Multiregion ultra‐deep sequencing reveals early intermixing and variable levels of intratumoral heterogeneity in colorectal cancer

Intratumor heterogeneity (ITH) contributes to cancer progression and chemoresistance. We sought to comprehensively describe ITH of somatic mutations, copy number, and transcriptomic alterations involving clinically and biologically relevant gene pathways in colorectal cancer (CRC). We performed multiregion, high‐depth (384× on average) sequencing of 799 cancer‐associated genes in 24 spatially separated primary tumor and nonmalignant tissues from four treatment‐naïve CRC patients. We then used ultra‐deep sequencing (17 075× on average) to accurately verify the presence or absence of identified somatic mutations in each sector. We also digitally measured gene expression and copy number alterations using NanoString assays. We identified the subclonal point mutations and determined the mutational timing and phylogenetic relationships among spatially separated sectors of each tumor. Truncal mutations, those shared by all sectors in the tumor, affected the well‐described driver genes such as APC, TP53, and KRAS. With sequencing at 17 075×, we found that mutations first detected at a sequencing depth of 384× were in fact more widely shared among sectors than originally assessed. Interestingly, ultra‐deep sequencing also revealed some mutations that were present in all spatially dispersed sectors, but at subclonal levels. Ultra‐high‐depth validation sequencing, copy number analysis, and gene expression profiling provided a comprehensive and accurate genomic landscape of spatial heterogeneity in CRC. Ultra‐deep sequencing allowed more sensitive detection of somatic mutations and a more accurate assessment of ITH. By detecting the subclonal mutations with ultra‐deep sequencing, we traced the genomic histories of each tumor and the relative timing of mutational events. We found evidence of early mixing, in which the subclonal ancestral mutations intermixed across the sectors before the acquisition of subsequent nontruncal mutations. Our findings also indicate that different CRC patients display markedly variable ITH, suggesting that each patient's tumor possesses a unique genomic history and spatial organization.

Intratumor heterogeneity (ITH) contributes to cancer progression and chemoresistance. We sought to comprehensively describe ITH of somatic mutations, copy number, and transcriptomic alterations involving clinically and biologically relevant gene pathways in colorectal cancer (CRC). We performed multiregion, high-depth (3849 on average) sequencing of 799 cancerassociated genes in 24 spatially separated primary tumor and nonmalignant tissues from four treatment-na€ ıve CRC patients. We then used ultra-deep sequencing (17 0759 on average) to accurately verify the presence or absence of identified somatic mutations in each sector. We also digitally measured gene expression and copy number alterations using NanoString assays. We identified the subclonal point mutations and determined the mutational timing and phylogenetic relationships among spatially separated sectors of each tumor. Truncal mutations, those shared by all sectors in the tumor, affected the well-described driver genes such as APC, TP53, and KRAS. With sequencing at 17 0759, we found that mutations first detected at a sequencing depth of 3849 were in fact more widely shared among sectors than originally assessed. Interestingly, ultra-deep sequencing also revealed some mutations that were present in all spatially dispersed sectors, but at subclonal levels. Ultra-high-depth validation sequencing, copy number analysis, and gene expression profiling provided a comprehensive and accurate genomic landscape of spatial heterogeneity in CRC. Ultra-deep sequencing allowed more sensitive detection of somatic mutations and a more accurate assessment of ITH. By detecting the subclonal mutations with ultra-deep sequencing, we traced the genomic histories of each tumor and the relative timing of mutational events. We found evidence of early mixing, in which the subclonal ancestral mutations intermixed across the sectors before the acquisition of subsequent nontruncal mutations. Our findings also indicate that different CRC patients display markedly variable ITH, suggesting that each patient's tumor possesses a unique genomic history and spatial organization.

