Batch effects in a multiyear sequencing study: False biological trends due to changes in read lengths

Abstract

High-throughput sequencing is a powerful tool, but suffers biases and errors that must be accounted for to prevent false biological conclusions. Such errors include batch effects; technical errors only present in subsets of data due to procedural changes within a study. If overlooked and multiple batches of data are combined, spurious biological signals can arise, particularly if batches of data are correlated with biological variables. Batch effects can be minimized through randomization of sample groups across batches. However, in long-term or multiyear studies where data are added incrementally, full randomization is impossible, and batch effects may be a common feature. Here, we present a case study where false signals of selection were detected due to a batch effect in a multiyear study of Alpine ibex (Capra ibex). The batch effect arose because sequencing read length changed over the course of the project and populations were added incrementally to the study, resulting in nonrandom distributions of populations across read lengths. The differences in read length caused small misalignments in a subset of the data, leading to false variant alleles and thus false SNPs. Pronounced allele frequency differences between populations arose at these SNPs because of the correlation between read length and population. This created highly statistically significant, but biologically spurious, signals of selection and false associations between allele frequencies and the environment. We highlight the risk of batch effects and discuss strategies to reduce the impacts of batch effects in multiyear high-throughput sequencing studies.

The presence and impact of reference bias on population genomic studies of prehistoric human populations

“…Differences in sequencing strategies (e.g. read length) and bioinformatic processing have been shown to generate batch effects and dramatically affect downstream analyses (Leek et al ., 2010; Leigh et al ., 2018; Shafer et al ., 2016; Mafessoni et al ., 2018). Another well known bias in population genetics is ascertainment bias which arises when the studied variants were ascertained in selected populations only, and can substantially impact measurements of heterozygosity and related methods (Albrechtsen et al ., 2010).…”

Wild GWAS—association mapping in natural populations

“…In particular, low variation in chromosome size, small sample size, low heritability, a skewed effect size distribution and clustering of loci can lead to both loss of power and bias and consequently altered inference from chromosome‐partitioning analyses. Another important cautionary tale is presented by Leigh, Lischer, Grossen, and Keller (), concerning a false signal of selection and environmental association which arose as a consequence of batch effects in the sequencing of individuals sampled from a number of populations of Alpine ibex ( Capra ibex ). They show that the incremental addition of populations to the study meant that many populations were sequenced at different times and using different sequencing technologies, with corresponding changes to read length.…”

“…Leigh et al. (), however, outline a number of bioinformatic approaches to test and account for batch effects, particularly when randomization of sample groups across batches may not be possible. Given the incremental addition of samples for long‐term wild studies in particular, careful consideration of the potential for batch effects is important in validating signals of association in natural populations.…”

“…Due to the random nature of which parts of the genome get adequately sequenced per individual, these approaches are typically characterized by high rates of missing data. The impact of missing data on association mapping will clearly be to reduce power by reducing sample size per marker tested, but bias may also result, for example, from data that are missing in a systematic manner across samples due to batch effects (see Leigh et al., ). Pool‐Seq is a promising approach which optimizes coverage per phenotype in the absence of high coverage sequencing sufficient to call individual genotypes (Micheletti & Narum, ; Schlötterer et al., ) but may lack power to detect small‐effect loci.…”

Uneven missing data skews phylogenomic relationships within the lories and lorikeets

“…We compared the amount of parsimony informative sites and missing data in samples with their age. Because we sequenced samples over multiple years, batch effects, or biases attributable to differences among sequencing run could also bias our results (Leigh et al 2018). To provide a qualitative assessment of batch effects, we provide plots of trees where tips have been colored according to one of three plates that they were sequenced on.…”