BSMS205 · Genetics

Data Types for
Alleles and Populations

Chapter 24 · Part IV · Population Genetics

Today's central question

How do you turn
a genome
into data?

The universal answer · VCF

Variant Call Format
Every variant caller outputs VCF
Every database stores VCF or a sibling
Every analysis tool accepts VCF

The universal language of genetic variation.

Roadmap for today

From sequencing to variant calling · the pipeline
VCF file structure · the core columns
The INFO field · population statistics
Genotypes · FORMAT and GT
Multi-sample VCF and joint genotyping
Scaling · dense vs sparse · the future

§ 1

From Sequencing
to Variant Calling

The pipeline

Extract DNA · sequence on Illumina → billions of short reads
Align to reference (GRCh38) → BAM / CRAM file
Variant call against reference → VCF file

Common tools: BWA · Bowtie2 · GATK · DeepVariant · FreeBayes

The pipeline is three stages. Stage one. You extract D N A from a sample and sequence it on a platform like Illumina. The output is millions or billions of short sequencing reads, each typically one hundred or one hundred fifty bases long. Stage two. You align those reads to a reference genome. For humans, the current standard is G R C h thirty-eight. Alignment is done by tools like B W A or Bowtie two. The output is a B A M file, or its more compressed version C R A M — a sorted list of every read and where it came from. Stage three. You compare the aligned reads to the reference at every position, and ask: does this sample differ from the reference here? That is variant calling. Tools like G A T K, DeepVariant, and FreeBayes perform this step. The output is a V C F file.

What the variant caller decides

Reference at chr1:12345 is A
Reads at this position show consistent T
Caller evaluates: read quality · mapping quality · strand bias · coverage
Decision: is this a real variant, or noise?

§ 2

VCF File Structure

A VCF is a plain text file

One line per variant
Each line: where the variant is, what changes, how confident
Tab-separated columns
Header lines start with #

Nine core columns

Column	Meaning
CHROM	Chromosome (chr1, chrX, ...)
POS	Position (1-based)
ID	Identifier (rsXXXX or `.`)
REF	Reference allele
ALT	Alternate allele
QUAL	Quality score
FILTER	PASS or failure reason
INFO	Site-level metadata
FORMAT	Per-sample field definitions

One line · read it out

chr1 1234567 rs12345 A T 99.0 PASS AC=3;AN=6;AF=0.5 GT:DP 0/1:32

chr1 pos 1,234,567 · ref A · alt T
Quality 99 · passes filters · rsID rs12345
3 alternate alleles in 6 total (AF = 50%)
This sample is heterozygous · 32-read depth

§ 3

The INFO Field ·
Population Statistics

Three fundamental fields

Field	Meaning
AC Allele Count	Number of alternate alleles observed
AN Allele Number	Total alleles successfully called · ideally 2 × N
AF Allele Frequency	AC ÷ AN

A concrete example

Sequence 100 people → 200 alleles total
10 alternate alleles observed

AC

AN

200

AF

0.05

Why these three numbers matter

Tell you how common or rare a variant is
Help decide pathogenic vs benign
Feed directly into GWAS, burden tests, population genetics

§ 4

Genotypes ·
FORMAT and GT

The FORMAT column

Defines what information appears for each sample
Colon-separated list of field names
Example: GT:DP:AD
Each sample column follows the same template

Genotype codes · the heart of VCF

Code	Meaning
`0/0`	Homozygous reference — no variant
`0/1`	Heterozygous — one ref, one alt
`1/1`	Homozygous alternate — both carry variant
`./.`	Missing — genotype not called

Phased vs unphased

Unphased · `/`

Know which alleles are present
Don't know which came from mom vs dad
The default from variant calling

Phased · `|`

Know which allele is on which chromosome
Essential for haplotype work
Requires extra inference or family data

Supporting fields · DP and AD

Field	Meaning	Example
DP	Total read depth	`DP=32`
AD	Allelic depth (ref, alt)	`AD=16,16`

0/1 + AD=16,16 → balanced heterozygote · consistent call
0/1 + AD=28,4 → suspicious · possible sequencing error

§ 5

Multi-Sample VCF
and Joint Genotyping

Multi-sample VCF (msVCF)

One row per variant site
One column per sample
INFO summarises population · each sample column gives individual genotype
Same file · population and individual views simultaneously

Three samples · one variant

Person	Genotype	T alleles
Alice	`0/1` · heterozygous	1
Bob	`1/1` · homozygous alt	2
Carol	`0/0` · homozygous ref	0

→ AC = 3 · AN = 6 · AF = 0.5

Variant ≠ Genotype

A variant is a site where an alternate allele exists in the population.
A genotype is an individual's specific allele combination.

