A pilot systematic genomic comparison of recurrence risks of hepatitis B virus-associated hepatocellular carcinoma with low- and high-degree liver fibrosis

For this RNA sequencing study, a total of 21 pairs of tumor and non-neoplastic liver tissue samples were selected from HBV-HCC patients who underwent primary surgical resection at the Mount Sinai Medical Center in New York, NY, USA, between 2008 and 2013. Prior to study initiation, all aspects of the research were approved by the Icahn School of Medicine Institutional Review Board. The study protocol conformed to the ethical guidelines of the 1975 Declaration of Helsinki.

The cohort of this RNAseq study is a subset of a cohort previously described [22]. Patients were assessed pre-operatively by dynamic axial imaging (three-phase computerized tomography with intravenous contrast or multi-phase magnetic resonance imaging with intravenous contrast). Liver resection was performed in patients with surgically resectable disease and well-preserved synthetic liver function as assessed by normal serum total bilirubin, albumin, and international normalized ratio. Patients with portal hypertension as evidenced by a platelet count < 100 × 10³/μL, peri-esophageal or peri-splenic varices on axial imaging, or a portal-systemic venous pressure gradient ≥ 10 mm Hg were excluded from liver resection. This cohort included only Child–Pugh A cirrhotic patients since patients with clinical evidence of Child-Pugh B–C cirrhosis were generally not suitable for liver resection surgery.

This RNAseq pilot study included patients who (1) had the largest tumor diameter smaller than 5 cm; (2) had either minimal liver fibrosis (Ishak stage 0–3) or end-stage liver fibrosis (Ishak stage 6) as determined by dedicated pathology review by a single liver pathologist [20]; and had (3) paired fresh frozen tumor and non-neoplastic liver tissue as well as (4) intrahepatic HBV viral DNA copy numbers available. Median follow-up of the survivors was 49 months (4–90 months). There were more males than females included in the study, which is consistent with the sex bias in HBV-HCC [23]. A summary of the clinical information of patients in this study is listed in Table 1. Note that no patient underwent liver transplantation prior to HCC recurrence. One patient, P16, had liver transplantation after HCC recurrence.

All tissue samples used for RNAseq were collected from the first surgical resection. Total RNAs (1–3 μg/sample) extracted from surgical resection specimens were submitted to the Mount Sinai Genomic Core Facility for quality control analysis. The RNA quality was assessed using the Agilent 2100 Bioanalyzer, and the RNA integrity numbers for all 21 pairs of samples were approximately 8.2 ± 0.7 (mean ± SD). The poly(A)-RNA was captured using oligo-dT beads and used for cDNA library preparation using the standard TruSeq RNA Sample Prep Kit v2 protocol (Illumina, CA, USA). Briefly, total RNA was poly(A)-selected and then fragmented. The cDNA was synthesized using random hexamers, end-repaired and ligated with appropriate adaptors for sequencing. The library then underwent size selection and purification using AMPure XP beads (Beckman Coulter, CA, USA). The appropriate Illumina-recommended 6-bp barcode bases were introduced at one end of the adaptors during the PCR amplification step. The size and concentration of the RNAseq library was measured by Bioanalyzer and Oubit fluorometry (Life Technologies, NY, USA) before loading onto the sequencer. The mRNA libraries were sequenced on the Illumina HiSeq 2500 System with 100 nucleotide paired-end reads, according to the standard manufacturer’s protocol (Illumina, CA, USA). Sequence reads were aligned to human transcript reference sequences from the ENSEMBLE database (Homo_sapiens.GRCh37.55.cdna.all.fa) for the expression analysis at gene/transcript levels using TopHat and HTSeq softwares [24, 25]. The raw fastq sequences and the normalized RPKM matrix were deposited in Gene Expression Omnibus database with accession number GSE94660. The HBV reference genome sequence, NC_003977.1, was downloaded from the NCBI database (https://​www.​ncbi.​nlm.​nih.​gov/​nuccore/​NC_​003977.​1) to map reads onto viral transcripts.

