ColoSeq Provides Comprehensive Lynch and Polyposis Syndrome Mutational Analysis Using Massively Parallel Sequencing

Abstract

Lynch syndrome (hereditary nonpolyposis colon cancer) and adenomatous polyposis syndromes frequently have overlapping clinical features. Current approaches for molecular genetic testing are often stepwise, taking a best-candidate gene approach with testing of additional genes if initial results are negative. We report a comprehensive assay called ColoSeq that detects all classes of mutations in Lynch and polyposis syndrome genes using targeted capture and massively parallel next-generation sequencing on the Illumina HiSeq2000 instrument. In blinded specimens and colon cancer cell lines with defined mutations, ColoSeq correctly identified 28/28 (100%) pathogenic mutations in MLH1, MSH2, MSH6, PMS2, EPCAM, APC, and MUTYH, including single nucleotide variants (SNVs), small insertions and deletions, and large copy number variants. There was 100% reproducibility of mutation detection between independent runs. The assay correctly identified 222 of 224 heterozygous SNVs (99.4%) in HapMap samples, demonstrating high sensitivity of calling all variants across each captured gene. Average coverage was greater than 320 reads per base pair when the maximum of 96 index samples with barcodes were pooled. In a specificity study of 19 control patients without cancer from different ethnic backgrounds, we did not find any pathogenic mutations but detected two variants of uncertain significance. ColoSeq offers a powerful, cost-effective means of genetic testing for Lynch and polyposis syndromes that eliminates the need for stepwise testing and multiple follow-up clinical visits.

Defects in mismatch repair (MMR) are responsible for hereditary nonpolyposis colorectal cancer, also known as Lynch syndrome. Inherited loss of function mutations in MLH1, MSH2, MSH6, PMS2, and EPCAM result in a 25% to 75% lifetime risk of colon cancer and up to a 60% lifetime risk of endometrial cancer in women.¹ The population prevalence of Lynch syndrome is estimated to be as high as 1 in 440,² making it the most common inherited cancer predisposition syndrome. Germline mutations in APC and MUTYH result in adenomatous polyposis syndromes that are also associated with a very high lifetime risk of colorectal cancer. Mutations in APC cause familial adenomatous polyposis, Gardner's syndrome, Turcot's syndrome, and attenuated familial adenomatous polyposis, whereas MUTYH mutations are the cause of autosomal recessive MUTYH-associated polyposis syndrome.³ Even with clinical guidelines such as the Revised Bethesda and Amsterdam II criteria,⁴ it can be challenging to distinguish Lynch syndrome from adenomatous polyposis syndromes on the basis of clinical features, particularly with regard to attenuated familial adenomatous polyposis and Lynch syndrome, which present at a similar median age of cancer onset (∼45 to 60 years), and with similar numbers of colonic polyps.^5,6 Clinicians are therefore frequently faced with the dilemma of which gene or genes to test first when ordering expensive genetic testing using standard Sanger sequencing.

A complex battery of tumor-based screening tests have been developed for Lynch syndrome, in part because of the high cost of conventional germline genetic testing. These tests include functional assessment of defective MMR genes either by demonstration of genomic microsatellite instability (MSI) or loss of MLH1, MSH2, MSH6, or PMS2 protein expression by immunohistochemistry (IHC), BRAF V600E mutational analysis, and MLH1 promoter hypermethylation analysis (screening algorithms are reviewed in Pritchard and Grady⁷). Loss of a specific MMR protein revealed by IHC can be helpful in suggesting which gene most likely harbors a germline mutation. MMR proteins are often lost as pairs (MLH1/PMS2 or MSH2/MSH6) because they function as heterodimers. However, mutations in PMS2 and MSH6 are more likely to result in isolated loss of the corresponding protein by IHC because they are minor partners in the heterodimer.⁸ No tumor-based screening tests are currently available for polyposis syndromes.

Massively parallel next-generation sequencing technology has dramatically increased throughput and reduced the cost per nucleotide sequenced compared with traditional Sanger methods, enabling cost-effective sequencing of multiple genes simultaneously in the clinical laboratory setting.^9–13 Target enrichment is generally required to achieve adequate read depth for accurate identification of the spectrum of mutations, including large genomic rearrangements, small insertions and deletions (indels), and SNVs.¹⁴ We recently reported a proof-of-principle study demonstrating the accuracy and feasibility of solution-based targeted capture and next-generation sequencing for 21 genes that are associated with breast and ovarian cancer risk.¹² Here we report the validation results of ColoSeq, a clinical diagnostic assay for hereditary colon cancer that detects single nucleotide, indel, and deletion/duplication mutations in MLH1, MSH2, MSH6, PMS2, EPCAM, APC, and MUTYH. The ColoSeq assay can be performed for approximately the same cost as performing sequencing and deletion/duplication analysis of a single gene by traditional methods, and with equivalent or better sensitivity and accuracy, eliminating the need for stepwise molecular genetic testing in patients with suspected Lynch or polyposis syndromes.

