocols by Trapnell et al., 2012 (see Additional file 1 for further details) [24]. The set of output files obtained by Cufflinks was inspected and explored using the R-Bioconductor package CummeRbund v2.16.0, which provides functions to read, subset, filter and plot results. We selected genes differentially expressed in each of the pairwise comparisons if the Benjamini Hochberg adjusted P-value (FDR) was under 0.05 and if the Log₂ transformed fold change was greater than + 0.5 (up-regulated) or lower than − 0.5 (down-regulated). The lists of these genes were used to query Pathway and Gene Ontology databases. The functional enrichment analysis tool Webgestalt (WEB-based GEne SeT AnaLysis Toolkit) was used to detect significant enrichments for Kyoto Encyclopedia of Genes and Genomes (KEGG) pathways. The enrichment analysis was performed using the following criteria: an hypergeometric test for statistical analysis, FDR ≤ 0.05 and 10 as minimum number of genes for a category. Data generated during this study are included in this manuscript and in Additional files 1, 2 and 3.

DNA extraction was carried out with the Wizard Genomic DNA Purification Kit (Promega, WI, USA), including a RNA removal step, according to the protocol provided by the supplier. The DNA was quantified with the Nanodrop and 1 μg was used for bisulfite modification using EZ-96 DNA Methylation™ Kit (Zymo Research CA, USA) with the modification step according to the recommendations for array processing of the samples. Control PCRs were carried out before array analysis to confirm successful modification of the DNA. The bisulfite-modified DNA (500 ng) was laid on the Infinium HumanMethylation450 BeadChips (Illumina), which determine the methylation levels of 485,000 CpG sites. The fluorescence signals were measured from the BeadArrays using an Illumina BeadStation GX scanner. The raw fluorescence images (IDAT files) were then analyzed using R and R-Bioconductor packages. The ChAMP package was used for data preprocessing, normalization and comparison between groups [25, 26]. Singular value decomposition analysis was performed to identify confounding factors and evaluate possible batch effects, while the SWAN (Subset Within Array Normalization) method was used for probe intensity normalization. After these steps the fraction of failed probes was about 0.003%.

ChAMP assigns a score called “β value” to each CpG site, which corresponds to the ratio between the fluorescence signal of the methylated allele (C) and the unmethylated (T) alleles. The β value, ranging from 0 to 1, represents the methylation status of each probe from totally unmethylated (β = 0) to totally methylated (β = 1). The software was used to calculate probes that were differentially methylated between groups. CpG sites were considered as differentially methylated, in a contrast, if Δβ (delta beta) was below − 0.2 (hypo-methylation) or above 0.2 (hyper-methylation) and the FDR was lower than 0.05. Data generated during this study are included in this manuscript and in Additional files 1, 2 and 3.

The expression levels of 13 genes were analyzed using real-time, quantitative PCR in SKNBE2 shHIF1A#A and shCTR cells. Total RNA extraction using TRIzol LS Reagent (Invitrogen) and cDNA retrotranscription using the High Capacity cDNA Reverse Transcription Script (Applied Biosistem) was performed according to the manufacturer protocol. The cDNA samples were diluted to 20 ng/μ l. Gene-specific primers were designed by using PRIMEREXPRESS software (Applied Biosystems) and primers sequences for each gene are listed in Additional file 1. Real-time PCR was performed using SYBR Green PCR Master Mix (AppliedBiosystems). All real-time PCR reactions were performed using the 7900HT Fast Real-Time PCR System (Applied Biosystems). The experiments were carried out in triplicate for each data point. The housekeeping gene β -actin was used as the internal control. Relative gene expression was calculated using the 2 ^−ΔΔCT method as described in our previous work [27], where the ∆CT was calculated using the differences in the mean CT between the selected genes and the internal control (β -actin). The mean fold change of 2 − (average ∆∆CT) was determined using the mean difference in the ∆CT between the gene of interest and the internal control.

