Kircher et al. BMC Genomics 2011, 12:382
`http://www.biomedcentral.com/1471-2164/12/382
`
`CO R R E SP O N D E N C E
`Open Access
`Addressing challenges in the production and
`analysis of illumina sequencing data
`Martin Kircher1, Patricia Heyn2 and Janet Kelso1*
`
`Abstract
`Advances in DNA sequencing technologies have made it possible to generate large amounts of sequence data
`very rapidly and at substantially lower cost than capillary sequencing. These new technologies have specific
`characteristics and limitations that require either consideration during project design, or which must be addressed
`during data analysis. Specialist skills, both at the laboratory and the computational stages of project design and
`analysis, are crucial to the generation of high quality data from these new platforms. The Illumina sequencers
`(including the Genome Analyzers I/II/IIe/IIx and the new HiScan and HiSeq) represent a widely used platform
`providing parallel readout of several hundred million immobilized sequences using fluorescent-dye reversible-
`terminator chemistry. Sequencing library quality, sample handling, instrument settings and sequencing chemistry
`have a strong impact on sequencing run quality. The presence of adapter chimeras and adapter sequences at the
`end of short-insert molecules, as well as increased error rates and short read lengths complicate many
`computational analyses. We discuss here some of the factors that influence the frequency and severity of these
`problems and provide solutions for circumventing these. Further, we present a set of general principles for good
`analysis practice that enable problems with sequencing runs to be identified and dealt with.
`
`Background
`Recent advances in DNA sequencing have changed the
`field of genomics making it possible to generate giga-
`bases of genome and transcriptome sequence data at
`substantially lower cost than was possible just ten years
`ago http://www.genome.gov/sequencingcosts/. The rela-
`tive affordability of these high-throughput sequencers
`and the potential to generate large amounts of sequence
`data at lower cost means that scientists outside of tradi-
`tional sequencing facilities are now faced with the chal-
`lenges associated with design of large-scale projects and
`analysis of the data generated. This poses significant
`challenges for many groups since the inherent limita-
`tions of these platforms, and particular artifacts asso-
`ciated with sequences generated on these platforms,
`need to be understood and dealt with at various stages
`of the project including planning, sample preparation,
`run processing and downstream analyses.
`We present here an analysis of challenges encountered
`in using the Illumina sequencing instruments. A
`
`* Correspondence: kelso@eva.mpg.de
`1Max Planck Institute for Evolutionary Anthropology, Department of
`Evolutionary Genetics Deutscher Platz 6 04103 Leipzig, Germany
`Full list of author information is available at the end of the article
`
`thorough description of the Solexa/Illumina sequencing
`technology as well as a comparison to other platforms is
`available elsewhere [1-6]. We revisit here only the
`aspects relevant for project design and data analysis.
`Figure 1 shows the steps from the DNA sample prepara-
`tion to sequence read outs with quality scores. Indepen-
`dent of the actual application, Solexa/Illumina sequencing
`requires that the molecules to be determined are con-
`verted into special sequencing libraries. This is achieved
`by adding specific adapter sequences on both ends of frag-
`mented DNA molecules; allowing molecules to be ampli-
`fied,
`immobilized and primed for sequencing. For
`sequencing, typically double-stranded DNA libraries are
`provided and melted using sodium hydroxide to obtain
`single stranded molecules. These are then immobilized
`(hybridization and amplification on a solid phase; bridge
`amplification [1,7]) in one or more channels of an 8-chan-
`nel flow cell. The immobilization and subsequent bridge-
`amplification creates randomly scattered clusters, consist-
`ing of more than one thousand copies of the original
`sequence in very close proximity to each other. After the
`bridge amplification step, cluster molecules are largely
`double stranded. One of the strands then has to be
`removed to obtain single stranded, identically oriented
`
`© 2011 Kircher et al; licensee BioMed Central Ltd. This is an Open Access article distributed under the terms of the Creative Commons
`Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in
`any medium, provided the original work is properly cited.
`
`GUARDANT – EXHIBIT 2020
`GUARDANT – EXHIBIT 2020(cid:0)(cid:3)
`
`Twinstrand Biosciences, Inc. v. Guardant Health, Inc. Twinstrand Biosciences, Inc. v. Guardant Health, Inc.
