The matched tumor and normal samples analyzed were from two Singaporean patients. One GC exhibited evidence of microsatellite instability (MSI) and active H. pylori infection (see Table S1 in Additional file 1 for other clinical characteristics). Each tumor and matched normal sample was sequenced to more than 30-fold average base pair coverage by Illumina SR sequencing (Materials and methods; Table S2 in Additional file 1), and to > 130-fold physical coverage using large-insert (approximately 10 kbp) DNA-PET sequencing 9 on the SOLiD platform (Materials and methods; Table S3 and Note 1 in Additional file 1). Single nucleotide variants (SNVs) and short insertions and deletions (indels) from tumor and normal genomes were combined to identify somatic variants (Table 1 and Materials and methods) and reliability of somatic calls was confirmed using targeted sequencing (validation rate of 90% for SNVs and 96% for indels; Materials and methods). SR and DNA-PET data were also used to identify somatic copy-number variations (CNVs) and structural variations (SVs) (validation rate = 81%; Materials and methods; Note 1 in Additional file 1).

We integrated the SR and DNA-PET sequence information to perform de novo assembly of the tumor and normal genomes. While complete de novo assembly of a tumor genome still poses significant technical challenges and has not been attempted before, we were able to use the SR/DNA-PET data to construct highly contiguous draft assemblies of median scaffold lengths (N50) in the range of 41 to 148 kb, with DNA-PET data assisting in tripling sequence contiguity of the assemblies (Materials and methods; Note 2 and Table S5 in Additional file 1). Importantly, performing de novo SR/DNA-PET assembly revealed several findings not observed using conventional analyses of the SR data. First, the de novo approach allowed for characterization of large-scale somatic structural variations at single base-pair resolution (SR libraries were unable to identify nearly half of the validated SVs and fusions genes; Note 1 in Additional file 1). For example, NGCII092 exhibited a focal genomic amplification on chromosome 12p11-12 in a region containing the wild-type KRAS gene, a genomic event frequently observed in GC 10. The combined SR/DNA-PET data (Materials and methods) enabled a detailed putative reconstruction of the evolutionary lineage of the amplified KRAS locus with concomitant deletion of a proposed tumor suppressor gene RASSF8 (as well as another focal amplicon at chromosome 6p) as described in the supplementary text (Figure 1; Figures S1 and S2 and Note 3 in Additional file 1). Reconstruction of the tumor genomes also allowed the prediction of fusion genes and complex rearrangements that resemble patterns created by replication coupled mechanisms 11 and are further described in the supplementary text (Note 4 and Figures S3 and S4 in Additional file 1 and Table S6 in Additional file 2).

Second, a combined SR/DNA-PET analysis allowed us to assemble sequences present in the tumor genome but not in the reference human genome. For example, in patient NGCII082 exhibiting active H. pylori infection, we detected approximately 2,000 short-sequence reads and > 600 DNA-PET tags corresponding to the H. pylori genome (the first such report for a bacterial pathogen from tumor sequencing), in addition to a tumor-associated microbiome (these were not seen in NGCII092; see Figure S5 and Note 5 in Additional file 1 for details). Note that, despite being fewer in number, the DNA-PET tags contributed significantly to the physical coverage and analysis of the genomes (Figure S5 and Note 5 in Additional file 1).

Third, the de novo assembly enabled annotation of human genes and variants in sequences absent in the reference genome. In total, we identified more than 3 Mbp of novel sequence (longer than 500 bp), containing several genes (including an ortholog to a cytokine receptor-like factor - CRLF2), and more than a 1,000 somatic and germline variants for each patient (Materials and methods; Note 2 and Table S5 in Additional file 1).

