Variant Allele Fraction (VAF)

VAF is defined as:

VAF = Alternate Allele Depth / Total Read Depth

• Numerator: Number of sequencing reads containing the variant allele
• Denominator: Total number of reads covering that genomic position
• Expressed as a fraction (0.0 to 1.0) or percentage (0% to 100%)
• A heterozygous germline variant in a diploid genome theoretically has a VAF of ~0.5 (50%)
• Somatic mutations in tumor samples often exhibit VAFs ranging from <0.01 to 0.5 depending on tumor purity and clonality

VAF provides critical insight into the clonal architecture of a sample:

• Germline Heterozygous: VAF ≈ 0.5 in normal tissue; present in all cells
• Germline Homozygous: VAF ≈ 1.0; both alleles carry the variant
• Clonal Somatic: VAF reflects tumor purity and copy number; a heterozygous mutation in a pure tumor with diploid genome yields VAF ≈ 0.5
• Subclonal Somatic: VAF < expected clonal fraction; indicates mutation present in only a subset of tumor cells
• Artifactual: Strand-biased or low-complexity region variants often show aberrant VAF distributions

Observed VAF in tumor samples must be adjusted for tumor purity (the fraction of cancer cells in the sample) and local copy number:

• Expected VAF = (p × C_mut) / (p × C_total + (1-p) × 2)
  • p = tumor purity
  • C_mut = number of mutated allele copies
  • C_total = total copies at the locus in tumor cells
• This correction is essential for accurate clonal deconvolution and distinguishing driver mutations from passengers
• Tools like PyClone and ABSOLUTE integrate VAF with copy number to infer clonal populations

VAF distributions help distinguish true variants from technical noise:

• Oxidative damage artifacts (8-oxoguanine): Characteristic G>T transversions with low VAF (<0.05) and strong strand bias
• FFPE-induced deamination: C>T transitions with low VAF in formalin-fixed samples
• PCR errors: Random, non-reproducible low-VAF variants without strand bias
• Cross-sample contamination: Unexpected heterozygous variants with VAF ≈ 0.5 from a different individual
• Deep learning variant callers like DeepVariant implicitly learn VAF-related features from pileup images

VAF monitoring is central to precision oncology:

• Minimal Residual Disease (MRD): Tracking ultra-low VAF mutations (<0.001) after treatment indicates residual cancer
• Treatment Response: Declining VAF of driver mutations correlates with therapeutic efficacy
• Resistance Emergence: Rising VAF of known resistance mutations (e.g., EGFR T790M) signals clonal expansion
• Liquid Biopsy: Circulating tumor DNA VAF is typically <0.01 and requires error-suppressed sequencing methods
• Clonal Hematopoiesis: Low VAF mutations in blood may represent age-related clonal expansion rather than malignancy

The reliability of a VAF estimate depends on read depth and binomial sampling error:

• Low coverage (<30x): High variance in VAF estimates; difficult to distinguish 0.4 from 0.5
• High coverage (>500x): Enables detection of variants with VAF as low as 0.01
• Confidence interval can be modeled using the binomial distribution:
  • Var(VAF) = VAF × (1 - VAF) / Depth
• Fisher's exact test or beta-binomial models are used to assess whether an observed VAF deviates significantly from an expected value
• Deep learning models incorporate read depth and base quality as explicit features to calibrate variant confidence

The total number of sequencing reads aligned to a specific genomic locus. VAF is calculated as the ratio of variant-supporting reads to the total read depth at that position.

• Low depth (<10x) produces unreliable VAF estimates due to sampling noise
• High depth (>100x) enables detection of low-frequency variants
• Poisson and negative binomial distributions model the stochastic nature of coverage

Without sufficient depth, true low-VAF somatic mutations become indistinguishable from sequencing errors.

VAF distributions directly reflect zygosity state in diploid genomes:

• Heterozygous germline variants: Expected VAF of ~50% (0.4–0.6 range) in normal tissue
• Homozygous germline variants: Expected VAF approaching 100%
• Somatic mutations: Varies widely based on tumor purity and clonal architecture

Deviation from the 50% expectation in matched normal samples flags potential copy number alterations or allelic imbalance.

The observed VAF of a somatic mutation is diluted by admixed normal cells in the tumor sample. The corrected cancer cell fraction is calculated as:

CCF = VAF_observed × (1 / tumor_purity)

• A mutation at 10% VAF in a sample with 50% purity represents a clonal event (CCF ≈ 20%)
• Subclonal mutations exhibit lower CCFs, indicating late emergence
• PyClone and ABSOLUTE are tools that jointly estimate purity and clonal composition

Circulating tumor DNA (ctDNA) analysis pushes VAF detection to extreme lower limits:

• Ultra-low VAF (<0.1%) requires unique molecular identifiers (UMIs) and error-suppression methods
• Digital droplet PCR can resolve VAFs down to 0.01%
• Deep learning denoising models are now used to distinguish true ctDNA signals from oxidative damage artifacts

VAF dynamics over serial blood draws serve as a real-time biomarker for treatment response monitoring.

Age-related clonal hematopoiesis of indeterminate potential (CHIP) introduces somatic variants into blood-derived normal samples:

• CHIP variants typically appear at low VAF (2–10%) in matched normal blood
• Can be misclassified as tumor-specific if not computationally filtered
• Common in genes like DNMT3A, TET2, and ASXL1

Failure to account for CHIP leads to false somatic mutation calls and incorrect tumor mutational burden estimates.

A technical artifact where one allele fails to amplify or capture, skewing VAF dramatically:

• Can cause a heterozygous variant to appear homozygous (VAF ≈ 100%)
• Often caused by SNPs under primer binding sites or probe hybridization regions
• Strand bias metrics and independent validation assays help identify dropout events

Deep learning callers like DeepVariant implicitly learn to recognize dropout signatures from pileup image patterns.