Novel significant stage-specific differentially expressed genes in hepatocellular carcinoma

Normalized and log₂-transformed Illumina HiSeq RNA-Seq gene expression data processed by the RSEM pipeline [29] were obtained from TCGA via the firebrowse.​org portal [6]. The patient barcode (uuid) of each sample encoded in the variable called ‘Hybridization REF’ was parsed and used to annotate the controls and cancer samples (Fig. 2). To annotate the stage information of the cancer samples, we obtained the clinical information dataset for HCC from firebrowse.​org (LIHC.Merge_Clinical.Level_1.2016012800.0.0.tar.gz) and merged the clinical data with the expression data by matching the “Hybridization REF” in the expression data with the aliquot barcode identifier in the clinical data. The stage information of each patient was encoded in the clinical variable “pathologic stage”. The pathologic stage is essentially the surgical stage, prior to any treatment received, determined with the tissue obtained at the time of surgery. This interpretation is reinforced in the TCGA HCC sample inclusion criteria as follows: “Surgical resection of biopsy biospecimens were collected from patients diagnosed with hepatocellular carcinoma (HCC), and had not received prior treatment for their disease (ablation, chemotherapy, or radiotherapy)” (The TCGA [41]). The availability of this unequivocal information enables the analysis of cancer stages. The substages (A,B,C) were collapsed into the parent stage, resulting in four stages of interest (I, II, III, IV). We retained a handful of other clinical variables pertaining to demographic features, namely age, sex, height, weight, and vital status. With this merged dataset, we filtered out genes that showed little change in expression across all samples (defined as σ < 1). Finally, we removed cancer samples from our analysis that were missing stage annotation (value ‘NA’ in the “pathologic stage”). The data pre-processing was done using R (www.​r-project.​org).

Linear modelling of expression across cancer stages relative to the baseline expression (i.e, in normal tissue controls) was performed for each gene using the R limma package [42]. The following linear model was fit for each gene’s expression based on the design matrix shown in Fig. 3a:

where the independent variables are indicator variables of the sample’s stage, the intercept α is the baseline expression estimated from the controls, and β_i are the estimated stagewise log fold-change (lfc) coefficients relative to controls. The linear model was subjected to empirical Bayes adjustment to obtain moderated t-statistics [34]. To account for multiple hypothesis testing and the false discovery rate, the p-values of the F-statistic of the linear fit were adjusted using the method od Hochberg and Benjamini [22]. The linear trend across cancer stages for the top significant genes were visualized using boxplots to ascertain the regulation status of the gene relative to the control.

The linear model in eqn. (1) would not be sufficient to identify genes with an ordered monotonic trend of expression across cancer stages. Addressing this question would also help assess whether monotonic changes of gene expression were observed with disease progression. Towards this end, we designed a model of gene expression where the cancer stage was treated as a numeric variable:

where X takes a value in [0,1,2,3,4] corresponding to the sample stage: [control, I, II, III, IV], respectively. It was noted the mean expression of a gene could show the following monotonic patterns across cancer stages:

To perform contrasts, a slightly modified design matrix shown in Fig. 3b was used, which would give rise to the following linear model of expression for each gene:

where the controls themselves are one of the indicator variables, and the β_i are all coefficients estimated only from the corresponding samples. Our first contrast of interest, between each stage and the control, was achieved using the contrast matrix shown in Table 1. Four contrasts were obtained, one for each stage vs control. A threshold of |lfc| > 2 was applied to each such contrast to identify differentially expressed genes (with respect to the control). We used the absolute value of the lfc, since driver genes could be either upregulated or downregulated. Genes could be differentially expressed in any combination of the stages or no stage at all. To analyze the pattern of differential expression (with respect to the control), we constructed a four-bit binary string for each gene, where each bit signified whether the gene was differentially expressed in the corresponding stage. For example, the string ‘1100’ indicates that the gene was differentially expressed in the first and second stages. There are 2⁴ = 16 possible outcomes of the four-bit string for a given gene corresponding to the combination of stages in which it is differentially expressed. This is illustrated in set-theoretic terms in Fig. 4. In our first elimination, we removed genes whose |lfc| < 2 for all stages. For each remaining gene, we identified the stage that showed the highest |lfc| and assigned the gene as specific to that stage for the rest of our analysis.

