Characterizing and measuring bias in sequence data.

Ross MG, Russ C, Costello M, Hollinger A, Lennon NJ, Hegarty R, Nusbaum C, Jaffe DB. · 2013

Read Characterizing and measuring bias in sequence data. by Ross MG, Russ C, Costello M, Hollinger A, Lennon NJ, Hegarty R, Nusbaum C, Jaffe DB. (2013) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/europmc-pmc4053816.txt.

Introduction

Michael G. Ross and seven colleagues measure how each major sequencing platform distorts the DNA it reads, in Genome Biology in 2013. Ideal sequencing would cover a genome evenly. Real machines favour some stretches and starve others, and the skew can hide genes or fake findings.

The team built assays for discovering and scoring bias, then ran them on four platforms using human and microbe DNA. How the sample library was made mattered as much as the machine. One platform's coverage was the most even, but all struggled at very high or low GC content and across long runs of one letter. A thousand human promoters proved especially hard to read, and the team catalogues them. Deep data on one person still left a small slice of the genome nearly bare, most of it explained by known bias motifs.

Catalog record

CREATORRoss MG, Russ C, Costello M, Hollinger A, Lennon NJ, Hegarty R, Nusbaum C, Jaffe DB.
YEAR2013
COLLECTIONPAPERS
FORMATTXT
FILE SIZE79 KB
LICENSECC BY 4.0
RIGHTSCC BY
SOURCEEurope PMC

