Hepatocellular Carcinoma (HCC) ranks fifth among all cancers and third in mortality, accounting for hundreds of thousands of deaths per year. One-year survival rates remain less than 50% in the United States, despite advances in therapy [1]. HCC is the primary cancer morbidity evolving over decades in underlying hepatitis C (HCV) liver pathology in North America and Japan. HCC development is generally thought to be a multistep process resulting from hepatocyte turnover, chronic inflammation, regeneration, oxidative stress, DNA damage, and cirrhosis, as well as direct viral injuries. Unfortunately, the specific molecular mechanisms underlying hepatocarcinogenesis remain unclear.

In the last ten years, microarray technology has been a powerful tool to study the molecular basis of disease. By measuring whole-genome transcript levels, expression patterns associated with liver dysfunction have been examined. However, HCC remains a difficult disease to study, with widely variable findings between studies and several proposed non-overlapping gene signatures [2,3,4,5,6,7,8,9,10,11,12]. This is likely due not only to the biological heterogeneity of HCC pathogenesis, but also reflects the varied clinical background of patients and variation in statistical technique. There are significant statistical challenges which plague the analysis and interpretation of microarray experiments. Differences in technique in every stage of data pre-processing have been demonstrated to dramatically affect the end results, including background correction [13], normalization [14,15], and probe set summarization [16].

In addition, traditional statistical approaches are not particularly well-suited to cancer genomic data. When simultaneously comparing many thousands of genes, multiple testing problems are considerable. Worse, because the assumption of independent tests is often violated, actual false positive rates can be much higher than estimated by standard methods [13,14,15]. This implies that, even using conservative multiple testing correction methods, cancer studies which generate thousands of significantly differentially expressed genes could have several hundred false positive results. To reduce this problem, we used a knowledge-driven approach, using what is already known about normal and disease processes to generate hypotheses that can be tested with a relatively small number of statistical tests [17,18,19].

Another difficulty stems from the heterogeneity of cancer processes, in which changes in the expression of important genes occur only in subsets of tumors. This results in skewed density curves (sometimes even bi-modal) that may not be easily detected by means-based tests. Most statistical tests in common use are based on comparing the magnitude of mean change relative to the variation. These tests also place focus on the largest magnitude changes which are often products of tumor behavior, such as increased metabolism and cell proliferation/turnover, rather than drivers that often have smaller fold-changes [20]. We suspect that there are modest changes in the expression of critical genes that may be difficult to distinguish from “noise” in the data, but may have a significant impact on tumor development [21,22]. To address this we applied techniques that find patterns in the data (such as dimension reduction techniques), focused on such genes, to identify those that drive tumor behavior even if they have modest expression changes or skewed distribution patterns.

Tumors are widely believed to arise through an accumulation of random mutations. Mutations and chromosomal instability have been demonstrated in several carcinomas, including non-Hodgkins lymphoma, ovarian, colorectal, and oral tissue types [23,24,25,26,27]. HCC arising from chronic HBV infection has also been associated with mutations via HBx gene interference with p53 binding, leading to faulty DNA repair mechanisms [28,29]. It would be natural to think that the widespread dysregulation of gene expression in HCV-HCC is also largely random. However, HCV-HCC may be unusual because hepatitis C is an RNA virus that codes proteins that have direct interaction with over thirty host proteins. Tumors emerge from an environment of decades of host response to infection and liver damage. Therefore we hypothesize that induction of HCC in chronic HCV liver pathology may depend more on host response to chronic infection and HCV-host interactions than on direct DNA damage. If this is true, the effects of the HCV virus will be seen in the perturbation of the “tools at hand”: gene expression changes that might be expected include modified expression of genes already in use in the liver (including genes expressed by activated hepatic stellate cells), target genes of host proteins that HCV proteins interact with, and genes used in the liver’s own life history. Such genes contain the specific transcription factor binding sites (TFBS) that are responsive to the transcription factors expressed in the liver, while genes that are not normally expressed in the liver are responsive to different promoters. For instance, the promoter region for FGF7 (expressed in the embryonic liver) contains binding sequences for ATF2, FOXD1, HNF3B, STAT3, and JUN which are all expressed in the liver and dysregulated in liver disease. This reasoning also implies that genes never expressed by a healthy liver would not be expected to be activated by HCV-induced tumors to the same extent as in HBV-HCC or other cancers.

