Clinicopathologic and gene expression parameters predict liver cancer prognosis

A total of 272 HCC patients were included in the initial training dataset. Resected tumor and adjacent non-tumor liver tissues were collected from patients who had undergone hepatectomy for curative treatment of HCC at Queen Mary Hospital, Pokfulam, Hong Kong between 1993 and 2007 [10]. Informed consents were obtained from patients regarding to the use of the liver specimens for research. Additional File 1, Table S1 summarized the demographic and clinicopathologic features of these patients. All samples came from individuals who provided written informed consent to make their samples available for scientific research. In addition, all of the samples and patient data were approved for use in this study by IRBs specific to each of the participating organizations. Experimental research reported in this paper was also approved by IRBs of each participating organizations. Research reported in this paper was in compliance with the Helsinki Declaration.

The clinicopathologic features of the patients analyzed were sex, age, tumor size, number of tumor nodules, cellular differentiation according to the Edmondson classification, venous infiltration without differentiation into portal or hepatic venules, tumor node metastasis stage (pTNM and AJCC), serum hepatitis B surface antigen (HBsAg) status, and background liver disease in non-tumorous liver tissue [12].

Fresh frozen tissue was placed in a chilled milling tube along with a stainless steel bead, dipped in a liquid nitrogen bath and loaded onto the QIAGEN TissueLyser for milling (30 Hz in 30 second intervals). Multiple cycles of milling were sometimes required to achieve complete pulverization of the tissue to a fine powder. Isolation of RNA was achieved using the following procedures. The milled tissue samples were homogenized in cryopreservation tubes using a Polytron with disposable rotostator probes. The tissue was homogenized in 750 to 1000 uL of 100% TRIzol. 100% Chloroform was added to the TRIzol/GITC lysate (1:5 ratio) to facilitate separation of the organic and aqueous components using the phaselock (Eppendorf) system. The aqueous supernatant was further purified using the Promega SV-96 total RNA kit, incorporating a DNase treatment during the procedure. Isolated total RNA samples were then assayed for quality (Agilent Bioanalyzer) and yield (Ribogreen) metrics prior to amplification.

Samples were amplified and labeled using a custom automated version of the RT/IVT protocol and reagents provided by Affymetrix. Hybridization, labeling and scanning were completed following the manufacturer's recommendations (Affymetrix). Sample amplification, labeling, and microarray processing were performed by the Rosetta Inpharmatics Gene Expression Laboratory in Seattle, WA. The expression data has been deposited in GEO (GSE25097, http://​www.​ncbi.​nlm.​nih.​gov/​geo/​).

Four recent studies were employed as independent validation. The first study consisted of 23 HCC tumor tissues collected in Singapore and assayed on the Affymetrix GeneChip HU133, resulting in the identification of a 57-gene signature associated with tumor recurrence [4]. The second study employed formalin-fixed paraffin-embedded non-tumor (i.e., adjacent normal) liver tissues from 82 Japanese HCC patients. A 186-gene signature was derived from this study using an Illumina array containing 6000 human genes [9]. The third study involved an HCC metastases cohort from Asia (N = 115 normal tissues) in which an Incyte 9, 180-reporter two channel array was used to profile the normal liver tissues [13]. Finally, a China-Belgium study (sample size N = 90) of mostly tumor tissues was profiled on a Qiagen 70-mer two channel array [14].

Tumor and adjacent non-tumor liver tissues were collected from 272 Chinese HCC patients who received curative surgery (referred to here as the HCC cohort). Additional File 1 Table S1 summarizes the characteristics of this cohort. Nearly 2/3 of the patients were right-censored (67.8%) and the other 1/3 (32.2%) of the patients were deceased (failure) upon data analysis. Half of the patients (51.1%) suffered from tumor recurrence during the follow-up period. The primary endpoints (overall survival and DFS) were found to be significantly associated with tumor size, AFP, ALBU, venous infiltration, pTNM and new AJCC stage, and NOTN. Previous analysis has shown that these clinicopathologic parameters can classify patients in the HCC cohort into two groups (denoted as good and poor prognosis groups) that give rise to distinct clinical outcomes [10]. For the survival endpoint, the good and poor survival groups contained roughly equal numbers of patients and were significantly different with respect to clinical outcome (log-rank test p-value of 6.0E-6; Figure 1 left panels). The linear predictor (h) shown in Figure 1 was derived using a leave-one-out (LOO) cross validation procedure to reduce biases that result from over fitting. Similarly, we classified the patients into good-DFS and poor-DFS groups (log-rank test p-value = 5.6E-9, Figure 1 right panel) [10]. Within each survival or DFS group, the clinicopathologic parameters could not further partition patients into subgroups of significantly different prognosis. In other words, the variation of prognosis within each group was not explained by clinicopathologic parameters.

