Category Archives: PCA - Page 3

Harappa Participants on 3-D PCA

sv wanted to see where he was on the South Asian 3-D PCA plot, so I obliged.

It's a quick and dirty method, but you should see a dropdown select box under the 3-D plot. Just select one of the participant IDs from there and that person's dot on the 3-D plot should increase in size so that it's easier to spot.

Ref 1 South Asians + Harappa PCA

I ran PCA on the South Asian populations included in Reference I dataset (excluding Kalash and Hazara) as well as 38 South Asian participants of Harappa Project. I excluded Kalash and Hazara because they usually dominate a South Asian PCA plot being so distinct.

The reference populations included are: Balochi, Bnei Menashe Jews, Brahui, Burusho, Cochin Jews, Gujaratis (divided into two groups), Makrani, Malayan, North Kannadi, Paniya, Pathan, Sakilli, Sindhi, and Singapore Indians.

Here's the spreadsheet showing the eigenvalues and the first 15 principal components for each sample.

I computed the PCA using Eigensoft which removed 26 samples as outliers. The Tracy-Widom statistics show that about 30 eigenvectors are significant.

Here are the first 15 eigenvalues:

1 3.874124
2 1.819077
3 1.663232
4 1.335721
5 1.293500
6 1.242984
7 1.230921
8 1.225775
9 1.222177
10 1.214539
11 1.212808
12 1.204000
13 1.198930
14 1.195450
15 1.192848

Here is a 3-D PCA plot (hat tip: Doug McDonald) showing the first three eigenvectors. The plot is rotating about the 1st eigenvector which is vertical. Also, I have stretched the principal components based on the corresponding eigenvalues.

Now here are plots of the first 14 eigenvectors. In this case, I have not stretched the principal components, so keep in mind that the first eigenvector explains 3.874124/1.819077=2.13 times variation compared to the 2nd eigenvector.

UPDATE: At the bottom of the 3-D plot, you can see a dropdown. Just select one of the project participants from there and that participant's dot in the plot with become bigger so they are easy to spot.

Singapore Indians

In the South Asian PCA plot, we saw that Singapore Indian samples from the SGVP dataset had a lot of diversity. Let's zoom into that plot so it's not dominated by the distinctiveness of the Kalash.

Eigenvector 1 explains 1.45 times the variation compared to eigenvector 2.

We see that Singapore Indians are spread in the whole region from Sindhis to North Kanaddi.

Now let's look at the individual admixture results (at K=12 ancestral populations) for the Singapore Indians. I have added some South Asian reference population averages so you can place them in context.

You can click on the legend to the right of the bar chart to sort by different ancestral components.

From these results, a majority of the Singapore Indian samples look South Indian but there are definitely a few from the northwest of the subcontinent (Punjabis or Sindhis?) There are also a few who could be from the Hindi belt.

There are 2-3 samples who have a significant amount of Southeast Asian. Could they be originally from Bengal? Or could they have partial Singapore Malay ancestry?

Chinese Samples

Mithra asked:

Almost all the Chinese are now around 50% SE Asian, didn’t see this before is it right.

So I decided to look at the Chinese samples in Reference I dataset.

I ran Admixture on the whole Reference I dataset for K=10 ancestral populations. The green component is what I call Southeast Asian, blue is Northeast Asian (highest among the Japanese) and violet is Siberian (highest among the Yakut).

Here is the plot for the 106 HapMap Chinese samples from Denver (label: us chinese):

HapMap US Chinese

For the 137 HapMap samples from Beijing, China (label: han chinese):

HapMap Han Chinese

For the 34 HGDP Han samples (label: han):


For the 10 HGDP Han samples from North China (label: han-nchina):

HGDP Han North China

As you can see, the "Southeast Asian" component goes down from the top group to the bottom one, which is as expected.

I wasn't satisfied with these results, so I decided to run Admixture on the East Asian samples in Reference I separately.

East Asian Admixture K=3

At K=3, the results are about the same as at K=10 for the whole reference I population. The Han all have a significant amount of blue component which is highest among the Southeast Asians.

East Asian Admixture K=4

At K=4, we get a Chinese ("East Asian") component. So we have Japanese, Chinese, Yakut and Southeast Asian components. This is what most of you were probably expecting.

Why did the Japanese become the modal population for the Northeast Asian component? I ran a PCA on the East Asian data to see how the different populations looked on a PCA plot. Remember that eigenvector 1 explains 1.49 times the variance of eigenvector 2 and 1.9 times the variance of eigenvector 3. Thus, eigenvector 2 explains 1.28 times the variation explained by eigenvector 3.

East Asian PCA eig1 vs eig2

East Asian PCA eig1 vs eig3

East Asian PCA eig2 vs eig3

As you can see, the Yakut are the far away, but the Japanese are also fairly well-separated from the Chinese populations.

If I didn't have the 141 Japanese samples in my reference dataset, the Northeast Asian component would be centered on the Han most likely, which is the case for Dodecad.

I think this shows that it is not correct to think of the ancestral components inferred from admixture as some pure ancestral population.

South Asian PCA

I used Eigensoft to create a PCA plot of the South Asians in our Reference I dataset (a total of 398 samples) along with the first batch of South Asian Harappa Project participants (HRP0001 to HRP0009).

The PCA software removed 2 Makranis, 1 Sindhi, 1 Balochi and 1 Brahui as outliers, thus leaving us with 402 samples to perform a PCA on.

Here are the plots for the first four eigenvectors. Click to see bigger images.

South Asian PCA eig1 vs eig2

South Asian PCA eig1 vs eig3

South Asian PCA eig2 vs eig3

South Asian PCA eig1 vs eig4

South Asian PCA eig2 vs eig4

South Asian PCA eig3 vs eig4

If you have seen the South Asian plot at 23andme, the first plot here isn't very different except that it seems rotated.

UPDATE: Eigenvectors 1 through 4 explain 1.12%, 0.77%, 0.71% and 0.44% of the total variance.