Quality control for SNP datasets

tidypopgen has two key functions to examine the quality of data across loci and across individuals: qc_report_loci and qc_report_indiv. This vignette uses a simulated data set to illustrate these methods of data cleaning.

Read data into gen_tibble format

library(tidypopgen)
data <- gen_tibble(system.file("extdata/related/families.bed",
  package="tidypopgen"), quiet = TRUE, backingfile = tempfile(),
  valid_alleles = c("1","2"))

Quality control for loci

loci_report <- qc_report_loci(data)
summary(loci_report)

##      snp_id         maf          missingness          hwe_p         
##  Min.   :  1   Min.   :0.0000   Min.   :0.00000   Min.   :0.004644  
##  1st Qu.:241   1st Qu.:0.1667   1st Qu.:0.00000   1st Qu.:0.340961  
##  Median :481   Median :0.2727   Median :0.00000   Median :0.523810  
##  Mean   :481   Mean   :0.2696   Mean   :0.03867   Mean   :0.507695  
##  3rd Qu.:721   3rd Qu.:0.3750   3rd Qu.:0.08333   3rd Qu.:0.739938  
##  Max.   :961   Max.   :0.5000   Max.   :0.25000   Max.   :0.796935

The output of qc_report_loci supplies minor allele frequency, rate of missingness, and a Hardy-Weinberg exact p-value for each SNP. These data can be visualised in autoplot :

autoplot(loci_report, type = "overview")

Using ‘overview’ provides an Upset plot, which is designed to show the intersection of different sets in the same way as a Venn diagram. SNPs can be divided into ‘sets’ that each pass predefined quality control threshold; a set of SNPs with missingness under a given threshold, a set of SNPs with MAF above a given threshold, and a set of SNPs with a Hardy-Weinberg exact p-value that falls above a given significance level.

The upset plot then visualises our 961 SNPs within their respective sets. The number above the first bar indicates that, of our total SNPs, 609 occur in all 3 sets, meaning 609 SNPs pass all of our automatic QC measures. The combined total of the first and second bars represents the number of SNPs that pass our MAF and HWE thresholds, here, 956.

The thresholds for each parameter, (level of missingness that is accepted etc) can be adjusted using the parameters provided in autoplot. For example:

autoplot(loci_report, type = "overview", miss_threshold = 0.03, maf_threshold = 0.02, hwe_p = 0.01)

To examine each QC measure in further detail, we can plot a different summary panel.

autoplot(loci_report, type = "all", miss_threshold = 0.03, maf_threshold = 0.02, hwe_p = 0.01)

We can then begin to consider how to quality control this raw data set. Let’s start by filtering SNPs according to their minor allele frequency. We can visualise the MAF distribution using:

autoplot(loci_report, type = "maf")

Here we can see there are some monomorphic SNPs in the data set. Let’s filter out loci with a minor allele frequency lower than 2%, by using select_loci_if. Here, we select all SNPs with a MAF greater than 2%. This operation is equivalent to plink –maf 0.02.

data <- data %>% select_loci_if(loci_maf(genotypes)>0.02)
count_loci(data)

## [1] 956

Following this, we can remove SNPs with a high rate of missingness. Lets say we want to remove SNPs that are missing in more than 5% of individuals, equivalent to using plink –geno 0.05

autoplot(loci_report, type = "missing", miss_threshold = 0.05)

We can see here that most SNPs have low missingness, under our 5% threshold, some do, however, have missingness over our threshold. To remove these SNPs, we can again use select_loci_if.

data <- data %>% select_loci_if(loci_missingness(genotypes)<0.05)
count_loci(data)

## [1] 609

Finally, we may want to remove SNPs that show significant deviation from Hardy-Weinberg equilibrium, if our study design requires. To visualise SNPs with significant p-values in the Hardy-Weinberg exact test, we can again call autoplot:

autoplot(loci_report,type = "significant hwe", hwe_p = 0.01)

Few SNPs in our data are significant, however there may be circumstances where we would want to cut out the most extreme cases, if these data were real, these cases could indicate genotyping errors.

data <- data %>% select_loci_if(loci_hwe(genotypes)>0.01)
count_loci(data)

## [1] 608

Once we have quality controlled the SNPs in our data, we can turn to individual samples.

#Quality control for individuals

individual_report <- qc_report_indiv(data)
summary(individual_report)

##     het_obs        missingness
##  Min.   :0.2206   Min.   :0   
##  1st Qu.:0.2227   1st Qu.:0   
##  Median :0.2347   Median :0   
##  Mean   :0.2351   Mean   :0   
##  3rd Qu.:0.2409   3rd Qu.:0   
##  Max.   :0.2591   Max.   :0

The output of qc_report_indiv supplies observed heterozygosity per individual, and rate of missingness per individual as standard.