We applied a four-pronged criteria to establish the significance of the stage-specific differentially expressed genes.

To obtain the above p-values (ii) - (iv), we used the contrast matrix shown in Table 2, which was then used an an argument to the contrastsFit function in limma.

Principal component analysis (PCA) were performed using prcomp in R. To choose 100 random genes, we used the rand function. Gene set enrichment analysis were performed on KEGG (https://​www.​genome.​jp/​kegg/​) and Gene Ontology [5] using kegga and goana in limma, respectively. In order to visualize outlier genes that are significant with a large effect size, volcano plots could be obtained by plotting the -log10 transformed p-value vs. the log fold-change of gene expression. Heat maps of significant stage-specific differentially expressed genes were visualized using heatmap and clustered using hclust. Novelty of the identified stage-specific genes was ascertained by screening against the Cancer Gene Census v84 [14].

Three C-type lectin domain proteins (CLEC4M, CLEC1B, CLEC4G) were detected in the top ten genes of linear model given by eqn. (1). Interestingly, this identical cluster of three genes was detected as the most significantly downregulated liver cancer-specific genes in a qPCR study of an independent cohort of 65 tumor-normal matched cases [21]. On screening the top 200 linear model (1) genes against cancer driver genes in the Cancer Gene Census, only four genes were found, namely BUB1B, CDKN2A, EZH2, and RECQL4. The top 200 genes of the linear model given by eqn. (2) overlapped with 111 genes of linear model (1) and yielded six genes from the Cancer Gene Census, namely BUB1B, EZH2, CDKN2C, CANT1, POLD1, and STIL. Both CDKN2A and CDKN2C are cyclin-dependent kinase inhibitors. CDKN2A was a member of the gene signatures for HCC prognosis independently proposed by Gillet et al. [16] and Yang et al. [59]. It was remarkable that GABRD stood out as the top gene in both the linear models, and with a monotonic order of expression with the cancer stage. GABRD is discussed further in the section on Stage-IV specific genes. A gene with a monotonicity of expression may be increasingly upregulated as the cancer initiates, progresses and metastasizes, signalling its oncogenic progression; or conversely, it may be increasingly downregulated with the cancer stages, signalling the loss of tumor suppressor activity. Screening the top 200 genes with monotonic expression against the Cancer Gene Census yielded a completely different set of six genes: HSP90AB1, ALDH2, ESR1, PPP2R1A, HIST1H4I, SEPT5. HSP90AB1, a heat shock protein and molecular chaperone, was a key result of Xu et al. [56] where it played a dual role, one in the set of 50 hub genes correlated with Barcelona Clinic Liver Cancer (BCLC) staging of HCC patients, and another, in the set of 13 hub genes correlated with overall survival of HCC patients. HSP90AB1 might have a significant role in the aetiology of HCC, given that its expression is known to be upregulated by hepatitis B virus encoded X protein [31]. The monotonic changes in HSP90AB1 might further facilitate its known roles in angiogenesis [19]. The top 200 genes with monotonic expression had 15 genes in common with the top 200 of linear model (1) and 16 genes in common with the top 200 of linear model (2). However, only six genes were common to the top 200 of all three (namely GABRD, PIGU, NDUFA4L2, CRHBP, FLVCR1, TTC13; Fig. 11). NDUFA4L2 has been identified as a target gene of HIF-1 (hypoxia-inducible transcription factor-1), and a key factor driving the metabolic reprogramming in hypoxic micro-environments [46]. Our findings established that not only was NDUFA4L2 significantly overexpressed in HCC (as noted in [27]), but its overexpression follows a significant monotonic pattern across cancer stages, a much stronger statement that would support the role of NDUFA4L2 in driving HCC progression. Similarly, the expression of CRHBP has been recently shown to be negatively associated with the tumor size in HCC [55]. Our study provides a more quantitative account of the significant monotonic downregulation of CRHBP with the HCC stage. Two proteins of the glycosylphosphatidylinositol (GPI) anchoring system, PIGU and PIGC, were top genes with respect to significant monotonic expression (Table 6); of these, PIGU is a known bladder cancer oncogene [18].

