Downloaded from
`
`genome.cshlp.org
`
`
`
` on February 14, 2022 - Published by Cold Spring Harbor Laboratory Press
`
`
`
`Method
`Cost-effective, high-throughput DNA sequencing
`libraries for multiplexed target capture
`Nadin Rohland1 and David Reich
`Department of Genetics, Harvard Medical School, Boston, Massachusetts 02115, USA; Broad Institute of Harvard and MIT,
`Cambridge, Massachusetts 02139, USA
`
`Improvements in technology have reduced the cost of DNA sequencing to the point that the limiting factor for many
`experiments is the time and reagent cost of sample preparation. We present an approach in which 192 sequencing libraries
`can be produced in a single day of technician time at a cost of about $15 per sample. These libraries are effective not only
`for low-pass whole-genome sequencing, but also for simultaneously enriching them in pools of approximately 100 in-
`dividually barcoded samples for a subset of the genome without substantial loss in efficiency of target capture. We
`illustrate the power and effectiveness of this approach on about 2000 samples from a prostate cancer study.
`
`[Supplemental material is available for this article.]
`
`Improvements in technology have reduced the sequencing cost per
`base by more than a 100,000-fold in the last decade (Lander 2011).
`The amount of sequence data that is needed per sample, for exam-
`ple, for studying small target regions or low-coverage sequencing of
`whole genomes is often less than the commercial cost of ‘‘library’’
`preparation, so that library preparation is now often the limiting cost
`for many projects. To reduce library preparation costs, researchers
`can purchase kits and produce libraries in their own laboratories or
`use published library preparation protocols (Mamanova et al. 2010;
`Meyer and Kircher 2010; Fisher et al. 2011). However, this approach
`has two limitations. First, available kits have limited throughput so
`that scaling to thousands of samples is difficult without automation.
`Second, an important application of next-generation sequencing
`technology is to enrich sample libraries for a targeted subsection of
`the genome (like all the exons) (Albert et al. 2007; Hodges et al. 2007;
`Gnirke et al. 2009), and then to sequence this enriched pool of DNA,
`but such experiments are expensive because of the high costs of
`target capture reagents. One way to save funds is to pool samples
`prior to target enrichment (after barcoding to allow them to be dis-
`tinguished after the data are gathered). Although the recently in-
`troduced Nextera DNA Sample Prep Kit (Illumina) together with
`‘‘dual indexing’’ (12 3 8 indices and two index reads) allows higher
`sample throughput for library preparation (Adey et al. 2010) and
`pooling of up to 96 libraries, the long indexed adapter may interfere
`during pooled hybrid selection (see below).
`We report a method for barcoded library preparation that al-
`lows highly multiplexed pooled target selection (hybrid selection
`or hybrid capture). We demonstrate its usefulness by generating
`libraries for more than 2000 samples from a prostate cancer study
`that we have enriched for a 2.2-Mb subset of the genome of interest
`for prostate cancer. We also demonstrate the effectiveness of li-
`braries produced with this strategy for whole-genome sequencing,
`both by generating 40 human libraries and sequencing them to
`fivefold coverage, and by generating 12 microbial libraries and
`sequencing them to 150-fold coverage. Our method was engi-
`neered for high-throughput sample preparation and low cost, and
`thus we implemented fewer quality-control steps and were willing
`
`1Corresponding author.
`E-mail nrohland@genetics.med.harvard.edu.
`Article published online before print. Article, supplemental material, and publi-
`cation date are at http://www.genome.org/cgi/doi/10.1101/gr.128124.111.
`
`to accept a higher rate of duplicated reads compared with methods
`that have been optimized to maximize library complexity and
`quality (Meyer and Kircher 2010; Fisher et al. 2011). Because of this,
`our method is not ideal for deep sequencing of large genomes (e.g.,
`human genome at 303), where sequencing costs are high enough
`that it makes sense to use a library that has as low a duplication rate
`as possible. However, our method is advantageous for projects in
`which a modest amount of sequencing is needed per sample, so that
`the savings in sample preparation outweigh costs due to sequencing
`duplicated molecules or failed libraries. Projects that fall into this
`category include low-pass sequencing of human genomes, micro-
`bial sequencing, and target capture of human exomes and smaller
`genomic targets.
