Somatic events detected by WGS

We extracted DNA from 22 frozen HBV-related HCC tissues and their matched normal lymphocytes, and sequenced their whole genomes by massive parallel sequencing. Clinical and pathological information is shown in Table S1 in S1 File. Average sequencing depths of the cancer and control (lymphocytes) genomes were 36.8× and 29.9x, respectively, after the removal of PCR duplications (Table S2 in S1 File). In total, 209,055 (3,597–19,063) somatic substitutions were detected (Table S4 in S1 File). Among them, 1,096 missense, 32 nonsense and 35 splice-site mutations in protein-coding regions, were identified, where splice-site mutations were defined to be those affecting splice acceptor and donor sites located at the first and last two bases of an intron sequence (essential splice-sites). Of the 5,725 indels identified in the whole genomes, 105 were located in protein-coding regions and 6 affected essential splice-sites. The inferred driver genes in order of statistical significance of recurrence were TP53, ARID2, BRD7, HNF4A and RPS6KA3 (P-value <0.001, Table S5 in S1 File). In addition, 2,254 SVs (15–577 per tumor) were detected, 1,168 of which affected annotated protein-coding genes. Furthermore, 86 (0–12 per tumor) HBV integrations were identified with 2 and 5 recurrent integrations at the TERT and MLL4 loci, respectively, which is consistent with previous studies [4], [6], [7], [8].

Splicing aberrations related with genomic mutations

Total RNAs extracted from the frozen 22 HCCs and their adjacent non-cancerous liver tissues were subject to RNA-Seq, and its summary is shown in Table S3 in S1 File.

First, we investigated the status of transcripts around essential splice-site mutations by manually checking the alignments of sequence reads. After excluding transcripts with no expression, we observed single or multiple splicing aberrations for 19 out of the remaining 24 essential splice-site mutations (3 splice-site slips, 6 exon skips, and 13 intron retentions, Figure S1 in S2 File, Table S6 in S1 File), indicating that the somatic mutations at essential splice-sites showed strong effects on splicing aberrations as expected. Affected genes included recurrently mutated genes from WGS analysis (TP53, ARID2, HNF4A and RPS6KA3) as well as AXIN1.

In order to obtain a comprehensive list of cancer-specific splicing aberrations, we systematically detected four types of splicing aberrations (splice-site slip, exon skip, pseudo-exon inclusion and intron retention, Fig. 1B) by using internally developed algorithms (see Materials and Methods). Overall, 292 splicing aberrations events (26 splice-site slips, 41 exon skips, 77 pseudo-exon inclusions and 148 intron retentions) were detected (Table S7 in S1 File). Since we performed non-directional RNA-Seq, discriminating between intron retention and cancer-specific antisense transcription was difficult, and thus the results of intron retentions should be carefully interpreted. This list included aberration events corresponding to the 10 essential splice-sites investigated above. PCR and Sanger sequencing validated 154 highly cancer-specific splicing events out of the 239 detected. For 72 events, we confirmed target splicing aberrations for both cancer and non-cancerous liver tissues.

In addition to essential splice-site mutations, we identified various types of mutations and short indels that appear to be the direct causes of the observed splicing aberrations, through changing the edit distances of splicing donor or acceptor motifs. Three mutations near exon-intron junctions (<10 bp), but not in the essential splice-site caused splicing aberrations (Fig. 2A, B, and C). Three mutations in coding regions, which likely generated novel splice-site donor motifs, caused splice-site slips (Fig. 2D, E, and F). One mutation affecting the LAMB2 gene was synonymous. Two pseudo-exon inclusion event affecting THRAP3 and TSC1 seemed to be triggered by somatic mutations deep within introns proximal to the new splicing junction point (Fig. 2G, and H). Furthermore, seven exon skips, whose affected genes included several tumor suppressor genes such as IQGAP2, ST7 and TP53, had long deletions between the junction points (Fig. 2I, J, and K). In addition, eight intron retentions had rearrangements within the corresponding introns, for example RB1 (Fig. 2L). They indicate that various types of splicing aberrations are frequently driven by not only essential splice-site mutations, but also mutations in coding regions, including synonymous silent mutations, deep intronic mutations, and SVs. On the other hand, a large part of splicing aberrations (239/292 = 81.8%) did not have proximal mutations (within 1 kb) or SVs (within 500 kb). Some of these are likely seemed to be caused by epigenetic changes [14], or expressional changes in anti-sense transcripts as noted above.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 2. Several examples of genomic changes other than essential splice-site mutations causing splicing aberrations obtained from our comparative whole genome and transcriptome sequencing analyses.