Source: https://europepmc.org/articles/PMC4053816

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Characterizing and measuring bias in sequence data By Michael G Ross, Carsten Russ, Maura Costello, Andrew Hollinger, Niall J Lennon, Ryan Hegarty, Chad Nusbaum, David B Jaffe Published in: Genome Biology DOI: 10.1186/gb-2013-14-5-r51 ABSTRACT Abstract Background DNA sequencing technologies deviate from the ideal uniform distribution of reads. These biases impair scientific and medical applications. Accordingly, we have developed computational methods for discovering, describing and measuring bias. Results We applied these methods to the Illumina, Ion Torrent, Pacific Biosciences and Complete Genomics sequencing platforms, using data from human and from a set of microbes with diverse base compositions. As in previous work, library construction conditions significantly influence sequencing bias. Pacific Biosciences coverage levels are the least biased, followed by Illumina, although all technologies exhibit error-rate biases in high- and low-GC regions and at long homopolymer runs. The GC-rich regions prone to low coverage include a number of human promoters, so we therefore catalog 1,000 that were exceptionally resistant to sequencing. Our results indicate that combining data from two technologies can reduce coverage bias if the biases in the component technologies are complementary and of similar magnitude. Analysis of Illumina data representing 120-fold coverage of a well-studied human sample reveals that 0.20% of the autosomal genome was covered at less than 10% of the genome-wide average. Excluding locations that were similar to known bias motifs or likely due to sample-reference variations left only 0.045% of the autosomal genome with unexplained poor coverage. Conclusions The assays presented in this paper provide a comprehensive view of sequencing bias, which can be used to drive laboratory improvements and to monitor production processes. Development guided by these assays should result in improved genome assemblies and better coverage of biologically important loci. BACKGROUND Ideal whole-genome shotgun DNA sequencing would distribute reads uniformly across the genome and without sequence-dependent variations in quality. All existing sequencing technologies fall short of this ideal and exhibit various types and degrees of bias. Sequencing bias degrades genomic data applications, including genome assembly and variation discovery, which rely on genome-wide coverage. Undercovered regions might lead to a missed SNP in an important region or cause an assembler to produce shorter contigs. For example, Figure 1 plots the coverage of the transcription start site and first exon of human gene NCS1, which encodes a neurotransmitter regulator [1], in whole-genome shotgun sequencing (data set A2). Despite 198-fold mean coverage of the genome, the first 72 bases of this exon are completely uncovered. This type of bias can reduce the effectiveness of biological and medical research. Recently published work on drug-resistant tuberculosis identified thousands of zero-coverage sites in an entire class of the bacterium's genes, despite sequencing to an average depth of 134× [2]. Alleviating gaps or dips in coverage through additional reads inflates sequencing costs, and may have limited effectiveness. For these reasons, improving our knowledge of sequencing bias is essential to improving the utility of DNA sequencing data. [Figure: Diagram illustrating the low coverage of NCS1 exon 1 in 198× Illumina HiSeq shotgun data. The first 72 bases of the first exon of human gene NCS1, including the transcription start site, were uncovered in a 198× whole-genome shotgun data set (#A2). The displayed 2,000 base region is chromosome 9:132,933,910-132,935,910. NCS1 encodes calcium-binding proteins that regulate neurotransmitter release [1].] Our goal in this work was to develop a rigorous method for discovering and monitoring coverage and error biases, then to apply it to data from a wide range of sequencing platforms (Illumina HiSeq and MiSeq, Ion Torrent PGM, Pacific Biosciences RS, and the Complete Genomics sequencing service). This study complements previous work in the field [3-7]. Bias manifests in multiple ways. Coverage bias is a deviation from the uniform distribution of reads across the genome. Similarly, error bias is a deviation from the expectation of uniform mismatch, insertion, and deletion rates in reads across the genome. This paper focuses primarily on coverage bias because it is the most damaging sequencing failure. Sequencing technologies are vulnerable to multiple sources of bias. Methods based on bacterial cloning and Sanger-chemistry sequencing [8] were subject to many coverage-reducing biases, notably at GC extremes, palindromes, inverted repeats, and sequences toxic to the bacterial host [9-17]. Illumina sequencing [18] has been shown to lose coverage in regions of high or low GC [19-22], a phenomenon also seen in other 'next-generation' technologies [3,6]. PCR amplification during library construction is a known source of undercoverage of GC-extreme regions [20,21] and similar biases may also be introduced during bridge PCR for cluster amplification on the Illumina flowcell [23]. Illumina strand-specific errors can lead to coverage biases by impairing aligner performance [24]. Ion Torrent [25], like 454 [26], utilizes a terminator-free chemistry that may limit its ability to accurately sequence long homopolymers [4,27,28], and may also be sensitive to coverage biases introduced by emulsion PCR in library construction. Complete Genomics [29] also uses amplification along with a complex library construction process. The Pacific Biosciences [30] process is amplification-free; therefore, one might expect it to exhibit lower levels of coverage bias than the other technologies. In addition to sources in the wet lab, bias can be introduced by any of the computational steps in the sequencing pipeline. Signal-processing and base calling limitations could result in under-representation or increased error rates in some locations, as can inaccurate alignment. An inaccurate reference or sample-reference differences can cause coverage or accuracy variations that may be misdiagnosed as sequencing bias. Therefore, detecting bias is only the first step and must be followed by more detailed experiments to assign responsibility to the library preparation, sequencing, or computational stages. We employ two methodologies for measuring bias. Per-base bias measurements, which rely on deep-coverage sequencing, are hypothesis-free and ideal for discovering new types of bias. Motif bias measurements, which require only shallow-coverage sequencing, are ideal for comparisons across experimental conditions and for monitoring ongoing sequencing pipeline performance at known bias-prone sequence contexts and locations. Bias motif monitoring plays a useful role by providing a critical metric in determining and ameliorating the sources of sequencing bias. Together these methodologies can be used to compare platforms, to measure the utility of combining data from multiple platforms, and to determine the extent to which coverage bias is described by the statistics of known motifs. RESULTS AND DISCUSSION Per-base bias We begin by defining the bias statistics. The fundamental statistic of coverage bias is 'relative coverage', which is defined as: [formula: coverageofagievenreferencebaseinagenomemeancoverageofallreferencebases.] The coverage of a given reference base is computed by counting the number of read bases mapped to it in an alignment (see Materials and methods). The mean coverage is computed by averaging this value across every base in the reference. Then the relative coverage for a particular base is computed as the ratio of these values. A relative coverage of 1 indicates that a particular base is covered at the expected average rate. A relative coverage above 1 indicates higher than expected coverage and below 1 indicates lower than expected coverage. Some reads cannot be mapped to a single locus, and the probability of ambiguous mapping increases as reads become shorter or less accurate. Ambiguous mapping is also more likely for reads that derive from repetitive or low complexity regions of the genome, including

Read the complete text