DNAseq [13] and RNAseq [26] data for nine paired HCC tumor and adjacent normal tissue samples in a BGI HCC study are publicly available. The WGS data were downloaded from the European Genome-phenome Archive under accession number ERP001196. RNAseq data were downloaded from NCBI Sequence Read Achieve under accession number SRA074279. We ran our pipeline on the DNA sequencing data of 11 N, 11 T, 22 N, 22 T, 30 N, 30 T, 70 N, 70 T, 82 N, 82 T, 180 N, 180 T, 200 N, and 200 T. At the same time, we ran our pipeline on RNAseq data of 18 samples separately (28 N, 28 T, 65 N and 65 T in extra). The integration sites detected from DNAseq and RNAseq data, as well as experimentally validated ones, were used to validate our pipeline and results. In addition, we downloaded RNAseq data of 21 pairs of HBV-positive HCC tumors and corresponding non-tumor tissues in the TCGA Liver Hepatocellular Carcinoma (LIHC) dataset (https://​gdc-portal.​nci.​nih.​gov/​legacy-archive/​search/​f). Among these patients 13, 5, and 2 were white, Asian, and African-American, respectively; the ethnicity of one patient was unknown. We also downloaded transcriptome sequencing data of 21 pairs of non-tumor and HBV-associated HCC [27] from the International Cancer Genome Consortium (ICGC, https://​icgc.​org). Detailed information of the TCGA and ICGC samples used in our study is shown in Additional file 2: Table S1. Additional RNAseq dataset from Chiu et al. [16] with 16 paired HCCs and non-tumorous livers (SRA ID: SRP062885) were also used for pathogenic SNP load analysis.

VirusFinder is an automated virus-host integration detection software package that can deal with virus-induced host genome instability and viral genome variability [21, 28]. It has been shown that VirusFinder performs better than other state-of-the-art virus integration detection pipelines such as VirusSeq [29] and VirusFusionSeq [30] in terms of both accuracy and time efficiency [28]. Our virus integration detection pipeline was based on VirusFinder, with several modifications. Firstly, more candidate sequences were analyzed through our pipeline. One of the main differences was the addition of a re-mapping and confirmation step after potential integration sites were identified to increase pipeline sensitivity and specificity in identifying HBV integration sites (detailed in Additional file 3: Supplementary Materials and Methods). Multiple and different simulation studies were performed to compare HBV identification accuracy between our pipeline and VirusFinder (Additional file 3: Supplementary Materials and Methods).

The procedure was described previously [22] (detailed in Additional file 3: Supplementary Methods). In brief, HBV DNA and cccDNA were amplified from genomic DNA extracted from surgically resected tumor or non-neoplastic liver specimens using the QIAamp DNA extraction kit (Qiagen) [31]. A spectrophotometric ratio of absorbance at 260 nm and 280 nm between 1.8 and 2.0 was assured in all genomic DNA samples. Quantitative PCR was standardized to the human albumin copy number in order to determine the viral DNA copy number/hepatocyte.

For each RNAseq sample, we inferred SNP variants based on RNAseq following the suggested workflow of GATK Best Practices (https://​software.​broadinstitute.​org/​gatk/​documentation/​article.​php?​id=​3891). This workflow is designed specifically for SNP calling based on RNAseq data by modifying original workflow for DNAseq [32]. The workflow consists of the following steps: (1) mapping raw RNAseq reads to reference based on STAR 2-pass alignment [33]; (2) adding read groups, sorting, marking duplicates, and indexing through Picard processing steps; (3) splitting reads into exon segments and hard-clipping any sequences overhanging into the intron regions, (4) base recalibration, and (5) variant calling and filtering using GATK tools. Every parameter was set as default presented in the guide. After inferring the genotype of each sample, tumor variants were compared with those of matching non-neoplastic liver to define somatic mutations for individual SNPs and somatic mutations called for each patient were compared with potential pathogenic SNPs curated in COSMIC mutation data [34]. Pathogenic mutations were defined by Functional Analysis through Hidden Markov Models, which predict the functional consequences of sequence variants [35].

A prognostic nomogram based on clinicopathologic data was developed to predict 2- and 5-year recurrence-free survival [37]. The nomogram scores were calculated for the Mount Sinai dataset and compared between patients with or without cancer recurrence. Higher scores were observed in patients with cancer recurrence (Fig. 2a). However, the nomogram scores for recurrence after 2 or 5 years significantly correlated with the recurrence status only in patients with end-stage fibrosis (Ishak, 6), but not in those with low liver fibrosis (Ishak, 0–3). This result suggests that different recurrence risk models are needed for HCC patients in early or late stage of liver fibrosis and that there may be different underlying mechanisms of tumor recurrence between the two groups.