Materials and Methods

DNA Samples

We tested a total of 82 unique DNA samples, including 23 peripheral blood DNA samples from patients with known mutations in MLH1, MSH2, MSH6, PMS2, EPCAM, APC, or MUTYH; 31 peripheral blood DNA samples from patients with a clinical history suggestive of Lynch or polyposis syndrome; 19 peripheral blood DNA samples from patients without a known family history of cancer; 6 publicly available DNA samples from the HapMap project;¹⁵ and 3 DNA samples from colon cancer cell lines known to harbor mutations in MMR genes and/or APC. Tumor cell lines (LoVo, HCT116, LS174T) were obtained from William M. Grady (University of Washington). HapMap samples (NA18507, NA18558, NA07019, NA07348, NA10857, NA10851) were obtained from Coriell Cell Repositories (Camden, New Jersey). Clinical specimens were obtained in accordance with the Declaration of Helsinki and the study was approved by the Human Subjects Division of the University of Washington (protocol 34173) and the University of Chicago Institutional Review Board.

Library Construction, Gene Capture, and Massively Parallel Sequencing

Three micrograms of DNA was sonicated to a peak of 200 bp on a Covaris S2 instrument (Covaris, Woburn, MA) in 1× low TE (10 mmol/L Tris/0.1 mmol/L EDTA) for 6 minutes using frequency sweeping mode with duty cycle, 10%; intensity, 5; cycles per burst, 200 at a temperature 4° to 7°C. After sonication, DNA was purified with AMPure XP beads (Beckman Coulter, Brea CA) and subjected to three enzymatic steps: end repair, A-tailing, and ligation to Illumina paired-end adapters as described in the SureSelectXT Target Enrichment for Illumina multiplexed sequencing that is available for free download. All liquid handling steps were performed in 96-well plates on a Bravo liquid-handling instrument (Agilent Technologies, Santa Clara, CA). The adapter-ligated library was amplified by polymerase chain reaction (PCR) for five cycles with Illumina primers 1.0 and 2.0 and quantified by a DNA1000 chip on a Bioanalyzer 2100 instrument (Agilent Technologies). Individual paired-end libraries (500 ng) were hybridized to a custom design of complementary RNA (cRNA) biotinylated oligonucleotides targeting 31 genes in 30 genomic regions (see Supplemental Table S1 at http://jmd.amjpathol.org). The 120-mer oligonucleotide baits were designed in Agilent's eArray web portal with the following parameters: centered tiling, 3x bait overlap and a maximum overlap of 20 bp into repetitive regions. The custom design targets a total of 1.1 Mb of DNA including 209 kb in MLH1, MSH2, MSH6, PMS2, EPCAM, APC, and MUTYH (Table 1). The BED file of probe sequences is available on request. After 24 hours of hybridization at 65°C, the library-bait hybrids were purified by incubation with streptavidin-bound T1 Dynabeads (LifeTechnologies, Carlsbad CA) and washed with increasing stringency to remove nonspecific binding. After capture, each library was amplified by PCR directly on the Dynabeads for 13 cycles with primers containing a unique 6-bp index (Table 2). After PCR amplification, the libraries were quantified by a high-sensitivity chip on a Bioanalyzer 2100 instrument (Agilent Technologies). Equimolar concentrations of 96 libraries were pooled to a final concentration of 10 pM, denatured with 3N NaOH, and cluster amplified with a cBot instrument on a single lane of an Illumina v3 flow cell. Sequencing was performed with 2 × 101-bp paired-end reads and a 7-bp index read using SBS v3 chemistry on a HiSeq2000 (Illumina, Inc, San Diego, CA).

Table 1.

Genes Validated for ColoSeq

Gene	Chromosome	Kb targeted⁎	Genomic region targeted (Hg 19)
			Start	End
MLH1	3	35	37029979	37097337
MSH2 + EPCAM	2	48	47595263	47715360
MSH6	2	19	48005221	48039092
PMS2	7	17	6007870	6053737
APC	5	77	112038202	112186936
MUTYH	1	14	45789914	45811142

A total of 30 genomic regions representing 31 genes related to hereditary cancer risk were targeted¹⁹, but here we focus on validation of these 7 genes.

^⁎Repetitive DNA elements were not targeted.

Table 2.

