1
$\begingroup$

Utilizing a reference panel, I want to assign most likely population label to each individual in the study. Following are the files I have:

  1. Reference panel population labels:

    HG00121 EUR
    HG00122 EUR
    HG00123 EUR
    
  2. Reference panel PC (Principal Component) coordinates:

    HG00121 -0.0108487      0.0259052       -0.0099605      0.0168545
    HG00122 -0.0115789      0.0264642       -0.01722        0.0157885
    HG00123 -0.00996687     0.0266983       -0.0155755      0.0155963
    
  3. Study individual PC coordinates:

    Study_individual1 -0.0108487      0.0259052       -0.0099605      0.0168545
    

How do I go about assigning a population label to the study individual based of the reference panel? Most of the softwares I have come across such as evaluate_check_ancestry(plinkQC), FRAPOSA rely on inputs provided to PLINK which is not helpful in my case.

$\endgroup$
6
  • 1
    $\begingroup$ I meant Principal Components (PC) coordinates. Edited question. $\endgroup$ Sep 13, 2023 at 16:04
  • 1
    $\begingroup$ Why start with PCA? It's a means of reducing dimensionality to retain the highest amount of variability, but those reduced dimensions may have nothing at all to do with differences between the categories. Suppose you find that the query study is very similar to Study X in PC1 and PC2 - that doesn't necessarily imply it's the same category as Study X, as PC1 and PC2 may totally irrelevant to the category. $\endgroup$ Sep 13, 2023 at 18:17
  • $\begingroup$ @NuclearHoagie In the era of machine learning PCA is a common starting point. It's classed as unsupervised learning and used to generate the hypothesis for supervised learning. Also in the world of RNAseq (which this isn't) its frontline approach. In pop gen ... generally its one of a number of starting points, but its not the end analysis. $\endgroup$
    – M__
    Sep 13, 2023 at 19:01
  • 1
    $\begingroup$ @M__ Right, I just don't think it's a particularly good starting point for this problem. If you want to predict some label based on feature values, you probably shouldn't start by throwing away most of the data without having any idea if what you're discarding is useful for prediction or not. If study label is easily determined by a low-variance feature surrounded by lots of high-variance noise, using only the top PCs will make the task all but impossible, when it could be easily accomplished by starting with a supervised method. There's no reason why "distance" in PC space should be "better". $\endgroup$ Sep 13, 2023 at 19:13
  • 1
    $\begingroup$ @NuclearHoagie it turns out that for human genetic data, the major principals components do align with major population groups and lower PCs strongly correlate with finer scale populations - it’s not an ‘impossible task’ at all. PCA is fine for this purpose if you need to do it quickly for lots of samples, particularly if you are just interested in continental-scale groups like the OP seems to be interested in $\endgroup$
    – user438383
    Sep 14, 2023 at 8:04

2 Answers 2

1
$\begingroup$

I apologise for being negative about this, but I am afraid you cannot use principle components for population inference other than assigning clusters within a bivariate or 3D graph. Eigen values are complex derivatives of data to achieve dimensionality reduction.

For population genetics you require a sample size => 10. I don't know the organism nor the question but goal is to derive allelic frequencies for each genome position. A common metric is F-statistics, for example Fst or Hardy-Weinberg equilibrium.

$\endgroup$
4
  • 1
    $\begingroup$ I do have study sample size of 10 (tried to make the question by including just 1). Thank you for the input. I will look into Fst and Hardy-Weinberg equilibrium $\endgroup$ Sep 13, 2023 at 18:08
  • 1
    $\begingroup$ Dear @user2998764 upvotes/accepts much preferred to thanks, but thanks. I'm glad this helped. 10 is an okay sample size and you can always ask further given your experimental aims for the most useful metric. $\endgroup$
    – M__
    Sep 13, 2023 at 18:10
  • $\begingroup$ ‘You cannot use PCA for population inferrence other than assigning clusters’ but isn’t that basically what the OP is trying to do? $\endgroup$
    – user438383
    Sep 14, 2023 at 8:05
  • $\begingroup$ Its application is for the data mining phase, not the final answer. At a minimum the OP would need to assess cumulative variance across PCs and it gets complicated. $\endgroup$
    – M__
    Sep 14, 2023 at 11:50
1
$\begingroup$

Here is a very simple way to get your started. Of course you can make it more complex:

Take the mean PC1 and PC2 coordinates for each 'reference panel population label' and then calculate the reference panel population which has the closest Cartesian distance to your target, i.e. $\sqrt{{(PC1_{r} - PC1_{t})^2 + (PC2_{r} - PC2_{t})^2}}$, where $PC1_{r}$ is the mean PC1 for a ref population and $PC1_{t}$ is PC1 for the target.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.