SKNBE2 NB cells have biochemical features of neurons and display NMYC amplification, a marker of NB malignant progression. SKNBE2 cells were depleted for HIF1A expression (shHIF1A) by the use of two short hairpin against HIF1A (SKNBE2 shHIF1A#A and SKNBE2 shHIF1A#B) and were grown in normoxia and hypoxia conditions (NX and HYP); unsilenced cells were used as control (SKNBE2 shCTR) (Fig. 1a). To evaluate the hypoxic status of the cells after their exposure to low oxygen conditions we tested the expression of known hypoxia targets (Additional file 2: Figure S1). To provide genes and pathways differentially regulated by HIF1A triplicates of silenced (SKNBE2 shHIF1A#B NX and SKNBE2 shHIF1A#B HYP) and unsilenced (SKNBE2 shCTR NX and SKNBE2 shCTR HYP) cells were subjected to RNA-seq experiment.

To get insights into the relationships between the experimental conditions, we performed hierarchical clustering of gene-based FPKM counts. Results clearly showed two main branches of the dendrogram separating silenced and unsilenced conditions (Additional file 2: Figure S2A). Although small changes of global expression levels were observed in the four conditions (Additional file 2: Figure S2B), principal component analysis (PCA) and multi-dimensional scaling (MDS) highlighted the important role of oxygen in separating shCTR HYP and shCTR NX whereas shHIF1A NX and shHIF1A HYP showed highly similar profiles in PCA analysis (Additional file 2: Figure S2C and D).

By comparing gene expression in shHIF1A NX vs shCTR NX and in shHIF1A HYP vs shCTR HYP, we obtained two gene sets. Raw calls of differentially expressed genes were subsequently filtered by fold change (Log₂ ≥ + 0.5 or ≤ − 0.5) and statistical significance (FDR ≤ 0.05) (Additional file 3: Tables S1 and S2). HIF1A silencing in normoxia affects the expression of much more genes (“shHIF1A NX vs shCTR NX” includes 2656 genes) than in hypoxia (“shHIF1A HYP vs shCTR HYP” includes 1886 genes) (Additional file 2: Figure S3A). KEGG pathway analysis showed that the most significantly enriched terms were “metabolic pathway” in shHIF1A NX vs shCTR NX and “axon guidance” in shHIF1A HYP vs shCTR HYP (FDR ≤ 0.05) (Additional file 2: Figure S3B).

By intersecting the above-cited “two gene sets”, we obtained three gene lists: 1) genes (n = 630) regulated “exclusively in hypoxia”; 2) genes (n = 1400) regulated “exclusively in normoxia” and 3) a list of genes (n = 1256) which are commonly regulated by HIF1A that we named HIF1A target genes (Fig. 1b). The expression trend of 1237 out of 1256 HIF1A target genes (98.88%) was concordant upon HIF1A depletion in NX and HYP (Fig. 1c). Conversely, 19 genes out of 1256 (less than 1.2%) had an opposite regulation in shHIF1A NX vs shCTR NX and shHIF1A HYP vs shCTR HYP, suggesting they are downstream targets of HIF1A related pathways. KEGG pathway analysis of the two “exclusive” gene sets revealed an enrichment of metabolic pathways in normoxia and an enrichment of axon guidance and pathways in cancer in hypoxia (FDR ≤ 0.05). KEGG pathway analysis of HIF1A target genes revealed an enrichment of MAPK signaling pathway, pathways in cancer and axon guidance (FDR ≤ 0.05) (Fig. 1d).

The reliability of RNA-seq data was assessed by RT-PCR in SKNBE2 shCTR and shHIF1A#A (Fig. 1e). We validated genes that have RNAseq log₂ fold change ranging from − 2 to 2, in shCTR HYP vs shCTR NX (Additional file 3: Table S3) and shHIF1A HYP vs shCTR HYP gene list. We found that expression levels measured by RNA-Seq were consistent with those obtained by RT-PCR.

Additionally, we confirmed these results by RT-PCR in SHSY5Y NB cells that have biochemical features of neurons but do not display high-risk marker as NMYC amplification. As described in Supplementary data, SHSY5Y cells were depleted for HIF1A expression (shHIF1A) and unsilenced cells were used as control (shCTR). The gene expression levels measured by RT-PCR in SHSY5Y cells were consistent with those obtained by RNA-Seq in SKNBE2 cells (Additional file 2: Figure S4).

