DAPAB W5D2: Curation of genotypes

Call rates (per SNP and per individual)
Call rate

Proportion of genotypes that are called (observed)

SNPs with low call rate may have poor genotyping quality

Call rates (per SNP and per individual)
Calling of SNP genotypes from SNP arrays (in diploids)

Call rates (per SNP and per individual)

Missing rate

Proportion of genotypes that are not called

Individuals with low call rate may have

Poor quality DNA

Contaminated DNA samples

Some (rare) disorder: chimera, aneuploidy

Call rates (per SNP and per individual)

Calculate proportion observed genotypes per SNP

Calculate proportion observed genotypes per individual

Determine call rate threshold

Typically, 95% or greater, depending on the genotyping quality

Plot the distribution of observed call rates

Determine a threshold based on discontinuity of the distribution

Minor allele frequency

SNPs have two alleles → two allele frequencies: p+q=1

MAF is the frequency of the least frequent allele; MAF=min(p,q)

Loci with (very) low MAF:

• May not be informative

• Removing too many may affect outcome subsequent analyses

• May be due to genotyping erros

Determining a MAF threshold

Very low MAF may be due to genotyping errors

• Expected error rate could be sensible MAF threshold

If an allele is observed more often, the probability that these are (only) due to genotyping errors decreases

So:

• Decide how often an allele should be seen to be credible

• Convert this number into a MAF threshold

MAF should always be >0

• To at least remove non-segregating SNPs (with MAF=0)

Consider the number of genotyped individuals

• BASE MAF on minimum number observations for an allele

Determing a MAF threshold - example

• We have 1000 genotyped (diploid) individuals

• A locus with

  • All animals having (true) genotype GG
    A genotyping error rate of 0.1%

  • Two individuals have a called genotype GA

For example:

• MAF = 100%*(2/2000) = 0.1%

For the example:

• Allele should be seen 5 times to be credible

  • 100%/(5/2000)=-0.4%

Hardy-Weinberg equilibrium

Example

(Departures from) Hardy-Weinberg equilibrium

Sometimes loci that are not in H-W are removed

Useful to remove awkward SNPS:

• Only 2 out of 3 genotyeps are observed

• Only heterozygote individuals are observed

• A liberal H-W threshold would remove those

But this is a bit tricky → Why could this be

• Inbred populations, no random matings

Comparison of genotype to pedigree data

(Mendelian Inconsistencies)

Two main (complementary) approaches

Genomic vs pedigree relationships

Mendelian inconsistencies

Genomic vs pedigree information

Genomic informaiton may replace pedigree information in selection

Is there any use for pedigree information?

• Can be used for prediction (i.e. single-step BLUP)

  • And for quality control

Pedigree yields expected relationships

• Full sib relationship are always 0.5 (or greater, with inbreeding)

Genomic information yields actual relationships

• Full sib relationship vary around 0.5

Genomic relationships capture more variance

• Can be simply visualized by plotting

How to clean the data

Detect discrepancies between pedigree and genomic data

• Mendelian inconsistencies

Duplication of samples (top left)

Monozygotic twins of duplication of samples (top middle)

Sample swaps or pedigree error (of full-sin or parent/offspring) right bottom

Mendelian inconsistencies (MI)

Identify animals with conflicting pedigree and SNP info

Identified by counting loci with opposing homozygotes between 2 animals

Identification is straightforward for parent-offspring pairs

• Expected number of loci with opposing homozygotes = 0

MI - parent-offspring pairs

MI - paternal half sibs (dams unknown)

If we do not know the sires genotype (and dams)

Determine threshold # opposing homozygotes

Observed (histogram)

Expected (allele frequencies) (red)
Thresholds (empirical) (green)

Possible causes of errors in the data

Wrong animal ID

Wrong parent ID(s)

• Mistake in registration

• Swapping straws of semen

Mislabelling/mixing up of DNA samples

• When taking DNA sample

• In genotyping process (i.e. swapping batches)

Algorithm to clean the data

1. Detect Mendelian Inconsistencies (starting: parent-offspring)

2. Detect which animal causes most inconsistencies

   1. Pedigree or genotype error?

3. Remove one of the following for this animal

   1. Genotypes

   2. Pedigree information

   3. Genotypes + pedigree

4. Repeat 1-3 untill all consistencies are removed

5. Evaluate A vs G relationships after cleaning

Use Mendelian inconsistencies to remove SNPs

Mendelian inconsistencies can also be used as quality control for genotype calling of individual SNPs

SNPs that show many inconsistencies in otherwise consistent parent-offspring pairs should be removed

• i.e. remove SNP with > 2% inconsistencies between parent and offspring

More powerful methods

SNP based tests are sufficient for close relationships

• Parent-offspring

More precise tests are required for more distant relationships

• Grandparent-grandoffspring

• Greatgrandparent-greatgrandoffspring

Based on counting long shared haplotypes

Long-shared haplotypes

Alternative when there is increasing distance between relatives

Mendelian inconsistencies (by counting opposing homozygotes) provide a …

straightforward method for quality control of genotype data

Power of the method decreases rapidly with

increasing distance between relatives

Pruning for (near) complete LD

Adjacent loci may have high LD (i.e. r² values close to 1.0)

How to remove this redundancy?

Approach

• Determine an r² threshold; usually 0.95 or 0.99

• Remove one SNP from adjacent pairs exceeding the threshold

  • Using a sliding window approach (see screenshot)

Dealing with (left-over) missing genotypes

Due to

• No genotype call

• Genotype removed due to mendelian inconsisteny

How to deal with those in subsequent analyses

• Randomly assign based on gentoype frequencies :(

• Assign the mean genotype :/

• Impute using family or linkage disequilibrium (LD) info :)

Imputation using

Family information (parent-offspring, or further)
Linkage disequilibrium information

Both

Imputation using family information

Find the parental haplotypes inherited by the offspring

Fill in (impute) the allele not observed in the offspring

Imputation using LD information

Principle the same as using family information

Now compare to all haplotypes observed in the population

Impute best mathc, or matching haplotype with highest frequency