Exome Capture Bias

The differential melting temperature (Tm) and binding energy of capture probes directly drive coverage non-uniformity. Probes with suboptimal GC content or those targeting regions with stable secondary structures exhibit reduced hybridization efficiency. This results in systematically lower read depth for high-AT or high-GC exons, creating a thermodynamic fingerprint that must be computationally normalized before variant calling.

A significant fraction of sequenced fragments map outside targeted exonic regions despite probe design specificity. This off-target capture includes intronic and intergenic sequences with partial homology to baits. The proportion of on-target reads typically ranges from 50-70%, directly reducing effective coverage depth for true exonic targets and introducing noise into copy number estimation pipelines.

Regions with extreme guanine-cytosine (GC) content exhibit characteristic coverage depletion. Low-GC regions (<30%) suffer from reduced probe binding affinity, while high-GC regions (>70%) experience incomplete denaturation and polymerase stalling during amplification. This creates a parabolic coverage profile that requires loess regression or hidden Markov model-based normalization to prevent systematic false-negative variant calls in affected exons.

Exonic regions overlapping low-complexity sequences or segmental duplications suffer from ambiguous read alignment. Short sequencing reads cannot be uniquely placed, causing apparent coverage drops that mimic true deletions. Capture bias correction must integrate mappability tracks and exclude paralogous sequences from variant calling to avoid false-positive structural variant calls driven by alignment artifacts rather than true biology.

Different exome capture kit versions and manufacturing lots introduce systematic coverage biases independent of biological signal. Probe design differences between Agilent SureSelect, Twist Bioscience, and IDT xGen platforms produce distinct coverage landscapes. Cross-batch normalization using principal component analysis or ComBat-seq is essential when combining samples processed with different reagents to prevent batch effects from being misinterpreted as copy number alterations.

The post-capture amplification step introduces optical and PCR duplicates that distort true library complexity. Over-amplified fragments create coverage spikes at specific loci, while under-amplified regions appear depleted. Deduplication using unique molecular identifiers (UMIs) or read-position-based algorithms is required to recover the true molecular coverage and prevent inflated variant allele fraction estimates in highly duplicated libraries.

The primary driver of exome capture bias where regions with extreme guanine-cytosine (GC) content (below 30% or above 70%) exhibit significantly reduced capture efficiency. Probes targeting GC-rich regions suffer from reduced hybridization kinetics due to stable secondary structures, while AT-rich regions have weaker probe binding. This creates a characteristic inverted-U shape in coverage plots, with optimal capture occurring at approximately 40-50% GC content. Modern correction tools model this relationship using locally weighted scatterplot smoothing (LOESS) regression to normalize read depth against GC percentage.

A statistical correction applied post-alignment to adjust read counts across targeted exonic intervals to a uniform baseline. The process involves:

• Binning the genome into non-overlapping windows (typically 100-1000 bp)
• Calculating log2 ratios of observed vs. expected coverage
• Applying LOESS or polynomial regression to model bias as a function of GC content, mappability, and probe density
• Scaling raw counts by the inverse of the predicted bias factor Tools like GATK's DenoiseReadCounts and CNVkit implement these corrections to prevent false positive copy number variant calls.

The upstream engineering strategy to minimize capture bias by optimizing the oligonucleotide probe library before sequencing. Modern bait design algorithms use thermodynamic models to select probes with uniform melting temperatures and avoid regions prone to secondary structure. Key strategies include:

• Tiling density adjustment: increasing probe overlap in historically low-coverage regions
• Degenerate base incorporation: using mixed bases at wobble positions to capture GC-rich variants
• Machine learning prediction: training models on historical capture data to predict and preemptively correct for bias-prone targets
• Spike-in controls: adding synthetic templates at known concentrations to calibrate capture efficiency

The direct consequence of exome capture bias on heterozygous variant detection, where unequal capture of the reference and alternate alleles skews the observed variant allele fraction (VAF) away from the expected 0.5. In regions with strong GC bias, one allele may be preferentially captured, causing:

• False negatives: VAF drops below the caller's detection threshold
• Genotype miscalls: heterozygous sites incorrectly called as homozygous reference
• Strand-specific artifacts: bias affecting only reads from one DNA strand Variant callers like DeepVariant learn to recognize these distorted pileup patterns, but explicit bias correction in the BAM file remains critical for sensitive detection.