SNPs with missing genotypes result in all-NA GWAS output

[xiaolongsong00]

Hello,

I am using FarmCPUpp with my own genotype (GD), genotype map (GM), and phenotype (Y) files.
When I run farmcpu(), I notice that some SNPs in the GWAS output have all results as NA:

From my observation, these NA results happen whenever the corresponding SNP in the GD file contains at least one missing value (NA).

My question

• Does FarmCPUpp currently support handling SNPs with missing values?
• Should missing values in the numeric GD file be represented as NA, -9, or something else?
• If missing data are not supported, what is the recommended pre-processing approach (e.g. imputation, filtering)?

Example of my workflow

library(bigmemory)
library(FarmCPUpp)

myY <- read.table("taxa.txt", header = TRUE, stringsAsFactors = FALSE)
myGM <- read.table("mdp_SNP_information.txt", header = TRUE, stringsAsFactors = FALSE)

myGD <- read.big.matrix("mdp_numeric.txt",
                        type = "double", sep = "\t", header = TRUE,
                        col.names = myGM$SNP, ignore.row.names = FALSE,
                        has.row.names = TRUE,
                        backingfile = "mdp_numeric.bin",
                        descriptorfile = "mdp_numeric.desc")

myResults <- farmcpu(Y = myY, GD = myGD, GM = myGM)



Could you please clarify:

What is the correct way to represent missing genotypes in the numeric GD file?

If missing values are not supported, is imputation required before running FarmCPUpp?

Thank you for your time and for maintaining this package!

[amkusmec]

It looks like this isn't stated clearly in the documentation. Sorry for the confusion.

• No, FarmCPUpp does not support missing genotype calls.
• Genotype values should be real-valued entries in the range $[0,2]$ (inclusive). The predictors are assumed to be the counts of the alternative alleles per Fisher's single-locus model.
• All missing values should be imputed with the expected allele count. The usual approach is to assume a binomial sampling model for the alleles where the minor allele frequency denotes the probability of success, $p$, and the ploidy of the organism is the number of trials, $n$. Then the imputed allele count is $np$ or $2p$ for the usual case of a diploid organism.

The following R snippet shows how to do this. You'll need to adapt it to the big.memory functionality:

myGD <- apply(myGD, 2, function(x) {
  non_miss <- sum(!is.na(x))
  p <- sum(x[non_miss])/(2*non_miss)
  x[is.na(x)] <- 2*p
  x
})