Met-regulated expression signature defines a subset of human hepatocellular carcinomas with poor prognosis and aggressive phenotype (original) (raw)
Characterization of the Met-regulated gene expression signature in primary hepatocytes. To identify HGF/Met–regulated genes, we performed expression microarray analysis after inducible activation of Met receptor in primary cultures of hepatocytes established from WT and Met conditional KO mice. Total RNA was isolated from untreated hepatocyte cultures as well as from cultures treated with 50 ng/ml of HGF for 0.5, 2, 12, or 24 hours. RNA collected from these experiments was converted to fluorescently labeled cDNA and used for hybridizations of oligonucleotide microarrays containing 21,997 features representing 19,140 unique mouse genes. After normalization of the data, 13,477 features with a sufficient number of valid expression values were selected for further analysis as described in Methods.
To define the set of HGF/Met–regulated genes, we compared experiments using a multivariate permutation t test at each time point. In total, 730 unique features showed significant (P < 0.001) and at least 1.5-fold expression differences between the 2 genotypes. As the only variable at these comparisons was the presence or absence of intact Met receptor, we could conclude that the expression of significant genes was regulated in a Met-dependent manner. The diagram in Figure 1 gives a summary of the data analysis strategy applied to select the significant HGF-regulated genes.
Diagram of data analysis. HGF/Met–regulated genes were identified by comparison of expression profiles from WT and Met KO hepatocytes. Expression of common orthologous HGF target genes was also assessed in 2 independent HCC data sets. The classifier was constructed from cross-species–conserved HGF/Met target genes to predict patient survival. Results were validated with multiple prediction algorithms on separate training and validation sets of HCC samples. FDR, false discovery rate.
The set of differentially expressed genes could be further divided into 2 major categories. The first category was represented by genes that showed permanent transcriptional changes in Met KO primary hepatocytes. Thus, expression of 60 genes was found to be altered in Met KO cultures at the 0 time point (after overnight incubation), and 57 of them remained differentially expressed during 24 hours of HGF exposure as compared with control cultures (P < 0.005) (Figure 2A). The presence of the permanent gene expression changes implies that in the absence of Met signaling, KO cells undergo a genotype-specific transcriptional adaptation.
Gene expression patterns of HGF-regulated genes in primary mouse hepatocytes. Gene expression ratios from duplicated dye-swapped hybridizations per sample were averaged before generation of the heat map. (A and B) Two matrices were constructed from normalized, log2-transformed expression ratios of genes with permanent (A) and HGF-induced (B) expression differences between the WT and Met KO cells. Red and blue columns at the top represent triplicate WT and Met KO samples, respectively, from the consecutive treatment points (0, 0.5, 2, 12, and 24 hours). Rows represent individual genes. (C) HGF/Met_–_dependent genes also formed clusters with different temporal induction patterns. Bar graphs show the mean expression differences ± SEM between the WT and Met KO samples in 6 gene clusters with expression peaks at early (C1, C4), late (C3, C6), or both early and late (C2, C5) HGF treatment points. Other gene clusters displayed permanently higher expression levels either in the WT or in the Met KO samples.
As expected, the majority of significant genes (672/730) were only detected in control hepatocytes after HGF treatment, since Met KO cells did not exhibit a specific response to HGF. The genes in this second category could be further separated on the basis of their temporal expression patterns. Up- or downregulated genes were divided into early and late target gene clusters, as they displayed maximal expression differences between the genotypes after a short-term (a half hour or 2 hours) or long-term (12 or 24 hours) HGF treatment. The heat map image created with the mean-centered log2-transformed expression ratios of the significant Met-regulated genes clearly demonstrates the presence of clusters with the distinctive temporal regulation and the reproducibility of the data in replicate experiments (Figure 2, B and C).
To validate the specificity of the Met targets, we also compared gene expression between HGF-treated and untreated control primary hepatocytes as well as between control cells treated at consecutive time points. This approach yielded 1,383 differentially expressed genes using the same selection criteria as in the previous comparisons. Notably, some of these genes did not show significant expression differences between the WT and KO hepatocytes, as they probably reflect common adaptive responses to the culture conditions with time. However, 353 from the previously determined Met targets were also identified with both selection strategies. In most cases, the timing and magnitude of the most significant responses overlapped in the horizontal and vertical comparisons, indicating that the majority of the differentially expressed genes represented a specific response to HGF induction.
Expression differences observed with microarray profiling were verified by quantitative RT-PCR analysis. Good correlation with microarray data was found for all 10 randomly selected significant genes (Figure 3).