Postindex Barcodes

ATCACG	AACAAA	AATAGG	ACTCTC
CGATGT	CACGAT	CCACGC	CGGAAT
TTAGGC	GATATA	GCTCCA	GTGGCC
TGACCA	TATAAT	TCGGCA	TGCTGG
ACAGTG	AACCCC	ACAAAC	ACTGAT
GCCAAT	CACTCA	CCCATG	CTAGCT
CAGATC	GATGCT	GGCACA	GTTTCG
ACTTGA	TCATTC	TCTACC	TGGCGC
GATCAG	AACTTG	ACATCT	AGAAGA
TAGCTT	CAGGCG	CCCCCT	CTATAC
GGCTAC	GCAAGG	GGCCTG	CGTACG
CTTGTA	ATAATT	ATCCTA	ATGAGC
AAACAT	AAGACT	ACCCAG	AGATAG
CAAAAG	CATGGC	CCGCAA	CTCAGA
GAAACC	GCACTT	GTAGAG	GAGTGG
TAATCG	TCCCGA	TGAATG	TTCGAA
AAAGCA	AAGCGA	ACCGGC	AGCATC
CAACTA	CATTTT	CCTTAG	CTGCTG
GAATAA	GCCGCG	GTCCGC	GGTAGC
TACAGC	TCGAAG	TGCCAT	TTCTCC
AAATGC	AAGGAC	ACGATA	AGCGCT
CACCGG	CCAACA	CGAGAA	CCGTCC
GACGGA	GCCTTA	GTGAAA	ATTCCT
AGGCCG	ATACGG	ATCTAT	AGGTTT

Mutation Analysis

Sequence alignment and variant calling were performed against the reference human genome (UCSC hg19). Sequencing reads were aligned using MAQ,¹⁶ and SNVs and insertions and deletions were detected as previously described.¹² Each variant was annotated with respect to gene location and predicted function in Human Genome Variation Society nomenclature. Deletions and duplications of exons were detected by depth of coverage analysis.¹⁷ All previously unidentified frameshift, nonsense, and splice site mutations predicted to be deleterious to protein function were confirmed by PCR amplification and Sanger sequencing. Exonic deletions and duplications were confirmed by multiplex ligation-dependent probe amplification or gap-PCR and direct sequencing.¹⁸

Results

ColoSeq Assay

The objective of our study was to evaluate the performance of targeted DNA capture and massively parallel sequencing for the detection of inherited mutations in colon cancer in the clinical laboratory setting. We designed oligonucleotides to target all exons, introns, and approximately 10 kb of 5′ and 3′ flanking genomic regions of the seven genes that are most commonly responsible for inherited risk of colon cancer (Table 1). Our capture panel also includes 24 other genes that rarely harbor mutations causing colon cancer, endometrial cancer, and other solid tumors,¹⁹ for a total of 31 captured genes and 1.1 Mb of captured DNA after removal of repetitive sequences. Here we focus on validation results of the seven genes listed in Table 1 that constitute the ColoSeq assay for Lynch and polyposis syndromes.

Workflow of the ColoSeq assay is summarized in Figure 1. We begin by fragmenting genomic DNA extracted from peripheral blood and preparing libraries with an average insert size of approximately 200 bp. ColoSeq genes are captured via solution hybridization with complementary oligonucleotide RNA (cRNA) baits (Agilent SureSelect). Each captured sample library is amplified by PCR to incorporate a unique index barcode, and up to 96 barcoded samples are pooled and run on a single lane of an Illumina HiSeq2000 instrument.

Figure 1.

ColoSeq workflow.

It is unusual to pool patient samples in the clinical setting because of the risks of sample misidentification and analytic interference. To control for specimen mix-up and for potential barcode-specific sequencing biases, each clinical sample is indexed with two distinct barcodes and run in duplicate. Barcodes were selected such that a single base change could not lead to one barcode being misidentified as another (Table 2). Less than 0.005% of sequencing reads were assigned to the incorrect patient because of barcode misidentification.

Paired-end 2 × 101-bp reads are generated at a median of 320-fold coverage per nucleotide across the entire targeted region (range of median coverage between samples 145-fold to 556-fold; Figure 2A). The median coverage specifically across the seven ColoSeq genes is 475-fold (Figure 2B). Raw sequence of 0.3 gigabases (Gb) (range 0.15 to 0.56 Gb) of high-quality (>Q30) sequence per sample is generated when the maximum 96 samples are pooled within a lane. On average, 62% of sequence reads are on target, mapping specifically to the captured regions. We align sequences to the reference genome (build Hg 19) and select variants for further analysis that meet the following criteria: i) variant is present in sequence reads from both strands (mean ± SD, 433 ± 97 variants/sample), ii) population frequency of the variant is less than 5% (44 ± 31 variants/sample remain), and iii) variant represents at least 5% of the sequence reads at a particular site (26 ± 21 variants/sample remain). We chose a threshold of 5% variant reads to ensure that variants in highly segmentally duplicated regions of genes would be detected. Frequencies of variants among individuals without colon cancer are based on an internal database of more than 1000 unique patients from different ethnic backgrounds who have had targeted capture and massively parallel sequencing performed for these genes.^12,19 By using an internal frequency database, we can quickly filter out both common benign variants and potential recurrent artifactual variant calls, leaving an average of 20 noncoding and only 6 rare coding or splice junction variants per patient across the seven genes. These filtered variants are then compared with the exome variant server (NHLBI Exome Sequencing Project [ESP], Seattle, WA), dbSNP, and public mutation databases such as InSight²⁰ (http://www.insight-group.org/mutations) and results reviewed by a laboratory director. Novel variants of uncertain significance (VUS) are analyzed by PolyPhen2,²¹ SIFT,²² and MutationTaster²³ to predict potential deleterious effects and are assessed for evolutionary conservation. Pathogenic mutations and VUS are confirmed by Sanger sequencing or multiplex ligation-dependent probe amplification (for deletions/duplications). Missense variants that are not well characterized are reported as VUS when they meet the following criteria: i) population frequency <5%, ii) evolutionarily conserved, iii) predicted to be at least possibly damaging by in silico prediction tools, and iv) confirmed by Sanger sequencing.