`Our method reduces costs and increases throughput by paral-
`lelizing the library preparation in 96-well plates, reducing enzyme
`volumes at a cost-intensive step, using inexpensive paramagnetic
`beads for size selection and buffer exchange steps (DeAngelis et al.
`1995; Lennon et al. 2010; Meyer and Kircher 2010), and automation
`(Farias-Hesson et al. 2010; Lennon et al. 2010; Lundin et al. 2010;
`Fisher et al. 2011). To permit highly multiplexed sample pooling
`prior to target enrichment or sequencing, we attach ‘‘internal’’
`barcodes directly to sheared DNA from a sample that is being se-
`quenced, and flank the barcoded DNA fragments by partial se-
`quencing adapters that are short enough that they do not strongly
`interfere during enrichment (the adapters are then extended after
`the enrichment step). By combining these individual libraries in
`pools and enriching them for a subset of the genome, we show that
`we obtain data that are effective for polymorphism discovery,
`without substantial loss in capture efficiency.
`
`Outline
`
`Our method is based on a blunt-end-ligation method originally
`developed for the 454 Life Sciences (Roche) platform (Stiller et al.
`2009), which we have extensively modified for the Illumina plat-
`form to reduce costs and increase sample processing speed, by
`parallelizing the procedure in 96-well format and automating the
`labor-intensive cleanup steps (Figs. 1, 2A; Methods; Supplemental
`Notes). Some of the modifications adapt ideas from the literature,
`such as DNA fragmentation on the Covaris E210 instrument in 96-
`well PCR plates (Lennon et al. 2010) or replacing the gel-based size
`selection by a bead-based, automatable, size selection (Lennon et al.
`
`22:939–946 Ó 2012, Published by Cold Spring Harbor Laboratory Press; ISSN 1088-9051/12; www.genome.org
`
`Genome Research
`www.genome.org
`
`939
`
`00001
`
`EX1016
`
`

`

`Downloaded from
`
`genome.cshlp.org
`
`Rohland and Reich
`
`
`
` on February 14, 2022 - Published by Cold Spring Harbor Laboratory Press
`
`
`
`Figure 1. Experimental workflow of the library preparation protocol for 95 samples for pooled hybrid capture.
`
`2010; Borgstrom et al. 2011). Another change is to replace a com-
`monly used commercial kit (AMPure XP kit) for SPRI-cleanup steps
`with a homemade mix. An important feature of our libraries
`compared with almost all other Illumina library preparation
`methods (Cummings et al. 2010; Mamanova et al. 2010; Meyer
`and Kircher 2010; Teer et al. 2010) is that we add a 6-bp ‘‘internal’’
`barcoded adapter to each fragment (Craig et al. 2008). These
`adapters are ligated directly to the DNA fragments, leading to
`‘‘truncated’’ libraries with 34- and 33-bp overhanging adapters at
`the end of each DNA fragment. Adapters at this stage in our library
`preparation are sufficiently short that they interfere with each
`other minimally during hybrid capture, compared with what we
`have found when long adapters are used (64 and 61 bp on either
`side, including the 6-bp internal barcode). The truncated adapter
`sites are then extended to full length after hybrid capture allowing
`the libraries to be sequenced (Fig. 2A). To assess how this strategy
`works in different-sized pools (between 14 and 95), we applied it to
`2.2 Mb of the genome of interest for prostate cancer, where it re-
`duces the capture reagent that is required by two orders of mag-
`nitude while still producing highly useful data. Sequencing these
`libraries shows that we can perform pooled target capture on at
`least 95 barcoded samples simultaneously without substantial re-
`duction in capture efficiency.