Our previous studies indicate that intrahepatic cccDNA count and HBV replicative activity were associated with overall survival [22, 31]. Herein, we compared cccDNA counts and HBV replicated activities with regards to HCC recurrence in low and high fibrosis groups (Fig. 2b, c). In general, cccDNA counts were lower and HBV replicative activities were higher in non-neoplastic liver tissues of HCC recurrence for both low and high fibrosis groups. However, the differences were not significant due to the small sample size. We next examined genomic features and underlying molecular mechanisms associated with tumor recurrence in patients with low and high stage of liver fibrosis.

In our previous study, we reported a set of differentially expressed genes in non-neoplastic liver between low and high Ishak staged patients [22]. Herein, the fibrosis stage signatures consistently overlapped with liver cancer survival or recurrence signatures, respectively (detailed in Additional file 3: Supplementary Results, Additional file 4: Table S2), suggesting a prognostic value of fibrosis stage. No significant gene expression change was found between groups with and without recurrence in low or high liver fibrosis in both non-neoplastic liver and tumor tissues. Existing prognosis signatures, including prognostic signatures from Hoshida et al. [38], failed to classify our samples into tumor recurrent or non-recurrent groups (detailed in Additional file 3: Supplementary Results, Additional file 5: Figure S2). This is not surprising given that our samples were specific for HBV-associated HCC with various stages of liver fibrosis. These results indicate that we need to explore other genomic features (e.g., HBV integration sites and SNP patterns) associated with tumor recurrence risk in low or high liver fibrosis groups.

After HBV infection, HBV can insert its genome into the human genome and induce multiple hepatocarcinogenesis events. The power to identify a HBV insertion event depends on the HBV insertion allele frequency (IAF) and sequencing depth and coverage [39]. To enhance the power to detect insertion events of low IAF we modified VirusFinder [21] in multiple steps and developed our own pipeline for HBV integration site detection (Fig. 3a, Methods). Our simulation studies (described in Additional file 3: Supplementary Materials and Methods) suggested that a large fraction of integration sites were not detected at 10× coverage of whole genome sequencing (Fig. 3b). When VirusFinder and our pipeline were applied to the same simulated datasets, our pipeline resulted in more accurate predictions for integrations with low IAFs than VirusFinder in both DNA and RNA sequencing data (Fig. 3c, d). To further validate our pipeline, we applied it to a publically available HBV-HCC dataset, referred to as the BGI dataset, which consists of both whole genome sequencing [13] and RNA sequencing data [26] of the same patients (Methods). Based on WGS data, our pipeline identified 90% (9/10) and 81% (26/32) of the HBV integration sites reported by Sung et al. [13] in normal and tumor tissues, respectively; a few of the integration sites reported by Sung et al. (1 and 6 in normal and tumor tissues, respectively) but not detected by our pipeline were due to low alignment qualities and regions with unknown sequences (Additional file 6: Figure S3, Additional file 3: Supplementary Materials). When applied to RNAseq data in the BGI dataset, our pipeline identified more integration sites than those identified based on WGS data. Additionally, more integration sites in adjacent normal tissues were identified than in tumor tissues based on both WGS and RNAseq data (Additional file 7: Table S3). Interestingly, 24 and 2 integration sites were identified based on both WGS and RNAseq data by our pipeline, but not by Sung et al. [13], in normal and tumor tissue, respectively, suggesting that our approach is sensitive in detecting true HBV integration sites. This observation is consistent with our simulation results that the low sequence depth in WGS is disadvantageous for detecting integration sites, especially in normal tissue, where a relatively lower HBV IAF is expected compared to tumor tissues with clonal expansion [40]. It is also supported by the fact that, generally, more integration sites were obtained from RNAseq than WGS since RNAseq is typically focused on transcript regions with more than tens or hundreds of millions of reads [39]. We also compared HBV integration sites in the TCGA dataset by our pipeline and those from a recent TCGA paper [41], with the results suggesting that our pipeline had greater sensitivity and specificity (Additional file 8: Table S4, Additional file 3: Supplementary Materials).