Exonic and intronic sequences are designated by capital and small letters, respectively. Red sequences are somatic mutations in HCCs. Blue and green numbers on the side of sequences are edit distances from splicing donor motif (AG|GTRAGT, [38]) and splicing acceptor motif (YYYYNCAG|G), respectively. Most somatic mutations changed the edit distance to splicing donor motifs so that the corresponding alteration can be enhanced.

https://doi.org/10.1371/journal.pone.0114263.g002

Fusion transcripts related with genomic SVs

To detect gene fusions with unannotated transcripts and/or viral sequences such as HBV, we used Genomon-fusion (see Materials and Methods). We detected 245 candidates of human-human fusion transcripts and 192 gene fusions after removing redundant splicing variants (Figure S2 in S2 File and Table S8 in S1 File), 66 of which involved transcripts without gene annotation (UCSC known genes, RefSeq, Ensemble), and 21 of which were un-spliced fusion transcripts sharing the breakpoints with their corresponding genomic SVs (Figure S3 in S2 File). RT-PCR followed by Sanger sequencing validated 113 (71.9%) of 157 fusion transcripts.

Through comparison with WGS data, 83 gene fusions were found to be supported by somatic SVs at the corresponding genomic locations (Figure S4 in S2 File). While some of gene fusions without observed corresponding SVs may be ascribed to either false positives for gene fusions in RNA-Seq analysis or false negatives for SVs in WGS analysis, ratios of expression values of fusion transcripts imply the existence of minor sub-clones with undetectable associated SVs (Figure S5 in S2 File). We also detected 147 gene fusions in non-tumor liver tissues (Table S9 in S1 File), many of which involved genes with extremely high expression values in liver tissues, such as ALB, HP, and TF, suggesting that detected fusion transcripts may also have originated from SVs harbored within minor sub-clonal liver cells (Figure S5 in S2 File).

Among them, NBEAP1 (BCL-8) fusion transcripts were recurrently detected and validated in two HCCs, with over-expression specific to both specimens (Fig. 3 and Figure S6 in S2 File). Rearrangements involving the BCL-8 locus with over-expression were reported to occur in about 4% of diffuse large-cell lymphoma [15]. Many fusion transcripts affecting chromatin modification pathway genes (CHD4, CTCF, KDM4C and HDAC4) were detected, and fusion transcripts with known tumor suppressor genes (TSC1 and SUFU), a component of the crucial NF-κB modulator (IKBKB), and a key meditator of the WNT signaling pathway (TCF7L1) were also validated [16]. Although no specific over-expression resulted from these gene fusions, we speculate several of them have a loss-of-function nature though the loss of physiologically important domains (Figure S7 in S2 File).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 3. The expression profiles of 22 HCCs and non-cancerous liver samples for eight over-expressed genes.

Blue and red bars show the FKPMs for HCCs and the corresponding non-cancerous liver, respectively, which is calculated by RNA-Seq data. Red circles indicate samples with HBV integrations on the loci of the overexpressed genes. Green circles indicate those with gene fusions and/or SVs that can drive gene over-expression.

https://doi.org/10.1371/journal.pone.0114263.g003

HBV integration and its effects on transcription

Overall, 33 HBV-human fusions were detected, including those affecting TERT (2 samples) and MLL4 (5 samples), and WGS could identify associated HBV integration sites for 23 of these 33 fusions (Table S10 in S1 File). HBV integrations with associated gene fusions tended to have breakpoints concentrated around the locus of HBx genes (1770 bp–1830 bp) in the positive direction (Figure S8 in S2 File), as reported previously [8].

Interestingly, 7 discrete HBV-TERT fusion transcripts were detected in one sample (RK010), which appeared to be derived from one HBV integration site (Fig. 4A). 4 out of the 7 variants are inferred to be in-frame. RK166 had HBV integration just before the transcription start sites of TERT, which generated the full TERT transcript directly connected to HBV sequence (Figure S9 in S2 File). Both samples showed a marked over-expression of TERT compared to other samples (Fig. 3). A HBV-CDK15 gene fusion detected in RK050 also had multiple fusion transcripts including one in-frame fusion which caused CDK15 over-expression (Figure S10 in S2 File). The breakpoints of spliced HBV-human fusion transcripts were concentrated at the HBV genome coordinate of 458bp. We call this position the HBV fusion splicing hotspot.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 4. HBV integrations and fusion events in 22 HCCs.