Figure 2.

Depth of coverage. Depth of coverage from a representative patient sample run on a single lane of a HiSeq2000 instrument in a 96-plexed ColoSeq run. A: Distribution of coverage for each targeted base. Each bar represents the number of base pairs (in 1000s) at a particular depth of coverage. Of the targeted bases, 98.5% have a minimum of 50-fold coverage. B: Median coverage across each ColoSeq gene is shown (average coverage of 475-fold). Areas with depth of coverage >1000-fold reflect capture of highly homologous genomic regions.

Sensitivity

To assess the sensitivity of the ColoSeq assay to detect pathogenic mutations, we blindly tested peripheral blood DNA samples from 23 cancer patients and 3 colon cancer cell lines (LS174T, HCT116, LoVo) with previously defined germline mutations in MLH1, MSH2, MSH6, PMS2, EPCAM, MUTYH, or APC. Some cell lines harbored more than one mutation. ColoSeq correctly identified all mutations (23/23 patient mutations, 5/5 cell line mutations, 100% sensitivity) including nonsense, missense, frameshift, in-frame deletions, splice site, and large deletions and duplications (Table 3, cohorts: known and cell line).^24–39

Table 3.

Mutations and Variants of Uncertain Significance Identified by ColoSeq

Cohort	ID	Gene	Mutation/VUS†	Effect	Chr	Start (Hg19)	End (Hg19)	Total reads‡	Variant reads‡
Cell line	HCT116	MLH1	c.755C>A20	S252⁎	3	37,056,000		244	244
Cell line	LoVo	APC	c.3340C>T24,25	R1114⁎	5	112,174,631		417	194
Cell line	LoVo	APC	c.4289delC25	M1431Cfs⁎42	5	112,175,580	112,175,581	174	97
Cell line	LoVo	MSH2	∼51kb deletion26	del exons 3-8	2	∼47,637,277	∼47,688,664	NA	NA
Cell line	LS174T	MLH1	c.350C>T27	T117M	3	37,045,935		325	323
Known	FO5	MSH2	∼31kb deletion28	del ex 9-16	2	∼47,688,689	∼47,719,344	322	160
Known	011	MSH2	c.513delG29	K172Afs⁎2	2	47,637,378	47,637,379	64	31
Known	012	MSH6	c.3103C>T30	R1035⁎	2	48,028,225		389	117
Known	013	MSH2	c.1609A>T	K537⁎	2	47,693,895		151	73
Known	014	MLH1	c.1668-1G>A28	Splice	3	37,083,758		180	91
Known	017	MLH1	c.1517 T>C31	V506A	3	37,070,382		227	114
Known	022	APC	c.221-2A>G	Splice	5	112,102,884		143	80
Known	023	MLH1	c.1381A>T32	K461⁎	3	37,067,470		139	75
Known	025	MSH2	c.380A>G	N127S (VUS)	2	47,637,246		64	35
Known	031	MSH6	c.694C>T	Q232⁎	2	48,025,816		152	76
Known	032	MSH2 + EPCAM	∼40kb deletion	del ex 1-2 + EPCAM	2	∼47,595,376	∼47,635,837	322	172
Known	034	MSH2	c.2501_2507delCTAATTT	N835Lfs⁎4	2	47,707,877	47,707,884	129	47
Known	035	PMS2	c.1927C>T33	Q643⁎	7	6026,469		97	48
Known	037	MLH1	∼18kb duplication	dup ex 6-12	3	∼37,050,170	∼37,067,982	351	509
Known	086	MSH6	c.3485_3487delCTG34	A1162del	2	48,032,092	48,032,095	134	58
Known	087	MSH2	∼20kb deletion	del exon 1-6	2	47,629,721	47,649,842	274	120
Known	088	MSH6	c.3956_3959delAAGC34	A1320Gfs⁎6	2	48,033,745	48,033,749	174	72
Known	090	MSH6	c.2731C>T34	R911⁎	2	48,027,853		130	61
Known	093	MSH6	c.2731C>T34	R911⁎	2	48,027,853		74	41
Known	094	MSH6	c.3939_3940ins1934	Q1314Sfs⁎5	2	48,033,728	48,033,728	166	20
Known	096	MLH1	∼3kb deletion35	del exon 14-15	3	∼37,081,516	∼37,084,593	345	177
Known	147	MUTYH	c.1187C>T36	G396D§	1	45797228		147	68
Known	148	MUTYH	c.1187C>T36	G396D§	1	45797228		102	32
Control	047	MLH1	c.2161T>C	Y721H (VUS)	3	37,092,034		220	137
Control	063	MLH1	c.1151T>A31	V384D (VUS)	3	37,067,240		167	77
Unknown	003	APC	c.1959G>A37	?Splice (VUS)	5	112,173,250		199	92
Unknown	004	MSH6	c.2147_2150delCAGT	V717Afs⁎18	2	48,027,269	48,027,273	100	41
Unknown	006	APC	c.426_427delAT38	L143Afs⁎4	5	112,111,329	112,111,331	143	66
Unknown	018	MUTYH	c.158-1G>A	Splice (VUS)	1	45,799,276		61	33
Unknown	036	PMS2	c.2192_2196delTAACT39	L731Cfs⁎3	7	6018,306	6018,311	255	64
Unknown	089	PMS2	c.2192_2196delTAACT39	L731Cfs⁎3	7	6018,306	6018,311	545	81
Unknown	040A	MSH6	c.2588A>G	E863G (VUS)	2	48,027,710		278	140
Unknown	041	APC	c.646C>T38	R216⁎	5	112,128,143		179	83
Unknown	080	PMS2	3.3kb deletion	del exon 8	7	6033,256¶	6036,607¶	409	192

