What Is Gene Frequency

How to Calculate Allele Frequencies in a Population

You pull a sample from a field population, genotype a handful of individuals at a single locus with two alleles, and then try to figure out what the allele frequencies actually are. That's the basic starting point. The math is straightforward enough that you can do it by hand for a small dataset, but the complications show up fast once you start working with real genomic data. Gene frequency, or allele frequency, is simply the proportion of a particular allele among all copies of a gene in a population. If you have a diallelic locus with alleles A and a, and you count 120 copies of A and 80 copies of a in a sample of 100 diploid individuals, the frequency of A is 0.6 and the frequency of a is 0.4. That's the definition. The useful part is what you do with those numbers. Hardy-Weinberg equilibrium gives you the shortcut. Under ideal conditions, genotype frequencies should be p², 2pq, and q² where p and q are the allele frequencies. You can estimate p directly by counting alleles, or you can infer it from observed genotype counts. The direct counting method is more robust because it doesn't assume equilibrium.

I ran into a problem last year working with a wild population of a Mediterranean plant species where we were tracking a resistance allele. We genotyped about 200 individuals using SNP arrays. When I calculated allele frequencies under Hardy-Weinberg assumptions, the expected heterozygosity didn't match the observed heterozygosity at all. The observed heterozygotes were way lower than p² plus 2pq predicted. I spent two days wondering if my genotype calls were bad before I realized we were dealing with a Wahlund effect. The sample wasn't one panmictic population. It was two subpopulations mixed together, each with different allele frequencies. Once I separated the samples by geographic origin and recalculated, the numbers made sense. The workaround was straightforward: use STRUCTURE or a similar clustering algorithm first to identify substructure, then calculate allele frequencies within each cluster instead of pooling everything together. This is the kind of thing that doesn't show up in introductory textbooks. You learn the formulas, you apply them, and then your results look wrong and you have no idea why.

The Mechanics of Calculation

Here's how the actual calculation works when you have raw genotype data. You need a VCF file or a simple spreadsheet with genotype calls at your locus of interest. For each individual, you record whether they're homozygous for allele 1, heterozygous, or homozygous for allele 2. Then you count. Every homozygote contributes two copies of that allele. Every heterozygote contributes one copy of each allele. Divide by the total number of allele copies in the sample, which is twice the number of diploid individuals, and you have your frequency. In R, this is almost trivial if you have a genlight object from the adegenet package. A single function call like freq() will give you allele frequencies across all loci. If you're working in Python, the cyvcf2 library lets you iterate through VCF files efficiently without loading everything into memory. For a dataset with thousands of SNPs and hundreds of samples, this matters because loading a large VCF into pandas will eat your RAM and slow everything to a crawl. One thing most people miss: missing data handling. If 20% of your genotypes are missing at a locus, you can't just ignore those individuals and recalculate the denominator. The effective sample size drops, and your frequency estimate gains variance. The standard approach is to exclude loci with high missingness, but the threshold depends on your study. For population genetics papers, I usually set a cutoff around 10% missing data per locus. Anything higher and the frequency estimates become unreliable, especially in small samples.

Another practical issue is batch effects. If you ran your samples across multiple sequencing lanes or genotyping arrays, the allele frequency estimates can drift between batches for technical reasons rather than biological ones. I've seen this in microbiome work and in plant population studies. The fix is to include batch as a covariate in your analysis or to normalize across batches before calculating frequencies. PopGenome in R has functions for this, though the documentation could be better. When you move beyond diploid organisms, the math changes slightly. For haploid organisms or for mitochondrial and chloroplast loci in plants, you don't multiply the sample size by two. Each individual contributes one allele. For polyploids, things get messier because you need to account for the dosage of each allele. Most standard software assumes diploidy, so if you're working with a tetraploid crop species, you'll need specialized tools like tetraploidSNPcall or custom scripts that handle multinomial genotype probabilities.

Common Pitfalls and What to Do Instead

The biggest mistake I see is treating allele frequency as a static number. It's not. It changes every generation through drift, selection, migration, and mutation. In a small population of fewer than a thousand individuals, genetic drift alone can shift allele frequencies by several percentage points in a single generation. If you're sampling a wild population once and reporting a single frequency value, you're giving a snapshot that may not represent anything stable. The confidence interval around your estimate, especially with modest sample sizes, can be surprisingly wide. A second pitfall is conflating allele frequency with genotype frequency. They're related but distinct. Saying an allele has a frequency of 0.3 doesn't tell you how many people are homozygous, heterozygous, or absent for that allele without additional assumptions. If you need genotype frequencies, you have to either observe them directly or assume Hardy-Weinberg equilibrium, which as I mentioned earlier, is often violated in real populations. FNA markers add another layer of complication. Microsatellites have multiple alleles per locus, sometimes dozens. Calculating frequencies is the same basic operation, but the interpretation changes. Expected heterozygosity becomes a more useful metric than single allele frequencies because no single allele dominates. In those cases, I usually report both the most common allele frequency and the Shannon diversity index, which captures the evenness of the distribution.

Software Options

For basic calculations, GenAlEx runs in Excel and is fine for small datasets and teaching purposes. For anything involving real research data, I'd recommend VCFtools for quick Linux-based allele frequency extraction from VCF files. The command vcftools --freq will output allele frequencies for every site in your VCF in seconds. For population-level summaries, PLINK is solid and handles large datasets well. If you're doing Bayesian clustering or need to account for uncertainty in allele frequencies, BayeScan or STRUCTURE are the standards, though they're computationally heavy. The download links for these tools are all publicly available. VCFtools is on SourceForge, PLINK is at the Broad Institute, and GenAlEx is available through the University of Sydney's website. None of them require paid licenses. One last thing that people overlook: allele frequency databases. If you're working with human populations, dbSNP and the 1000 Genomes Project give you reference frequencies that you can compare against. For non-model organisms, these resources don't exist, so you're building your own baseline from scratch. That's fine, but you need to be clear about it in your methods section. Reviewers will ask where your reference frequencies came from.