(A) Seven HBV-TERT fusion transcripts were detected in RK010. One transcript was an un-spliced transcript having the same breakpoint as the genomic integration breakpoint. The others existed in spliced forms and GT-AG splicing motifs were observed at the breakpoints of all but one. In addition to HBV fusion splicing hotspot (458 bp), 3 fusion transcripts were spliced at the coordinate of 1634 bp coordinates in HBV sequences. One fusion transcript included a newly generated 87 bp pseudo-exon sequence as well as subsequence exonic sequences. (B) HBV integrations in the MLL4 loci and their resultant fusion transcripts in five samples. Green triangles on the genome sequence show the HBV integration sites. Most fusion transcripts shared breakpoints with those of genomic HBV integration coordinates for both sides, and thus, they appear to exist in un-spliced forms. The fusion transcripts for RK141 and RK159 were validated to be concatenated (Figure S11). (C) HBV-FN1 fusion transcripts for 7 adjacent non-cancerous liver samples. Almost all the fusion transcripts had the breakpoint at the HBV fusion splicing hotspot. The other fusion transcripts which had breakpoints at intronic regions appear to be un-spliced transcripts around the integration sites.

https://doi.org/10.1371/journal.pone.0114263.g004

On the other hand, gene fusions involving MLL4 showed different patterns. While most observed gene fusion contained HBV on the 5′ end, we detected two types of HBV-MLL4 gene fusions: those with HBV on the 5′ side and those with HBV on the 3′ side (Fig. 4B). PCR with subsequent Sanger sequencing validated that these were parts of concatenated unspliced fusion transcripts of MLL4-HBV-MLL4 for at least two samples (Figure S11 in S2 File). No evidence of splicing was obtained for these fusion transcripts except for one in RK141. Although slightly increased expression of MLL4 was observed in HCC samples, these out-of-frame HBV-MLL4 fusion transcripts suggest that HBV integrations on MLL4 loci may lead to loss-of-function.

In non-cancerous liver tissues, 161 HBV-human fusion transcripts were detected (Figure S12 in S2 File and Table S11 in S1 File). Notably, HBV-FN1 gene fusions were recurrently observed in 7 non-cancerous liver tissues, and most of them had multiple splicing variants including in-frame fusion transcripts with the HBV fusion splicing hotspot as in the HBV-TERT fusion (Fig. 4C).

Over-expression caused by somatic SVs

There are several documented examples of chromosomal SVs leading to ectopic expression of downstream oncogenes with and without the formation of gene fusions [11]. Here, we examined the relationships between somatic SVs and over-expression of genes adjacent to the breakpoints. Through iterative application of the Grubbs-Smirnov test, we identified 3,735 over-expression events in 3017 genes in 22 HBV-related HCCs. The list of over-expression events included TERT and CDK15 over-expression driven by HBV integration. Proximal SVs and HBV integrations (within gene regions or 500 kb up-stream from transcription start sites) could be associated with 63 and 5 over-expression events, respectively (Table S12 in S1 File). Of those breakpoints, 44 events occurred within the promoter regions, suggesting mechanisms of over-expression other than gene fusions are common phenomena. Although it is difficult to confirm each observed over-expression event is truly the consequence of the identified SVs, statistical significance by permutation test (P-value <0.0001, Figure S13 in S2 File) indicates many over-expressing events could be driven by somatic SVs.

Interestingly, over-expression events of WNT ligands were recurrently observed in two HCCs. Over-expression of WNT1 and WNT10B (Fig. 3) in RK107 had associated SVs. Although a gene fusion involving WNT10B was observed, this did not seem to be the direct cause of over-expression because WNT10B was the upstream side of fusion transcripts (Figure S14 in S2 File). For WNT3A overexpression in RK010 (Fig. 3), we detected fusion transcripts involving WNT3A and a few supporting read pairs for them in WGS data, suggesting complex rearrangements around the WNT3A locus may drive its over-expression. Furthermore, the over-expression of c-KIT with associated SV could be detected in RK092 (Fig. 3). The above findings indicate that SVs can play an important role in over-expression of oncogenes and molecular target genes.