`IPR2022-01400(cid:0)(cid:3)
`IPR2022-01400
`
`

`

`Kircher et al. BMC Genomics 2011, 12:382
`http://www.biomedcentral.com/1471-2164/12/382
`
`Page 2 of 14
`
`Sample
`
`(a) Library preparation
`
`Flowcell preparation
`
`(b) Immobilization
`and solid phase
`amplification
`
`TTACTATGCCGCTGGTGGCTCTAGATGTGAGAAAGGGATGTGCTGCGAGAAGGCTAGANNNNNNNNNNNNNNNNNNCTAGCCTTCTCGAGCATACGGCAGAAGACGAAC
`
`ACACTCTTTCCCTACACGACGCTCTTCCGATCT
`
`TTACTATGCCGCTGGTGGCT
`
`TCGTATGCCGTCTTCTGTTG
`
`TCGTATGCCGTCTTCTGTTG
`
`TTACTATGCCGCTGGTGGCT
`
`TCGTATGCCGTCTTCTGTTG
`
`(c) Linearization, 3' blocking,
`sequence primer hybridization
`
`TTACTATGCCGCTGGTGGCTCTAGATGTGAGAAAGGGATGTGCTGCGAGAAGGCTAGANNNNNNNNNNNNNNNNNNCTAGCCTTCTCGAGCATACGGCAGAAGACGAAC
`
`ACACTCTTTCCCTACACGACGCTCTTCCGATCT
`
`TTACTATGCCGCTGGTGGCTCTAGATGTGAGAAAGGGATGTGCTGCGAGAAGGCTAGANNNNNNNNNNNNNNNNNNCTAGCCTTCTCGAGCATACGGCAGAAGACGAAC
`
`ACACTCTTTCCCTACACGACGCTCTTCCGATCT
`
`TTACTATGCCGCTGGTGGCT
`
`TCGTATGCCGTCTTCTGTTG
`
`TCGTATGCCGTCTTCTGTTG
`
`TTACTATGCCGCTGGTGGCT
`
`TCGTATGCCGTCTTCTGTTG
`
`TTACTATGCCGCTGGTGGCTCTAGATGTGAGAAAGGGATGTGCTGCGAGAAGGCTAGANNNNNNNNNNNNNNNNNNCTAGCCTTCTCGAGCATACGGCAGAAGACGAAC
`
`ACACTCTTTCCCTACACGACGCTCTTCCGATCT
`
`(d) Reversible terminator chemistry
`and read out of incorporated dyes
`
`A
`ACACGACGCTCTTCCGATCT
`TGTGCTGCGAGAAGGCTAGATCATGGCTGAA...
`
`A
`ACACGACGCTCTTCCGATCT
`TGTGCTGCGAGAAGGCTAGATCATGGCTGAA...
`
`T
`ACACGACGCTCTTCCGATCT
`TGTGCTGCGAGAAGGCTAGAACGTTGCAGGAGCATTGCACTAGCCT
`
`T
`ACACGACGCTCTTCCGATCT
`TGTGCTGCGAGAAGGCTAGAACGTTGCAGGAGCATTGCACTAGCCT
`
`C
`ACACGACGCTCTTCCGATCT
`TGTGCTGCGAGAAGGCTAGAGACAGGCGATT...
`
`C
`ACACGACGCTCTTCCGATCT
`TGTGCTGCGAGAAGGCTAGAGACAGGCGATT...
`
`G
`ACACGACGCTCTTCCGATCT
`TGTGCTGCGAGAAGGCTAGACATAGCGAGGA...
`
`filter A
`
`A
`
`G
`ACACGACGCTCTTCCGATCT
`TGTGCTGCGAGAAGGCTAGACATAGCGAGGA...
`
`filter C
`
`C
`
`A
`ACACGACGCTCTTCCGATCT
`TGTGCTGCGAGAAGGCTAGATCATGGCTGAA...
`
`A
`ACACGACGCTCTTCCGATCT
`TGTGCTGCGAGAAGGCTAGATCATGGCTGAA...
`
`T
`ACACGACGCTCTTCCGATCT
`TGTGCTGCGAGAAGGCTAGAACGTTGCAGGAGCATTGCACTAGCCTT
`
`T
`ACACGACGCTCTTCCGATCT
`TGTGCTGCGAGAAGGCTAGAACGTTGCAGGAGCATTGCACTAGCCTT
`
`C
`ACACGACGCTCTTCCGATCT
`TGTGCTGCGAGAAGGCTAGAGACAGGCGATT...
`
`G
`
`C
`ACACGACGCTCTTCCGATCT
`TGTGCTGCGAGAAGGCTAGAGACAGGCGATT...