Chr, chromosome; NA, not applicable.

^⁎Note: Asterisk (*) refers to stop codon (per Human Genome Variation Society standard nomenclature).

^†Reference sequences: MLH1 NM_000249.3; MSH2 NM_000251.1; MSH6 NM_000179.2; PMS2 NM_000535.5; EPCAM NM_002354.2; APC NM_000038.5; MUTYH NM_001128425.1.

^‡Numbers reflect perfect matches only. For CNVs, total reads refers to the normalized read depth and variant reads to the read depth in the sample.

^§This mutation is commonly referred to as G382D (NM_001048171.1:c.1145G>A).

^¶Exact breakpoints were determined for the PMS2 exon 8 deletion, which we detected in patient 080.

To assess the analytic sensitivity of detecting all heterozygous single base pair variants across these genes, we evaluated six DNA samples that had been genotyped by the HapMap Project¹⁵ (http://www.hapmap.org, last accessed December 6, 2011). HapMap genotype data were available for a total of 1388 sites in aggregate across the six samples, including 226 heterozygous SNVs (38 ± 23 per sample), 144 homozygous variants (27 ± 15 per sample), and 1018 sites that matched the reference genome (wild type, 170 ± 23 per sample). We focused on defining the sensitivity of heterozygous variant detection because pathogenic mutations in Lynch and polyposis syndromes are almost always heterozygous, except in MUTYH. ColoSeq identified 222/226 heterozygous SNVs in the seven genes captured by the assay. Two of the four discrepant variants were found by Sanger sequencing to be errors in the HapMap annotation, for a sensitivity of 99.4% (222/224) (Table 4). The two HapMap variants that were missed by ColoSeq corresponded to a single intronic SNV (rs3771280) that was missed in two different samples. This SNV is in close proximity to two other SNVs (rs3771278, rs3771279), which were both either heterozygous or homozygous for the rarer allele in the two samples, resulting in the reads mapping to this region being discarded because of poor alignment with the reference genome. ColoSeq correctly identified all HapMap variants within exons and splice junctions (27/27; 100% sensitivity). Among homozygous SNVs, all were correctly identified except rs3771280.

Table 4.

Analytic Sensitivity, Specificity, and Accuracy

	Sensitivity	Specificity	Accuracy
All variants (with introns)	99.4% (222/224)	99.4% (1012/1018)	99.4% (1234/1242)
Exons only	100% (27/27)	100% (198/198)	100% (225/225)

Calculations are based on heterozygous variant detection in 6 HapMap samples.