Introduction
Colorectal cancer (CRC) is the third most common cancer worldwide (Ferlay et al., 2014). Large-scale collaborative sequencing projects such as The Cancer Genome Atlas (TCGA) have catalogued genetic alterations across patients with CRC (interpatient heterogeneity) (Cancer Genome Atlas Network, 2012). Heterogeneity can also exist in a single tumor, as subpopulations of cancer cells with distinct genomic alterations may exist across different regions (intratumor heterogeneity or ITH). Analysis of ITH allows the reconstruction of the phylogenetic tree of subpopulations within a tumor and provides insights into the timing and pervasiveness of the genomic events that contributed to cancer development. At the same time, ITH poses significant challenges to cancer therapy by limiting our ability to personalize therapy based on a single biopsy and by contributing to therapeutic escape, drug resistance, and metastasis. Understanding ITH in CRC thus has important biological and clinical implications.
Intratumor heterogeneity has recently been characterized by multiregion sequencing in several major cancer types, including cancers of the prostate, breast, kidney, brain, ovary, liver, and lung (Boutros et al., 2015;de Bruin et al., 2014;Friemel et al., 2015;Gerlinger et al., 2012Gerlinger et al., , 2014Hoogstraat et al., 2014;Liu et al., 2009;Martinez et al., 2013;Navin et al., 2010;Sottoriva et al., 2013;Yates et al., 2015;Zhang et al., 2014). These studies described a variety of different evolutionary branching patterns, including both 'palm-like' patterns, in which most mutations in likely driver genes were present in all regions of the tumor, and more branched, 'oak-like' patterns, with many mutations in likely driver genes found in only one or a few regions (Table 1).
In CRC, early studies were based on low-throughput sampling of a few loci, for example, by sequencing PCR products from a few selected sites (Baisse et al., 2001;Losi et al., 2005;Naxerova et al., 2014;Thirlwell et al., 2010), while recent studies have examined ITH by whole exome sequencing of multiple tumor regions (Hardiman et al., 2016;Kim et al., 2015;Sottoriva et al., 2015;Uchi et al., 2016). These exome studies reported varying levels of ITH and both 'palm'-and 'oak'-like phylogenetic structures (Table 1). However, these exome-based sequencing studies were performed at relatively shallow depths, which might fail to detect somatic mutations with low variant allele frequencies. This could also overestimate ITH, if mutations that were actually shared across the regions were not detected in some of them, because of a low proportion of malignant cells or because the mutations were subclonal in some regions . This could also underestimate ITH, if a private subclonal mutation were missed altogether.
To address the need for understanding ITH in CRC comprehensively and accurately, we used targeted deep next-generation sequencing (mean coverage 3849) of 799 genes to characterize the genetic profiles of four CRCs. We analyzed five primary tumor sectors and a matched normal tissue from each patient. Next, we performed ultra-high-depth amplicon sequencing (17 0759) to assess the presence or absence of each alteration in the various sectors. We also performed NanoString gene expression profiling and NanoString cancer copy number analysis.

Subjects, samples, and consent
Fresh frozen primary tumors and matched normal tissues were harvested at surgery. None of the patients had undergone any systemic treatment prior to surgery. Two patients were in stage IV with liver metastasis and were undergoing palliative resection of the primary tumors, and the other two were in stage I and stage IIIb, respectively. Table S1 provides further clinicopathologic details. A pathologist carefully selected five spatially separated,~1-cm 3 sectors from each tumor mass and one from adjacent normal mucosa. The samples were snap-frozen and stored in liquid nitrogen. The study protocols were approved by the Institutional Review Board of Singapore Health Services (IRB approval number: 2011/439/B and 2011/ 110/B). Written informed consent for the use of tissue specimens for research was obtained from all participants. Short reads from our study have been deposited in the European Genome-phenome Archive (EGA) under the study accession ID EGAS00001001720.

Targeted deep multiregion sequencing of 799 cancer-associated genes
We performed the deep targeted multiregion hybridcapture sequencing with a custom Agilent SureSelect panel that focused on exons of 799 cancer-related genes, as described by Tan et al. (Tan et al., 2015). This panel was derived from a comprehensive literature and database survey and comprises genes biologically and clinically relevant to cancer. DNA was extracted from the samples using the QIAamp Blood and Cell Culture DNA Mini Kit (Qiagen) according to the manufacturer's instructions. The quality and yield of the DNA samples were assessed by Quant-IT PicoGreen dsDNA Assay (Invitrogen Life Technologies). The DNA was sheared using a Covaris S2 (Covaris Inc) to a size distribution of 150-200 bp. Sequencing libraries were prepared, and sequencing was carried out on HiSeq 2000 sequencers, and reads were aligned to the reference human genome. Average sequencing depth was 3849 (range: 2549-7099, Table S2).
We called variants with two algorithms: Genome Analysis Toolkit (GATK)-based algorithm (McKenna et al., 2010) and MuTect (Cibulskis et al., 2013) as previously described (Tan et al., 2015). The variants were curated by manually inspecting the sequencing reads corresponding to each variant using the Integrated Genomics Viewer (IGV) (Thorvaldsd ottir et al., 2013). We used the CCDS (Pruitt et al., 2009a), RefSeq (Pruitt et al., 2009b), Ensembl (Flicek et al., 2010), and UCSC known genes (Hsu et al., 2006) databases to annotate transcripts and amino acid changes. Detected somatic mutations included nonsynonymous and synonymous single-nucleotide substitutions and small insertions/deletions. We classified mutations as one of the 'truncal' (common to all sectors), 'branched' (shared among ≥ 2, but not all, sectors), and 'private' (unique to one sector). There was no a priori VAF threshold for calling mutations, but all somatic mutation VAFs called by both callers were ≥ 2.5%.

