GC Bias Correction

The most common implementation uses locally estimated scatterplot smoothing (LOESS) to model the relationship between GC content and read depth. For each genomic bin, the algorithm:

• Calculates the GC percentage of the reference genome within that window
• Fits a non-linear regression to the observed coverage across all bins
• Computes a correction factor that normalizes coverage to the genome-wide average This removes the systematic bias where high-GC regions are over-amplified and low-GC regions are under-represented during library preparation.

Effective GC bias correction depends on specific data characteristics:

• Bin size: Typically 10-100 kb non-overlapping windows across the reference genome
• Mappability filtering: Exclusion of repetitive regions where reads cannot be uniquely aligned
• Sufficient coverage: Requires at least 0.1-1x genome-wide depth for stable LOESS fitting
• Matched controls: Ideally, a panel of normals from healthy individuals processed with identical library preparation
• GC content range: Bins with extreme GC values (<20% or >80%) are often excluded due to sparse data and high variance

Without GC correction, copy number profiles exhibit wavy artifacts that obscure true amplifications and deletions. Correction provides:

• Reduced noise: Variance in log2 ratio decreases by 40-60% in well-calibrated assays
• Improved segmentation: Circular binary segmentation algorithms produce fewer false-positive breakpoints
• Lower limit of detection: Subclonal copy number alterations at <10% tumor fraction become resolvable
• Cross-sample comparability: Normalized profiles from different sequencing runs can be directly compared Failure to correct GC bias is a leading cause of false-positive 1p/19q codeletion calls in glioma liquid biopsy.

Modern pipelines employ more sophisticated approaches than simple LOESS:

• GAM (Generalized Additive Models): Handles overdispersion in low-coverage regions better than LOESS
• Conditional quantile normalization: Corrects for GC bias while preserving biological variation at specific quantiles of the coverage distribution
• Deep learning models: Convolutional neural networks trained on matched tumor-normal pairs can learn complex, multi-factorial bias patterns including fragment length and mappability interactions
• Simultaneous correction: Joint modeling of GC content, mappability, and replication timing removes confounding between these covariates These methods are particularly important for ultra-low pass whole genome sequencing (<0.1x coverage).

Post-correction QC ensures the normalization was effective:

• GC-corrected coverage distribution: Should show no visible trend when plotting corrected coverage vs. GC content
• Autocorrelation: The autocorrelation function (ACF) of corrected log2 ratios should decay rapidly, indicating removal of long-range GC-driven waves
• Median absolute pairwise difference (MAPD): A measure of bin-to-bin noise; values below 0.2-0.3 indicate adequate correction for copy number calling
• Control sample concordance: Corrected profiles from normal samples should show flat log2 ratios near zero across all chromosomes
• Known CNV recovery: Spiked-in controls or cell line mixtures with known copy number states should be accurately recovered

GC bias correction directly impacts clinical decision-making:

• HER2 amplification detection: Accurate copy number calling in plasma cfDNA guides trastuzumab therapy in metastatic breast cancer
• AR amplification in prostate cancer: Enzalutamide resistance monitoring requires precise copy number estimation from ctDNA
• Tumor fraction estimation: ichorCNA and similar tools rely on GC-corrected coverage to estimate the circulating tumor DNA fraction from shallow WGS
• MRD monitoring: Minimal residual disease assays using copy number signatures require extremely low noise floors achievable only with proper GC correction
• Multi-cancer early detection: Pan-cancer screening algorithms integrate GC-corrected copy number features with fragmentomic and methylation signals