We characterized mutational signatures in the GC genomes based on 14,856 somatic SNVs (11,738 indels) in NGCII082 and 17,473 somatic SNVs (2,486 indels) in NGCII092 that were identified from the whole-genome data (Table 1). This accounts for an average mutation frequency of 5 per megabase and included > 100 SNVs in protein coding regions for each tumor (Table 1; Note 6 in Additional file 1). Note that we identified more than five times the number of somatic variants uncovered in earlier sequencing studies 67 that were restricted to exomes (5,588 SNVs and 2,347 indels identified from 37 exomes), highlighting the statistical advantage of whole-genome analysis for studying mutational signatures. Overall, NGCII082, an MSI-positive tumor, displayed an excess of SNVs in protein coding regions (P-value < 0.02, χ² test) and a striking seven-fold higher frequency of micro-indels (Figures 2 and 3d) but a lack of large-scale SVs and amplifications or deletions (Figure 2 and Table 1). In contrast, NGCII092 exhibited a complex copy number profile of extensive focal amplifications and deletions, and a mutated TP53 gene, consistent with the presence of chromosomal instability (CIN) in the tumor genome (Figure 2). These results agree with the mutual exclusivity seen in MSI and CIN pathways for inducing mutations in other cancers as well 12.

The clear excess of micro-indels in the MSI-positive GC (Figure 3d; Figure S10 in Additional file 1) was characterized by a pattern of single base-pair thymine deletions in mononucleotide repeats (79%). In contrast, there were a comparable number of insertions in both the MSI-positive and CIN-positive GC, and a similar deletion-specific pattern has also been noted before 13. Also, non-thymine and non-mononucleotide repeat deletions were not found to be in excess. The correlation between MSI phenotype and the specific deletion signature identified here was further confirmed from previous exome-sequencing data 7 (four MSI-positive exomes), though this aspect was not noted in the previous work. In terms of genomic location, the deletions were randomly scattered throughout the genome and occurred in proportion to the regional presence of thymine mononucleotide repeats (that is, 85% of homopolymers > 5 bp). Thus, despite the bias towards thymine deletions, there seems to be an absence of a targeting mechanism on the genome for the MSI-associated signature.

Despite exhibiting very different somatic alteration patterns (MSI or CIN), the mutational frequencies of both GCs at the single nucleotide level were highly similar, being significantly biased towards C > A and T > A alterations compared to normal genomes (P-value < 10^-16, χ² test; Figure 3a). These alterations likely represent mutations caused by reactive oxygen and nitrogen species (ROS and RNS), which are known to produce C > A and T > A mutations 14. Also, a likely trigger is H. pylori infection, which has been shown to cause chronic inflammation and ROS/RNS production in gastric epithelial cells 14. The C > A mutations observed were associated with highly significant sequence-selectivity, being marked by an excess at CpCpT (NGCII082, odds ratio (OR) = 3.2, P-value < 10^-16, χ² test) or TpCpA sites (NGCII092, OR = 1.7, P-value < 10^-16, χ² test) and extensions of these motifs (Materials and methods; Note 6 and Figure S6 in Additional file 1 and Table S14 in Additional file 6). This pattern is distinct from the C > A signature seen in smoking-associated small-cell lung cancer where an excess was seen in CpG dinucleotides outside CpG islands, suggesting a link with methylation status 23. Further work is required to identify the mechanistic basis of sequence selectivity in this genome-wide GC-specific signature.

The overall impact of the mutational signatures identified here on gastric tumorigenesis is a complex question influenced by several factors, including the nature of mutations, the function of genes that are frequently impacted as well as genetic background and selection processes. We aimed to provide an initial assessment using two approaches: (i) by characterizing the proportion of genes affected by various mutational classes; and (ii) by identifying recurrently mutated genes in subtypes of GC defined by mutational processes.