`The fact that we are using internal strategy in which barcoded
`oligonucleotides are ligated directly to fragmented DNA is a non-
`standard strategy, which deserves further discussion. First, when
`combined with indexing (introducing a second barcode via PCR
`after pooling) (Fig. 2A; Meyer and Kircher 2010), an almost un-
`limited number of samples can be pooled and sequenced in one
`lane. We are currently using this strategy in our prostate cancer study
`to test library quality and to assess the number of sequence-able
`
`molecules per library prior to equimolar pooling for hybrid cap-
`ture. Second, a potential concern of our strategy of directly ligating
`barcodes is that differences in ligation efficiency for different
`barcodes in principle could cause some barcodes to perform less
`efficiently than others. However, to date, we have used each of 138
`barcodes at least 15 times and have not found evidence of particular
`barcodes performing worse than others as measured by the number
`of sequenced molecules per library. Third, the blunt-end ligation
`used in our protocol results in a loss of 50% of the input DNA be-
`cause two different adapters have to be attached to either side. This
`is not a concern for low-coverage and small-target studies using
`input DNA amounts of 500 ng or higher but is not an ideal strategy
`for samples with less input material. Fourth, chimeras of blunted
`DNA molecules can be created during blunt-end ligation. In our
`protocol, the formation of chimeras is reduced by using adapter
`oligonucleotides in such vast excess to the sample DNA that the
`chance of ligating barcodes to the DNA is much higher than ligating
`two sample molecules (while the adapters can form dimers, these are
`removed during bead cleanup). Fifth, when using our internal
`barcodes, it is important to pool samples in each lane in such a way
`that the base composition of the barcodes is balanced, because the
`Illumina base-calling software assumes balanced nucleotide com-
`position especially during the first few cycles. This is of particular
`importance when only a few barcoded samples are being pooled. To
`prevent base-calling problems in such unbalanced pools, a PhiX li-
`brary can be spiked into the library to increase diversity.
`We performed a rough calculation breaking the cost for our
`method down into (a) reagents, (b) technician time, and (c) capital
`equipment (Table 1). The reagents and consumables cost is about $9
`per sample without taking into account discounts that would be
`available for a project that produced large numbers of libraries. The
`
`940 Genome Research
`www.genome.org
`
`00002
`
`

`

`Downloaded from
`
`genome.cshlp.org
`
`
`
` on February 14, 2022 - Published by Cold Spring Harbor Laboratory Press
`
`
`
`Libraries for highly multiplexed target capture
`
`Figure 2.
`(A) Schematic overview of the library preparation procedure using the Illumina PE adapter (internal barcode in red). After a cascade of
`enzymatic reactions and cleanup steps, enrichment PCR can be performed to complete the adapter sites for Illumina PE sequencing (Rd1 SP, Rd2 SP are PE
`sequencing primers). Alternatively, libraries can be pooled for hybrid selection (if desired), and then enrichment PCR can be performed after hybrid
`selection. To achieve an even higher magnitude of pooling for sequencing, ‘‘indexing PCR’’ can be performed instead of ‘‘enrichment PCR,’’ whereby
`unique indices (in purple) are introduced to the adapter, and a custom index sequencing primer (index-PE-sequencing-Primer) is used to read out the
`index in a separate read. Finished libraries that have all the adapters necessary to allow sequencing are marked with an X. (B) Schematic figure of ‘‘daisy-
`chaining’’ during pooled solution hybrid capture, which may explain why a large proportion of molecules are empirically observed to be off-target when
`using long adapters. Library molecules exhibiting the target sequences hybridize to biotinylated baits, but unwanted library molecules can also hybridize
`to the universal adapter sites. The adapters of our ‘‘truncated’’ libraries (including barcode: 34 and 33 bp) are about half the length of regular ‘‘long’’
`adapters (64 and 61 bases), and thus may be less prone to binding DNA fragments that do not belong to the target region.
`
`cost for technician time is $3 per sample, assuming that an in-
`dividual makes 480 libraries on five plates per week. Capital costs
`are difficult to compute (because some laboratories may already
`have the necessary equipment), but if one computes the cost of
`a Covaris LE220 instrument, a PCR machine, and an Agilent
`Bravo liquid handling platform, and divides by the cost of
`100,000 libraries produced over the 2–3-yr lifetime of these in-
`struments, this would add about $3 more to the cost per sample.