CA9 is a member of carbonic anhydrases, which are a large family of zinc metalloenzymes that catalyse the reversible hydration of carbon dioxide. Its expression in clear cell Renal carcinoma, but not in functional kidney cells has gained attention for its use as a pre-operative biomarker [30]. The WNT7B protein is part of the Wnt family, a family of secreted signalling proteins. Elevated WNT7B in pancreatic adenocarcinoma has been found to mediate anchorage independent growth [4]. Surprisingly, both CA9 and WNT7B are downregulated in HCC, most so in stage-I, contrary to their role in other cancers. A concrete interpretation of the role of these genes in HCC awaits appropriately designed experimental studies.

It is pertinent to ask the following question here: which genes are essential for the initiation of HCC? Clearly these genes would be differentially expressed in stage I relative to control. All significantly differentially expressed genes with maximal |lfc| in stage-I would be the best candidates for genes involved in the initiation of HCC. These 122 genes are provided in the Supplementary File S3.

APOBEC3B, a DNA cytidine deaminase, is a known cancer driver gene in the Cancer Gene Census, but there are no literature reports of its stage-specificity in any cancer. It is known to account for half the mutational load in breast carcinoma, and its target sequence context was found to be highly mutated in Bladder, lung, cervix, neck, and head cancers as well [7]. Further studies have attributed specific hypermutation signatures across all cancers to the APOBEC family, including APOBEC3B [1]. Here APOBEC3B is upregulated, increasing its capacity to inflict the hypermutator phenotype, and highlighting an intriguing stage-specificity in its action. FAM186A polymorphisms have been reported in GWAS and SNP studies on colorectal cancer patients and shown to have a significant odds ratio in risk heritability [48]. FAM163A was a component of the 8-gene signature used for the risk stratification of HCC patients [39].