Overall, a majority of mutated genes in NGCII082 were due to SNVs (77%) while CNVs and SVs played a dominant role in NGCII092 (82%) (Table 1). In total, we identified 107 SVs that affected genes by truncation, fusion, deletion, tandem duplication or rearrangements within the gene body. Ninety-six (90%) of these were identified in the CIN phenotype exhibiting tumor NGCII092, illustrating the genic burden from this mutational process. In contrast, small insertions and deletions (indels) were seen in few genes, even in the tumor with MSI phenotype (despite indels being roughly as common as SNVs genome-wide; Table 1), though their ability to cause frameshifts is likely to impact gene function more often than SNVs. Among SNVs, even though the deamination-related C > T signature is only seen in a small fraction of the genome, it plays a larger role in GC due to its targeted impact on genes. More than 48% of the non-synonymous mutations seen (48% in NGCII092 and 59% in NGCII082) in the two tumors were due to C > T mutations, compared to less than 19% for C > A mutations (Table 1). Among recurrently mutated genes in GC (Table S7 in Additional file 1 and Table S9 in Additional file 3), non-synonymous mutations in the tumor suppressor genes TP53 (mutated in 50% of samples) and PTEN (18% of samples), and oncogenes PIK3CA (13%; 8% have PTEN and PIK3CA mutations) and CTNNB1 (10%) were often in the form of C > T mutations (29%). This was also seen in several novel recurrently mutated genes such as AQP7, SPTA1 and RP1L1 (mutated in > 10% of tumors; Table S7 in Additional file 1).

Pathway analysis of mutated genes revealed that the two most enriched sets were β1-integrin mediated cell-surface interactions and signaling events mediated by class III histone deacetylases, a refinement of previous analysis 7 (Table S10 in Additional file 4). Furthermore, we identified genes implicated in RAC1 regulation to be mutated in 83% of H. pylori positive samples (P-value < 0.05 Fisher's exact test). RAC1 is a member of the Rho GTPase family known to play diverse oncogenic roles 18, shown to regulate the H. pylori virulence factor VacA, and known to promote vacuole formation in epithelial cells 19. Mutations in the RAC1 pathway could thus simultaneously promote H. pylori infection as well as gastric tumorigenesis.

Finally, to further characterize the impact of mutational processes on genes in GC, we considered two specific subtypes for identifying recurrently mutated genes, MSI-positive GC and TP53-wild-type GC (Tables S11 and S13 in Additional file 1 and Table S12 in Additional file 5). We used TP53-wild-type status as a surrogate marker for tumors without the CI phenotype as TP53 is known to suppress chromosomal instability 20. In this class of GCs, in addition to the tumor suppressor gene PTEN and TTK that interact with TP53, we identified PAPPA, a marker for pregnancies with aneuploid fetuses 21, as being recurrently mutated (Table S13 in Additional file 1; note that the average mutation rate for the whole-genome sequencing (WGS) samples in an approximately 2 Mbp window surrounding PAPPA is similar to the genome-wide rate, that is 5.3 versus 5.2 mutations/Mbp). A screen of an additional 94 gastric cancer/normal pairs confirmed the frequency of PAPPA mutations as being 6% among all GC samples (Table S12 in Additional file 5) and 20% among TP53 wild-type GCs (with mutations in key functional domains; Figures S13 and S14 in Additional file 1), highlighting it as a potential driver gene in this subtype.

In MSI-positive GCs, ACVR2A, RPL22, LMAN1, and STAU2 were observed to have recurrent single base thymine deletions in poly(T) regions (Table S11 in Additional file 1) and this was confirmed in a screen of an additional 94 gastric cancer/normal paired samples (9 MSI-positive; Table S12 in Additional file 5 and Figure S9 and Note 9 in Additional file 1). In total, ACVR2A was mutated in a region of 8 thymines in 86% of MSI-positive GCs tumors, RPL22 in a region of 8 thymines in 64%, LMAN1 in a region of 9 thymines in 50% and STAU2 in a region of 8 thymines in 29%. Based on the average frequency of mutations in homopolymer regions in the MSI-positive tumors (4.5% of 8 thymine stretches (n = 778) and 4.8% of 9 thymine stretches (n = 183), respectively, in exomic regions), mutations in ACVR2A, RPL22 and LMAN1 were in significant excess (Bonferroni-corrected P-value ≤ 0.0003, exact binomial test). In each gene, all the deletions occurred in the same homopolymer tract containing thymines, a pattern linked to the MSI phenotype, and none of the MSI-negative GC tumors carried these mutations. In contrast, mutations in the recently reported MSI-associated putative driver gene ARID1A were not restricted to deletions or MSI-positive tumors 7. Interestingly, ACVR2A (encoding a TGF-β super-family differentiation factor) has been described to be recurrently mutated in MSI-positive colorectal cancer 22. Also, the frequency of mutations seen here is comparable to the previously reported frequency in MSI-positive colorectal cancer 2324 and emphasizes the importance of ACVR2A and TGF-β signaling in MSI-positive GC, while unraveling the oncogenic roles of RPL22 and LMAN1 requires further investigation.