`
`T
`
`G
`ACACGACGCTCTTCCGATCT
`TGTGCTGCGAGAAGGCTAGACATAGCGAGGA...
`
`G
`ACACGACGCTCTTCCGATCT
`TGTGCTGCGAGAAGGCTAGACATAGCGAGGA...
`
`filter G
`
`filter T
`
`(e) Image registration and intensity extraction
`
`(f) Base calling and quality scoring
`
`Cycle 1
`
`Cycle 2
`
`Cycle 3
`
`A
`
`A
`
`A
`
`...
`@SOLEXA-GA03_0001_PEi_SG:5:1:1033:5267
`AGACAGACACAGAGNAAGACCCAGTCCGCCACACAGGCAAACTCA
`+SOLEXA-GA03_0001_PEi_SG:5:1:1033:5267
`4--'-(/.23/044!51/+//.400/-/1-62/.6021834///6
`...
`
`Figure 1 Illumina sample preparation and sequencing. Illumina sequencing requires that a DNA sample (a) is converted into special
`sequencing libraries. This can be achieved by shearing DNA to a designated size and adding specific adapter sequences on both ends of the
`DNA molecules (b). These adapters allow molecules to be amplified and immobilized in one or more channels of an 8-channel flow cell (c).
`Immobilization and solid-phase amplification create randomly scattered clusters, consisting of a few thousand copies of the original molecule in
`very close proximity to each other. One of the DNA strands is removed to obtain single stranded, identically oriented copies, 3’ ends of the DNA
`are blocked and a sequencing primer hybridized on the adapter sequences. Afterwards, the reversible terminator chemistry is performed (d).
`Here, four differently labeled nucleotides are provided and used for extension of the primers by DNA polymerases. The polymerase reaction
`terminates after the first base incorporation since the nucleotides used are not only labeled, but also 3’-blocked. After washing away free
`nucleotides, the nucleotides incorporated are readout by piece-wise imaging of the flow cell. Then, the terminator and fluorophore are removed
`and another incorporation cycle started. The four images are overlaid (registered) and light intensities extracted for each cluster and cycle using
`a cluster position template obtained from the first instrument cycles (e). Resulting intensity files serve as input for base calling, the conversion of
`intensity values into bases and quality scores (f).
`
`copies of the starting molecule. This is achieved by selec-
`tive cleavage of base modifications of oligonucleotides on
`the flowcell. After the free 3’ ends of the DNA have been
`blocked, the copies can be sequenced by hybridizing a
`sequencing primer onto the adapter sequences and start-
`ing the reversible terminator chemistry. Here, four differ-
`ently labeled nucleotides are provided and used for
`extension of the sequencing primers by DNA polymerases.
`The DNA polymerase reaction terminates after the first
`base incorporation since the nucleotides used are not only
`labeled, but also 3’-blocked (i.e. they carry a terminator
`
`group at the third carbon atom of the sugar, which pre-
`vents further extension). After free nucleotides are washed
`away, the nucleotides being incorporated are read by cap-
`turing the light signal of the fluorophore labels after laser
`excitation. Imaging of the flow cell is carried out in so-
`called tiles which are the units in which the flowcell is
`imaged and data processed. The terminator and fluoro-
`phore are then removed and another incorporation cycle
`started [1].
`Initially, the number of sequencing cycles, and thereby
`the length of the sequence reads, was limited to 26
`
`

`

`Kircher et al. BMC Genomics 2011, 12:382
`http://www.biomedcentral.com/1471-2164/12/382
`
`Page 3 of 14
`
`cycles because of steeply increasing sequencing error.
`Between 2008 and 2010 there were several technical
`updates to the Genome Analyzer (GA) platform includ-
`ing improvements in mechanics, chemistry and software.
`Even though sequencing error still increases with each
`cycle, up to 150 sequencing cycles are currently per-
`formed with reasonable error profiles (average error
`below 1%, and up to 10% in the final cycles). Further,
`flow cell cluster densities were increased from 5-12 mil-
`lion clusters to about 35-60 million clusters per lane
`(and twice that for HiSeq instruments; where clusters
`on the top and bottom of the flow cell are read). A
`technical update made sequencing of the reverse strand
`of each molecule possible. Using this “paired-end
`sequencing” approach for determining the reverse
`strand, doubles the amount of sequence data generated.
`Known insert size, and thereby a known distance separ-
`ating the paired reads obtained, provides additional
`information for later assembly or mapping [8]. This
`technical update of the Genome Analyzer in 2008, the
`Paired End (PE) module, also allowed the hybridization
`of further sequencing primers in the same strand orien-
`tation, making it possible to sequence a sample index (i.
