Base Quality Recalibration (BQR)

BQR constructs a covariate model to estimate the empirical probability of a base being an error. Instead of trusting the sequencer's raw quality score, the algorithm groups bases by shared characteristics that correlate with inaccuracy.

• Key Covariates: Machine cycle, dinucleotide context, and raw reported quality.
• Mechanism: A Gaussian mixture model or binomial regression is trained on a known reference (e.g., dbSNP) to learn the true error rate for each covariate group.
• Output: A recalibrated quality score that reflects the actual probability of mismatch.

The algorithm stratifies every base call into a multi-dimensional bin based on its technical context. This grouping is essential for detecting systematic biases that raw quality scores ignore.

• Cycle Context: Errors increase in later sequencing cycles due to phasing and signal decay.
• Dinucleotide Context: Specific preceding and current nucleotide pairs (e.g., AC, GT) have distinct error profiles due to polymerase preferences.
• Quality Bin: The original predicted quality score is used as a prior to refine the posterior error probability.

Recalibration directly reduces the transition/transversion (Ti/Tv) ratio of false positive calls, moving it closer to the expected biological ratio.

• False Positive Reduction: Systematic errors often masquerade as rare variants. BQR collapses these inflated quality scores.
• Sensitivity Preservation: By adjusting scores down for likely errors and up for likely true bases, BQR increases the power to detect low-frequency variants.
• Liquid Biopsy Relevance: In ctDNA analysis, where variant allele frequencies are often <1%, distinguishing a true mutation from a sequencer artifact is impossible without accurate base quality scores.

BQR is fundamentally a supervised learning problem applied to signal processing. The model learns the conditional probability P(Error | Covariates).

• Algorithm: Typically uses a binomial logistic regression or a naive Bayes classifier to estimate the posterior error probability.
• Training Data: The reference genome acts as ground truth for non-variant sites.
• Feature Vector: The combination of read group, quality score, cycle, and context forms the input feature vector for the recalibration model.

BQR treats each read group (a lane/flowcell combination) as an independent source of systematic bias. This is critical because technical artifacts are often specific to a physical sequencing run.

• Run-Specific Artifacts: Optical duplicates and flowcell-specific noise are modeled separately.
• Platform Compatibility: Works across Illumina, BGI, and other sequencing-by-synthesis platforms.
• Data Fusion: Merges data from multiple lanes without cross-contaminating the error models, preserving the integrity of each technical replicate.

Recalibration relies on a catalog of known polymorphic sites to establish a baseline of truth. The process:

• Excludes known variant sites to isolate likely sequencing errors.
• Calculates the empirical error rate for every covariate bin.
• Computes the mismatch rate between the aligned read and the reference genome.
• Adjusts the Phred-scaled quality score to match the observed error frequency. This ensures that a reported quality score of Q30 actually corresponds to a 1 in 1000 error probability.

The fundamental unit manipulated by BQR is the Phred quality score (Q-score). It is logarithmically related to the error probability:

• Q10 = 1 in 10 error probability (90% accuracy).
• Q20 = 1 in 100 error probability (99% accuracy).
• Q30 = 1 in 1000 error probability (99.9% accuracy).
• Q40 = 1 in 10,000 error probability (99.99% accuracy). BQR adjusts these scores so they are well-calibrated, meaning the empirical error rate matches the predicted rate.

BQR corrects for systematic biases inherent to specific sequencing chemistries and hardware:

• Illumina platforms: Corrects for phasing/pre-phasing errors and signal decay toward the 3' end of reads.
• Pacific Biosciences (PacBio): Models stochastic kinetic errors in circular consensus sequencing.
• Oxford Nanopore: Addresses context-specific biases in current signal interpretation. Each platform exhibits a unique error signature that machine learning models can learn and neutralize, making BQR essential for cross-platform data harmonization.