To advance our understanding of the molecular processes and to extend the predictive power of the clinicopathologic variables, we explored the relationship of gene expression to outcome. Tumor and adjacent normal tissues were profiled using the Affymetrix whole genome expression array. The experiment surveyed 37, 585 unique transcripts, among which 23, 788 were mapped to well-annotated genes. Using the gene expression traits to predict group membership, we observed similar predictive power as was achieved using the clinicopathologic parameters alone.

In order to identify biomarkers that further enhance prognosis prediction based on the clinicopathologic parameters, a stratified analysis using univariate gene expression feature screening was carried out within each prognosis group (Figure 1&2). Under the null hypothesis the pvalues generated by this screening should follow a uniform distribution, substantial enrichment of low pvalues indicated a large number of prognosis-associated genes that might offer predictive power, shown Additional File 2 Figure S1. However, many of the prognosis-associated genes captured the same information as the clinical parameters, and contributed no extra prediction power. In the stratified analysis, we were able to identify genes associated with survival or DFS in each prognosis group. For example, many genes in the tumor tissue were associated with survival in the good-survival group but not in the poor-survival group (Additional File 3 Figure S2), suggesting that tumor gene expression may further enhance prognosis prediction in the good-survival group but not the poor-survival group. Interestingly, there were many normal tissue gene expression traits associated with DFS in the good DFS group, while many tumor tissue gene expression traits associated with DFS in the poor DFS groups (Additional File 3 Figure S2), again suggesting that the gene expression data may help stratify patients according to DFS beyond what could be achieved by the clinicopathologic parameters alone.

Motivated by these results, we combined the gene expression data with the clinicopathologic parameters to construct models to predict HCC prognosis and compare to those models constructed from the clinicopathologic data alone. We again employed a LOO cross validation strategy in each subset of patients to minimize over-fitting. For example, in the good-survival group, the LOO procedure was performed on the 113 patients with available normal tissue gene expression data (Figure 3, left panel).

The normal tissue expression data improved the prediction for the good-survival group (p-value = 0.011, Figure 3 left panel). The blue line in Figure 3 represented an excellent survival function (over 90%) for a group of patients carrying both favorable clinicopathologic and gene expression profiles. The same procedure was conducted on the 117 patients in the poor-survival group with available normal tissue gene expression data. As expected, the expression data could not further improve the predictive power of the model (Figure 3, right panel). Prediction in the good-DFS (p-value = 0.0027), but not the poor-DFS group was also further improved by incorporating normal tissue expression data into the model (Figure 4).

We carried out the same analysis just described using the tumor tissue expression data. It was noted that the sample sizes of the tumor tissue analysis was different from that of normal tissue (Figure 1&2). Figure 5 shows that the prediction in the good-survival (p value = 9.0E-4) and poor-survival (p-value = 0.0023) groups were further improved by the incorporation of the tumor expression data. Curiously, the pattern was opposite for DFS compared to what we observed in the normal tissues. For the poor-DFS group, tumor expression significantly further partitioned patients into good and poor DFS subgroups (p-value = 5.5E-8; Figure 6, right panel). No prediction improvement for the good-DFS group was observed. Since both normal and tumor tissues had predictive power in the good-survival group (Figure 3 and 5), we were motivated to explore predictive models incorporating expression traits from both tissue types. The number of patients with both tissues available was N = 110. The combined data resulted in a further improvement of the prediction performance (Additional File 4 Figure S3).