`e. barcode) as part of the ligated adapter [9,10]. Such an
`index read allows for multiple samples to be sequenced
`in one lane (multiplexing). These can later be computa-
`tionally separated based on the sequence of their index.
`During progression of the sequencing run, or when
`images for all cycles have been collected (depending on
`the setup and version), the four images captured per tile
`are overlaid (registered) and light
`intensities are
`extracted for each cluster and cycle [1]. The resulting
`cluster position template is then aligned with images of
`all cycles and the intensities minus the surrounding
`background in the four different images extracted.
`Resulting intensity files serve as input for base calling -
`the conversion of intensity values into bases. Base call-
`ing on the Illumina platform is complicated by at least
`two effects: (1) a strong correlation of the A and C
`intensities as well as of the G and T intensities due to
`similar emission spectra of the fluorophores used and
`their limited separation by optical filters, and (2) depen-
`dence of the signal for a specific cycle on the signal of
`the cycles before and after, known as phasing and pre-
`phasing, respectively. Phasing and pre-phasing describe
`the loss of synchrony in the readout of the sequence
`copies of a cluster. Phasing is caused by incomplete
`removal of the 3’ terminators and fluorophores as well
`as sequences in the cluster missing an incorporation
`cycle. Pre-phasing is caused by the incorporation of
`nucleotides without effective 3’-blocking. The proportion
`of sequences in each cluster which are affected by phas-
`ing and pre-phasing increases with cycle number; ham-
`pering correct base identification [11-14].
`
`From this whole process, the Illumina user typically
`obtains sequences and per base quality scores. The set
`of sequences for each lane is usually quality filtered and
`the user gets a summary report for judging run quality.
`Finally, the Illumina CASAVA package provides addi-
`tional tools and an interface to the visualization routines
`in Illumina’s Genome Studio. Different commercial as
`well as free programs are available that replace some
`parts of the processing such as image analysis [15], base
`calling [11-15], quality assessment (e.g. TileQC [16] or
`FastQC [17]), mapping [18-21], as well as downstream
`data analysis and processing [8,22-25]. There is a large
`community of users and developers for the Illumina
`platform; the http://seqanswers.com website is an excel-
`lent resource when starting to explore the variety of
`programs available for analyzing the data generated.
`
`Results
`We present each stage of a sequencing project from the
`generation of sequencing libraries to the base-calling of
`the sequences. For each step we discuss the potential
`effects on data analysis and final data quality. Where
`possible we offer suggestions and guidelines on how to
`avoid specific artifacts that arise during sequencing.
`
`Sequencing libraries, minimum insert size and adapter
`artifacts
`The most important requirement for a DNA library to
`be sequenced on the Illumina platform is the presence
`of specific outer adapter sequences, complementary to
`the oligonucleotides on the flow cell used for cluster
`generation, the so-called “grafting sequences”. As differ-
`ent sequencing primers can be used (see below), the rest
`of the library design is very flexible and various library
`preparation protocols with partially distinct adapter
`sequences are used for specific applications. Library
`adapters can be added by single strand ligation (e.g. Illu-
`mina small RNA protocol), double strand blunt-end
`ligation (e.g. for a multiplex protocol [10]), double
`strand overhang ligation (e.g. A-overhang for Illumina
`genomic library protocols, and restriction enzyme over-
`hangs in the Illumina DGE protocols), or by extension
`from overhanging primers (e.g. multiplex PCR or mole-
`cular
`inversion probes
`[26,27]). Each of
`these
`approaches has a different susceptibility to the creation
`of library adapter dimers, chimeric sequences and other
`library artifacts. Each therefore requires a different
`approach to enrich for only those molecules with cor-
`rectly added adapters, and to remove short/no insert
`molecules and molecules which are too long (> 800nt)
`from the library before sequencing. While short-insert
`molecules will, as described below, directly impact data
`analysis, longer molecules will perform differently in
`flow cell generation and generate more wide-spread and
`
`

`

`Kircher et al. BMC Genomics 2011, 12:382
`http://www.biomedcentral.com/1471-2164/12/382
`
`Page 4 of 14
`
`less dense clusters. If not accounted for by modified
`cluster generation protocols, these will result in lower
`quality reads.