We next evaluated the performance of ColoSeq in a cohort of 31 prospectively collected blood samples from patients with a clinical history suggestive of Lynch or polyposis syndrome, but without a previously determined mutation. In this cohort, we identified six patients with pathogenic mutations and three additional patients with VUS; the mutations were all confirmed using alternative methods (Table 3; cohort: unknown). Most of the patients in this cohort had previous clinical testing performed by outside laboratories for one or more of the ColoSeq genes, with negative results. Even though the group was enriched for patients who had negative test results for one or more of the ColoSeq genes, our mutation detection in 6 of 31 patients was similar to published rates among patients meeting Bethesda criteria,⁴⁰ highlighting the utility of a comprehensive Lynch and polyposis panel.

Specificity and Accuracy

We determined analytic specificity by analyzing the 6 HapMap characterized samples at 1018 known nonvariant (reference sequence) sites in the seven ColoSeq genes. ColoSeq correctly identified 1012/1018 sites as nonvariant from the reference genome for an analytic specificity of 99.4% (Table 4). The six false-positive variants that we detected but that were not reported as present in HapMap samples were in MLH1 intron 11 (rs3774339, detected in one sample), and in segmentally duplicated exons 11 and 14 of PMS2 (rs1805321, detected in three samples; rs1059060, detected in two samples). Analytic specificity was 100% within single-copy exons and splice junctions (198/198). We calculated analytic accuracy as [number of true positives + number of true negatives]/[total number of HapMap genotyped sites]. Accuracy was 99.4% and 100% for all variants and exons only, respectively (Table 4).

To assess clinical specificity we selected 19 control patients without a personal or family history of colon cancer or other Lynch syndrome–associated cancers. Patients were selected to represent diverse ethnic backgrounds including African American, Pacific Islander, white, Indian, Cambodian, Vietnamese, Iranian, Hispanic, Alaskan Native, Chinese, Laotian, and Filipino (see Supplemental Table S2 at http://jmd.amjpathol.org). We did not find any pathogenic mutations in this cohort, for a specificity of 100% (19/19). However, we identified a VUS that met our criteria for clinical reporting in two control samples; one each from an African-American and a Vietnamese control subject (Table 3, cohort: control), suggesting that about 10% of nonwhite patients without Lynch or polyposis syndromes will have a VUS detected (2/17).

Deletion/Duplication Analysis

The very high fold coverage (mean, 320-fold) allowed accurate detection of large deletions and duplications using normalized depth of coverage and split-read analysis (Figure 3).^12,17 Exact breakpoints could not be determined for most large deletions and duplications because breakpoints are commonly in Alu or other repetitive DNA elements that are not captured by our design. However, the assay demonstrated at least exon-level resolution for all large deletions and duplications in all cases, which was comparable or better than the resolution of traditional approaches to deletion/duplication analysis such as multiplex ligation-dependent probe amplification. ColoSeq correctly identified 6/6 known large deletions/duplications in blinded challenge specimens (Table 3, cohort: known). In addition, in DNA from a patient with microsatellite instability–positive endometrial cancer and isolated loss of PMS2 protein expression (by IHC) who had not had previous genetic testing, we detected a PMS2 exon 8 deletion. We confirmed the exact breakpoints of this mutation by gap-PCR and Sanger sequencing (Table 3, patient 080).

Figure 3.

Deletions and duplications detected by normalized depth of coverage. Three examples of genomic deletions (MLH1 exons 14-15, MSH2 exons 1-6, and PMS2 exon 8) and one example of a genomic duplication (MLH1 exons 6-12) detected by ColoSeq. Read depth is normalized across the 96 samples on a run. Significant deviations less than or greater than a normalized ratio of 1 across adjacent baited regions reflect genomic deletions (red) and duplications (blue). Genomic coordinates shown on the x axis are hg19.

Reproducibility and Cutoff

To determine the between-run assay reproducibility, we ran 75 samples on at least two independent runs. We detected 39/39 pathogenic mutations and VUS (Table 3) in replicate runs, yielding 100% between-run reproducibility of mutation detection. We next assessed the between-run reproducibility for all variant calls in unique exons and splice junctions (exons only, mean = 39 ± 9 variants per sample), or for the entire captured region, including deeply intronic, nonunique, and repetitive DNA elements (with introns, mean = 451 ± 70 variants per sample). We separated SNVs and indel variants for this analysis. Not surprisingly, assay reproducibility increased as a function of the number of variant sequence reads, particularly for indel variants (Figure 4). We selected a minimum threshold of 15 variant sequencing reads to assess assay reproducibility based on the following two criteria: all pathogenic mutations and VUS were reliably detected with at least 15 variant reads, and the threshold of 15 variant reads substantially improved assay reproducibility without sacrificing sensitivity (all detected HapMap variants had at least 15 reads). We did not use receiver operating characteristic analysis to select the cutoff because the analytic specificity was >99% at all cutoff values (based on the 1018 HapMap known wild-type SNVs) (see Supplemental Figure S1 at http://jmd.amjpathol.org). For exons only, for which the performance of the assay is most critical, between-run reproducibility was 99.3% for SNVs and 98.9% for indels at a cutoff of 15 variant reads (Figure 4). Reproducibility improved to 100% for indels and SNVs at cutoff values higher than 40, which is representative of the majority of variant calls (66% of indels and 86% of SNVs had 40 or more variant reads). Reproducibility was lower with introns, at 93.9% for SNVs and 92.5% for indels at the cutoff of 15 reads (see Supplemental Figure S2 at http://jmd.amjpathol.org). Variant calls that did not replicate on repeated runs were most prevalent in introns of PMS2, which has a family of highly homologous pseudogenes on chromosome 7.⁴¹ Importantly, we were able to reproducibly detect all SNV, indel, and deletion mutations in exons of PMS2, even in regions of the gene that are segmentally duplicated (Table 3).