Ultra-deep amplicon sequencing to validate somatic mutations and assess their presence across all sectors
Using ultra-deep amplicon sequencing, we sought to validate all candidate somatic mutations in Patients 1, 2, and 4 and to validate 42 of the 261 candidate mutations in the hypermutated tumor of Patient 3. The purpose of this was to confirm the presence of these somatic mutations and also to sensitively test for their presence in every sector of the tumor. The 42 candidate nonsynonymous mutations selected from Patient 3 comprised 4 truncal variants, 15 of 70 nonsynonymous branched variants, and 5 or 6 private variants in    Table S5).
The following stringent filters were applied to candidate somatic mutations in the ultra-deep amplicon sequencing data to calculate ITH: (a) We only considered the sites with ≥ 1009 coverage of nucleotides with base qualities ≥ 30 in normal and all tumor sectors and (b) we only counted as true-positive sites lacking evidence of the mutation in both the normal sample and HapMap controls. This assessment was based on the mean and standard deviation of variant reads in the normal sample and controls (Z score of all variant reads in the tumor sectors > 3) and on visual examination in IGV (Thorvaldsd ottir et al., 2013). Data for five variants did not meet the criterion (a), leaving 116 primer pairs for which we could assess the false-positive and false-negative rates in the initial targeted sequencing. In total, 580 mutation sites were assayed by ultra-deep sequencing. There was no a priori VAF threshold for calling mutations in the ultra-deep sequencing, but all somatic mutation VAFs validated by ultra-deep sequencing were ≥ 0.07%.

Identification of copy number alterations from sequencing data
For each tumor sector (paired with the appropriate normal sample), we estimated the proportion of malignant cells and identified the broad genomic regions of copy number gain or loss by applying the allele-specific copy number analysis of tumors (ASCAT) software (Van Loo et al., 2010) to the single-nucleotide polymorphism allele fractions and the relative read depths extracted from the exome sequencing data.

Subclonality analysis
Using the estimates of VAFs from the ultra-deep sequencing data, we determined which truncal or branched nonsynonymous somatic mutations were likely subclonal. Using the estimated malignant cell fraction and genomic copy number at the site of a given mutation (both obtained from ASCAT analysis), we estimated the minimum expected VAF of that mutation. We estimated that a mutation was clonal if the observed VAF was (a) similar to or greater the minimum expected VAF and (b) similar to that of other mutations of the same group (i.e., truncal or branched) in the same sector. Conversely, if the observed VAF was substantially lesser than expected based on these criteria, we concluded that the mutation was likely subclonal. Details for each mutation analyzed in this way are in Table S4.

Phylogenetic analysis
We used the 'discrete-characters Wagner parsimony' method in PHYLOGENY INFERENCE PACKAGE (PHYLIP) version 3.695 (Felsenstein, 2005) to generate phylogenetic trees with the matched nonmalignant tissue as the outgroup (input to PHYLIP). The trees were drawn using the DrawTree tool under the PHYLIP package.
2.7. Detection of copy genomic number alterations and mRNA profiling with NanoString About 300 ng of purified genomic DNA was mixed with NanoString nCounter Cancer Copy Number Variation CodeSets (NanoString Technologies, Seattle, WA, USA). This assay digitally measures the genomic copy numbers of 87 important genes that are often amplified or deleted in cancer. We digested the DNA with AluI and then hybridized with the CodeSets according to the manufacturer's protocol. The resulting signals were analyzed with the NCOUNTER DIGITAL ANALYZER software. CodeSets with average counts > 3 were used for further analysis. The NanoString nCounter platform was used to quantify mRNA levels. About 1 lg of total RNA was extracted from 20 to 25 mg of frozen tissue and hybridized to the NanoString Pan Cancer Panel CodeSets. The NSOLVER software tabulated the raw counts, checked the quality, and normalized. Due to low malignant cell content, we excluded one sector from the analysis for Patient 3. We conducted a gene pathway analysis based on the log2-transformed counts using the generally applicable gene-set enrichment for pathway analysis (GAGE) R package (Luo et al., 2009). Pathway dysregulation scores and the associated P-values were calculated for each of the 13 cancer pathways represented by the NanoString Pan Cancer Panel CodeSets, using the patient's nonmalignant tissue as the baseline reference. Pathways with adjusted P-values < 0.1 were selected for further analysis. Hierarchical clustering and plotting based on the dysregulation scores of the pathways was carried using the heatmap.2 function in R using default parameters.