`Failure to perform an enrichment during library pre-
`paration has two potential effects: (i) the artifact
`sequences may have a negative impact on the image
`analysis and base-calling which are both challenged by
`an overrepresentation of one sequence population (see
`below) and (ii) sequencing of large numbers of such
`artifacts is uneconomical and lowers the potential num-
`ber of informative sequences that can be generated per
`run. Libraries prepared from small amounts of input
`material tend to suffer from a higher fraction of library
`artifacts due to the relative abundance of adapter oligo-
`nucleotides compared to insert molecules.
`It is possible to computationally post-process sequen-
`cing data in cases where enrichment has not been
`
`performed. Figure 2, exemplifies for the Illumina NlaIII
`DGE protocol (a protocol for digital gene expression tag
`profiling) that adapter chimeras might be created which
`are of comparable length as the targeted library mole-
`cules and thus may not be removed by selecting a speci-
`fic library insert-size (e.g. by gel length selection, silica
`column purification or Solid Phase Reversible Immobili-
`zation (SPRI) purification [28]). In this case, a program
`like TagDust [29] can be used with the original adapter
`and primer oligonucleotide sequences to identify such
`artifacts in a library (Figure 2B). This program can be
`either used to directly remove these sequences or, for a
`representative lane, its results can be clustered and the
`most frequent ones used with other software tools.
`Inappropriate size selection during library preparation
`may also complicate analysis due to partial sequencing
`of the adaptor at the sequence ends. Thus, when
`
`A: Digital gene expression / SAGE experiment
`
`B: Most frequent TagDust filtered sequences
`
`mRNA enrichment with polyT beads
`
`NlaIII
`Poly-A-Tail
`recognition site
`NNN...NNNCATGXXXXXXXXXXXXXXXNNN...NNNAAAAAAAAAAA
`17nt sequence
`mRNA
`NNN...NNNGTACXXXXXXXXXXXXXXXNNN...NNNTTTTTTTTTTT
`
`5'
`
`3'
`
`cDNA 1st and 2nd strand synthesis
`NNN...NNNCATGXXXXXXXXXXXXXXXNNN...NNNAAAAAAAAAAA
`NNN...NNNGTACXXXXXXXXXXXXXXXNNN...NNNTTTTTTTTTTT
`
`1st Digestion
`NlaIII
`NNN...NNNCATGXXXXXXXXXXXXXXXNNN...NNNAAAAAAAAAAA
`NNN...NNNGTACXXXXXXXXXXXXXXXNNN...NNNTTTTTTTTTTT
`
`Ligation GEX adapter 1
`...TCTACAGTCCGACATGXXXXXXXXXXXXXXXNNN...NNNAAAAAAAAAAA
`...AGATGTCAGGCTGTACXXXXXXXXXXXXXXXNNN...NNNTTTTTTTTTTT
`
`2nd Digestion
`
`MmeI
`recognition site
`...TCTACAGTCCGACATGXXXXXXXXXXXXXXXNNN...NNNAAAAAAAAAAA
`...AGATGTCAGGCTGTACXXXXXXXXXXXXXXXNNN...NNNTTTTTTTTTTT
`
`Bead
`
`Bead
`
`Bead
`
`Bead
`
`Bead
`
`Ligation GEX adapter 2
`
`Sequencing primer site
`...CAGAGTTCTACAGTCCGACATGXXXXXXXXXXXXXXXTCGTATGCCGTCTTCTGCTTG
`CAAGTCTCAAGATGTCAGGCTGTACXXXXXXXXXXXXXNNAGCATACGGCAGAAGACGAAC
`
`GEX Adapter 2.1
`TCGTATGCCGTCTTCTGCTTG
`TCGTATGCCGTCTTCTGCTTG
`35191 GTATGCCGTCTTCTGCTT
`35191 GTATGCCGTCTTCTGCTT
`4733 AA-TCGTATGCCGTCTTCT
`4733 AA-TCGTATGCCGTCTTCT
`3109 TCGGAC-TCGTATGCCGTC
`3109 TCGGAC-TCGTATGCCGTC
`2963 G-TCGTATGCCGTCTTCTG
`2963 G-TCGTATGCCGTCTTCTG
`2875 TCGGACTGTAGAATCGTA
`2875 TCGGACTGTAGAATCGTA
`2818 ATGGC-TCGTATGCCGTCT
`2818 ATGGC-TCGTATGCCGTCT
`2339 AGGAG-TCGTATGCCGTCT
`2339 AGGAG-TCGTATGCCGTCT
`2307 TC-TCGTATGCCGTCTTCT
`2307 TC-TCGTATGCCGTCTTCT
`2108 TCGGACTGTAGAACTCTT
`2108 TCGGACTGTAGAACTCTT
`1936 A-TCGTATGCCGTCTTCTG
`1936 A-TCGTATGCCGTCTTCTG
`1886 AGGAGT-TCGTATGCCGTC
`1886 AGGAGT-TCGTATGCCGTC
`1880 CGTATGCCGTCTTCTGCT
`1880 CGTATGCCGTCTTCTGCT
`GTATGCCGTCTTCTTCTT
`1667
`GTATGCCGTCTTCTTCTT
`1667
`1527 AG-TCGTATGCCGTCTTCT
`1527 AG-TCGTATGCCGTCTTCT
`1509 CCAG-TCGTATGCCGTCTT
`1509 CCAG-TCGTATGCCGTCTT
`1366 GTGA-TCGTATGCCGTCTT
`1366 GTGA-TCGTATGCCGTCTT
`1238 GTG-TCGTATGCCGTCTTC
`1238 GTG-TCGTATGCCGTCTTC
`1209 TCGGACTGTAGA-TCGTAT
`1209 TCGGACTGTAGA-TCGTAT
`TCGGACTGTAGAACTCTGAAC
`TCGGACTGTAGAACTCTGAAC
`GEX Adapter 1.2
`
`Matches human NlaIII restriction
`site chr20:+1868416-1868436
`1907 GTTTCAGGAGTTTATTTT
`1907 GTTTCAGGAGTTTATTTT