These data can also be visualised using autoplot:

autoplot(individual_report)

Here we can see that most individuals have low missingness. If we wanted to filter individuals to remove those with more than 3% of their genotypes missing, we can use filter.

data <- data %>% filter(indiv_missingness(genotypes)<0.03)
nrow(data)

## [1] 12

And if we wanted to remove outliers with particularly high or low heterozygosity, we can again do so by using filter. As an example, here we remove observations that lie more than 3 standard deviations from the mean.

mean_val <- mean(individual_report$het_obs)
sd_val <- stats::sd(individual_report$het_obs)

lower <- mean_val - 3*(sd_val)
upper <- mean_val + 3*(sd_val)

data <- data %>% filter(indiv_het_obs(genotypes) > lower)
data <- data %>% filter(indiv_het_obs(genotypes) < upper)
nrow(data)

## [1] 12

Next, we can look at relatedness within our sample. If the parameter kings_threshold is provided to qc_report_indiv(), then the report also calculates a KING coefficient of relatedness matrix using the sample. The kings_threshold is used to provide an output of the largest possible group with no related individuals in the third column to_keep. This boolean column recommends which individuals to remove (FALSE) and to keep (TRUE) to achieve an unrelated sample.

individual_report <- qc_report_indiv(data, kings_threshold = 0.177)
summary(individual_report)

##     het_obs        missingness  to_keep              id       
##  Min.   :0.2206   Min.   :0    Mode :logical   Min.   : 1.00  
##  1st Qu.:0.2227   1st Qu.:0    FALSE:2         1st Qu.: 3.75  
##  Median :0.2347   Median :0    TRUE :10        Median : 6.50  
##  Mean   :0.2351   Mean   :0                    Mean   : 6.50  
##  3rd Qu.:0.2409   3rd Qu.:0                    3rd Qu.: 9.25  
##  Max.   :0.2591   Max.   :0                    Max.   :12.00

We can remove the recommended individuals by using:

data <- data %>% filter(id %in% individual_report$id & individual_report$to_keep == TRUE)

We can now view a summary of our cleaned data set again, showing that our data has reduced from 12 to 10 individuals.

summary(data)

##        id          population     genotypes        
##  Min.   : 1.00   Min.   : 1.00   Length:10         
##  1st Qu.: 3.25   1st Qu.: 3.25   Class :character  
##  Median : 5.50   Median : 5.50   Mode  :character  
##  Mean   : 5.80   Mean   : 5.80                     
##  3rd Qu.: 7.75   3rd Qu.: 7.75                     
##  Max.   :12.00   Max.   :12.00

Linkage Disequilibrium

For further analyses, it may be necessary to control for linkage in the data set. tidypopgen provides LD clumping. This option is similar to the –indep-pairwise flag in plink, but results in a more even distribution of loci when compared to LD pruning.

To explore why clumping is preferable to pruning, see https://privefl.github.io/bigsnpr/articles/pruning-vs-clumping.html

LD clumping requires a data set with no missingness. This means we need to create an imputed data set before LD pruning, which we can do quickly with gt_impute_simple.

imputed_data <- gt_impute_simple(data, method = "random")

In this example, if we want to remove SNPs with a correlation greater than 0.2 in windows of 10 SNPs at a time, we can set these parameters with thr_r2 and size respectively.

to_keep_LD <- loci_ld_clump(imputed_data, thr_r2 = 0.2, size = 10)
head(to_keep_LD)

## [1] FALSE FALSE FALSE FALSE  TRUE FALSE

loci_ld_clump provides a boolean vector the same length as our list of SNPs, telling us which to keep in the data set. We can then use this list to create a pruned version of our data:

ld_data <- imputed_data %>% select_loci_if(loci_ld_clump(genotypes, thr_r2 = 0.2, size = 10)) 

Save

The benefit of operating on a gen_tibble is that each quality control step can be observed visually, and easily reversed if necessary.

When we are happy with the quality of our data, we can create and save a final quality controlled version of our gen_tibble using gt_save.

gt_save(ld_data, file_name = tempfile())

## 
## gen_tibble saved to /tmp/RtmpgkbW6s/file1c0a625ce38b.gt

## using bigSNP file: /tmp/RtmpgkbW6s/file1c0a15e14c15.rds

## with backing file: /tmp/RtmpgkbW6s/file1c0a15e14c15.bk

## make sure that you do NOT delete those files!

## to reload the gen_tibble in another session, use:

## gt_load('/tmp/RtmpgkbW6s/file1c0a625ce38b.gt')

## [1] "/tmp/RtmpgkbW6s/file1c0a625ce38b.gt" 
## [2] "/tmp/RtmpgkbW6s/file1c0a15e14c15.rds"
## [3] "/tmp/RtmpgkbW6s/file1c0a15e14c15.bk"

Grouping data

For some quality control measures, if you have a gen_tibble that includes multiple datasets you may want to group by population before running the quality control. This can be done using group_by.

First, lets add some imaginary population data to our gen_tibble:

data <- data %>% mutate(population = rep(c("A", "B"), each = 5))

We can then group by population and run quality control on each group:

grouped_loci_report <- data %>% group_by(population) %>% qc_report_loci()
head(grouped_loci_report)

## # A tibble: 6 × 4
##   snp_id   maf missingness hwe_p
##    <int> <dbl>       <dbl> <dbl>
## 1      2  0.3            0 1    
## 2      3  0.2            0 1.11 
## 3      5  0.35           0 0.127
## 4      8  0.3            0 0.127
## 5     10  0.5            0 0.476
## 6     11  0.1            0 1

The loci report that we receive here will calculate Hardy-Weinberg equilibrium for each SNP within each population separately, providing a Bonferroni corrected p-value for each SNP.

Similarly, we can run a quality control report for individuals within each population:

grouped_individual_report <- data %>% group_by(population) %>% qc_report_indiv(kings_threshold = 0.177)
head(grouped_individual_report)

## # A tibble: 6 × 4
##   het_obs missingness to_keep    id
##     <dbl>       <dbl> <lgl>   <int>
## 1   0.223           0 TRUE        1
## 2   0.228           0 TRUE        2
## 3   0.223           0 TRUE        3
## 4   0.233           0 TRUE        4
## 5   0.236           0 TRUE        5
## 6   0.222           0 TRUE        6

This is important when we have related individuals, as background population structure can affect the filtering of relatives.