Base Quality Score Recalibration (BQSR)

BQSR applies a Gaussian mixture model or table-based recalibration to analyze covariates that correlate with sequencing errors. Key covariates include:

• Read group identity: Accounts for machine-specific or lane-specific biases.
• Reported quality score: The raw Phred score assigned by the sequencer.
• Machine cycle: The position within the read, as error rates often increase toward the 3' end.
• Dinucleotide context: The preceding and current base, as certain sequence motifs (e.g., AC, GT) are prone to specific substitution errors. By binning bases according to these covariates, the algorithm calculates an empirical mismatch rate against the reference genome, adjusting the final quality score to be globally accurate.

The core function of BQSR is to replace the subjective, vendor-generated quality score with an empirical quality score. The algorithm calculates the actual probability of error for every base in a given covariate bin. For example, if a sequencer reports a quality score of 30 (implying 0.1% error) but the empirical mismatch rate for that specific dinucleotide context and cycle is 1%, BQSR will downgrade the quality score to 20. Conversely, it can upgrade scores that are overly pessimistic. This calibration is essential for downstream variant callers that rely on these scores to compute genotype likelihoods.

To avoid conflating true biological variation with sequencing error, BQSR requires a database of known polymorphic sites (e.g., dbSNP, 1000 Genomes). The algorithm masks these loci during the initial training phase so that genuine single nucleotide polymorphisms (SNPs) are not counted as mismatches. Without this masking step, recalibration would incorrectly penalize high-quality bases at common variant sites, leading to false negative variant calls in downstream analysis. A second pass is often performed after an initial variant call to refine the mask.

Recalibration directly improves the sensitivity and specificity of variant calling. By correcting inflated quality scores, BQSR reduces the number of false positive variant calls caused by systematic sequencing errors that appear with high confidence. It also improves the accuracy of genotype likelihood calculations, which depend on the Phred-scaled probability of each base being correct. Studies show that applying BQSR can reduce the number of false positive SNPs by over 50% in certain genomic contexts, significantly increasing the precision of the final variant call set.

A critical prerequisite for BQSR is the correct assignment of read groups in the alignment file. A read group identifies the specific library preparation, flowcell lane, and sample barcode for a set of reads. BQSR treats each read group as an independent source of technical error because different physical runs of a sequencer exhibit distinct error profiles. Failing to properly stratify data by read group will cause the algorithm to average distinct error modes, resulting in suboptimal recalibration and residual systematic bias in the output data.