`
`GEX Adapter 1.1
`ACAGGTTCAGAGTTCTACAGTCCGACATG
`ACAGGTTCAGAGTTCTACAGTCCGACATG
`1268 GCCACCCTCTACAG-CCGA
`1268 GCCACCCTCTACAG-CCGA
`
`Figure 2 Adaptors and adaptor chimeras are a common sources of sequence artifacts. Specific outer adapter sequences, complementary
`to the grafting sequences on the flow cell are essentially the only requirement for sequencing a DNA library on the Genome Analyzer platform.
`As different sequencing primers can be used, library design is very flexible and various protocols with partially distinct adapter sequences have
`been established. The Illumina NlaIII DGE tag protocol illustrated here (a protocol for digital gene expression tag profiling) uses short adapters
`which are not compatible with paired end sequencing and are added by overhang ligation (A). For this protocol the majority of adapter dimers
`are removed by a gel excision step after library preparation. However, the protocol may also create adapter chimeras with a length comparable
`to the targeted library molecules. The resulting chimera sequences also show the sequences required for cluster generation as well as the
`necessary priming site, causing them to be sequenced together with the real DGE tags. A program like TagDust [29] can be used with the
`original adapter and primer oligonucleotide sequences to identify such artifacts (B). Shown are the twenty most frequent identified artifacts from
`one lane with human DGE tags, as well as the oligosequences they might be based on. One of the 20 sequences seems to be a real DGE tag
`that was incorrectly identified as an artifact.
`
`

`

`Kircher et al. BMC Genomics 2011, 12:382
`http://www.biomedcentral.com/1471-2164/12/382
`
`Page 5 of 14
`
`selecting for insert-size, it should be considered that
`current experimental methods generally do not provide
`precise length cutoffs. The lower cutoff selected should
`therefore be well-above the desired sequencing length.
`For sequence reads where part of the adapter sequence
`is included, the position in the sequence read at which
`the adapter sequence begins has to be identified and the
`read trimmed appropriately. Unfortunately, this is not
`part of the standard Illumina data processing and also
`non-trivial for short adapter fragments, especially given
`the increasing sequencing error at the end of reads. If
`reads are not filtered for known chimeras and trimmed
`for adapter sequences, these may interfere with map-
`ping/alignment and reads will either be incorrectly
`excluded or placed incorrectly. In both cases down-
`stream data analysis will be affected.
`In order to test how Illumina’s ELAND mapper as
`well as the widely used mapping program BWA [20] are
`impacted by incipient adapter sequence, we simulated
`101-cycle reads of an Illumina Paired End genomic
`library with 10,000 reads for every adapter start point
`between 1 to 350nt and the error profile observed for
`an actual run of this length (Figure 3). Considering that
`both mapping programs implement very different
`approaches (seed alignment versus semi-global align-
`ment of the whole read respectively), the performance
`of Illumina’s ELAND mapper is expected to be different
`from BWA. Since ELAND requires only a fixed seed in
`the beginning of the read (typically of 32nt length) adap-
`ters starting after this seed region should not affect
`ELAND’s mapping. Indeed, ELAND maps 98% of all
`simulated reads of at least 30nt insert size (2nt of adap-
`ter sequence being compensated by 2 mismatches being
`allowed in the seed), while BWA only reports 98% suc-
`cessful mappings for reads with an insert size of at least
`97nt. More relevant for many analyses, however, is the
`number of mappings reported to be uniquely placed and
`whether they are mapped at the correct position in the
`genome. ELAND reports a uniquely placed 20nt-insert-
`size read, but it is placed incorrectly (as are all uniquely
`placed reads reported up to an insert size of 67nt).