Figure 4.

Between-run and within-run reproducibility. Between-run and within-run reproducibility are shown as a function of minimum threshold for indel variants and calling single nucleotide (SNV). Variants were considered reproducible if they met or exceeded the specified cutoff value in the index sample and appeared in the replicate sample regardless of cutoff value. Reproducibility is shown for all variants in single-copy regions of exons and splice junctions. A threshold of 15 variant reads was chosen for the assay. Black squares, between-run; grey circles, within run.

Within-run reproducibility was determined by running 12 samples in duplicate on the same run. The within-run reproducibility for exons only was 100% for both SNVs and indels at the cutoff of 15 reads (Figure 4). Within-run reproducibility with introns was 95.0% for SNVs and 92.1% for indels at the cutoff of 15 variant reads, which was qualitatively similar to the between-run results (see Supplemental Figure S2 at http://jmd.amjpathol.org).

Discussion

Next-generation sequencing technology is rapidly changing the landscape of genetic testing. During a plenary session at the 2011 Association of Molecular Pathology annual meeting, Dr. Stephen Kingsmore suggested that molecular pathologists will look back to this time as a golden age in the field of molecular diagnostics because of the new possibilities afforded by massively parallel sequencing. Much of the excitement surrounding this technology is directed at the possibility of clinical exome and whole-genome sequencing as a one-stop, all-inclusive genetic test. Clinical laboratories such as Ambry Genetics and Baylor Whole Genome Laboratory are already offering exome sequencing.^42,43 However, exome and genome sequencing currently have important limitations compared with targeted approaches with regard to both cost-effectiveness and analytic performance in the clinical setting. Exome sequencing often does not achieve adequate depth of coverage to detect structural rearrangements such as large deletions and duplications that are a significant component of the mutational spectrum for many inherited disorders. Considering that up to 15% of disease-causing mutations in the Lynch syndrome are due to large deletions,² it is clinically relevant if this class of mutation is missed. An important advantage of the targeted approach we have taken with ColoSeq is that it can readily detect structural rearrangements through a very high depth of coverage.

Additional advantages of our targeted capture approach for clinical diagnostic testing include the following: i) For scalability, we were able to multiplex and pool 96 index barcoded samples onto one lane of an Illumina HiSeq flow cell, facilitating cost-effectiveness and allowing samples to be run in duplicate to control for sample mix-up and analytic variability. ii) Clinical validity was achieved by targeting and validating only genes that, when mutated, are well-established causes of hereditary colon cancer; it is therefore straightforward to incorporate the results of ColoSeq testing into clinical decision-making. iii) Limiting VUS, which can cause stress for both patients and providers, was achieved by our focus on a relatively limited number of genes and the existence of well-annotated hereditary colon cancer mutation databases such as InSight,²⁰ and MMRUV⁴⁴; we found a reportable VUS in only about 1 in 10 patients. iv) Clinical expertise is strengthened by focusing on just a handful of genes, making it reasonable for a molecular pathologist to gain familiarity with their mutational spectrum and to consider the data at each variant position individually before signing out a case.

At least one previous study has evaluated targeted capture and massively parallel sequencing for the detection of Lynch and polyposis syndromes.⁴⁵ In that study, solid-phase NimbleGen 385K Custom Sequence Capture Arrays were used to examine 22 genes that included MLH1, MSH2, MSH6, and APC. One of four samples with a known Lynch syndrome indel mutations was missed using both Roche 454 (Roche, Branford, CT) and Illumina GAII technology, along with many other neutral variants that were detectable by Sanger sequencing. The reason for the relatively poor performance could be attributed to inefficient/off-target capture, limitations in sequencing homopolymer regions on the 454 platform, and difficulties calling indels on the Illumina GAII platform due to short (36 bp) single-end read length. Both gene capture and sequencing technology has matured in the past 2 years, allowing us to overcome these limitations through more efficient capture and much longer 101-bp paired-end reads.