Patient samples and clinical information on tissue and blood samples were obtained from patients who had undergone surgery for gastric cancer at the National University Hospital, Singapore, and Tan Tock Seng Hospital, Singapore. Informed consent was obtained from all subjects and the study was approved by the Institutional Review Board of the National University of Singapore (reference code 05-145) as well as the National Healthcare Group Domain Specific Review Board (reference code 2005/00440). Clinical information for the two patients whose samples were analyzed by whole-genome sequencing is provided in Table S1 in Additional file 1 and additional information for the 94 gastric tumors used for targeted screening is provided in Table S12 in Additional file 5.

For WGS sequencing, genomic DNA isolated from tumor and blood samples was randomly fractionated using a Roche Nebulizer following the manufacturer's instructions (Madison, Wisconsin, USA). Fractionated DNA was then end-repaired, A-tailed at the 3' end, ligated with Illumina paired end adaptors, PCR amplified followed by gel-selection of a range of 400 to 600 bp fragments as templates and sequenced by Illumina GA from both ends to obtain 76 or 101 bp reads at each end (Table S2 in Additional file 1). DNA-PET libraries were constructed as described elsewhere 9 and were sequenced by the Applied Biosystems SOLiD system (Carlsbad, California, USA, Table S3 in Additional file 1). Exome sequencing was performed as described earlier using SureSelect Human All Exon Kit v1 (Agilent Technologies, Santa Clara, California, USA) and sequencing on two lanes of Illumina GA-IIx sequencer using 76 bp paired-end reads 6.

Paired-end Illumina reads were mapped to the reference human genome (UCSC hg18) using ELAND (Illumina Inc.) and reads that failed pass-filter were removed from further analysis. SNVs and indels were called for each sample separately using SAMtools 25 (v0.1.7-6, SNP-quality threshold = 20, consensus-quality threshold = 30) (Table S4 in Additional file 1). Identical variant calls in tumor and matched normal samples were used to identify germline variants. Variant calls unique to the tumor, where the normal genotype called by SAMtools was different and where less than two reads of the variant genotype were seen in the normal sample, provided the list of somatic variants. Illumina reads from exome sequencing were analyzed using this pipeline after BWA 26 mapping (Table S8 in Additional file 1). As a control, we noted that germline SNV frequencies were nearly identical across all exomes from WGS and exome sequencing datasets (Figure S7 in Additional file 1). Somatic SNV frequencies and neighborhoods were compared to germline frequencies to assess enrichment. A neighborhood of up to 2 bp surrounding an SNV was used to identify enriched motifs. Somatic indel calls were required to be supported by at least 20% of the reads, by reads on both strands, with a minimum of 10 reads overlapping the position in the tumor and no indel calls in the normal sample. Somatic SNVs and indels in protein-coding regions and introns were confirmed by Sanger sequencing to have a high validation rate (83 SNVs, validation rate = 90%; 72 indels, validation rate = 96%). SNV neighborhood analysis was done by extracting 5 bp sequences upstream and downstream of mutations. Germline and somatic copy number variants were identified using the program RDXplorer 27 with default parameters.

