GC Bias Correction

GC bias originates during library preparation when DNA fragments with extreme GC content (high or low) amplify less efficiently than fragments with balanced base composition. This creates a systematic, non-biological correlation between local GC percentage and observed read depth. The bias is particularly severe in whole-genome bisulfite sequencing and single-cell assays where amplification cycles are extensive. Without correction, GC-rich promoters and GC-poor intergenic regions exhibit artificially inflated or deflated coverage, confounding copy number variant detection and peak calling.

The most widely implemented correction method fits a locally weighted scatterplot smoothing (LOESS) curve to the relationship between GC content and read count in non-overlapping genomic windows. The algorithm:

• Divides the genome into fixed-size bins (typically 100 bp to 10 kb)
• Computes the GC fraction and log-transformed read count per bin
• Fits a robust LOESS regression to model the expected coverage as a function of GC content
• Subtracts the fitted bias from observed values to produce GC-normalized coverage

This approach is implemented in tools like HMMcopy and GATK for copy number analysis.

An alternative to regression-based correction, modal normalization identifies the most frequent read count (the mode) for each GC-content stratum and scales all bins to match a global reference. This method assumes that the majority of the genome is copy-neutral, making the mode a robust estimator of baseline ploidy. It is particularly effective in cancer genomics where large-scale copy number aberrations can skew mean-based corrections. Tools like Control-FREEC employ this strategy to simultaneously correct GC bias and infer tumor purity.

GC bias correction must account for mappability—the uniqueness of a genomic region in the reference genome. Low-mappability regions (repetitive elements, segmental duplications) exhibit artificially low coverage independent of GC content. Advanced correction pipelines apply a two-dimensional LOESS that models read depth as a joint function of GC content and mappability score. The ENCODE consortium's uniform processing pipeline mandates this dual correction to prevent false-positive copy number calls in repetitive regions of the genome.

Uncorrected GC bias is a leading cause of false-positive segmental duplications and deletions in shallow whole-genome sequencing. A 1% deviation in GC content can produce a 2-3 fold change in normalized coverage in extreme cases. After correction, the coefficient of variation in read depth across the genome typically drops from 0.15–0.30 to below 0.10, enabling reliable detection of single-copy alterations at sequencing depths as low as 0.1–1x. This is critical for non-invasive prenatal testing and liquid biopsy applications.

Recent methods replace LOESS with convolutional neural networks that predict expected coverage directly from DNA sequence context, capturing higher-order sequence features beyond simple GC percentage. Models like Basenji and Enformer implicitly learn to normalize coverage by training on diverse epigenomic assays. For targeted applications, GCcorrect uses a shallow neural network trained on diploid control samples to predict per-bin bias, achieving superior normalization in regions with complex nucleotide composition where linear GC models fail.

The primary source of GC bias occurs during library preparation when DNA fragments undergo polymerase chain reaction (PCR) amplification. DNA polymerases exhibit sequence-dependent efficiency, preferentially amplifying fragments with balanced GC content (40-60%) while under-amplifying extreme AT-rich or GC-rich regions. This creates a systematic, non-linear relationship between true fragment abundance and observed read counts.

• Mechanism: Differential denaturation efficiency and polymerase processivity
• Impact: Can introduce >2-fold coverage variation in GC-extreme regions
• Mitigation: PCR-free library preparation protocols eliminate this artifact entirely

A locally weighted scatterplot smoothing technique that models the non-linear dependence of read coverage on GC content. For each genomic bin, a weighted linear regression is fit using neighboring bins in GC space, producing a sample-specific correction curve. The observed count is then divided by the predicted value to yield a normalized signal.

• Window size: Typically 1% GC content bins across 0-100% range
• Output: Corrected read counts with GC dependence removed
• Limitation: Assumes uniform true copy number across the genome, which is violated in tumor samples with aneuploidy

An extension of standard quantile normalization that corrects GC bias while preserving biological signal heterogeneity. Rather than forcing all samples to identical distributions, this method normalizes within GC-content strata independently, then reconstructs the full distribution. This is critical for cancer genomics where copy number alterations create genuine coverage variation.

• Stratification: Bins grouped by GC percentage before normalization
• Preservation: Maintains true copy number variants while removing artifact
• Application: Widely used in copy number variation (CNV) detection pipelines

A confounding factor often corrected alongside GC bias. Mappability refers to the uniqueness of a genomic region—repetitive sequences produce ambiguously mapped reads that are typically discarded, creating artificial coverage depressions. The mappability score (0-1) for each base represents the inverse of how many locations a k-mer of that length aligns to in the reference genome.

• Calculation: Pre-computed using tools like GEM or Umap
• Integration: Combined with GC correction in a joint model
• Critical for: Centromeres, telomeres, and segmental duplications

Neural network approaches that learn the complex, non-linear mapping between sequence context and observed coverage directly from data. Unlike LOESS, these models can capture higher-order sequence features beyond simple GC percentage, including k-mer frequencies, DNA shape, and local repeat density. Models like Basenji and Enformer implicitly correct for GC bias through their multi-task training on normalized epigenetic profiles.

• Input: Raw DNA sequence or k-mer spectrum
• Architecture: Convolutional or transformer-based
• Advantage: Captures sequence-dependent amplification biases missed by GC-only models

A related artifact where insert size distribution varies with GC content. GC-rich fragments tend to produce shorter library inserts due to differences in transposase or sonication efficiency. Since shorter fragments cluster more tightly in coverage profiles, this creates a secondary bias that correlates with but is distinct from GC bias. Modern correction methods model coverage as a function of both GC content and fragment length.

• Measurement: Inferred from paired-end read mapping distances
• Correction: Bivariate LOESS or generalized additive models (GAMs)
• Relevance: Critical for ATAC-seq and ChIP-seq where fragment size carries biological meaning