Comparison of gene expression patterns of selected Met-regulated genes from microarray and real-time PCR experiments. Gene expression levels in real-time PCR experiments were normalized to β2-microglobulin expression, and the average expression ratios between WT control and Met KO hepatocytes were calculated from triplicate experiments at the different treatment points. Each bar represents the log2-transformed mean expression ratios ± SEM.
Functional analysis of target genes confirms the role of Met as an essential regulator of cell motility. In accordance with previous studies, we detected significant changes in the expression levels of known Met target genes, including Hmga1, Spp1 (21), Itg_β_1 (22), Egr1 (23), and Cldn2 (24). Consecutive functional analysis of the Met target genes allowed a more detailed insight into the cellular machinery associated with the Met-induced phenotype. A significant number of genes induced at 12 and 24 hours were involved in cell motility (Cxcl10, Capn2, Spp1, Fn1), angiogenesis (Vcam1, Anptl4, Ctgf, Neo1, Robo1), cell adhesion (Cldn2, Tjp3, Cdh17), and cytoskeletal organization (Hspa5, Arpc1b, Cap1, Nck2, Tpm2, Msn, Mid1, Vim, Dnm3, Tubb3, Tubb6, Krt2-8, Tuba1). Some of these significant genes, arranged by their postulated functions, are listed in Table 1. We also observed an early induction of several transcription factors (Hmga1, Egr1, JunB, MafF) after HGF treatment. A number of these immediate early targets of the Met pathway could regulate the expression of other differentially expressed genes at the later treatment points. This type of multistep regulation is well documented in the case of the HGF-Egr1-fibronectin (Fn) sequence (25).
Functional classification of selected HGF/Met target genes
HGF treatment had an especially prominent effect on the expression of genes involved in actin cytoskeleton organization and lamellipodium formation. Thus, HGF significantly upregulated Arpc1b and Nck2, a member and an important activator, respectively, of the Arp2/3 complex, which is involved in the regulation of the actin polymerization, particularly at the leading edge of moving cells (26). Furthermore, genes such as the Ras-responsive adenylyl cyclase–associated protein (Cap1), a key regulator of actin and cofilin turnover (27), as well as moesin (Msn), which connects actin filaments to the cell membrane (28), were induced by HGF treatment. Similarly, upregulation of tubulin-α1, -β3, and -β6 demonstrated that microtubular elements are transcriptional targets of Met signaling. The peak expression of the cell motility–related genes occurred at 12 and 24 hours coincidently with the onset of HGF-induced scattering in hepatocyte cultures. Differential expression of genes involved in actin cytoskeleton and microtubular organization as well as cell adhesion was consistent with the phenotypic differences revealed by immunofluorescence staining of WT and KO hepatocytes with antibodies against F-actin, α-tubulin, and vinculin after 24 hours of HGF treatment (Supplemental Figure 1, A–F; supplemental material available online with this article; doi:10.1172/JCI27236DS1).
We also found that osteopontin (Spp1), a secreted glycoprotein, was upregulated by HGF in primary hepatocytes, in agreement with published data (21). Previously, CD44v6, a surface receptor for osteopontin, was also identified as a Met target gene (23). Although CD44 was not differentially expressed in our model, we detected a concomitant induction of several integrin family members, including integrin-αV (Itg_α_V), integrin-α3 (Itg_α_3), and integrin-β1 (Itg_β_1) (29). Interestingly, αvβ3 integrin binding to osteopontin forms a complex that facilitates angiogenesis and tumor cell migration (30). Thus our data provide genetic evidence that transcriptional induction of genes controlling migratory and angiogenesis programs represents an essential part of Met signaling.
Expression profiling reveals a novel regulatory function of the Met pathway in oxidative stress response. Among the most striking observations was a profound misregulation of genes involved in antioxidative stress response and glutathione metabolism in Met KO hepatocytes. Notably, the transcription factor nuclear factor (erythroid-derived 2)–like 2 (Nfe2l2), as well as numerous oxidative stress response genes (Aldh1a1, Aldh1a7, Adh1, Ephx1, Ephx2), glutathione-S_-transferase isotypes (Gsta1, Gsta3, Gstm6, Mgst1, Gstm2, Gstm3), and glutamate-cysteine ligase (Gclc), a regulator of glutathione metabolism (31), showed significant overexpression in the Met KO hepatocytes. The majority of these genes are well-documented targets of the basic helix-loop-helix transcription factor Nfe2l2 (32),_ a key regulator of a detoxifying pathway activated by increased oxidative or xenobiotic stress in cells. Consistent with the microarray data, increased nuclear levels of Nfe2l2 protein were detected both in cultured hepatocytes and in intact livers from Met KO mice (our unpublished observations). In contrast, nuclear dimerization partners and possible antagonists of Nfe2l2, the nuclear factors MafF and MafK (33), were more abundant in the control hepatocytes. Upregulation of the antioxidant genes may reflect the altered redox homeostasis of the KO cells. This was also evidenced by decreased oxidized/reduced glutathione ratios (Supplemental Figure 1G) as well as by increased staining with the oxidation-sensitive probe 2′,7′-dichlorofluorescin (Supplemental Figure 1H) in the Met KO hepatocytes.