`This accounting does not include administrative overhead, space
`rental, process management, quality control on the preparation of
`reagents, bioinformatic support, data analysis, and research and
`development, all of which could add significantly to cost.
`
`Application 1: Enrichment of more than 2000
`human samples by solution hybrid capture
`
`For many applications, it is of interest to enrich a DNA sample for a
`subset of the genome; for example, in medical genetics, a candidate
`region for disease risk, or all exons. The target-enriched (captured)
`sample can then be sequenced. To perform studies with statistical
`power to detect subtle genetic effects with genome-wide signifi-
`cance, however, it is often necessary to study thousands of samples
`(Kryukov et al. 2009; Lango Allen et al. 2010), which can be pro-
`hibitively expensive given current sample preparation and target
`enrichment costs. We designed our protocol with the aim of
`allowing barcoded and pooled samples to be captured simulta-
`neously. Specifically, our libraries have internal barcodes that are
`tailored to pooled hybrid capture, whereas most other libraries
`have external barcodes in the long adapters. It has been hypoth-
`esized that hybridization experiments using libraries that already
`
`have long adapters do not work efficiently in pooled hybridiza-
`tions because a proportion of library molecules not only hybridize
`to the ‘‘baits’’ but also catch unwanted off-target molecules
`with the long adapter (‘‘daisy-chaining’’) (Mamanova et al. 2010;
`Nijman et al. 2010), thus reducing capture efficiency (Fig. 2B).
`In the Supplemental Notes (‘‘Influence of Adapter Length in
`Pooled Hybrid Capture’’), we present experiments showing that
`the number of reads mapping to the target region increased from
`29% to 73% when we shortened the adapters (Supplemental
`Table S1), providing evidence for the hypothesis that interference
`between barcoded adapters is lowered by short adapters. Our re-
`sults show empirically that short adapters improve hybridization
`efficiency.
`To investigate the empirical performance of our libraries in the
`context of target capture, we produced libraries for 189 human
`samples starting from 0.2–4.8 mg of DNA (98%<1 mg for fragmen-
`tation), prepared in two 96-well plates as in Supplemental Figure S1.
`We combined the samples into differently sized pools of libraries (14,
`28, 52, and 95) and then enriched the pooled libraries using a cus-
`tom Agilent SureSelect Target Enrichment Kit in the volume rec-
`ommended for a single sample (the target was a 2.2-Mb subset of the
`genome containing loci relevant to prostate cancer). We sequenced
`the three smaller pools together on one lane of the Genome Ana-
`lyzer II instrument (36 bp, single reads) and the 95-sample pool on
`one lane of a HiSeq2000 instrument (50 bp, paired-end reads). We
`aligned the reads to the human genome using BWA (Li and Durbin
`2009), after removing the first six bases of the first read that we used
`to identify the sample. We removed PCR duplicates using Picard’s
`(http://picard.sourceforge.net) MarkDuplicates and computed
`hybrid selection statistics with Picard’s CalculateHsMetrics. For
`
`Genome Research
`www.genome.org
`
`941
`
`00003
`
`

`

`Downloaded from
`
`genome.cshlp.org
`
`
`
` on February 14, 2022 - Published by Cold Spring Harbor Laboratory Press
`
`
`
`Rohland and Reich
`
`Table 1. Cost and time assumptions for library preparation
`
`Task
`
`Item
`
`Covaris shearing
`Cleanup
`Blunt end repair
`Barcoded adapter
`ligation
`Nick fill-in reaction
`Amplification
`Copy number
`determination
`Consumables
`
`Plate
`Beads and ethanol
`Kit
`Kit and oligonucleotides
`
`Enzyme and buffer
`Kit and oligonucleotides
`qPCR reagents or
`sequencing costa
`Plates and pipette tips
`Subtotal
`
`Technician salary
`
`Capital equipment
`