DNA-PET tags were mapped individually to the reference human genome (UCSC hg18) in color space allowing two color code mismatches per tag by the SOLiD System Analysis Pipeline Tool Corona Lite (Applied Biosystems Inc.). Contigs of the reference sequence with unresolved location (random_chr) and alternative MHC haplotypes were excluded from the reference for mapping. Individually mapped tags were paired by Corona Lite. In cases where one or both tags had multiple mapping locations, a process termed 'rescuing' favored the creation of concordant PETs (both tags are on the same chromosome, same strand, same orientation, correct 5' → 3' order and in the expected distance to each other).

SVs, based on clusters of non-concordant PETs, were called using the GIS DNA-PET pipeline 9 with refined quality control criteria: (i) PET clusters of size < 6 were excluded; (ii) the regions to which the 5' and 3' tags of a cluster mapped had to be at least 1 kbp in size each; (iii) PET clusters that had a supercluster (connected component of overlapping clusters 9) size > 100 required a higher cluster size of 10; and (iv) PET clusters with high sequence similarity between the two fused regions (BLAST score > 2,000 for 20 kbp windows around the predicted break points) were excluded. To distinguish between germline and somatic SVs, paired normal and tumor samples were compared as described previously 9. Further filtering of known germline SVs and PCR validation are described in Note 1 in Additional file 1.

Contig assembly, scaffolding and gap-filling of the Illumina sequencing data were done using the assembler SOAPdenovo 28. DNA-PET reads were mapped to the SOAPdenovo assembly with Bowtie 29 and the resulting linking information was used to produce larger scaffolds based on the optimal scaffolder Opera 30. Scaffolds and contigs were refined further with the gap-filling module in SOAPdenovo, employed for bridging scaffold gaps, where feasible. Using the SR reads alone, we obtained 12 kb scaffold N50 for both tumors. The DNA-PET reads allowed for improvement of assembly connectivity to a N50 of 65 kb and 41 kb for NGCII082 and NGCII092, respectively. Assemblies were compared to the reference human genome (UCSC hg18) using the MUMmer package 31 and alignments longer than 1 kbp were used to identify deletions and insertions larger than 20 bp. Overall, 12,861 deletions and 143 insertions were found in NGCII082 and 9,274 deletions and 108 insertions in NGCII092 of which 3 events > 2 kbp missed by DNA-PET analysis were identified in each sample. Fusion genes were validated and breakpoints were confirmed by using the gap-filling module in SOAPdenovo to bridge scaffolds constructed around the breakpoint. Sequences missing in the reference human genome were identified based on the criteria that they should be > 500 bp long and have no match to the reference genome with > 90% identity. Reads were mapped to the novel sequences using Bowtie to identify regions with no read coverage in the middle of a scaffold that could indicate a potential mis-assembly.

Reads with a putative microbial or viral origin were identified by mapping reads with no mapping to the human genome, to a database of complete bacterial and viral genomes in NCBI (using Bowtie 29). Matches were filtered for low-complexity sequences (more than three matches of any 5-mer) and the remaining reads were used to estimate the abundance for each species (pooling reads mapped to different strains of a species). Each species was checked for multiple distinct read matches to its genome (> 4 distinct regions, where the genome was segmented in 1 kbp windows) and the presence of unique read matches (using the unique option in Bowtie). The small fraction of reads of putative bacterial origin in the matched blood samples (possibly reagent contamination) were used as control and read matches to the corresponding species were excluded in determining the tumor associated microbiome. Concentration of H. pylori cells in relation to tumor cells was estimated based on the assumption of uniform coverage of both cell types, where coverage = k × Number of cells × Size of genome, for a constant k and the populations are assumed to be clonal.

For all samples, SNV and indel calls were annotated using the SeattleSeq server 32 and SIFT 33, respectively. Pathway analyses were performed based on non-synonymous SNVs and indels using the Pathway Interaction Database 34 (sample pfg005T from Wang et al. 7 was excluded as it only had four somatic mutations).

Sequencing data for this publication have been deposited in NCBI's Gene Expression Omnibus 35 and is accessible through GEO Series accession number GSE30833.