Genetic ITH varies across tumors
We used deep targeted sequencing of 799 cancer-associated genes in 24 spatially separated tissue sectors from four patients (five tumor sectors and one matched nonmalignant tissue per patient) to comprehensively catalogue somatic mutations in different regions of the tumors. Figure 1 summarizes the general workflow of our study.
We validated the mutations with ultra-deep amplicon sequencing, which also detected the presence of some of the mutations at low VAFs in sectors in which they were not identified by the initial targeted deep sequencing. Ultra-deep sequencing confirmed 289 of the 307 candidate somatic mutations detected in the initial targeted sequencing, for a true-positive rate of 94% (289/307). Of the 18 false positives, 17 appeared to be germline variants that were seen in the adjacent normal tissue or HapMap controls. The remaining variant was not observed in any sample in the ultradeep sequencing. In addition, ultra-deep sequencing detected 76 mutations that had been missed by the initial targeted sequencing, yielding a false-negative rate of 21% (76/(289 + 76) in the original targeted sequencing. Of these, 72 were probably missed by the initial targeted sequencing because of low VAFs and four were missed due to low coverage of the targeted capture sequencing data in the tumor sector.
Patient 1 had stage IIIb rectosigmoid carcinoma without distant metastases (Fig. 2, Table S1). This carcinoma was highly homogeneous, with 85% of mutations being truncal, as indicated by the long trunk in Fig. 2. These truncal mutations affected the known CRC drivers, including TP53, FAT4, and BRAF (Hisamuddin and Yang, 2006). The only nonsilent branched mutation was in CDH11, which is not widely considered a driver in CRC. Therefore, we postulate that in this tumor, all the mutations required for tumorigenesis occurred before the last clonal expansion (Fig. 2). Unless the cells from the last clonal expansion actually replaced preexisting tumor cells, this suggests that these mutations occurred when the tumor was still small. Patient 2 had stage IV transverse colon carcinoma with liver-limited metastasis (Fig. 3, Table S1). Of the mutations, 61% were truncal, including those in KRAS and APC (Fig. 3). A majority of these truncal mutations were likely clonal, except for an insertion in RAF1 and a deletion in DNMT3A that were subclonal in one sector and four sectors, respectively (Fig. 3, Table S4). Interestingly, we also observed clonal mutations that were branched or private. This suggests the possibility that, despite being acquired later, these mutations became clonal because they conferred a selective advantage. For example, FBXW7, a known driver gene of CRC (Rajagopalan et al., 2004 , 2007), was clonally mutated in Sectors A, B, and E. Patient 3 had microsatellite unstable stage IV primary adenocarcinoma in the ascending colon and liver-limited metastasis (Fig. 4, Table S1). The tumor was hypermutated (56.7 somatic mutations per Mb, Table 2), showed the mutation signature of mismatch repair deficiency (Alexandrov et al., 2013), and had MSH-2 and MSH-6 protein losses as determined by immunohistochemistry (Fig. 4, Tables 2 and S1). There was no evidence of somatic mutations in MSH2 or MSH6 in this patient. Microsatellite instability is often due to epigenetic silencing of MLH1. This tumor had the highest ITH: Only 33% of the mutations were truncal (Fig. 4, Table 2). The extensive heterogeneity stemmed from the high overall mutation rate (56.7 mutations per Mb).
Differences in genetic ITH are likely contributed by the variation in mutation rates in the tumors. The hypermutated tumor with the highest mutation rate of 56.7 mutations per Mb exhibited significantly greater ITH than the rest (Table 2).