`Total assuming
`480/weekb
`Amortized over
`100,000 librariesc
`
`Price per
`sample
`
`$ 0.04
`$ 0.54
`$ 0.75
`$ 3.30
`
`$ 0.48
`$ 1.58
`$ 0.67
`
`$ 1.40
`$ 8.76
`
`$ 3.00
`
`$ 3.00
`
`Total for library
`preparation
`
`$ 14.76
`
`Sample
`processing time
`for 192 samples
`
`Technician
`hands-on time
`for 192 samples
`
`44 h
`4 h
`1.5 h
`1.2 h
`
`1 h
`1–2 h
`
`2 h
`2 h
`0.5 h
`0.5 h
`
`0.5 h
`0.5 h
`
`6 h
`
`To demonstrate that pooled target
`capture using our libraries is amenable
`to an experiment on the scale that is rel-
`evant to medical genetic association
`studies, we used the library preparation
`method to prepare 2152 DNA samples
`from one population (African-Americans)
`in the space of 2 mo. We normalized these
`samples to the lowest concentrated sam-
`ple in each pool, combined them into 15
`pools of between 138 and 144 samples,
`and enriched these 15 pools for the 2.2-
`Mb target. We sequenced the captured
`products on a HiSeq 2000 instrument
`using 75-bp paired-end reads to an aver-
`age coverage of 4.1 in nonduplicated
`reads (data not shown). The duplication
`rate of the reads was on average 72%,
`an elevation above the levels reported
`in Table 2 and Supplemental Table S2
`that we hypothesize is due to dilution to
`the lowest-complexity library within the
`pools. We were able to solve this problem
`by replacing the dilution with a cherry-
`picking approach that combines samples
`of similar complexity. We tested this ap-
`proach by pooling 81 prostate cancer li-
`braries with similar complexity (allowing
`no more than a 53 difference in molecule count per library),
`resulting in a duplication rate of 24% on average at 73 coverage
`(Supplemental Table S2e).
`The experiment was highly sensitive for detecting polymor-
`phisms in the targeted regions. After restricting to sites with at least
`one-fourth of the average coverage, we discovered 35,211 poly-
`morphisms at high confidence (10,000:1 probability of being real
`based on their quality score from BWA). This is more than double
`the 16,457 sites discovered by the 1000 Genomes Project in 167
`African ancestry samples over the same nucleotides (February 2011
`data release) (The 1000 Genomes Project Consortium 2010). Ex-
`ploring this in more detail, we found that we rediscovered 99.7% of
`sites in the 1000 Genomes Project with minor allele frequency >5%
`and 83% of 1000 Genomes Project sites with lower frequency in the
`African samples. As a second measure of the quality of our data, we
`
`aqPCR for two measurements per sample, or sequencing one lane SR36 and indexing read, divided by
`2152 libraries.
`b$3/sample for personnel time (assuming salary and benefits of $70,000 per year and processing five 96-
`well plates/week).
`c$3/sample for capital equipment (assuming purchase of a Covaris LE220 instrument, a PCR machine,
`and an Agilent Bravo liquid handling platform, and dividing over 100,000 libraries).
`
`the 95-sample pool (unnormalized before hybrid capture), f2 = 93%
`of samples had a mean target coverage of within a factor of 2 of the
`median, f1.5 = 67% within a factor of 1.5 of the median, and the
`coefficient of variation (standard deviation divided by mean cov-
`erage) was CV = 0.40. For the three smaller pools where normali-
`zation was performed, coverage was in general more uniform: For
`the pool of 14, f2 = 93%, f1.5 = 86%, CV = 0.66; for the pool of 28, f2 =
`100%, f1.5 = 96%, CV = 0.19; and for the pool of 52, f2 = 100%, f1.5 =
`94%, CV = 0.19 (Supplemental Table S2). In the 95-sample experi-
`ment, the percentage of selected bases, defined as ‘‘on bait’’ or
`within 250 bp of either side of the baits, was 70%–79% across
`samples (Table 2; Supplemental Table S2), comparable to the liter-
`ature for single-sample selections (Supplemental Table S3). Results
`on the 95-sample pool are as good as the 14-, 28-, and 52-sample
`pools.