The identification of prognosis-associated gene signatures in HCC is an active field of research, with previous reports indicating that different gene signatures may be harvested from tumor and adjacent non-tumor tissues [4, 9, 13, 14, 16]. To capture information from the expression data that was not yet covered by the clinicopathology, the clinicopathologic parameters should be conditioned on during the search for expression signatures, as we have done. Unfortunately, many of the reported studies did not follow such a strategy. Ideally, our gene signatures would be tested in independent HCC cohorts. However, raw clinical outcome data was not available for any of the published studies [4, 9, 13, 14]. Therefore, we compared our gene signature, which showed predictive value towards HCC prognosis, with published signatures. We would not expect perfect concordance among the signatures given the different patient recruitment criteria, experimental conditions, array platforms, and statistical methods across the studies. For example, five different array platforms were used in studies listed in Table 1 and 2: Affymetrix whole genome custom array (current study), Illumina 6000-gene human array [9], Qiagen 70-mer two channel array [14], Incyte 9, 180-reporter two channel array [13], and Affymetrix GeneChip HU133A&B arrays [4]. More importantly, since the sample sizes for the published studies were all small (n < 200), the power to capture prognosis-associated genes was low. Hence, we applied liberal p-value cutoffs (e.g. 5E-4) in order to capture a majority of the HCC prognosis-associated genes and in turn enhance the comparison to the published signatures (Table 1, 2 and Additional File 5 Table S2). The survival- and DFS-signatures we identified significantly overlapped with all published gene lists (Table 1). For example, Lee et al reported a comprehensive HCC gene signature on a Chinese-Belgium dataset that overlapped our results in a highly significant manner (4.8 fold enrichment; Fisher Exact Test p-value = 1.4E-12 for the tumor tissue DFS signature, Table 1 and 2). As one would expect, when the study design of a published gene signature matches our design, we observed a more extensive overlap with the signatures. For example, Hoshida et al. [9] and Budhu et al. [13] studied adjacent non-tumor liver tissues in HCC patients. Their gene lists overlapped with our normal tissue signature more significantly than with the tumor signature. Alternatively, Lee et al. reported a gene list that overlapped heavily with our tumor tissue signatures. It is noteworthy that the hypergeometric test p-values were driven by both magnitudes of overlaps and the size of the published gene lists. The overlap among the various gene lists were summarized in Additional File 6, 7, 8 and 9, Table S3. The first part of Additional File 5, 6, 7, and 8, Table S3 highlighted genes that reached genome-wide significance in our HCC data. There was no adjustment for clinicopathologic parameters in order to retain consistency with other published signatures. A strict cutoff (Cox p-value ≤ 2e-6) was applied to the identified genes that were genome-wide significant after a Bonferroni correction. The second part of Additional File 6, 7, 8 and 9, Table S3 summarized genes that appeared in at least two other gene lists.

To explore the predictive genes in the context of a normal-to-tumor network reconfiguration, we looked at whether the genetic perturbations of the predictive genes were associated with HCC prognosis. Two sets of genes were explored in this way: 1) genes differentially connected between the normal and tumor tissue co-expression networks (gene-set-1), as detailed by Lamb et al [17]; and 2) genes whose expression levels were significantly associated with copy number abnormalities (CAN) in the tumor tissue (gene-set-2).

To explore whether these gene sets were enriched for eSNPs that associate with HCC prognosis, we genotyped DNA isolated from adjacent normal tissues using the Illumina 650Y SNP arrays and characterized the genetic architecture of gene expression based on a previously described method [18]. This genome-wide association study of gene expression resulted in 1, 296 cis expression QTLs (eQTLs) [18] at a 10% FDR (translating to p-value cutoff = 9E-6). We termed the significant SNPs (associated with the corresponding expression trait at a p < 9E-6) underlying the eQTLs as eSNPs. There were 707 and 2, 840 eSNPs for gene-set-1 and gene-set-2, respectively. Further, we examined if the eSNPs linked to gene-set-1 and -2 were enriched for SNPs that associated with clinical endpoints (referred to as clinical SNPs or cSNPs). In brief, the cSNPs were identified using Cox models focusing on survival or DFS. The model was adjusted for clinicopathologic parameters.

At an α level of 0.01, we identified 6, 399 cSNPs associated with HCC survival (referred to as survival-cSNPs). Interestingly, the eSNPs underlying gene-set-1 were 2.1-fold enriched for survival-cSNPs (p = 7E-4); and the eSNPs underlying gene-set-2 were 1.7-fold enriched for survival-cSNPs (p = 2E-5). In parallel, at an α level of 0.01, the SNP screening using a Cox model yielded 5, 246 SNPs associated with DFS (referred to as DFS-cSNPs). We found the eSNPs underlying gene-set-1 were not enriched for DFS cSNPS; in contrast, eSNPs for gene-set 2 were 1.9-fold enriched for DFS cSNPs (p-value = 5E-7). For comparison, we randomized the clinical endpoints and repeated the Cox modeling, and at α level of 0.01, we yielded pseudo-cSNPs, which arose purely from type I errors. The eSNPs underlying gene-set-1 and -2 were not enriched in the pseudo-cSNPs.