Methods and guidelines to rigorously validate clinical diagnostic assays using next-generation sequencing technology are just beginning to emerge. The Association of Molecular Pathology is leading an effort to establish formal guidelines for the validation of tests using this technology. This includes a statement from the Association of Molecular Pathology in June 2011 that outlines aspects of method validation and ongoing quality control, including sensitivity and specificity, and recommends that precharacterized control samples such as HapMap DNAs be analyzed on each run (http://www.amp.org/documents/AMPCommentsNGS_June2011_Final.pdf, last accessed December 1, 2011). Additional groups such as the Centers for Disease Control–led Nex-StoCT working group (http://www.cdc.gov/osels/lspppo/Genetic_Testing_Quality_Practices/Nex-StoCT.html, last accessed December 1, 2011) and the College of American Pathologists are also working to establish guidelines. We have taken a multifaceted approach to validate the ColoSeq assay that includes measurements of both analytic and clinical sensitivity and specificity, between- and within-run reproducibility, and selection of cutoffs (analytic measurement range). In addition to using blinded challenge specimens with defined mutations at a single locus, we determined analytic sensitivity using HapMap samples that had between 50 and 100 defined variants each in the seven ColoSeq genes. Through this process, we discovered a false-negative intronic SNV (rs3771280) and, importantly, were alerted to the mechanism by which this SNV was missed.

We were particularly concerned with defining the VUS detection rate of the assay, given the large amount of DNA sequenced and the potential consequence that the assay would cause undue worry among patients and their families if a high number of VUS were reported. We were pleased to find a reportable VUS for only two nonwhite patients of 19 patients without Lynch or polyposis syndrome, even among patients of diverse ethnic backgrounds in which missense variants are not well characterized. Still, we will undoubtedly encounter VUS more frequently than in standard single-gene assays. Better in silico tools to predict the relative pathogenicity of VUS will be valuable, such as gene-specific algorithms⁴⁶ and aggregate pathogenicity scores⁴⁷ that can integrate established algorithms such as PolyPhen2,²¹ SIFT,²² Genomic Evolutionary Rate Profiling,⁴⁸ and MutationTaster.²³

We designed the seven-gene ColoSeq panel to be focused only on genes that have a well-established role in clinical decision making for patients with Lynch or polyposis syndromes. Mutations in genes such as MLH3, MSH3, PMS1, and EXO1 that have been reported to cause Lynch syndrome in rare cases² are not included in the ColoSeq panel because of lack of knowledge regarding their clinical significance. Our design targets additional genes for rare hereditary cancer syndromes that may predispose to gastrointestinal cancers, including Cowden syndrome (PTEN), hereditary diffuse gastric cancer (CDH1), Peutz-Jeghers syndrome (STK11), and Li-Fraumeni syndrome (TP53), and we are optimistic about validating these genes in a future expanded version of the ColoSeq panel.

A limitation of the assay is the long turnaround time necessitated by the 9-day run time on the HiSeq2000 instrument. ColoSeq is amenable to transition to a smaller, more clinically oriented instrument, with faster turnaround time such as the MiSeq (Illumina) or Ion Torrent (Guilford, CT). However, an important drawback of these more rapid instruments is that only about 1 to 2 Gb of sequence data are generated per run, compared with 30 to 40 Gb for a single lane of a HiSeq2000. We estimate that only about three to five index barcoded samples could be pooled per run on the MiSeq using 100-bp or 150-bp paired-end reads to achieve the same depth of coverage as our current 96-plexed assay on the HiSeq2000. The assay workflow and turnaround time might also be improved through newer, more streamlined library preparation strategies such as transposase-based methods.⁴⁹

In conclusion, we have developed and validated a comprehensive and cost-effective test for hereditary colon cancer syndromes that uses solution-based targeted capture and next-generation sequencing. The assay detected 100% of known mutations in challenge specimens, including all indel mutations and large deletions and duplications. We believe that ColoSeq will streamline and simplify genetic testing in patients with suspected hereditary colon cancer syndromes, sparing patients from needing to have multiple rounds of expensive genetic testing to establish a diagnosis.

Supplementary data

Supplemental Figure S1

Receiver operating characteristic analysis. The sensitivity of the assay to detect HapMap-characterized single nucleotide polymorphisms is plotted versus 1-specificity at different minimum fold-coverage (number of reads) cutoffs. Even at the lowest reads cutoff, the specificity was >99%, which is why 1-specificity does not go to 0.

Supplemental Figure S2

Assay reproducibility for introns, nonunique, and repetitive DNA. Between-run (black squares) and within-run (gray circles) reproducibility is shown as a function of minimum threshold for calling SNV and indel variants. Variants were considered reproducible if they met or exceeded the specified cutoff value in the index sample and appeared in the replicate sample regardless of cutoff value. All variants (including in deeply intronic, nonunique, and repetitive DNA) were included in this analysis. Nonreproducible variant calls were not in clinically important regions and were most prevalent in segmentally duplicated introns of PMS2.