`
`Table 2. Sequencing results
`
`Application
`
`Number
`of
`libraries
`
`Input
`DNA
`(mg)
`
`Normalization
`strategy
`
`PF reads
`per
`library
`
`% Reads aligning
`to reference
`genome
`
`% Duplicated
`reads
`(removed)
`
`Mean target
`coverage
`per librarya
`
`% Selected
`basesb
`
`% Target
`with 23
`coveragea
`
`Human hybrid selectionc
`Human hybrid selectionc
`Human hybrid selectionc
`Human hybrid selectiond
`Human hybrid selectiond
`Human whole-genome
`shotgune
`Microbial sequencingd
`
`14
`28
`52
`95
`81
`40
`
`12
`
`2.8 3 105
`0.6–0.9 Dilution
`3.3 3 105
`0.2–0.9 Dilution
`2.7 3 105
`0.3–0.9 Dilution
`1 3 106
`0.2–4.8 Unnormalized
`0.6–2.6 Cherry picking 5.6 3 105
`7.1 3 107
`0.75
`
`1
`
`7.2 3 106
`
`73
`72
`74
`89
`92
`95
`
`97
`
`53.6
`56.4
`51.1
`37.5
`24.4
`14.4
`
`1
`
`0.9
`1.1
`1.1
`7.4
`7.1
`5.4
`
`147
`
`78
`76
`78
`74
`92
`n/a
`
`n/a
`
`23
`31
`29
`79
`87
`n/a
`
`n/a
`
`aTarget for the hybrid selection experiment is defined as the regions where baits were designed.
`b‘‘Selected bases’’ is defined as in Picard as 250 bp on either side of the bait (target).
`c36 cycles of single-read sequencing on GAII.
`d50 cycles of paired-end sequencing on HiSeq2000.
`e100 cycles of paired-end sequencing on HiSeq2000; four libraries were prepared for each of 10 samples.
`
`942 Genome Research
`www.genome.org
`
`00004
`
`

`

`Downloaded from
`
`genome.cshlp.org
`
`
`
` on February 14, 2022 - Published by Cold Spring Harbor Laboratory Press
`
`
`
`compared n = 1642 African-American samples that had previously
`been genotyped on an Illumina 1M array at 1367 SNPs that over-
`lapped between that array and the 2.2-Mb target region of the
`capture experiments. We found that 99.77% of the mapped de-
`duplicated reads are consistent with the ‘‘gold standard’’ results
`from genotyping. As a third measure of data quality, we checked for
`a potential reference bias by counting the reads matching the ref-
`erence and variant allele at the 1367 SNPs where we knew the true
`genotypes. As shown in Supplemental Figure S2, there is a slight bias
`(Nref/Ntot = 1,289,080/2,537,488 = 50.8%) for the reference allele,
`which is sufficiently small that we do not expect it to cause a major
`problem for most applications such as identification of heterozy-
`gous sites.
`
`Application 2: Whole-genome sequencing of 40
`human libraries to 53 coverage
`
`Whole-genome shotgun sequencing (WGS) of mammalian ge-
`nomes to high coverage (e.g., 303) is still a process that is dominated
`by sequencing costs. However, lighter sequencing is of interest for
`some applications. For example, Genomewide Association Studies
`(GWAS), which have discovered more than 1300 associations to
`human phenotypes (Manolio 2010), cost hundreds of dollars per
`sample on SNP arrays, which is less than commercial costs of library
`preparation, and hence sequence-based GWAS are not economical.
`However, the situation would change if library production costs
`were lower. If libraries were inexpensive, sequencing the genome to
`light coverage followed by imputing missing data using a reference
`panel of more deeply sequenced or genotyped samples, in theory
`would allow more cost-effective GWAS (Li et al. 2011). With suffi-
`ciently low library production costs, sequencing may begin to
`compete seriously with SNP array–based analysis for medical genetic
`association studies, as is already occurring in studies of gene ex-
`pression analysis, where RNA-seq is in the process of replacing array-
`based methods (Majewski and Pastinen 2010).
