Recently I was looking at a 3-D PCA animation which Zack generated from the Harappa Ancestry Project data set. Click the link and come back. Notice the outlier clusters? The Burusho are straightforward, they seem to have low levels of Tibetan admixture. But what about the Gujarati cluster? Again, we see what we’ve seen before, the fractioning out of the Gujaratis in PCA into two groups, one a tight cluster, and the other relatively widely distributed. This prompted me to look more closely at the HapMap Gujarati sample. Today I was exploring the question with Plink’s identity-by-descent feature. First I’ll start out with a smaller data set, my family (father, mother, sibling 1, sibling 2, and myself), and an Indian (from Uttar Pradesh) and Pakistani as unrelated individuals. I merged out 23andMe derived genotypes, and with ~900,000 markers calculated pairwise IBD:
./plink --bfile IBDControl --genome
Here are the relevant results:
|Individual 1||Individual 2||Z0||Z1||Z2||PI_HAT||DST||PPC||RATIO|
You can infer some things without even knowing what the columns mean. Notice that there are differences between parent-child, sibling-sibling, and unrelated comparisons. The distance measure, DST, is basically exactly the same as the genome-wide comparison in 23andMe. Either the web app is running Plink, or, it’s using the exact same algorithm. Z0 = IBD 0, Z1 = IBD 1, and Z2 = IBD 2. Notice that with my siblings I have a fair amount of IBD 2, but far less with my parents. That’s because parents give you one copy, but you can share zero, one, or two, of a gene with your siblings. In contrast, with our parents there is hardly any IBD = 0, since they’re guaranteed to give you one copy. I assume that the IBD = 2 in that case is population wide fixation of a variant. Notice in the last column that there are different values for unrelated individuals (~2), siblings (~10), and parent-children (~500).
I ran a similar test among the Gujaratis. Remember that I’ve labeled them Gujarat_A and Gujarati_B based on PCA clusters, where the latter form a tight population cluster, and the former do not. Here are the mean pairwise DST values with the groups of pairs:
Mean of all: 0.746
Mean of Gujarati_A only: 0.744
Mean of Gujarati_B only: 0.749
Mean of Gujarati_A and Gujarati_B pairs only: 0.745
Gujarati_B are marginally closer to each other than Gujarati_A. I’m not sure these DST values are totally comparable to the ones from the 23andMe files. I’ll show you why. I constrained the pairs to those where the RATIO was > 2.5. Here’s what I found:
|Individual 1||Individual 2||Z0||Z1||Z2||PI_HAT||DST||PPC||RATIO||PopX||PopY|
Notice that Z2 ~ 0, in contrast to the calculations above. I assume someone reading this knows that there’s a simple reason for this, so do tell. The IBD estimates for 23andMe always struck me as too high. In any case, to my surprise the definitely related individuals seem to be in the Gujarati_A cluster! What’s going on there? My first thought is that I messed up the data, or, I coded something incorrectly. I assume that this was double-checked before it got into the HapMap data set. Has anyone else seen this weird result? If not, I assume I made an error (that’s kind of my working model right now actually).