The above results clearly highlight that HIF1A has a diverse role in normoxia than in hypoxia. Our hypothesis is that HIF1A driven response depends on the epigenetic reprogramming caused by low oxygen levels that may shape chromatin state and give HIF1A accessibility to HRE DNA regions previously closed. Furthermore, chromatin may remodel in regions not comprising HIF1A targets and allow HIF1A and other transcription factors access to new active DNA regions. To deeply investigate which genes are HIF1A dependently and HIF1A independently expressed, gene set shHIF1A HYP vs shCTR HYP (n = 1886 gene, Additional file 3: Table S2) and gene set of shCTR HYP vs shCTR NX (n = 3263 genes, Additional file 3: Table S3) were crossed. We found that 674 genes were regulated in both gene sets (Fig. 2a). Interestingly, 420 out of 674 genes show the same regulation in both gene sets (Log₂); the expression of these genes named “Hypoxia targets” is affected by low oxygen concentrations and not affected by HIF1A (Fig. 2b). KEGG pathway analysis shows that “Hypoxia targets” are enriched in pathway that regulate cytoskeleton, ligand-receptor interaction and axon guidance (FDR ≤ 0.05). By contrast, 254 out of 674 genes have an opposite regulation in the two lists (Log₂). These genes might be direct targets of HIF1A (here named: “HIF1A direct-targets”) in hypoxia because when HIF1A is depleted their regulation is inverted (Fig. 2c). KEGG pathway analysis shows that “HIF1A direct-targets” are enriched in metabolic and cancer pathways similar to pathways affected by HIF1A silencing “exclusively in hypoxia” (FDR ≤ 0.05). These findings suggest that NB cells adapt to hypoxia by HIF1A-dependent and HIF1A-independent driven response.

Genome-wide methylation analysis using Infinium HumanMethylation450 BeadChips was performed in triplicate as described for RNAseq. To get an overview of the methylation patterns in the normalized data, hierarchical clustering of the most variable probes was performed. The analysis separated samples into four clusters, one for each experimental condition, within which replicates are grouped (Fig. 3a). Probes hypo or hyper-methylated with a Δβ-value greater than 0.2 (20%) in shHIF1A HYP vs shCTR HYP (named HIF1A probes) and in shCTR HYP vs shCTR NX (named Hypoxia probes) were selected. The sets of HIF1A probes and Hypoxia probes include 1078 (Additional file 3: Table S4) and 260 (Additional file 3: Table S5) differentially methylated CpG sites respectively. A global hypermethylation status of Hypoxia probes (Δβ ≥ 0.2) and hypomethylation status of HIF1A probes (Δβ ≤ − 0.2) was observed (Fig. 3b). Both probe sets cluster close to each other (Fig. 3a) whereas probes differentially methylated in shHIF1A NX vs shCTR NX were not observed.

We measured the genomic distribution of Hypoxia probes and HIF1A probes in relation to CGI centric annotation. Hypoxia probes were overrepresented in regions of low CpGs (open sea, > 4Kb from the CGI) and underrepresented in CGIs (island). Once HIF1A is depleted (HIF1A probes) open sea and island regions became under- and over-represented, respectively, (Fig. 3c). CGI shelves (>2Kb from the CGI) and shores (<2Kb from the CGI) showed a non-random distribution between the two probe sets. Hypoxia probes and HIF1A probes were mapped in relation to gene positions. In both contrasts, a strong overrepresentation of probes in intergenic regions (IGR) and gene body as well as an overall underrepresentation of probes located in the first exon, 3′ and 5’UTR and in the upstream region of transcription site (TSS) was observed (Fig. 3d). Hypoxia probes and HIF1A probes were crossed and 150 probes were found methylated in both probes lists. The methylation status (Δβ) of 150 common probes in hypoxia (shCTR HYP vs shCTR NX, Hypoxia probes) is reverted once HIF1A is depleted (shHIF1A HYP vs shCTR HYP, HIF1A probes) (Fig. 3e) and a strong overrepresentation of common probes in IGR is observed (Fig. 3f). Overall, these results suggest DNA methylation status is strictly correlated with oxygen storage and HIF1A control of DNA methylation of IGR, gene body and TSS probes could occur only upon hypoxia induced epigenetic reprogramming.