Complementary detection of somatic mutations and cancer-specific RNA-editing events

We investigated cancer-specific SNVs (single nucleotide variant) and short indels in the RNA-Seq data using the EBCall algorithm [17], and detected 6,024 candidates at RNA levels, including 1,205 nonsynonymous or splice-site mutations. Among them, WGS analysis integrated with RNA-Seq data (See Materials and Methods) indicated evidences of 1,912 somatic mutations (545 nonsynonymous or splice-sites). There is a certain level correlation between the allele frequencies of somatic mutations found to be highly confident in WGS and RNA-Seq (correlation efficiency  = 0.466, P-value  = 2.2×10⁻¹⁶ by Pearson's product-moment correlation, Figure S15 in S2 File), but the amount of correlation is not very strong. This may be because changes of post-transcriptional processes such as non-sense medicated decay or mRNA stabilization brought by somatic mutations. Of these, 417 (112 nonsynonymous or splice-sites) were not called as somatic mutations in initial WGS analysis without considering RNA-Seq data, including several known driver genes such as CTNNB1 and TSC2. Many of the somatic mutations detected by this integrative analysis were confirmed by Sanger sequencing of cancer DNAs (74/83 = 89.1%), and some of the unconfirmed mutations may be below the detection limit of Sanger sequencing owing to their low clonal proportion. Therefore, some of the false-negative somatic mutations resultant from the low sequencing coverage in WGS analysis can be rescued by complementary RNA-Seq analysis.

Finally, with no evidence of supporting variant reads in neither tumor nor normal WGS data, we identified 464 cancer RNA-specific events that are candidates for RNA-editing [18]. Although it is difficult to confirm the authenticity of each editing event, the mutation profile was abundant in A:T>G:C patterns and occurred in 3′ UTR regions, indicating that many of them are likely to be caused through RNA-editing by ADARs (adenosine deaminases) at the post-transcription stage [18] (Fig. 5A). The number of candidate RNA-editing events varied widely among the samples. We found a significant correlation (P-value = 2.38 ×10⁻⁷ by Wilcoxon rank sum test) between the number of A:T>G:C events and ADAR expression levels (Fig. 5B), implying the existence of new cancer subtypes determined by the amount of somatic RNA-editing.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 5. RNA editing candidates in 22 HCCs.

(A) The number of cancer-specific RNA mutation events (RNA editing candidates) and their substitution patterns for each sample. (B) Scatter plot between the number of A:T>G:C RNA-editing events and ADAR expression value (FKPM) calculated by whole transcriptome sequence data. There is a significant correlation (P-value  = 2.38×10⁻⁷ by Wilcoxon rank sum test) between the number of A:T>G:C events and ADAR expression levels.

https://doi.org/10.1371/journal.pone.0114263.g005

Landscape of genomic and transcriptomic disruptions in HBV-related HCCs

Considering the fact that many transcriptomic aberrations are caused by genomic changes, the detection of transcriptomic alterations, such as splicing aberrations and fusion transcripts, raises high probability of existence of proximal genomic changes. From this aspect, by complementary use of RNA-Seq to WGS analysis, we further rescued 64 combinations of genomic mutations and associated transcriptional aberrations (see Materials and Methods). In total, 252 genomic mutations causing transcriptional aberrations (GMTAs) were detected.

Through this integrated analysis of WGS and RNA-Seq, we could obtain comprehensive profiles of genomic and transcriptomic changes such as point mutations, indels, structural variations, splicing aberrations and gene fusions for each affected gene. Here, a critical problem is to discriminate cancer drivers from merely passenger events by using these profiles. In fact, although recurrent SVs (> = 3 HCCs) were seen in 12 genes (C10orf11, CIT, CLTC, CNTNAP2, DSCAML1, EYS, FHIT, LARGE, LRP1B, MACROD2, MAGI3, TTC28), many of them are located on or very proximal to common fragile regions [19] and actually were not found to have any influence on transcription (Figure S16A in S2 File), implying most of them are passenger SVs which occurred in unstable genomic regions. In order to remove passenger events, we only considered SVs evidenced by associated transcriptional aberrations, which is also helpful for removing false-positive detections in WGS analysis. On the other hand, recurrent gene fusions (> = 3 HCCs) were identified in 6 genes (ALB, CES1, FGA, SEPP1, SERPINA1, and TF). WGS analysis did not detected any SV associated with these fusions (Figure S16B in S2 File), implying that these fusions seem to come from minor sub-clonal cells or artifacts, and may not be driving forces for clonal expansion of cancer cells. These observations also support the importance of combinations of transcriptional aberrations and associated genomic mutations.