`BWA reports the first three uniquely placed fragments
`(mapping quality above 20) for an insert size of 83nt (2
`of them are correctly placed). If we require that 98% of
`the reads are correctly placed, ELAND achieves this for
`insert sizes of 83nt and above (14nt of adapter), while
`BWA can only compensate with mismatches for 4nt of
`adapter sequence (97nt insert size). However, BWA pro-
`vides a lower total number of false positive placements
`due to the inclusion of adapter sequence (8490 vs.
`6308). Moreover, for an insert size of at the least read
`length, BWA reports 99.999% of uniquely placed reads
`(94.2% of all reported alignments) at the correct geno-
`mic positions, while ELAND only reports 98.757% of
`
`the uniquely placed reads (83.8% of all reported align-
`ments) at the correct genome coordinates. BWA there-
`fore provides a more accurate mapping of these reads
`for downstream analysis.
`While length selection and dimer removal are impor-
`tant for the cost-effective sequencing of a library and
`downstream data analysis, experimental methods to
`achieve these generally consume sample material and
`may bias molecule representation. It is therefore often
`only practical to apply a minimum of these purification
`steps in order to maintain library quantity and complex-
`ity. In such cases, downstream sequence filtering prior
`to data analysis becomes extremely important.
`
`Short-insert libraries in paired-end sequencing
`experiments
`When libraries containing inserts shorter than the sum
`of forward and reverse read cycles are created, these can
`be sequenced from both ends to obtain higher quality
`sequence information for the overlapping sequence part.
`For such paired-end reads the correct identification of
`the adapter is eased by maximizing autocorrelation of
`the two reads as well as requiring identical adapter start
`positions for both reads [30-32]. This strategy is more
`powerful than alignment(-like) approaches used for
`identifying adapter starts in single reads, which fre-
`quently remove sequence from the read ends that match
`the adapter by chance, or which do not identify real
`adapter sequence due to the higher sequencing error at
`the end of reads. Thus, for short insert libraries, paired
`end sequencing is preferable. As previously reported
`[30], the read merging performed for these short-insert
`libraries considerably decreases the number of errors
`and creates sequences reflecting the original outer mole-
`cule length (e.g. of interest for authenticity of ancient
`DNA samples [31]). Applying this merging approach to
`the simulated data set described above, but this time
`using both paired-end reads, we see a factor of 5 reduc-
`tion in the error rate of all merged sequences (average
`error of 0.24% reduces to 0.05%; Additional File 1). For
`sequences shorter or equal to the read length a reduc-
`tion by a factor of about 21 (0.146% to 0.007%) is
`observed.
`
`Library contamination
`Sample contamination during library preparation from
`other DNA/RNA sources might be an important issue
`for some types of analyses and applications. Contamina-
`tion may be introduced by the experimenter or may
`stem from lab chemicals and equipment. Library pre-
`parations starting from low amounts of sample DNA
`and protocols using single strand ligation procedures
`can be considered the most prone to contamination.
