Lecture 2 of the Molecular Genomics series. How genotypes are produced, how they are coded, and where the errors come from.
The previous lecture ended with the claim that SNP markers are cheap enough to score in large numbers. This lecture is about how that is done, because the details determine what the data look like, which errors occur, and why the first steps of any analysis are quality control.
The dominant technology is the microarray, marketed by Illumina as the BeadArray or BeadChip and referred to universally as the SNP chip. A chip carries hundreds of thousands to millions of microscopic silica beads held in etched wells. Each bead is coated with many copies of a short oligonucleotide probe, and each probe is designed to be complementary to the sequence immediately adjacent to one particular SNP in the genome.
Which SNPs appear on a chip is a design decision made once by the manufacturer, in consultation with the research community. This has an important consequence.
Each assayed position has exactly two possible alleles, conventionally called A and B, and therefore exactly three possible genotypes, AA, AB and BB, plus the possibility of a failed call. This is a property of the assay, not of the genome. Positions with three segregating alleles exist but are not put on chips, because the two-allele design is what makes the readout simple.
It also means a chip can only ever report on variation that was known when the chip was designed. A SNP private to an African indigenous breed will not be on a chip built from European dairy cattle. This is ascertainment bias, and it matters a great deal for diversity work; it returns in Lecture 7.
The assay is a single-base extension reaction, read optically.
The output at this stage is not a genotype but a pair of intensities. Converting intensities to genotypes is a statistical step, and it is where errors enter.
Plot every animal for a single SNP, with one intensity channel on each axis. If the assay works, the points fall into three clouds: one where only the A signal is present (genotype AA), one where only B is present (BB), and one intermediate cloud where both are present at roughly equal intensity (AB). Calling genotypes means assigning each point to a cloud.
The calling algorithm fits clusters automatically and attaches a confidence score to each call, the GenCall score in Illumina's implementation. Calls below a threshold, conventionally 0.25, are left missing rather than guessed. A SNP whose clusters do not separate produces many low-confidence calls and should be removed from the dataset entirely.
This is the origin of two of the standard quality-control filters. Call rate per SNP removes markers that failed to cluster; call rate per animal removes samples whose DNA was poor. Both are applied before any analysis, and both are covered practically in the Genomics Boot Camp.
When genotyping is finished the machine produces a text file, the final report. It is unglamorous and it is where every downstream analysis begins. From this point on the work is computational: the file is converted into the format expected by PLINK or another package, filtered, and analysed.
Two practical points follow. Datasets of this size cannot be inspected by hand, so some programming ability is not optional. And because many programs exist and each expects its own format, a large fraction of real analysis time is spent converting between them rather than doing statistics.
The same genotype can be written in several ways, and confusing them is a common source of error when combining datasets.
| Scheme | Homozygote 1 | Heterozygote | Homozygote 2 |
|---|---|---|---|
| Nucleotide | AA | AG | GG |
| A/B | AA | AB | BB |
| Numeric allele | 11 | 12 | 22 |
| Genotype count | 0 | 1 | 2 |
The last row is the one that matters statistically. Coding a genotype as 0, 1 or 2 counts the number of copies of a chosen reference allele carried by the animal. That count is a number, so it can enter a regression directly, and it is the form in which genotypes appear in every model in Lecture 6 and Lecture 10.
Chips are species-specific and come in several densities. For cattle the established tiers are roughly seven thousand markers (low density), fifty-four thousand (medium, the workhorse, usually written 50K), and seven hundred and seventy-seven thousand (high density, HD). Custom chips targeting particular regions also exist.
| Species | Approximate SNP count on a common chip |
|---|---|
| Human | 906,000 |
| Cattle | 54,000 (medium), 777,000 (high density) |
| Sheep | 54,000 |
| Pig | 60,000 |
| Horse | 70,000 |
| Dog | 230,000 |
| Mouse | 140,000 |
| Wheat | 660,000 |
| Strawberry | 95,000 |
Density is chosen by trading cost against information. More markers give finer resolution and better coverage of rare variants, but multiply the per-animal cost. For most breeding applications a medium-density chip is sufficient, because the missing markers can be reconstructed computationally rather than bought. That reconstruction is imputation, and it is Lecture 4.