In this WGS analysis we found that GMTAs were concentrated on significantly mutated genes (three in TP53, two in HNF4A and RPS6KA3, one in ARID2), indicating their implication in cancer pathogenesis (Fig. 6). Among the above eight GMTAs, four were SVs and could not be detected by sole investigation of coding regions, suggesting that combination of WGS and RNA-Seq analysis is effective to detect candidates driver genes. Thus, HNF4A is likely to be a novel driver gene for liver cancer, as well as ARID2 [4] and RPS6KA3 [20]. HNF4A plays a critical role in the regulation of multiple metabolic pathways in the liver as well as hepatocyte differentiation, and down-regulation of HNF4A has been shown to be associated with HCC [21], [22]. Furthermore, key genes in the WNT signaling pathway (APC, AXIN1, CTNNB1, TCF7L1, TCF7L2 and WNT ligands) were frequently mutated (11 mutations) in nine HCCs, six of which affected their transcriptional consequences as GMTAs.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 6. The status of genomic and transcriptomic alterations of representative genes, detected by WGS and RNA-Seq of 22 HBV-related HCCs.

The list of genes were extracted by (1) significantly mutated genes in WGS analysis, (2) having no less than 3 mutations (point mutations or indels in coding regions, or GMTAs), (3) having no less than 2 GMTAs and registered in cancer gene census [17]. (4) involved in WNT signaling pathway, (5) TERT or MLL4.

https://doi.org/10.1371/journal.pone.0114263.g006

Clinical samples

The clinical and pathological features of the 22 HBV-related HCCs are shown in Table S1 in File S1. WGS analysis on RK001, RK010, RK023, and RK034 were reported in our previous study [4], Hepatitis B surface antigen (HBsAg) in the serum of all samples except for RK001. In RK001, HBV-DNA was detected by sensitive PCR in serum, indicating occult HBV infection. All subjects had undergone partial hepatectomy, and pathologists estimated that the percentage of viable tumor cells in each sample was at least 80%. High molecular weight genomic DNA was extracted from the fresh-frozen tumor specimens and blood. Total RNA was also extracted from the tumor tissues and non-cancerous liver tissues by Trizol (Invitrogen) and their quality and quantity were evaluated by Bioanalyzer (Agilent). All subjects provided their written informed consent to participate in the study following ICGC guidelines [32]. Ethical committees at RIKEN, The Institute of Medical Science The University of Tokyo, Hiroshima University School of Medicine, Wakayama Medical University, Osaka Medical Center for Cancer and Cardiovascular Diseases, and Tokyo Women's Medical University approved this work.

WGS analysis

Illunima library with 500-bp insert were prepared from DNAs of the tumors and lymphocytes. Sequence data was generated on the Illumina HiSeq2000 platform with paired reads of 101 bp, according to manufacturer's instruction. Mapping and identification of point mutation and somatic indels were carried out as described previously [4]. In brief, read pairs were mapped to the human reference genome using BWA [33] and possible PCR duplications were removed by samtools [34]. Point mutations and indels were identified by an in-house mutation caller, and significantly mutated genes were identified as described previously [4]. SVs were detected by the following method. Anomalous read pairs were identified and read pairs mapped within 500 bp of each other were considered to support the same rearrangement. We identified rearrangement candidates in tumor (support read pairs ≥4) and lymphocyte (support read pairs ≥1) samples, and tumor specific rearrangement candidates were selected. To exclude mapping errors, we performed a blast search of read pairs that support rearrangements against the reference genome. If a paired read was mapped with correct orientation, distance (≤500 bp) and an E-value <10⁻⁷ to a second location, we excluded that read pair. Reads mapped with more than two mismatches were also discarded. After filtering, candidates supported by ≥4 read pairs and at least one perfect match pair were considered as somatic rearrangements. The false positive rate was estimated to be 8.6% (15/175) by PCR verification. Since HBV genomes are polymorphic, we selected best reference HBV genomes for each sample as previously described [8]. First, we gathered unmapped reads, mapped them to 73 HBV genomes independently, and selected the HBV genome sequence with the highest number of mapped reads as the reference genome for each sample. Then, using read pairs that were mapped to both the human genome and HBV sequence, we identified integration sites supported by ≥3 read pairs [4].