The correlation between the differential expression and differential methylation was explored for each gene-probe pair (p-value ≤0.05). We searched for changes in opposite directions (eg. Up-regulation of the gene expression and hypo-methylation of the related CpG probe). In the shHIF1A HYP vs shCTR HYP comparison, we selected 31 gene-probe pairs (p ≤ 0.05, Fig. 4a, Table 1), whereas in the shCTR HYP vs shCTR NX comparison 18 gene-probe pairs were kept (p ≤ 0.05, Fig. 4b, Table 1). These probes are located on regulatory regions as UTR, TSS200, TSS1500 and in gene bodies (Additional file 2: Figure S5).

We explored gene-probe pairs correlation in 105 NB tumors for which matched methylation and gene expression data were available (GEO accessions: GSE73515 and GSE73517, respectively) and restricted our analysis to Low risk (n = 40) and High risk (n = 56) tumors as defined by Henrich et al. [28]. We found that the correlations for KIF26B, EFCAB2, BCL2L11, VAV2 and SORBS2 gene expression with methylation status were validated (Additional file 2: Figure S6A-E; P < 0.05). Additionally, in an independent set of NB tumors (GSE16476), we found that the expression of these genes was also associated to NB patient’s survival (Additional file 2: Figure S6F).

The two gene signatures generated from gene-probe pairs were named “HIF1A signature in hypoxia” and “Hypoxia signature”. To assess the prognostic potential of these signatures, we used Low risk (n = 40) and High risk (n = 56) tumors from the GSE73517 gene expression data set [28]. Hierarchical clustering based on Euclidean distances of expression levels, showed that the genes in “HIF1A signature in hypoxia” (Fig. 4c) clustered the 35.7% (20/56) of High risk and the 77.5% (31/40) of Low risk patients, in two separate groups (P < 1.0 × 10^− 4) according to their Risk category. In contrast, “Hypoxia signature” (Fig. 4d), clustered the 50% (28/56) of High risk and the 57.5% (23/40) of Low risk patients according to their Risk category (P < 1.0 × 10^− 4). Indeed, we verified that our gene signatures correctly classified a discrete portion of both High and Low risk patients.

The strong enrichment of hypoxia differentially methylated sites in IGR suggests that the changes of methylation pattern mainly occur at putative regulatory regions distant from target genes (as shown in Fig. 3d-f). To further select which hypoxia differentially methylated sites (Hypoxia probes n = 260) are located in NB putative regulatory regions we re-analysed DNase hypersensitivity assay and Chip-Seq histone acetylation (H3K27ac) data deposited in Gene Expression Omnibus database (GSE65664, Additional file 3: Table S6) for additional SKNBE2, CHP134 and SHSY5Y cell lines. We obtained that 113 out of 260 probes (43.5%) were located in regulatory active regions and 14 probes were annotated in all cell lines with at least 4 epigenetic markers (Additional file 3: Table S6). We searched the genes distant 1 Mb up- or down-stream from the 14 aforementioned probes in the RNAseq data (shCTR HYP vs shCTR NX, Log₂ ≥ 0.2, ≤ − 0.2) and we listed the candidate targets of these regulative regions in Table 2 (and Additional file 3: Table S7). We observed that the methylation status of 9 out of 14 probes is affected by oxygen levels (Δβ Hypoxia; Hypoxia probes) but not by HIF1A expression (Δβ HIF1A; HIF1A probes). Putative targets show gene expression levels affected by oxygen levels (Log FC Hypoxia, in shCTR HYP vs shCTR NX, Log₂ ≥ 0.2, ≤ − 0.2) and not by HIF1A expression (Log FC HIF1A, in shHIF1A HYP vs shCTR HYP, Log₂ ≥ 0.2, ≤ − 0.2). Conversely, the methylation status of the remaining 5 probes is inversely affected in both conditions (Δβ Hypoxia and Δβ HIF1A) as the expression of putative targets (Log FC Hypoxia and Log FC HIF1A). The best NB candidate targets are represented by genes (highlighted in Table 2) whose expression correlates with NB patient’s event-free survival (Additional file 2: Figure S7).