To further target our hypotheses, we reviewed the current knowledge of processes involved in HCC. For instance, it has recently been noted that there appear to be pathways common to both cancer and embryonic development in HCC and other cancers [30,31]. In the context of the hypothesis of non-random response to HCV as described above, this led us to question whether any developmental genes involved in HCC are specific to liver development, and if paralog genes (similar in structure and function in other tissues) remain dormant. In this paper we demonstrate that HCV-induced liver cirrhosis and HCC do indeed show a general pattern of differential expression of liver development genes compared to paralog genes that have similar roles in the development of other tissues. Many of these developmental genes are up- or down-regulated in cirrhotic livers in a coherent way (clustering closely together), then degenerating into widely variable expression patterns in tumors. Some of the genes identified in this manner are already associated with HCC, while others appear to be novel. We also observed that some of these important embryonic signals are secreted from mesodermal tissues during development. These same signaling molecules may be secreted from mesodermally-derived stellate cells in adults. However, these cells comprise less than five percent of adult liver volume, which may result in an observed low signal that may have been difficult to distinguish from noise in previous studies.

Liver development is a multi-stage process orchestrated by nearly 200 master regulators, growth factors, and their receptors. Growth factors secreted externally and from within the developing liver bind receptors on the surface of liver cells, which transduce signals to transcription factors (TFs) within the nucleus. These transcription factors, either individually or as co-factors, regulate a complex program of inducing or repressing access to gene transcription by a number of activities including chromatin decompaction, recruitment of chromatin remodeling complexes, and histone marker methylation, demethylation, or acetylation, as well as by physically blocking or recruiting RNA polymerase. For example, some of the earliest TFs that induce hepatic fate from the endoderm (FOXA1-3, GATA4/6) open the chromatin structure around early hepatic markers such as albumin and alpha fetoprotein (AFP), while β-catenin (CTNNB1), LEF1, and TCF3 recruit chromatin remodeling complexes, and ARID5B and ATF2 modify histone markers of target genes. See Si-Tayeb et al. for a recent review [32] of embryonic liver development.

In the neonatal period, many of these regulators are repressed in most parts of the liver, while others maintain some level of activity throughout adult life either playing roles in metabolism and homeostasis or in maintaining niches of hepatic stem cells.

Microarray studies were obtained over a 10 year period from 180 samples of cirrhotic tissue and tumors collected from 140 patients with chronic HCV infection. As described in the Experimental Section, stringent quality control criteria were applied minimize technical artifacts, leaving us with 30 HCV-cirrhosis (CIR) and 49 HCV-HCC (HCC) tumors. These were compared to a control group of 12 samples taken from non-diseased, deceased donor livers (NOR). Tissue was obtained from 31 early HCC (stage T1 and T2) and 18 late HCC (T3 and T4). Twenty-nine HCC patients were transplanted, six died on the waiting list and 14 were never listed for transplant due to age, stage of cancer, or other co-morbidities.

We hypothesized that genes not normally expressed in adult livers are less likely to be transcriptionally activated in HCV-HCC. To test this, we identified a set of genes with zero counts in an RNA sequencing study on a normal liver sample. 1,399 were represented on the Affymetrix U133Av2 genechip. We tested for expression in HCC compared to normal samples in our dataset using a one-sided Kolmogorov-Smirnov (K-S) test of identical distribution. Of the 1,399 genes, none were expressed in the HCC samples (α < 0.001).

More specifically, we wished to determine whether the general developmental pathways altered in HCC were using genes specific to liver development, or whether any member of the gene family might be engaged. To examine this, we identified 26 paralog genes that have highly related developmental functions in other tissues that are not expressed in normal healthy livers (based on the RNA sequencing data and verified with a literature search). In our data, no paralogs were expressed in HCC compared to normal samples (α < 0.001) (Supplementary Table S1, column A). We validated this to an independently collected HCV-HCC dataset from Wurmbach et al. (see Experimental Section), which also had no expression of these paralog genes (Supplementary Table S1, column B). Next, we compared the expression patterns of liver genes to their non-liver paralogs in tumors. Most (31/33, 94%) of the liver genes had significantly different expression distributions compared to their paralog (Supplementary Table S1, column C). Figure 1 shows examples of the typically observed patterns of expression and poor regulation. Even when a liver gene is not differentially expressed in HCC, it often seems to be more poorly regulated with a wider distribution pattern than is seen in normal tissue.