RNA-Seq analysis and alignment

High-quality total RNA was subject to polyA+ selection and chemical fragmentation, and the 100–200base RNA fraction was used to construct cDNA libraries according to Illumina's protocol. RNA-Seq was performed on the HiSeq2000 platform using the standard paired-end 101 bp sequencing protocol. All sequencing reads were aligned to the known transcript sequences of UCSC known gene database (http://hgdownload.cse.ucsc.edu/goldenPath/hg19/database/knownGene.txt.gz) using Bowtie [35], with -a –best –strata -m 20–v 3 options, and the coordinates of the aligned reads were converted to the human reference genome (hg19). Unaligned reads were then aligned to the entire human reference genome (hg19) and as well as the HBV genome (AP011098) using blat [36], with -stepSize = 5 -repMatch = 2253, and aligned reads by Bowtie or blat were combined together. For each short read, the alignment with the highest number of matched bases was adopted, and the mapping quality was assigned as follows: For a location a, let B(a) denote the number of matched bases and let a_best denote the best location selected arbitrarily from those with the highest number of matched bases. The mapping quality for the read was assigned as follows:

min{100, -10×log10(1 – 1/(Σ_a0.02^∧(B(a_best)-B(a))))}.

Finally, sorting and PCR duplicate removal of short reads were performed using Picard (http://picard.sourceforge.net/).

Quantification of expression values from RNA-Seq data

To quantify expression values, we have used modified version of RKPM (reads per kb of exon per million mapped reads) measures [37]. After removing improperly aligned or low quality sequencing reads (mapping quality <60), the depth of coverage for each base in the exonic region of each RefSeq gene was tallied. Then, the numbers of bases were normalized as per kb of exon and per 100 million of aligned bases. Finally, expression value of each gene was determined by choosing the maximum of multiple RefSeq genes (if any) corresponding to the gene symbol.

Splicing aberration detection

We detected exon skip, splice-site slip, and pseudo exon inclusion events through identification of cancer-specific spliced junctions. Spliced junctions supported by at least 4 read pairs in cancer samples less than 2 read pairs in matched liver samples, and without any annotation in RefSeq, Ensemble Gene predictions nor USCS known genes were collected. Next, the ratio of sequencing depth to the number of support read pairs for each spliced junction was compared between cancer and the matched liver samples for both sides by Fisher's exact test, and those with P-value <0.05 on either side were extracted. Finally, for accurately removing polymorphic splicing events, splicing junctions with the ratio between the number of supporting read pairs and the sequencing depth greater than 0.01 for at least one non-cancerous liver samples were filtered. Remained spliced junctions were categorized as follows (Fig. 1B): exon skip, if both junction sides were on annotated spliced junctions; splice-site slip, if one junction side was consistent with an annotated spliced junction and the other was in an intronic region; and pseudo exon inclusion, if one junction side was consistent with an annotated spliced junction and the other was in an exonic regions. Those with junctions in both sides were not consistent with annotated spliced junctions were discarded.

We detected intron retention events by evaluating the number of reads aligned on exon-intron boundaries (boundary reads), defined as those included at least 8 bp of both the exon and the intron sides of the boundaries. Among all the annotated exon-intron boundaries (RefSeq genes, Ensemble Gene predictions, and USCS known genes), we identified the candidates of intron retentions that satisfy all the following conditions: (1) The ratio between the number of boundary reads and the total reads was greater than 0.1 in the cancer sample. (2) The number of boundary reads in the cancer sample was more than 3. (3) The number of boundary reads in the non-cancerous liver sample was less than 4. (4) The ratio between the number of boundary reads and the total reads was significantly different between cancer and non-cancerous liver samples (P<0.05 by Fisher's exact test). (5) The ratio between the number of boundary reads and the total reads in cancer sample was significantly deviated from the beta-binomial distribution (P<0.0001) with parameters fitted by utilizing the number of boundary reads and the total reads for all the 22 non-cancerous liver samples at the same boundary. (6) The number of boundary reads was significantly deviated from the negative-binomial distribution (P<0.0001) with parameters fitted by utilizing the number of boundary reads for all the 22 non-cancerous liver samples at the same boundary.