We applied our pipeline to the RNAseq data for the 21 pairs of non-neoplastic liver and tumor tissues from Mount Sinai (Methods). A total of 407 and 118 unique integration sites within 374 and 106 unique host genes with HBV integration were identified in normal and tumor tissues, respectively (Table 2). All of identified HBV integration sites for non-neoplastic liver and tumor tissues are listed in Additional file 9: Table S5. It is worth noting that the number of host transcripts with HBV S ORF integrated in both non-neoplastic liver and tumor tissues was significantly correlated with serum HBsAg levels (Additional file 10: Figure S4A). Further, the trends were similar for the number of all host transcripts with HBV integration (Additional file 10: Figure S4B), suggesting that fusion transcripts with HBV S ORF may partially contribute to HBsAg levels in serum.

A more than three-fold HBV integration was observed in non-neoplastic tissue compared to tumor tissue, indicating that HBV integration patterns in non-neoplastic tissues are more diverse, consistent with recent results by Chiu et al. [16]. While most HBV fusion transcripts were detected only in one sample, 30 host transcripts with HBV fusion were detected in more than one sample (recurrent integration), and 18 of them were detected in both tumor and non-neoplastic liver tissue (Additional file 11: Table S6). A comparison of HBV integration in tumor versus paired non-neoplastic liver tissues showed a higher number of host transcripts with HBV integration and transcripts with recurrent HBV integration in non-neoplastic liver tissues (Wilcoxon test P = 0.002 and 0.03, respectively, as shown in Fig. 4a). Consistently, more host transcripts with HBV integration were identified in non-neoplastic liver tissues than in the paired tumor tissues when our pipeline was applied to BGI, TCGA, and ICGC HBV-HCC RNAseq datasets (Additional file 12: Figure S5).

To check whether preferential integration sites exist for HBV integration, the breakpoints of integration were counted in both the human and HBV genomes. HBV X gene transcript was more dominantly fused with human genome than other HBV transcripts, especially in normal samples (Fig. 4b), consistent with previous reports [11, 16]. More precisely, the breakpoint in the HBV genome preferentially occurred around nucleotides at nt1818 (Additional file 13: Figure S6A), consistent with previous reports [12, 13, 15]. In the human genome, HBV integration occurred mainly in the gene promoter and intron regions in non-neoplastic liver, while the intron region was the preferential integration site in tumor (Fig. 4c). Only 5–16% of all sequencing reads in each sample were mapped to intronic regions (Additional file 13: Figure S6B), consistent with ratios observed in other studies [42, 43]. However, HBV integrations preferentially occurred in promoter and intronic regions (Fig. 4c), suggesting regulatory roles of HBV integration in fusion gene expression. Moreover, Chiu et al. [16] reported that intronic HBV integrations have oncogenic properties. This pattern of HBV integrations preferentially occurring in gene promoter and intronic regions was also identified in the BGI and TCGA LIHC datasets (Additional file 13: Figure S6C), which was consistent with previously reported studies based on transcriptome sequencing [12, 16]. HBV integration sites were observed across entire chromosomes, while chromosome 1, 2 and 4 contained more than 30 fusion transcripts in non-neoplastic liver tissues (Fig. 4d).

HBV fusion transcripts identified in the Mount Sinai dataset were compared with integration results identified in other datasets or reported in previous studies [12, 15, 16] (Fig. 4e). Our results significantly overlapped with the HBV host transcripts identified based on RNAseq data of BGI (Fisher’s exact test (FET) P =1.8 × 10^-21 and 4.2 × 10^-15 for non-neoplastic liver and tumor tissues, respectively), TCGA LIHC dataset (FET P = 8.2 × 10^-5 and 3.9 × 10^-5 for non-neoplastic liver and tumor tissues, respectively), and ICGC HBV-HCC RNAseq dataset (FET P = 2.3 × 10^-8 and 0.0001 for non-neoplastic liver and tumor tissues, respectively). Individual HBV integration sites identified in these dataset are listed in Additional file 14: Table S7 and were also consistent with previously reported HBV fusion transcripts in several previous studies (Fig. 4e) [12, 15, 16]. While some fusion transcripts were commonly found in both tumor and normal tissues across different datasets, several HBV fusion transcripts were restricted to normal or tumor tissues. For example, some known oncogenes, such as KMT2B and TERT, were dominantly observed in tumor while fusion transcripts with CYP3A5, SERPING1, and WDR72 were only found in normal tissue. The most frequently identified fusion transcript in our dataset was FN1 (8/42, 19%); however, the frequency was biased towards normal samples (7 and 1 occurrence in normal and tumor tissues, respectively). This was consistent with previous studies indicating that FN1 is frequently targeted for HBV integration at the transcript level [44], but that it is not a cancer driver gene.

