Hey friend,

I have to thank you for posting this question, as it has been a very interesting problem. The problem with pvclust is that it does not allow you to supply a pre-existing distance matrix, so, I thought that one could perhaps edit the pvclust code in order to trick it into accepting a distance matrix. I tried that this morning but the function would require major overhaul.

I then thought about 'reverse engineering' the distance matrix in order to produce original data. Your original data is genotype data, but instead we could produce numerical data that approximately equaled your genotype data. I then produced the following code:

#Create a 20 x 20 sample distance matrix of random numbers
MyData <- matrix(rexp(400, rate=0.5), ncol=20)
MyDist <- dist(MyData, method="euclidean")

#'MyDist' can also be produced from other programs, like PLINK on genotype values

#'Reverse engineer' the distance matrix in order to recover approximate co-ordinates that could have produced it
iSampleNumber <- 20
x <- cmdscale(MyDist, k=iSampleNumber-1)

pdf("pvclustWithExistingDistMatrix.pdf", width=7, height=3)
    par(mfrow=c(1,3), cex=0.5)
    plot(hclust(MyDist, method="ward.D2"), main="Using supplied distance matrix")

    plot(hclust(dist(x, method="euclidean"), method="ward.D2"), main="Using recovered co-ordinates\nfrom supplied distance-matrix")

    require("pvclust")
    plot(pvclust(t(x), method.dist="euclidean", method.hclust="ward.D2"), main="pvclust of recovered co-ordinates\nfrom supplied distance-matrix")
dev.off()

In your case, you would start with your distance matrix MyDist. As you can see, this code (using cmdscale) has reverse-engineered the distance matrix in order to produce a matrix of numbers that could have approximately produced your distance matrix.

image host

Source of inspiration: http://personality-project.org/r/mds.html