C12orf48, also known as PARI, participates in the homologous recombination pathway of DNA repair, and its overexpression has been reported in pancreatic cancer[35]. Further PARI was recently identified as a transcriptional target of FOXM1 [62], which is a well-validated upregulated gene in HCC [21]. DLG5 is a cell polarity gene and its downregulation has been implicated in the malignancy of breast [32], prostate [49] and bladder cancers [65]. It has been recently found that lower DLG5 expression is correlated with advanced stages of HCC and essential for invadopodium formation, an event critical to cancer metastasis [26]. It is surprising that our study has identified a stage-III specific upregulation in DLG5. Interestingly, evidence is emerging to lend support to our finding that DLG5 might be tumor-promoting. In a very recent review, Saito et al. [43] reinterpreted published results on cell polarity and cancer, and advanced an alternative perspective on the role of polarity regulators in cancer biology. They argued that both cellular and subcellular polarity would be regulated by DLG5 and related polarity proteins. Subcellular polarity might improve the cellular fitness for proliferation and stemness, thereby causing tumor promotion. Hence cell polarity regulation is anti-tumorigenic and subcellular polarity regulation is pro-tumorigenic, and our analysis has uncovered the pro-tumorigenic upregulated activity of DLG5. ECT2 encodes a guanine nucleotide exchange factor that remains elevated during the G2 and M phase in cellular mitosis. ECT2 is found to be upregulated in lung adenocarcinoma and lung squamous cell carcinoma [66], as well as in invasive breast cancer [52]. NCAPG2 is a component of the condensing II complex and involved in chromosome segregation during mitosis. NCAPG2 level were found to be increased in non-small cell lung cancer, and its over-expression was found to be correlated with lymph node metastasis, thus enabling the use of NCAPG2 as a poor prognostic biomarker in lung adenocarcinoma [61]. GNMT is a methyltransferase that catalyses conversion of S-adenosine methionine to s-adenosyl cysteine. In the absence of GNMT, S-adenosine methionine causes hypermethylation of DNA, which represses GNMT levels and is found in HCC samples [24]. This is an epigenetic mechanism for loss of function of tumor suppressors and our study here confirmed the downregulation of GNMT expression. PRR11 is found to be over-expressed in lungs, and its silencing using siRNA resulted in cell cycle arrest and apoptotic cell death, followed by decreased cell growth and viability [64]. A similar knock out experiment of PRR11 in hilar cholangiocarcinoma cell lines resulted in decreased cellular proliferation, migration, and tumor growth [9]. WDHD1 is a key post-transcriptional regulator of centromeric, and consequently genomic, integrity [23] and its overexpression has been identified as biomarker of acute myeloid leukemia [53], and lung and esophageal carcinomas [44]. C15orf42 has been implicated in nasopharyngeal carcinoma [3]. ORC6L overexpression has been identified as a prognostic biomarker of colorectal cancer possibly by enhancing chromosomal instability [54]. XRCC2 was found to increase locally advanced rectal cancer radioresistance by repairing DNA double-strand breaks and preventing cancer cell apoptosis [40]. XRCC2 was also highlighted in the gene signature for HCC prognosis advanced by Gillet et al. [16].

GABRD, which was the top gene in the linear models as well, encodes for the delta subunit of the gamma-amino butyric acid receptor. The GABA receptor family was found to be frequently downregulated in cancers, except for GABRD, which was found to be up-regulated. Gross et al. [17] proposed that the GABA receptor gene family might play a role in the proliferation independent differentiation of cancer cells. GBX2 is part of the GBX gene family, which are homeobox containing DNA binding transcription factors. GBX2 is overexpressed in prostate cancer and studies show that expression of GBX2 is required for malignant growth of human prostate cancer [15]. PECAM1 overexpression has been linked to peritoneal recurrence of stage II/III gastric cancer patients [47]. CEND1 has been identified as a cell-cycle protein [50]. PGAM2 is a glycolytic enzyme whose upregulation is essential for tumor cell proliferation [57]. NR1I2 downregulation has been used in constructing a prognostic 9-genes expression signature of gastric cancer [51]. GDF5 has been shown to be a downstream target of the TGF-beta signaling pathway [33], stimulating angiogenesis required for the growth and spread of the cancer. GPR1 has been reported to be involved in promoting cutaneous squamous cell carcinoma migration [12]. Two other stage-IV specific genes, namely the downregulated CXCR2P1, which is a C-X-C motif chemokine receptor 2 pseudogene 1, and LOC25845, are minimally documented in the literature in the context of HCC, other cancers or any other condition. It is worth mentioning however that CXCR2, a member of the GPCR protein family binding the interleukin IL8, has been reported as an effective non-invasive blood based biomarker for HCC [45]. It is notable that ARHGAP42, a Rho GTPase activating protein, was another key result of Xu et al. [56], finding a place both in their set of 50 hub genes correlated with the BCLC staging of HCC patients, and in the set of 13 hub genes correlated with overall survival of HCC patients.Most of the stage-IV specific genes show contra-regulation (i.e, no clear trend) across cancer stages, and only 15 of the 35 genes revealed a monotonic pattern of expression (highlighted in Table 8). The other 20 genes could be unique to the hallmarks of stage-IV cancer, e.g., processes related to lymph node involvement and/or metastasis.