We then identified patterns of expression for liver development genes in HCV-CIR and HCV-HCC. To capture changes in either mean or variation, we assessed significance with a combined p-value from both t and F tests. We were particularly interested in whether the genes associated with progression from cirrhosis to cancer were those particular to a certain developmental stage. However, differentially expressed genes were found from all stages of development, and PCA plots of stage-specific genes all showed discrimination between normal, cirrhotic, and tumor samples (data not shown).

Twenty-nine developmental genes had a pattern of higher magnitude over-expression in cirrhosis, then declining values in HCC (Table 1). These include several extra-cellular matrix (ECM) genes and members of the TGFβ/BMP signaling pathway. Transcription factors following this expression pattern include SOX9, GATA6, HAND2, and IRS2. The only genes differentially expressed between early (T1 and T2) and late-stage (T3 and T4) tumors were EPCAM and tumor suppressor KLF6. Nine genes were down-regulated in cirrhosis and remained low in tumors: transcription factors FOXA1, FOXA2, GATA4, HNF1A and STAT3; growth factor receptors ACVR2B, RXRA; signaling molecule NRTN, and MMP15, which degrades intracellular matrix proteins.

Sixteen genes were differentially expressed uniquely in the tumors (Table 2). Magnitude of change was modest, however six of the 16 genes were transcription factors or activators (TBX3, HHEX, ATF2, FOXM1, PROX1 and STAT3) which might be expected to induce large effects with small expression changes. Thirty-five more genes were differentially expressed in cirrhosis and either had similar expression in tumors (25 genes) or had larger magnitude changes in HCC (10 genes). For example, GPC3 was up-regulated slightly in cirrhosis (×2), and greatly up-regulated in both early (×7.2) and late (×10.4) tumors.

Overall, 98 of the 179 (55%) genes critical to liver development had altered expression patterns in cirrhosis and early stage tumors. Only five of the genes were uniquely changed in late-stage tumors: ACVR1, HMGA2, IGF2, CP and YAP1. A complete list of all 179 liver development genes tested, and relative expression in each disease group can be found in Supplementary Table S2.

From these significant liver development genes we have been able to identify some specific functional groups. Genes related to extra-cellular matrix (ECM) maintenance or remodeling demonstrated major changes in both cirrhosis and tumors. PCA of the significantly changed genes demonstrate that these genes also independently discriminate between normal, cirrhosis, and tumor samples (Figure 2A). The BMP signaling pathway is also highly dysregulated in HCV-cirrhosis and HCC. BMP2, its receptors, and BMP inhibitors are all differentially expressed in cirrhosis and HCC. Embryonically, BMP signaling is antagonistic to FGF signaling and this balance is controlled by the DAN family of BMP antagonists from mesenchymal cells and GPC3 expressed by hepatocytes. BMP2, BMPR1A, FGF7, FGFR2 and ID3 were more highly expressed in cirrhosis than HCC, while the BMP inhibitors were more highly expressed in tumors. At least one of the inhibitors GPC3, GREM1, FSTL3, and/or FST were over-expressed (FC > 1.5) in 100% of tumor samples (Supplementary Table S3). A PCA plot demonstrates the gene set’s ability to discriminate most tumors from normal and cirrhosis samples (Figure 2B). Leave-one-out cross-validation (LOOCV) of ROC curves assessed predication sensitivity of 95.9% and specificity of 83.3%.

These findings were validated against the Wurmbach dataset by applying the PCA loadings generated from our data to their dataset, and we observed similar patterns of separation between normal, cirrhosis, and HCC tissues (Figure 2A,B). The ECM and BMP genesets were well described with the first two principal components.

Since 1997, HCV patients diagnosed with cirrhosis and HCC have been evaluated and treated at the Hume-Lee Transplant Center at VCUHS according to an Institutional Review Board approved study protocol [57]. Informed consent was obtained from all patients. After staging, HCC patients had their tumors ablated and were evaluated for liver transplant according to the United Network for Organ Sharing criteria. Tissue samples were collected from biopsies and explanted livers according to protocols established by the Liver Tissue Cell Distribution System (Richmond, Virginia, funded by NIH Contract #N01-DK-7-0004/HHSN267200700004C). Control liver samples were obtained from explanted donor livers. Donor livers were shown to have normal function and were negative for hepatitis C virus antibodies.