Host genes with HBV integration in non-neoplastic liver tissues were enriched for biological processes such as cell adhesion (P = 0.0002) and Wnt receptor signaling pathway (P = 0.005), whereas those in tumor tissues were enriched for platelet degranulation and activation (P = 4.9 × 10^-5) (Additional file 15: Table S8). Detailed results of functional analysis for the host genes with HBV integration are reported in Additional file 3: Supplementary Materials and Methods. Host genes with HBV integration detected in non-neoplastic tissues were significantly enriched for tumor suppressor genes [45] (P = 0.004; Fig. 5a, Additional file 16: Table S9). In addition, the host genes with HBV integration significantly overlapped with COSMIC cancer census genes [46] (P = 0.03 and 0.02 for non-neoplastic and tumor tissues, respectively), suggesting that cells with these HBV integrations likely resulted in a growth advantage during clonal expansion. HBV-human gene fusion events may alter the host gene expression (Additional file 3: Supplementary Materials and Methods). For example, KMT2B expression level was higher in tumor tissues in which HBV-KMT2B fusion transcripts were detected (Additional file 17: Figure S7A).

Host transcripts with HBV integration identified in non-neoplastic liver tissues in HCC recurrence groups were significantly enriched for tumor suppressor genes [45], while those in non-recurrence groups were not (Fig. 5a, Additional file 18: Table S10). The number of host transcripts with HBV insertion identified in non-neoplastic liver tissues in recurrence groups was less than that identified in non-recurrence groups for both low and high fibrosis (Fig. 5b, left), but the differences were not significant. In conjunction, these results suggest that there are selective clonal expansions in non-neoplastic liver tissues with a high risk for HCC recurrence.

Similarly, the number of host transcripts with HBV integration identified in tumor tissues in recurrence groups was lower than that identified in non-recurrence groups for both low and high fibrosis (Fig. 5b, right), and the difference in the low fibrosis group being statistically significant (P = 0.04). This further suggests that the tumorigenesis mechanisms for low and high fibrosis groups are likely different and therefore the exact tumorigenesis mechanism for each group needs further investigation.

To investigate what factors determine the number of host transcripts with HBV integration we compared these with HBV cccDNA count and HBV replicative activity (Additional file 2: Table S1). The larger number of HBV integration events was significantly associated with higher HBV cccDNA counts in non-neoplastic liver tissues (Wilcox test P = 0.004, Fig. 5c); this was also the trend in tumor tissues. There was a similar pattern between the number of HBV integration events and HBV replicative activity, but the association was not statistically significant (Fig. 5d).

Chronic inflammation induced by HBV infection may trigger somatic mutations. Therefore, we investigated whether the number of potential pathogenic mutations in cancer census genes (defined as pathogenic SNP load, Methods) is associated with liver fibrosis stage and tumor recurrence. In order to ensure a fair comparison between normal liver and tumor tissues, we also randomly selected 20 normal liver tissue samples from the GTEx dataset [47] and compared pathogenic SNP loads called for non-neoplastic liver and tumor samples in the Mount Sinai, BGI, TCGA, ICGC, and Chiu et al. [16] datasets. After SNPs were inferred for each sample, we selected those overlapping with pathogenic SNPs curated in the COSMIC dataset [35, 46] (Methods). The pathogenic SNP load was associated with tissue type and increased in the order of normal liver (GTEx), non-neoplastic liver tissues, and tumor (Fig. 6a). The pattern in the Mount Sinai dataset was consistent with results from the BGI, TCGA, ICGC, and Chiu et al. [16] datasets. The pathogenic SNP loads in TCGA non-neoplastic liver tissues were close to the pathogenic SNP loads in normal liver tissues. It is worth noting that HBV integrations were identified in only 7 of 21 pairs of samples in the TCGA HBV-HCC dataset. When considering only pathogenic SNPs in these seven samples (*TCGA in Fig. 6a), the pathogenic SNP load was significantly higher than that in normal liver (Wilcox P = 0.005). Genes with pathogenic mutations (Methods) were significantly overlapped with genes with HBV integration in non-neoplastic liver tissues across all datasets (FET P = 0.0001, 0.0009, 0.009, and 0.008 for the Mount Sinai, BGI, TCGA, and ICGC dataset, respectively; Additional file 19: Table S11), but not in tumor tissues, suggesting that HBV integrations in non-neoplastic liver tissues and functional somatic mutations target the same set of genes important for tumorigenesis.