`While avoidance of contamination is the most desirable
`
`

`

`Kircher et al. BMC Genomics 2011, 12:382
`http://www.biomedcentral.com/1471-2164/12/382
`
`Page 6 of 14
`
`BWA v0.5.5
`
`ELAND v1.5
`
`Read length (101nt)
`
`1.0
`
`0.8
`
`0.6
`
`0.4
`
`0.2
`
`0.0
`
`1.0
`
`0.8
`
`0.6
`
`0.4
`
`0.2
`
`0.0
`
`Fraction sequences reported mapped
`
`Fraction correctly placed
`
`0
`
`50
`
`100
`
`150
`
`200
`
`250
`
`300
`
`350
`
`0
`
`50
`
`100
`
`150
`
`200
`
`250
`
`300
`
`350
`
`1.0
`
`0.8
`
`0.2
`Fraction correct / placed
`
`0.4
`
`0.6
`
`D
`
`0.0
`
`C
`
`D
`
`0
`
`50
`
`100
`
`200
`150
`Molecule length
`
`250
`
`300
`
`350
`
`50
`
`60
`
`90
`80
`70
`Molecule length
`
`100
`
`110
`
`Figure 3 Effects of adapter sequence inclusion on mapping. Untrimmed adapter sequence at the read ends can interfere with alignment/
`mapping. We simulated 101-cycle human genomic shotgun reads for an Illumina Paired End library with 10,000 reads for every adapter starting
`point between 1 to 350nt, and the error profile observed for an actual run of this length. On this data set, we tested how ELAND and BWA are
`affected by inclusion of adapter sequence: (A) ELAND requires only a fixed seed (here 32nt) in the beginning of the read. Adapters beginning
`after this seed region may therefore have no effect on the output. ELAND reports 98% successful mappings for all simulated reads of at least
`30nt insert size (2nt of adapter sequence being compensated by 2 mismatches allowed in the seed), BWA only reports 98% successful mappings
`for reads with an insert size of at least 97nt. (B) Frequently only uniquely placed molecules are considered in data analysis. ELAND reports the
`first uniquely placed fragment for 20nt insert size. BWA reports the first three uniquely placed fragments (mapping quality above 20) for an
`insert size of 83nt. (C) All uniquely placed reads reported by ELAND up to an insert length of 67nt are placed incorrectly (when comparing to
`the coordinates the sequence was extracted from), as is one of the 3 reported by BWA for an insert size of 83nt. When requiring 98% correct
`placements, ELAND handles up to 14nt of adapter (83nt insert size), while BWA can only compensate with mismatches for 4nt of adapter
`sequence (97nt insert size). (D) For analysis purposes, BWA shows the better performance due to the lower number of false positive placements.
`Moreover, for an insert size of at least the read length (i.e. no adapters interfering with the alignment), BWA reports 99.999% of uniquely placed
`reads (94.2% of all reported alignments) at the designated genomic positions, while ELAND only reports 98.757% of the uniquely placed reads
`(83.8% of all reported alignments) at the correct position.
`
`D
`
`B
`
`1.0
`0.0
`Fraction sequences reported uniquely placed
`
`0.4
`
`0.2
`
`0.8
`
`0.6
`
`A
`
`

`

`Kircher et al. BMC Genomics 2011, 12:382
`http://www.biomedcentral.com/1471-2164/12/382
`
`Page 7 of 14
`
`approach, it has been suggested (e.g. [33]) that reads can
`be filtered by the alignment to the putative contaminant
`sequence before data analysis. However, such filtering
`may introduce biases in the data, especially if sequences
`are short and/or the evolutionary distance between con-
`taminant and sample is low. This is a frequent problem
`in ancient DNA studies of early modern humans, and
`Neandertals where contamination with even small
`amounts of modern DNA can quickly dominate sequen-
`cing output. Here the fraction of contamination can be
`deduced from informative sites (i.e. sites of known fixed
`differences between species/populations) and the frac-
`tion of contaminant molecules determined [31,32,34].
`This ratio can then be used in statistical models during
`data analysis. If no informative sites are known, esti-
`mates of contamination may be obtained from biallelic
`or triallelic sites in haploid/diploid sequences. Hence,
`also the sex of the sample may be exploited by counting
`Y chromosomal alignments in female samples and
`determining × chromosomal heterozygosity for males
`[35,36].
`Cross-contamination after library preparation (e.g.
`during preparation of the sequencing run) can be easily
`identified and filtered from the final sequencing data
`when sample-specific barcodes are included in the
`libraries and determined during sequencing [9,10]. A
`different problem may arise when pools of barcoded
`libraries are amplified simultaneously; in such experi-
`ments “jumping PCR” may cause barcodes to be trans-
`ferred between samples [37-40]. High error in the
`sequencing, amplification or synthesis of barcodes, espe-
`cially for those with limited sequence distance from one
`another, as well as barcode contamination during hand-
`ling or mixed clusters/small physical cluster distance on
`the flow cell can lead to sample misidentification. In
`projects for which data analysis is susceptible to any of
`these types of contamination, appropriate measures have
`to be taken to minimize contamination and to estimate
`its impact on analysis results.
`
`Alternative sequencing primers
`The sequencing primers used in sequencing of libraries
`constructed for different applications may differ. During
`flow cell preparation, different sequencing primers can
`be hybridized in each lane. However, in contrast to the
`first read prepared on the Cluster Station/cBot, it is not
`possible to use lane sp