Detection of fusion transcripts

We developed Genomon-fusion algorism, which can detect fusion genes involving un-annotated transcripts and chimeric transcripts fused with viral sequences. Briefly, Genomon-fusion detects candidate fusion transcripts by utilizing ‘soft-clipping’ information, the unmatched parts of the partially aligned reads, along with a number of rigorously designed filters to exclude false positives often generated by ambiguous alignments to numerous repetitive or homologous sequences. First, Genomon-fusion searches for candidate fusion transcripts through genome-wide screening of the mutually nested breakpoints of soft clipped sequences. At least 2 soft clipped sequences at each breakpoint are aligned to within 10 bp of the other corresponding breakpoint, while the consistency of positions and directions of read pairs including the soft clipped sequences were investigated. Then, for each candidate fusion transcript, a set of read pairs surrounding the corresponding fusion boundary is assembled using CAP3 to constitute a contig sequence, and extract the candidate having >2 supporting read pairs that are properly aligned on that contig. Finally, for each remaining candidate, the pair of the contig split up at the fusion boundary are aligned to the human reference sequence including unplaced sequences (such as in chr1_gl000191_random and or chrUn_gl000211) by blat, and then removed the candidate if one of the two divided contigs aligned to other genomic locations with less than 3 mismatches or aligned within 1 kb of the other corresponding breakpoint.

Detection of over-expressing genes

First, we calculated the processed expression value (PEV) for each gene, which is defined as the log₂ of the expression values with 0.5 pseudo counts. Then, we excluded genes whose maximum PEVs among 22 cancer samples was below log₂(1.5) or within 3 sigma from the average PEVs among 22 liver samples. Next, for each remaining gene, a Grubbs-Smirnov test for a set of PEVs among 22 cancer samples was repeatedly performed until no outliers were detected (P-value <0.05). The detected outliers for each gene and sample in the above procedure were identified as over-expressed genes.

Mutation and RNA-editing detection from RNA-Seq and WGS data

Cancer-specific mutations in RNA-Seq are detected by using EBCall software [17], which can sensitively discriminate genuine mutations from sequencing errors through identification of discrepancies between allele frequencies of the candidate mutations and the distribution of sequencing errors estimated from a set of non-matched reference samples. We used the RNA-Seq data of the 22 non-cancerous liver samples as normal reference samples. We identified somatic mutations by checking the evidence in WGS data: sequencing depth ≥8 for both tumor and normal sample, allele frequencies in tumor ≥0.1, allele frequencies in normal ≤0.02, number of variant reads in tumor ≥2 and number of variant reads in normal ≤1. Furthermore, for extracting RNA editing events, we required: allele frequencies in tumor ≥0.1, allele frequencies in normal ≤0.02, and sequencing depth ≥15 for both tumor and normal samples.

Complementary detection of GMTAs by WGS and RNA-Seq data

For rescuing point mutations or indels causing transcriptional aberrations given cancer-specific splicing aberrations detected by RNA-Seq, we searched for the variants satisfying the following. (1) The edit distance to splicing donor/acceptor motifs was changed consistent to causing the corresponding splicing aberrations. (2) The sequencing depths of tumor and normal samples were more than 9. (3) The allele frequencies of the variant were more than 10% for the tumor sample, and less than 5% for the normal sample. (4) The numbers of variant reads were no less than 3 for the tumor sample and no more than 2 for the normal sample.

For rescuing exon skips caused by SVs given SVs detected by WGS, we searched for the exon skips satisfying the following. (1) The junction points were located next or 2^nd next exons to the breakpoints. (2) The number of supporting reads is no less than 3. (3) The number of supporting reads for the target sample was 5 folds more than the maximum of the other samples.

For rescuing intron retentions caused by SVs detected by WGS, we searched for the intron retentions satisfying the following (1) The boundary of exon and intron was located next to the breakpoints. (2) The ratio between the number of boundary reads and the total reads was greater than 0.1 in the target cancer sample and 3 folds more than the maximum of the other samples.