Allele And Gene Frequency

Hard Data or Soft Approximation

I learned population genetics from a textbook and then tried to apply it to real sequencing data. The textbook worked perfectly. The data was a mess. Allele And Gene Frequency is the simple act of asking how common a variant is in a population. At the surface level, it is basic arithmetic. At the practical level, it is one of the most fragile calculations in genomics because every assumption you make quietly changes the result.

What Allele And Gene Frequency Actually Means

A gene is a locus. An allele is what sits at that locus. Gene frequency (more commonly called allele frequency) is the proportion of all chromosomes in a sample that carry a specific allele. If you have a biallelic SNP and your sample has 200 diploid individuals, you have 400 alleles total. If 60 of those carry the alternative allele, the frequency is 60 divided by 400, which is 0.15. That is the math. The difficulty is in the steps before and after.

Where It Breaks in Practice

I spent three weeks debugging a frequency pipeline once because the variant callers were not handling low-complexity regions correctly. A simple homopolymer run in a bacterial genome was being split into two adjacent variant calls instead of one indel. The frequency estimate doubled. Not close. Double. The fix was not a better formula. It was filtering out sites where mapping quality dropped below 30 and alt alleles mapped with a read bias above 0.75. That alone cut false variants by about 60 percent in those regions. Still not perfect. Better. People skip the boring quality filters because they want answers now. The answers are wrong. It takes ten minutes to add the filters and three days to realize you built something on sand.

The Standard Workflow

Start with a VCF or BCF file from your caller of choice. bcftools query is the fastest way to extract raw counts without loading everything into R. A command like this pulls genotype depth and allele depth for each sample: bcftools query -f '%CHROM\t%POS\t%REF\t%ALT\t%AD\n' variant.vcf.gz Sum the alternate allele depth across all samples, divide by the sum of total depth, and you get a rough allele frequency. This works fine for high-coverage germline data. Do not use it on somatic tumor samples without adjusting for purity and copy number. Tumor purity of 40 percent will make a heterozygous variant look like it has a frequency near 0.12 instead of 0.5. Anyone who tells you otherwise has never touched clinical sequencing data.

For population-level work, GATK's GenotypeGVCFs or bcftools +fill-tags give you genotype-level frequencies directly. The output includes INFO fields like AF, AC, and AN. AC is the alternate allele count. AN is the total number of alleles. AF is AC divided by AN. That is the whole pipeline if your input is clean.

Common Pitfalls Nobody Warns You About

The first issue is reference bias. Variant callers prefer reads that match the reference genome. If your population has a high frequency of the alternative allele at a locus, the caller will under-call the alt allele and your frequency estimate will be systematically too low. This is not a minor effect. I have seen frequencies shift by 0.08 to 0.12 purely from reference bias in diverse cohorts. The second issue is batch effects. Run your samples on different flow cells, different library prep kits, or even different dates, and allele frequency estimates drift. Not dramatically. Just enough to matter when you are comparing cases to controls and your p-values are borderline. I solved this once by regressing allele frequency on batch as a covariate before doing any association test. The coefficient for batch was tiny but significant at 0.004 per unit. Removing that signal changed three variants from significant to not significant. Worth knowing.

Allele And Gene Frequency in Structured Populations

If your population has structure, raw frequency estimates are misleading for association studies. Use FST or principal components to account for stratification. PLINK --freq gives you basic allele frequencies, but it does not correct for population structure. Eigenstrat or PCA-first approaches are the standard workaround. It adds about twenty minutes to a typical analysis but prevents false positives that would otherwise waste months of validation. Pseudogenes and segmental duplications are hard stops for standard frequency estimation. Reads map ambiguously. The variant caller assigns them randomly. Your frequency is noise. There is no clean workaround except dropping those regions entirely or using a specialized mappability mask from GEM or UCSC. Low-coverage whole genome sequencing below 4x has a different failure mode. Genotype calling becomes probabilistic. The allele frequency from hard-called genotypes is biased toward the extremes. Use imputation with a reference panel like TOPMed or UK Biobank instead. The output is still an estimate, but it is a calibrated one.

A Quick Practical Example

Say you have a cohort of 500 individuals and you want the frequency of a specific SNP. You call variants with GATK HaplotypeCaller, joint-genotype with GenotypeGVCFs, then run: bcftools query -f '%CHROM\t%POS\t%REF\t%ALT\t[%GT]\n' cohort.vcf.gz | awk '{for(i=4;i=NF;i++){if($i=="0/1")c++; if($i=="1/1")c+=2} n=(NF-3)*2} END{print c/n}' This gives you a raw frequency. Filter out sites with mean depth below 10 or above 150, remove sites with missingness above 5 percent, and recalculate. The frequency will change slightly. Usually by less than 0.01 for common variants. For rare variants below 0.01, the change can be larger because each missing sample matters more.

Summary of What Matters

The math is trivial. The data is not. Filter aggressively. Check for reference bias. Account for batch effects. Drop unmappable regions. Use imputation for low-coverage data. The workflow takes longer than you think because the cleanup is the work. Anyone who says otherwise is either working with simulated data or has not published yet.