Analyzing and minimizing bias in Illumina sequencing libraries

Abstract

Although Illumina shot-gun reads cover most genomes almost completely, sequences with extreme base compositions are often underrepresented or missing. Bias can potentially be introduced at any step during the library construction in the lab, on the Illumina instrument, in data processing or at the sequence analysis stage. Here we set out to evaluate sources of bias and ameliorate the effects.To dissect the library construction process, we developed a panel of qPCR assays for loci ranging from 6% to 90% GC that work well in a pool of three microbial DNA samples of different base composition: Plasmodium falciparum (19% GC), Escherichia coli (51% GC) and Rhodobacter sphaeroides (69% GC). We also developed qPCR assays for loci in the human genome that represent four categories of underrepresented sequence motifs as well as GC-rich promoters known to be underrepresented or missing in 'whole' genome sequencing data sets.We tracked the relative abundance of these loci throughout the standard Illumina library protocol and saw no significant introduction of bias in the initial steps including shearing, end repair, adaptor ligation and size selection. However, GC-rich and extremely GC-poor sequences were depleted during the subsequent PCR-enrichment step. Using qPCR as a readout, we tested different PCR enzymes, the addition of betaine and/or DMSO, and thermocycling profile variations. The choice of PCR instrument itself and the ramp rate had a significant effect on the GC profile of the PCR product, especially when using the recommended amplification conditions (Phusion HF and 10s denaturation per cycle).Our optimized conditions produce PCR-amplified libraries that display little systematic bias between 15% and 80% GC that resulted during sample preparation. We saw significantly improved representation of challenging human sequence motifs both in the PCR-amplified library (qPCR assay) and in the final Illumina reads. Our conditions are also more reliable and robust because they minimize the effect of PCR instrument and ramp rate. These conditions are currently being implemented in the Sequencing Platform at the Broad Institute. Finally, we still observe some bias in the sequencing readout, which is introduced by steps subsequent to sample preparation, including cluster generation and sequencing. These sources of bias are the object of ongoing investigations.

Assessment of the latest NGS enrichment capture methods in clinical context

“…It has for long been demonstrated that regions of low (<30%) or high (>60%) GC content usually result in poor DOC 11 15 . The low coverage of these regions may be due to a bias in PCR amplification (regions with a neutral GC content are more efficiently amplified) or during the hybridization step (these regions may have a reduced capture effectiveness) 26 27 28 29 . We clearly showed a smaller dispersion observed with the NimbleGen library, while NRCCE is more efficient than the others in low GC content context, and SureSelect QXT in high.…”

Combining 16S rRNA gene variable regions enables high-resolution microbial community profiling

“…Multiple regions tend to 'average' PCR bias: PCR amplification may have a different efficiency for different bacteria [15,16], which results in considerable differences in profiling depending on primer choice [17][18][19][20]. Moreover, Gohl et al [21] have shown that the same primer pair may yield different profiling results depending on the specific library preparation protocol.…”

ALE: a generic assembly likelihood evaluation framework for assessing the accuracy of genome and metagenome assemblies

“…The depth d i is ideally Poisson-distributed (Lander and Waterman, 1988). However, most next-generation sequencers and library preparation techniques can bias GC-rich areas of a genome (Aird et al, 2011). This bias affects the observed depth in specific areas.…”