`To test if our method can produce libraries appropriate for
`whole-genome sequencing, we prepared 40 libraries using an earlier
`version of our protocol that used microTUBES for shearing instead of
`plates and a slightly different enrichment PCR procedure (Supple-
`mental Fig. S3). (A more up-to-date protocol, which involves shear-
`ing in plates and which we used to produce libraries for the prostate
`cancer study, further reduces costs by about $5 per sample.) Table 2
`and Supplemental Table S4 show the results of sequencing these li-
`braries to an average of 5.43 coverage using 100-bp paired-end reads
`on 58 lanes on Illumina HiSeq 2000 instruments. A high proportion
`(95%) of the reads align to the human reference genome (hg19) using
`BWA (Li and Durbin 2009), and duplicates were removed. We found
`that 99.86% of the mapped reads are concordant with the ‘‘gold
`standard’’ SNP array data previously collected on these samples (Li
`et al. 2008) (sequences with quality $30 for the 40 libraries were
`compared at 585,481 SNPs). Thus, we have demonstrated that our
`protocol can produce libraries that are useful for low-pass whole-
`genome human sequencing.
`
`Application 3: Sequencing of 12 Escherichia coli strains
`to 1503 coverage
`
`An important application of high-throughput sequencing is the
`study of microbial genomes, for example, in an epidemiological
`context where it is valuable to study strains from many patients to
`study the spread of an epidemic, or in the same individual to study
`
`Libraries for highly multiplexed target capture
`
`the evolution of an infection. Microbial genomes are small so that
`the required amount of sequencing per sample can be small, and
`thus the limiting cost is often sample preparation. To explore the
`utility of our library preparation protocol for microbial sequenc-
`ing, we produced libraries for 12 E. coli strains for a project led by
`M. Lajoie, F. Isaacs, and G. Church (whom we thank for allowing us
`to report the data) (Isaacs et al. 2011). We produced these libraries
`as a single row on a 96-well plate with an input DNA amount of
`1 mg together with human libraries that we were producing for
`another study following the protocol in Supplemental Figure S4.
`Table 2 and Supplemental Table S5 report the results of the se-
`quencing of these 12 libraries on a single lane of a HiSeq 2000 (50-
`bp paired-end reads). We analyzed the data after separating
`the libraries by sample using internal barcodes and mapping to
`the E. coli reference (strain K12 substrain MG1655, Refseq
`NC_000913) using BWA (Li and Durbin 2009). Overall, 97% of
`reads mapped, with an average of 147-fold coverage and 1% du-
`plicated reads.
`
`Discussion
`
`We have reported a high-throughput library preparation method
`for next-generation sequencing, which has been designed to allow
`an academic laboratory to generate thousands of barcoded libraries
`at a cost that is one to two orders of magnitude less than the
`commercial cost of library preparation. These libraries are appro-
`priate for whole-genome sequencing of large and small genomes. A
`particularly important feature of these libraries is that they are ef-
`fective for pooling approximately a hundred samples together and
`enriching them for a subset of the genome of interest. We have
`proven that the method is practical at a scale that is relevant to
`medical genetics by generating more than 2000 libraries for a
`prostate cancer study, enriching them for more than 2 Mb of in-
`terest, and obtaining sequencing data that are concordant with
`previously reported genotype calls.