One variant · many genotypes.
Variant = population-level · genotype = individual-level.

Joint genotyping

Call variants across all samples simultaneously
Even sites where most people are 0/0 get an explicit call
Prevents missing data from corrupting AC / AN / AF
Standard for gnomAD, UK Biobank, All of Us

The attendance analogy

Without joint genotyping

Only write down who answered present
Don't know who was absent
Incomplete information

With joint genotyping

Full roster of who is present or absent
Complete information
Correct AF calculations

§ 6

Scaling ·
Dense vs Sparse

The scaling problem

UK Biobank · sequenced ~500,000 people
Dense VCF: one genotype entry per sample per site
For a rare variant: 499,990 samples are 0/0
Storing all those 0/0s is wasteful

Sparse VCF

Store only non-reference genotypes
0/0 entries are omitted, implied from absence
Population stats (AC, AN, AF) still in INFO — no information loss
For 500k samples · rare variant stores ~10 entries, not 500,000

Dense vs sparse · a comparison

Aspect	Dense	Sparse
Storage	All genotypes explicit	Only non-ref genotypes
File size	Very large, superlinear growth	Linear with carriers
Best for	10s–1000s of samples	Biobank scale (100k+)
Computational feasibility	Breaks at biobank scale	Enables modern genomics

§ 7

The Future of VCF

The fundamental challenge

Adding new samples means updating every existing row
File size grows superlinearly with sample count
gnomAD export in VCF = petabyte range
Doubling gnomAD → infeasible in standard VCF

GA4GH Future of VCF Working Group

Global Alliance for Genomics and Health · stewards the VCF standard
Future of VCF group formed 2019, meets monthly
Interoperability with htsget · Beacon · CRAM / SAM / BAM

Three approaches to scaling

Specification tweaks · SAV, SVCR, spVCF · keep the text format, compress smartly
New binary format · engineered for large scale and targeted queries
API-based hybrid · query protocol returns data in any VCF form

Three broad approaches are on the table. First, specification tweaks — modifications to the V C F format itself that make it scale linearly or even sublinearly. Examples include Sparse Allele Vectors, or S A V, Scalable Variant Call Representation, or S V C R, and sparse Project V C F, or s p V C F. These keep the familiar text format but compress smartly. They are easy to adopt but depend on the community coalescing on one choice. Second, a completely new binary file format, engineered from scratch for large-scale data and targeted queries. Harder to adopt because it requires updating every library that reads or writes V C F. Third, an A P I-based hybrid — a query protocol over a backing store, returning data in whatever V C F flavour the client asks for. This third approach is considered the most promising path forward.

Why this scaling effort matters

Enables reuse of large population-genomics datasets
Supports virtual cohorts — combining studies to gain statistical power
Especially valuable for rare disease research

§ 8

Why VCF Matters

Before VCF · the wild west

Every lab used its own format
Cross-study comparison was a nightmare
Custom converters for every pair of tools
Reproducibility was hard

After VCF · the universal standard

Directly compare variants across labs and technologies
gnomAD, ClinVar, dbSNP distribute in VCF
Clinical reporting · research · everything
The same format scales from one exome to one million genomes

§ 9

Summary

What to take away

Pipeline: reads → BAM → VCF
VCF = plain text · nine core columns · one line per variant
INFO: AC · AN · AF — population-level statistics
FORMAT + GT: 0/0 · 0/1 · 1/1 · ./.
Joint genotyping = complete data across all samples
Sparse VCF = future-proof scaling to biobank size

The bigger picture

VCF is how genomes become data — how biological molecules become numbers that computers can analyse, compare, and share.

From a single A→T change
…to biobanks with millions of variants across hundreds of thousands of people
Encoded as 0/0, 0/1, 1/1

End of Part IV

Allele frequency · population structure ·
linkage · recombination · VCF

Next: Part V · Functional Genetics

This is also the end of Part Four, Population Genetics. Let me summarise the arc of the whole part. Chapter twenty taught us allele frequency — the fundamental statistic of population genetics. Chapter twenty-one showed why the same variant can have different frequencies in different populations, and introduced the concept of population structure. Chapter twenty-two traced the history from Mendel to Morgan, and showed how linkage was discovered. Chapter twenty-three dove into the modern understanding of recombination, including non-cross-overs, hotspots, and haplotypes. And today, chapter twenty-four, closed the loop by showing how all this information is encoded in V C F files. You now have the full conceptual and computational toolkit of population genetics. In Part Five, we move into functional genetics — how individual genes are studied, edited, and engineered. See you then.