Copy number and mRNA profiles offered additional perspectives on ITH
Based on the targeted sequencing data, we identified the copy number alterations in a genomewide scale. We identified the losses in chromosome arms 3p and q, 4p and q, and 17p and gains in chromosome arms 6p, 7p, 7q, 13q, and 20p and q, all in at least two patients. These have been previously reported (Beroukhim et al., 2010;Cancer Genome Atlas Network, 2012;Zack et al., 2013) (Table S8). Copy number ITH was highest in Patient 2: There were 12 copy number alterations, of which only two were truncal (Fig. 3, Table S8). Copy number ITH was also high in Patient 1: Of 34 copy number alterations, only seven were truncal (Fig. 2, Table S8). Copy number ITH was lower, but still substantial, in Patient 4: Of 28 copy number alterations, 13 were truncal (Fig. 5, Table S8). Patient 3's tumor, consistent with the previous reports of genome stability in microsatellite unstable tumors (Cancer Genome Atlas Network, 2012), had a largely diploid genome with only two copy number alterations and thus very low copy number ITH (Fig. 4, Table S8).
NanoString nCounter Cancer analysis of copy number alterations at 87 genes that are commonly amplified or deleted in cancer was in general consistent with the analysis based on sequencing data and ASCAT. The NanoString data supported the general observation that the copy number alterations were largely similar across sectors (Figs 2-5). Hierarchical clustering of dysregulated pathways appeared to mirror the phylogenetic relationships in Patient 1, but for Patients 2, 3, and 4, a robust relationship between dysregulated pathways and phylogeny was absent (Figs 2-5). This is not surprising, as, in addition to reflecting the genetic heterogeneity of malignant cells per se, gene expression ITH likely also reflects the variation in the proportions of nonmalignant as well as malignant cell populations across sectors.
We studied the relationships between copy number ITH and gene expression for the 42 genes assayed by both the NanoString copy number and mRNA panels. Supporting the hypothesis that ITH in gene expression is sometimes driven by copy number ITH, in Patient 1, we identified three genes, VEGFA, ITGB4, and GRB2, with genomic amplification and overexpression only in Sectors A, C, and E (Figs 2 and S3). In addition, in Patient 4, there was a truncal genomic copy number gain associated with high expression in genes involved in the PI3K/Akt pathway (AKT2, BCL2L1, and CCNE1) and there was a truncal copy number loss associated with low expression of PTEN and PIK3CA (Figs 5 and S4).

Targeted deep sequencing followed by ultradeep sequencing improved ITH analysis
The depth (3849) of the initial sequencing that focused on cancer-associated genes ensured the sensitive initial detection of somatic mutations and made ultra-deep amplicon sequencing of the identified somatic variants across every tumor sector manageable. With the targeted sequencing, we may miss diversity in regions of the genome that we did not target and whose functional significance is not currently known. However, the targeted sequencing detected adequate numbers of mutated sites to reconstruct tumor phylogenies across the sectors.
Subsequent ultra-deep sequencing dramatically improved the precision of the analyses of ITH and tumor phylogenies. Ultra-deep sequencing detected the a b c d Fig. 4. Patient 3 in stage IV with metastasis to the liver. This patient had a hypermutated tumor. Panels (a) through (d) as in Fig. 3.
mutations that had very low VAFs, either because of a low proportion of malignant cells, because the mutations were present only in subclones, because the mutations were present on single chromosomes in hyperploid genomic regions, or because of a low proportion of malignant cells in the tumor sample (Table S7). This was true for some tissue samples (Figs 2-4) with low tumor purity, which partly reflects the initially observed genetic ITH. However, validation by ultra-deep sequencing further confirmed the presence of mutations that were missed by the initial sequencing. Thus, many mutations that appeared to be branched in the initial targeted sequencing data were revealed to be truncal by ultra-deep sequencing ( Table 3, Fig. S1). Similarly, many mutations that appeared to be private in the targeted sequencing data were revealed to be branched in the ultra-deep sequencing data. Indeed, heterogeneity (defined as the proportion of nontruncal mutations in the tumor) estimated in the ultra-deep sequencing was 4.0% to 38.4% lower than the heterogeneity estimated in the initial targeted sequencing data (Table 3). Ultra-deep sequencing was also critical for distinguishing the subclonal from the clonal mutations. For example, in Patient 2, the targeted sequencing identified the canonical KRAS G12D mutation in four of five sectors; ultra-deep sequencing showed that the mutation was truncal and clonal (Fig. S1). In summary, ultra-deep sequencing was critical for the identification of somatic mutations that were present across all sectors of the tumor and for an accurate assessment of ITH. Another technique that can detect the variants with very low VAFs is digital PCR (Day et al., 2013). However, this technique requires allele-specific primers or probes for each variant (Vogelstein and Kinzler, 1999), which for some variants are difficult to design. Furthermore, each digital PCR reaction involves partitioning into hundreds or thousands of replicate reactions, which is infeasible (costly) even with current microfluidics or droplet methods (Huggett et al., 2015). By contrast, the ultra-deep sequencing used in the present study required only a single sequencing run post-PCR for all of the 24 sectors and still detected the variants with very low VAFs.