It is worth stating plainly at this early stage: in the age of cheap genotypes, the phenotype is the limiting resource. Genotyping is now routine and inexpensive relative to the cost of measuring milk composition, feed intake, methane emission or disease incidence on the same animals.
Every method in this course estimates the relationship between genotype and phenotype. None of them can extract more information than the phenotypes contain. A large genotyped population with poorly recorded traits will give a weaker answer than a smaller population with careful phenotypes, and no amount of marker density repairs that.
Reading a cluster plot. For a given SNP, 480 animals form three tight clusters but 40 animals fall between the AB cluster and the BB cluster and receive GenCall scores near 0.2. What happens to those 40 calls under the standard threshold, and what should you check before deciding the SNP is fine?
With a threshold of 0.25 those 40 calls are set to missing rather than assigned. The SNP therefore has a call rate of \(480/520 = 92.3\%\), which fails the commonly used 95% filter and would normally be dropped.
Before accepting that, check whether the 40 animals are a random subset or a group. If they are the same 40 animals failing at many SNPs, the problem is the samples, not the markers: poor DNA quality, a bad plate, or a processing batch. Removing those animals first may rescue a large number of SNPs that would otherwise be discarded. If instead the 40 are different animals at each SNP, the marker itself is unreliable and should go.
The general point is that per-SNP and per-animal call rates are not independent, and the order in which the filters are applied changes the result.
Merging two datasets. Two laboratories genotype the same breed on the same 50K chip and send you files. In file 1, SNP rs123 is coded AA/AG/GG; in file 2 it is coded 0/1/2. You merge them and later find that this SNP has a large effect of opposite sign in the two halves of the data. What is the most likely explanation, and how would you detect it systematically?
The most likely explanation is that the two files count different reference alleles. If file 1 counts copies of G and file 2 counts copies of A, then an animal that is 0 in one coding is 2 in the other. Regressing a phenotype on a variable that has been reversed reverses the sign of the estimate, which is exactly what is observed.
The systematic check is allele frequency. Compute the frequency of the counted allele separately in each file. For a shared breed these should agree closely; markers where they are approximately \(p\) and \(1-p\) are flipped. Standard merge tools apply this check automatically.
The residual difficulty is A/T and C/G SNPs, where flipping the strand produces a valid-looking allele pair and the frequency check fails when \(p \approx 0.5\). These are usually excluded from merged datasets rather than resolved.
Choosing a density. A breeding programme in Ethiopia can genotype 2,000 animals on a 50K chip or 400 animals on a 777K chip for the same budget. Which would you recommend for estimating genomic breeding values, and what would change your answer?
For breeding value estimation, choose the 2,000 animals on 50K. The accuracy of genomic prediction is driven far more by the size of the reference population than by marker density, once density is high enough that markers are in reasonable linkage disequilibrium with the causal variants. In cattle, 50K is generally sufficient for within-breed prediction; going to HD typically buys a small percentage in accuracy while cutting the reference population by a factor of five.
Three things would change the answer. If the population has unusually short-range linkage disequilibrium, as many African indigenous populations do because they have larger effective population sizes and no history of intense selection, then 50K may not blanket the genome and higher density becomes necessary. If the goal is fine-mapping a causal variant rather than prediction, resolution is the whole point and density wins. And if a high-density reference already exists in a related population, the 2,000 animals can be genotyped at 50K and imputed up, which gets both.
Ascertainment. A commercial cattle chip was designed using variation discovered in Holstein, Angus and a few other European breeds. You use it to compare genetic diversity between a European breed and an East African zebu population, and find the zebu population is less diverse. Should you believe this?
No, not without further work. The chip contains SNPs that were polymorphic in the discovery panel. Variants private to zebu populations were never candidates for inclusion, so a real source of zebu diversity is invisible by construction. Meanwhile every marker on the chip is guaranteed to be variable in European taurine cattle.
The expected consequence is precisely the one observed: apparently reduced heterozygosity and fewer segregating markers in the non-discovery population, regardless of the true diversity. Since African cattle populations generally carry more sequence-level diversity than commercial European breeds, the result is likely to be the reverse of the truth.
Defensible options are to use sequence data, which has no ascertainment step; to use a chip designed with the population in question represented in discovery; or to restrict comparisons to statistics that are less sensitive to ascertainment and to state the caveat explicitly. This problem is a recurring obstacle in African animal genetic resource work.
