Allele Frequency

Allele frequency is the fraction of chromosomes in a defined population that carry a given allele at a given position. An allele is one of the possible sequences at a site: at a SNP where the reference base is C and some people carry T, C and T are the two alleles. You have two copies of most of your genome, so you carry two alleles per autosomal site. In a population of N diploid people there are 2N chromosomes at that site, and the frequency of the T allele is (count of T chromosomes) / 2N.

How it works

Count, divide, done. If 1,200 of 10,000 sampled chromosomes carry T, the frequency of T is 0.12 and the frequency of C is 0.88. Frequencies at a site sum to 1 across all observed alleles. A frequency of 1.0 means the allele is fixed: every sampled chromosome carries it and there is no variation left at that site in that sample. A frequency of 0 means the allele was never seen, which for a finite sample means “absent or rarer than roughly 1/2N”.

The number is always relative to a population, and that qualifier carries most of the information. A variant at 0.3 in Finns can sit at 0.02 in Yoruba. Large surveys show that the bulk of common variation is shared across continental groups while rare variants are heavily geographically restricted, which is why frequency lookups without a matched reference panel mislead you ¹. Earlier catalogs like ALFRED were built specifically to store frequencies per population rather than as a single global number ².

Minor allele frequency (MAF) is the frequency of the less common allele at a site, so it runs 0 to 0.5. Convention splits variants roughly at MAF 0.05 (common) and 0.01 (rare), though the cutoffs are arbitrary and you should pick them for your analysis rather than inherit them.

Frequencies move over time. Drift moves them randomly, with variance per generation of p(1−p)/2N, so small populations lose variation fast. Selection moves them directionally, and with time-series sampling you can estimate selection coefficients from the trajectory itself ³. Sites with obviously damaging predicted effects tend to sit at low frequency because selection removes them, which is the empirical basis for treating high-frequency variants as unlikely to be severely deleterious ⁴.

On the inbreeding question people ask alongside this: there is no single magic number. The genetic rule of thumb used in conservation work is an effective population size in the tens for short-term avoidance of inbreeding depression and in the hundreds to low thousands for maintaining adaptive variation. Effective size is usually well below census size.

In your own data

Your small variant calls arrive as VCF or gVCF. Two different frequency-like numbers live there and confusing them is the most common mistake:

• AF in the INFO field after annotation with gnomAD, ClinVar, or dbSNP means population allele frequency. Field names vary by pipeline: gnomAD_AF, AF_popmax, gnomad41_genome_AF_nfe.
• AF or VAF in the FORMAT/sample column means the fraction of your reads supporting the alt allele in you. For a germline heterozygote it should sit near 0.5, for a homozygote near 1.0.

Annotate with a specific tool and specific flags rather than a black box:

bcftools annotate -a gnomad.genomes.v4.1.sites.vcf.bgz \
  -c INFO/AF:=INFO/AF_joint,INFO/AF_grpmax \
  -Oz -o mine.annotated.vcf.gz mine.vcf.gz

Then filter. A rare-variant pass on 4–5 million germline calls typically looks like: keep FILTER=PASS, GQ >= 20, DP >= 10, and gnomAD AF < 0.001 in the population closest to your ancestry, plus a predicted protein consequence. That takes you from millions of sites to a few hundred.

Things that go wrong:

• Missing AF annotation is not AF = 0. Sites absent from gnomAD may be absent because of low mappability, not rarity. Check bcftools query -f '%CHROM %POS %INFO/AF\n' and count the dots.
• Using a global frequency when your ancestry is underrepresented. AF_grpmax (the highest frequency across groups) is the safer filter for ruling variants out.
• Reference-allele-is-minor sites. At several hundred thousand positions the GRCh38 reference carries the rare allele, so “matches reference” does not mean “common”.
• Comparing frequencies across builds or callers without liftover and left-alignment. Run bcftools norm -f GRCh38.fa -m -any first.

Limitations

Frequency tells you how often an allele appears, not what it does. Common variants can carry real effect, rare variants are frequently neutral, and the correlation between frequency and consequence is a tendency rather than a rule ⁵. Reference databases are sample-dependent: they skew toward European ancestry, exclude some clinical cohorts, and have uneven coverage in repeats and segmental duplications. Frequency estimates also assume the sample is a random draw from the population, which breaks when the cohort is ascertained for disease ⁶. Interpretation of any specific variant in a health context belongs with a clinician or a certified genetics lab, not with your filter chain.

Woolf Software builds longitudinal molecular profiles of individuals: whole-genome sequencing, RNA sequencing, proteomics, blood biomarkers, and continuous glucose data, integrated into one model of you. Build your profile.