Evidence for early intermixing of truncal mutations
Multiregional sequencing studies provide insights into mutational timing, with the truncal mutations likely to have occurred earlier than the branched or private mutations. Interestingly, in Patients 2 and 4, some of these truncal mutations were subclonal in more than  one sector, despite being likely early mutations (Figs 3  and 5). This suggests intermixing of subclones bearing different mutations during early stages of cancer development and could be consistent with the recently proposed 'Big Bang model' (Sottoriva et al., 2015). In this model, subclonal mutations appear early, producing a spatial mix of clones, and not only selection but also the timing of mutations contributes to ITH and the relative clonal compositions of various regions of a developing tumor. Although the sizes of our tumor sectors (1 cm 3 ) may be too large to definitively support the Big Bang model, our observations highlight that some, but not all, CRCs exhibit this intermixing phenomenon. Alternatively, these observations may also be explained by the existence of neutral mutations that occur early and that, absent selection, remain at subclonal levels in an expanding malignancy (Uchi et al., 2016;Williams et al., 2016). Examples of subclonal truncal mutations include a small deletion in DNMT3A in Patient 2 and mutation in GRM8 in Patient 4 (Table S4).

Varying levels of ITH across tumors and therapeutic implications
Our study demonstrated that the degree of ITH varies across different CRC tumors. This is important because ITH is thought to be a major determinant of a cancer's aggressiveness, resistance to therapy, and affects the patient's clinical course . A different genomic history characterizes each cancer and in some but not all cancers; imprints of early intermixing are observed. These divergent genomic histories and ITH among CRC tumors point out that interpatient heterogeneity is not just in the complement of mutations but also the degree of intratumoral heterogeneity and genomic history within each patient, adding an additional layer to the concept of personalized medicine. We also note that sequencing to identify the potential therapeutic vulnerabilities based on one biopsy might be sufficient for some patients, but not for others. This highlights the need for large prospective studies with multiregional profiling and serial tissue sampling to understand how each patient's tumor evolves across space and time and how this translates to the patient's response to treatment. These studies may be critical for identifying potentially poor responders that share the same ITH profiles and developing better therapeutic interventions for them. Such studies are already ongoing in lung cancer  with other studies being planned for in different cancers including CRC.

Conclusions
This study provided a thorough assessment of ITH in CRC, made possible by deep targeted next-generation sequencing followed by ultra-deep amplicon sequencing, which reduced the false negatives and enabled the detection of subclonality. We observed the variable patterns of ITH, with the proportion of nontruncal mutations varying from 15% to 67%. The tumors also differed in that two had evidence of early mixing, in the form of subclonal, truncal mutations, while two did not. Tumor evolution and ITH will have an impact on tumor biology and clinical behavior. Larger prospective studies may determine how the patterns of ITH could be used to help predict disease progression and inform therapy selection.

Supporting information
Additional Supporting Information may be found online in the supporting information tab for this article: Fig. S1. Comparison of somatic mutations called by targeted hybrid-capture sequencing and ultra-deep amplicon sequencing. Fig. S2. ASCAT profiles of twenty tumor sectors belonging to four patients. Fig. S3. Correlations between genomic copy number and expression of ITGB4, GRB2 and VEGFA. Fig. S4. Elevated copy number and higher expression of AKT2, BCL2L1 and reduced copy number and lower expression of PTEN and PIK3CA. Table S1. Summary of clinicopathologic and treatment details for the four patients. Table S2. Summary of sequencing metrics of targeted hybrid-capture sequencing. Table S3. List of primer sequences used to validate mutations with ultra-deep amplicon sequencing. Table S4. List of somatic SNVs and indels that are subclonal (in red). Table S5. Summary of sequencing metrics of ultradeep amplicon sequencing. Table S6. Summary of somatic mutations validated by ultra-deep amplicon sequencing and their variant allele frequencies. Table S7. ASCAT-estimated malignant cell content and average ploidy of each tumor sector. Table S8. Amplified and deleted regions across four patients based on ASCAT data.