I am experimenting with a new way to model the NorthWest Europeans in Global 25.
The reason is that I am getting dissatisfied with the role of the labels of the samples. Modern samples are labelled by geography ('English_Cornwall'), ancient samples are often labelled by culture ('Bell_Beaker').
In both cases the correlation of the labels with the DNA features is not 100%, to put it mildly. The number of different samples in the dataset is growing nearly by the month, and so is the labeling problem.
The labeling problem is especially pronounced in regions without natural barriers, like for instance the North European plain. As a result of a long period of mixing, neighboring population groups are genetically nearly identical, while every subgroup wants to emphasize its own uniqueness.
My dissatisfaction with the labels has reached the level that I prefer to model the samples without the labels altogether.
When working with raw genetic markers there is the choice of using the labels with qpAdm, or not using the labels with ADMIXTURE.
When working with PCA scores like in Global 25, one can use nMonte, which uses the labels. But there is no ADMIXTURE like tool which doesn't use the labels for clustering.
So I had to invent my own ADMIXTURE look-alike clustering.
Compared with raw data, the PCA scores have lost a lot of information. Therefore a high-K ADMIXTURE look-alike will be not be possible.
But a K=3 mixture turned out satisfying results. At K=4 the results were less satisfying, apparently because separation of the components became too hard.
As an example of the three-way mixture I present the results for the Dutch AG-members:
A B C
Radboud 0.993 0.006 0.001
FinnMom 0.987 0.000 0.013
FinnDad 0.985 0.004 0.011
Huijbregts 0.785 0.204 0.011
Miremont 0.663 0.320 0.016
As I will explain in my next post, 'A' is a Northern component which represents the core NW-European component, 'B' is a Southern component with especially French (but not French_South and Basque), and 'C' represents Eastern German and Czech.
All five of these samples have 'A' as the dominant component, but the two South Dutch samples also have 20%-30% component B, which the three Northern samples are missing. A nice result, that I have never seen before.
Another example with other AG members:
A B C
Jessie 1.000 0.000 0.000
Trelvern 0.718 0.282 0.000
Helgenes 0.136 0.861 0.004
DewslothDad 0.003 0.997 0.000
In the three-way model Jessie has only the core NW-European component. Trelvern and Helgenes have a varying proportion of the French component.
DewslothDad has such a high percentage of component B that I wonder whether his ancestry isn't Iberian or Italian. This model uses only NW-European reference samples, so using it out of this region isn't advisable.
To demonstrate why I am disappointed by the labels, I present the classic ADMIXTURE bar graph of the Dutch and French samples in Global 25.
ADMIXTURE.jpg
Note the many samples with DNA proportions that are very different from typical samples under the same label.
But both populations have a majority component, A for the Dutch and B for the French.
Here is the ADMIXTURE plot for the Germans:
ADMIXTURE_German.jpg
The German samples don't even have a majority component.
Particular striking is great number of samples with a very high proportion of the 'French' component B.
The samples German2,53,48,14,7,72,70,61 are even more than 0.999 on component B; which is more then Hallstatt_Bylany_DA111 (0.990).
I presume that these component B Germans were born in Southern Germany. Maybe Generalissimo can check this.
In my next post I will more systematically explore the relation between labeled populations and the DNA components within NW-Europe.
Reply With Quote
Originally Posted by Huijbregts
He's (in decreasing amounts) German>British>Belgian>Ashkenazi.