`From an engineering point of view, our method was designed
`with a different set of goals than have driven most previous library
`preparation methods. In most methods, the emphasis has been on
`producing libraries with maximal complexity (as measured by the
`number of unique molecules) and length uniformity (as measured
`by the tightness of the distribution of insert sizes) given the large
`amount of sequencing that was planned for each library. Our goal
`is different: to increase throughput and decrease reagent cost,
`while building libraries that are appropriate for pooled target
`capture. In this study, we empirically show that the human li-
`braries produced by our method are complex enough that when
`shotgun-sequenced to a coverage of around 53, they give dupli-
`cation rates of 9%–20%. This duplication rate is somewhat higher
`than some published protocols, and the problem of duplication
`becomes greater as coverage increases, so that for deep-sequencing
`studies (e.g., whole-genome sequencing at 303) in which thou-
`sands of dollars are invested per sample, it may be more econom-
`ical to use a more expensive library preparation protocol that
`minimizes duplication rates. One reason for an increased dupli-
`cation rate in our libraries is our distribution of fragment insert
`sizes. Because size selection with beads is not as tight as gel-based
`size selection, fragment insert sizes of the libraries produced with
`our protocol are variable. Longer fragments are more prone to
`duplicated reads (‘‘optical duplicates’’), in which the Illumina
`software identifies one cluster as two adjacent clusters. Another
`reason for an increased duplication rate is the low input DNA
`amount per ligation reaction (0.75 mg for each of the four ligation
`
`Genome Research
`www.genome.org
`
`943
`
`00005
`
`

`

`Downloaded from
`
`genome.cshlp.org
`
`Rohland and Reich
`
`
`
` on February 14, 2022 - Published by Cold Spring Harbor Laboratory Press
`
`
`
`reactions per sample), much less than the recommended 3–5 mg for
`standard whole-genome sequencing library protocols; we also lose
`complexity because 50% of molecules are lost during blunt-end li-
`gation due to wrong adapter combinations. Coverages of 10-fold or
`less, a level where our libraries have reasonable duplication rates,
`have been shown to be highly effective for SNP discovery and
`genotype imputation (The 1000 Genomes Project Consortium
`2010), and thus our libraries are valuable for most medical genetic
`applications. The high duplication rate for our prostate cancer
`target capture enrichment study (72% at about 43 coverage) arose
`from the normalization strategy of diluting to the lowest complex
`library within each pool. We were able to lower the duplication
`rate to 24% at about 73 coverage when we pooled similarly
`complex libraries and hope to be able to lower this even further in
`the future.
`The method we have presented is tailored to paired-end se-
`quencing using Illumina technology but is easy to adapt to mul-
`tiplexing (we recently switched to the Multiplexing-P7 adapter)
`and to other technologies, for example, 454 Life Sciences (Roche),
`Applied Biosystems SOLiD (Life Technologies), and Ion Torrent
`(Life Technologies). While these technologies are different at the
`detection stage, they are similar in sample preparation, in that
`technology-specific adapters are attached to DNA fragments, and
`the fragments are subjected to enrichment PCR to complete the
`adapter sites, allowing clonal amplification of the libraries and
`subsequent sequencing-by-synthesis. Thus, a method for one
`technology can be modified for use with the others. Although we
`only used the Agilent SureSelect platform for hybrid selections,
`we expect that similar hybridization-based target enrichment sys-
`tems, such as the Illumina TruSeq Enrichment kits (Clark et al.
`2011), the Roche/NimbleGen SeqCap EZ Hybridization kits, and
`array-based hybridization (Hodges et al. 2007), would enrich
`multiplexed samples as efficiently as the Agilent system if the li-
`braries are prepared with short adapters.
`There are several potential improvements to our method,
`which should make it possible to produce libraries at even higher
`throughput, and to further improve library quality. A bottleneck at
`present is the machine time required for sample shearing. On the
`Covaris E210 instrument, 21 h are required to shear to a mean
`insert size of 200–300 bp for a plate of 96 samples (although this
`takes negligible technician time), and thus two instruments would
`be required to produce enough sheared samples for a full-time
`technician. However, this bottleneck could be eliminated by a re-
`cently released instrument, the Covaris LE220, which is able to shear
`eight samples simultaneously. The number of samples that can be
`pooled per lane is 159 with our 6-mer 59-barcodes, but may not be
`enough if, for example, the target size is small and the desired cov-
`erage is low. When combining the barcoding strategy with indexing
`via PCR, a much greater number of samples can be pooled. Another
`way to increase the number of samples that can be pooled is to either
`extend the number of barcode nucleotides or to ligate two different
`adapters on either side of the molecule. Further improvements to
`the protocol and quality-control steps are important directions,
`which should improve the usefulness of these libraries even further.
`
`Methods
`
`We discuss each of the steps of t