Comparative functional genomic analysis identifies a subgroup of human HCCs with a prominent Met gene expression signature. Based on the previous reports (5, 34), we hypothesized that a considerable part of the Met gene expression signature may be conserved between mouse and human hepatocytes. Therefore, the expression signature generated using mouse hepatocytes could be applied to identify human HCCs with a prominent activation of the Met pathway. To test this hypothesis, we directly compared the expression profiles of the Met-regulated genes in mouse hepatocytes with those from 242 human HCC samples. Based on a list of curated homologous UniGene clusters (http://ncbi.nlm.nih.gov/UniGene/), we collected available human orthologs of the 730 Met-dependent mouse genes in 2 human HCC data sets. The first set, composed of 139 HCC samples (35, 36), has been previously analyzed in our laboratory (LEC [Laboratory of Experimental Carcinogenesis, National Cancer Institute] set), while the second set, containing expression profiles of 103 HCCs and 7 liver metastases, was obtained from the Stanford University microarray database (37).
Expression values were standardized for each gene by adjustment of SD to 1 and mean to 0 separately across all samples independently in the 3 different platforms as previously described (38). Next, we constructed 2 composite mouse-human data sets. One contained the expression profiles of 440 common genes from the LEC HCCs and from the mouse hepatocyte samples, whereas the second data set included 303 orthologous genes from the Stanford HCC tumors and from the mouse samples. Cluster analysis identified 2 distinctive clusters in both independent mouse-human data sets (Figures 4A and 5A), which divided HCC samples into 2 subgroups based on the presence (Met+) or absence (Met–) of Met gene expression signature. Tumors in the Met+ group showed an expression pattern highly similar to that of HGF-induced control hepatocytes. Strikingly, all 7 extrahepatic liver metastases included in the Stanford data set also exhibited a clear Met activation pattern (Figures 4B and 5B). These results indicate that the Met gene expression signature identified in the mouse primary hepatocytes may successfully discriminate a significant subset of human HCC.
Expression profiles of the Met-regulated genes in HCCs from the LEC data set. (A) Hierarchical cluster analysis of mouse hepatocyte samples and 139 HCCs from the LEC data set. Clustering was performed with 440 common orthologous genes that showed HGF/Met–regulated expression pattern in mouse hepatocytes. Normalized log2-transformed expression ratios are presented in a matrix where columns and rows represent individual genes and samples, respectively. (B) Dendrogram of cluster analysis indicates the presence of an HCC subset (Met+ group) in which the expression pattern of Met-regulated genes is similar to that in the HGF-treated WT hepatocyte samples. HCC samples in the Met– group and Met KO hepatocytes do not share the same gene expression pattern of Met activation. The figure also shows distribution of HCC samples between the previously described (37) worse-prognosis (cluster A) and better-prognosis (cluster B) tumor groups.
Expression profiles of the Met-regulated genes in HCCs from the Stanford data set. (A) Hierarchical cluster analysis of mouse hepatocyte samples together with 103 HCC and 7 liver metastases from the Stanford data set. Clustering was performed with 303 common orthologous genes that showed HGF/Met–regulated expression pattern in mouse hepatocytes. Normalized log2-transformed expression ratios are presented in a matrix where columns and rows represent individual genes and samples, respectively. (B) Dendrogram of cluster analysis shows that several HCC samples and all metastatic tumors (Met+ group) share a similar expression pattern with the HGF-treated control hepatocytes (WT). HCC samples in the Met– group and Met KO hepatocytes do not share the same gene expression pattern of Met activation.
Prevalence of Met-regulated gene expression signature is associated with aggressive phenotype in human HCC. Kaplan-Meier plots and log-rank survival statistics showed that patients with tumors from the Met+ expression subgroup had a significantly shortened mean survival time (35.1 ± 7.15 months) compared with other HCC patients (70.3 ± 9.67 months) (Figure 6A). Hierarchical cluster analysis also revealed significant overlap between the Met+ group and the previously identified (35) bad-prognosis HCC group (cluster A) as shown in Figure 4B. Distribution of other clinicopathological variables among the Met+ and Met– clusters is summarized in Table 2.