When Mount Sinai samples were further separated based on liver fibrosis and tumor recurrence status, there was a significant association between the number of potential pathogenic SNPs and liver fibrosis in non-neoplastic liver tissues (Fig. 6b). Further, pathogenic SNP loads were higher in patients with end-stage fibrosis than in other patients. Pathogenic SNPs and somatic mutations identified in Mount Sinai, TCGA, and ICGC samples with low and high liver fibrosis were significantly overlapped (Additional file 20: Figure S8A, P values for overlap are listed in Additional file 21: Table S12). Even though more pathogenic SNPs were identified in tumor tissues, a higher percentage of pathogenic SNPs identified in non-neoplastic liver were common across the three datasets than in tumor tissues in both low fibrosis and cirrhosis groups, suggesting that pathogenic SNPs in non-neoplastic tissues are important in tumorigenesis. Genes with common pathogenic SNPs or mutations were compared with GO biological processes (Additional file 20: Figure S8B). The genes with common pathogenic mutations identified in the non-cirrhosis group were significantly enriched for the biological process response to DNA damage (P = 0.0035), but the ones identified in the cirrhosis group were not (P = 0.23), suggesting potentially different mechanisms of tumorigenesis in non-cirrhotic and cirrhotic liver.

While the pathogenic SNP load itself was not associated with tumor recurrence status, the number of pathogenic mutations measured by comparing genotype between non-neoplastic liver and tumor tissues was significantly different between tumor recurred and non-recurred patients in both low and high liver fibrosis groups (Fig. 6c), and the number difference of pathogenic mutations between recurred and non-recurred patients was much larger in the low compared to the high fibrosis group, suggesting that different recurrence risk models are needed for patients of low and high fibrosis. We also tested whether the potential pathogenic SNPs and somatic mutations were associated with cccDNA or HBV replicative activity, but no clear differences were observed (Additional file 22: Figure S9, Additional file 3: Supplementary Materials and Methods). Further investigation of pathogenic mutations at gene level identified 10 and 16 genes that were preferentially mutated in the recurrence groups of low and high liver fibrosis, respectively (Fig. 6d, Additional file 23: Table S13). The significance of the bias pattern was assessed by permutations. Several of the genes with mutations that preferentially occurred in the recurrence groups are known for their association with HCC. For example, COL21A1, mutated in all four samples in the low fibrosis group, was reported as somatically mutated in two out of a nine intrahepatic metastatic samples in a HBV-HCC cohort [48]. The same study also reported somatic mutations in CSMD1, CDC27, SEH1L, and ATXN1 in their intrahepatic metastatic samples. HOXA7, mostly mutated in the high liver fibrosis group, was reported to promote metastasis of HCC with activation of Snail [49], while decreased expression of GATA2 was correlated with poor prognosis of HCC [50]. In addition, somatic pathogenic mutations related to tumor recurrence in low and high fibrosis identified in the Mount Sinai cohort also occurred in the TCGA dataset (Fig. 6d). For example, three out of five patients of non-cirrhosis with tumor recurrence had pathogenic somatic mutations in COL21A1, ITPR1, and SCAI. However, the information in the TCGA dataset was not sufficient to assess the significance. Considering all of the above, our results suggest that the extent of pathogenic SNPs and/or somatic mutations could provide potential information for HCC recurrence.