An independently published dataset of HCV-cirrhosis and HCC was also obtained for verification of results [58] (Wurbach et al., NCBI GEO database accession GEO6764). In addition, the absolute expression levels of target genes in a normal adult human were obtained from the BodyMap gene expression database [59] and mapped with Burrows-Wheeler Alignment tool (BWA) [60] with default parameters. In cases where BodyMap results were inconclusive (counts of 0–40), a literature search was performed to confirm adult expression of target genes.

Pre-transplant biopsies and explanted livers were sectioned and grossly examined. Samples from tumors and cirrhotic liver tissue (according to diagnosis and pathological examination) were freshly snap-frozen and processed in the Hume-Lee Transplant Center Molecular Diagnostic Laboratory. Liver tissue samples were collected in liquid nitrogen or RNAlater solution (Ambion, Austin, TX, USA) and stored at −80 °C until use. Explanted livers were sliced at intervals of 4–5 mm, and all nodules suspicious for HCC processed for light microscopy. Only tumor samples with more than 85% tumor cell content were used for the microarray studies. Normal and necrotic tissues were macro-dissected from the sample.

With minor modifications, the sample preparation protocol follows the Affymetrix GeneChip Expression Analysis Manual (Affymetrix, Santa Clara, CA, USA). After hybridization and scanning, the microarray images were checked for major chip defects or abnormalities in the hybridization signal. Total RNA quality and integrity of each sample were analyzed using the Agilent 2100 Bioanalyzer (Agilent Technologies), and products of cDNA synthesis and in vitro transcription (IVT) were tested before being considered for microarray analysis using the Agilent 2100 Bioanalyzer (cDNA synthesis 1.5 kb < cDNA < 5.0kb; IVT 1.0kb < cRNA < 4.5 kb).

All chips in the study were examined with several quality control tests [61,62]. Any chip that fell well outside the recommendations for any of the quality assessment tests was excluded from further analysis.

Data files were read into the R (version 2.13) programming environment using updated probe annotations from the BrainArray project (version 14.1.0, HGU133A2_Hs_REFSEQ), which have been shown to improve accuracy of probe—gene mapping over the standard Affymetrix annotation [63].

Robust Multichip Average (RMA) pre-processing is broadly accepted as robust, easy to implement, and widely applicable. However, we were concerned that the assumptions of RMA may not be met (that only a small proportion (1–5%) of genes are differentially expressed, and that about the same number of genes are over- vs. under-expressed). Therefore, we processed a test dataset using RMA and carefully examined differential expression results. Comparison of group contrasts identified 25–45% genes that were called as differentially expressed at an FDR < 0.05, and 2/3 of those genes were over-expressed in tumors. This suggested to us that quantile normalization may not be an appropriate method for this dataset. Instead, we first performed background correction, then normalized the data using non-parametric, distribution-free regression on technical covariates of probes (GC content, melting temperature, and probe location) to estimate and correct for systematic bias [64].

We used scaled Principal Components Analysis (PCA) on specific gene sets to explore the relationship between those genes and disease behavior. We used the function “prcomp” with scale = TRUE in the R environment [65].

Normally regulated genes generally have a narrow distribution with small variation, while poorly regulated genes may have broad, flat distribution of expression values or long thick tails and high variability. Genes that are differentially expressed in a subset of tumor samples will have a skewed distribution. Either of these situations violate the assumption of normal distribution and can be difficult to identify with a t-test. Instead, we used F-test of variance to identify high-variance genes and the non-parametric two-sample Kolmogorov-Smirnov Test to test differences in both location and shape of the distributions.

In this study, we investigated the hypothesis that genes critical to the development of the liver are poorly regulated or preferentially activated in HCV-cirrhosis and HCV-HCC. There are five main stages of liver development: hepatic fate specification, hepatoblast migration, liver bud growth, hepatocyte/biliary differentiation, and maturation. We identified 179 genes (via literature review) that are experimentally determined to be critical regulators necessary for normal liver development (See Supplementary Table S2 for a complete list). We also wished to determine whether liver development genes are more likely than their non-liver paralogs (which have highly similar functions in the development of other tissue types) to be dysregulated in liver tumors. We identified 26 such genes that are not normally expressed in either embryonic or adult liver tissue (Supplementary Table S1).