Association between the different clinicopathological variables and the Met+ and Met– HCC subsets defined by hierarchical cluster analysis of the LEC samples. Kaplan-Meier plot (A) and log-rank statistics demonstrate the difference in the overall survival between HCC patients from the Met+ and from the Met– groups. (B and C) Bar graphs represent vascular invasion rate (B) and average MVD (C) in HCC samples from Met+ and Met– sets. MVD was determined by quantitative analysis of CD34+ vascular features. (D and E) Microscope images show CD34 staining of endothelial cells in representative samples from Met+ (D) and Met– (E) tumor sets. Pictures were taken with ×200 original magnification.
Clinicopathological variables in the LEC HCC set
Since Met signaling has long been regarded as a promoter of tumor invasion and angiogenesis, we determined the vascular invasion status in HCCs available for histological analysis. As expected, the vascular invasion rate was significantly higher in the HCCs with prominent Met expression signature (χ2 = 4.01, P ≤ 0.05) than in the rest of the tumors (Figure 6B). Previous studies found a good correlation between the expression levels of HGF or Met receptor and microvessel density (MVD) in various human carcinomas (39, 40). Indeed, when MVD was assessed by CD34 immunohistochemistry in the representative HCC samples from each group, the results showed a significant correlation between the presence of Met signature and increased MVD. Accordingly, the average OD of CD34+ vascular features was significantly higher (P < 0.001) in the Met+ (90.78 ± 6.71) than in the Met– (44.55 ± 6.16) HCC subgroup (Figure 6, C–E).
Notably, in the LEC data set, average expression level of the Met was not significantly different between the Met– and Met+ clusters as detected by either microarray analysis or immunohistochemistry (data not shown). However, at least 2-fold upregulation of the Met receptor was found more frequently in the Met+ tumors (5/54) compared with the Met– group (2/85) in the LEC set. These data overlap well with the expression profile–based classification and suggest that, in some HCCs, overexpression of the Met receptor is the driving force behind the Met-dependent expression signature.
The Met expression signature predicts survival of HCC patients. To test the predictive value of the Met expression signature regarding prognosis of HCC patients, the expression patterns of the human homologs of the mouse Met-regulated genes were used to construct a classifier with 6 different supervised prediction algorithms, including the compound covariate predictor (CCP), nearest neighbors 1 and 3 (NN1 and NN3), nearest centroid (NC), support vector machine (SVM), and linear discriminator analysis (LDA) methods. Since survival data were only available for the patients from the LEC group, we randomly divided these samples into a training set (60 samples) and a validation set (79 samples). Next, we selected the common target genes with the matching expression pattern between tumors displaying the high and low Met gene expression signature in the primary hepatocytes and in the training set. Using these genes and all 6 algorithms, classifiers were built according to a leave-one-out cross-validation (LOOCV) strategy. The optimal classifier producing the highest correct classification rate in the training set contained 111 genes. When the classifier was applied to the validation set, all 6 algorithms could identify the subgroups of Met+ tumors. Moreover, membership of the Met+ group showed little fluctuation using different statistical methods (Table 3). Kaplan-Meier survival curves and results of the log-rank tests with all predictors showed that HCC patients with tumors harboring prominent Met gene expression signature have a worse survival rate compared with other patients (Figure 7, A–E). We also applied the prediction algorithms to the Stanford arrays, using the same LEC training set and only genes that were represented in both platforms. In the Stanford data set, the prediction rate of the metastatic liver lesions was 100% with 5 of 6 algorithms.
Survival analysis based on the predicted Met activation in the LEC validation set. Log-rank test results and Kaplan-Meier plots demonstrate the overall survival of HCC patients from the LEC validation set. Patients were stratified into 2 groups based on the expression pattern of preselected classifier genes using the CCP (A), NN1 (B), NN3 (C), NC (D), SVM (E), and LDA (F) algorithms. In the tumors from the Met+ patients, higher activation of the Met signaling pathway is predicted compared with that in the Met– group.
Summary of the class prediction results
Beyond predicting the disease outcome in HCC patients, the classifier genes may also represent the most conserved cross-species part of prominent HGF/Met–regulated expression signature that plays a critical role in the Met-induced cellular transformation. Several of these genes either were previously defined as important contributors to metastasis formation, including HIG2 (41), EPHA2 (42), MAPK3 (43), _P85_α (44), ITG_α_V (29), and ITG_β_1 (29), or could be related to cell motility and invasiveness by their postulated functions (CAP1, ARPC1B, NCK2)(Table 4) (26, 27, 45).
Expression of selected classifier genes in HCC subclasses