Population Structure

[CarmenRamoser]

Hi,
I’ve tried Neural Admixture on the UKBiobank, but running it in the unsupervised mode got a lot of structure in the White Europeans instead of the continental ancestries like in your paper. Do you know what could be the issue? Here is what I tried:
neural-admixture train --k 6 --data_path 500k_sampleQC_LDpruned_allchr.bed --save_dir sample500k/outputs --name sample500k --initialization kmeans --num_cpus 24

neural-admixture train --k 6 --epochs 50 --batch_size 400 --learning_rate 1e-5 --data_path 500k_sampleQC_LDpruned_allchr.bed --save_dir sample500k/outputs --name sample500k_Test2 --initialization kmeans --num_cpus 24

The PCAs look fine.
Thanks for your help!

[jcgrenier]

According to some experiments I've ran myself, the training relies on the PCA, which is by itself prone to biases if an unbalanced dataset is given as an input. I would guess that a solution for this would be to either downsample the european individuals or train on a reference dataset, and infer the UKBiobank afterwards. Would it make sense?

[CarmenRamoser]

Yes, I will try now to train on a reference, it makes sense that the UKBiobank is just too unbalanced. Was just curious as it seems to have worked in the paper in an unsupervised way but couldn't figure out which parameters were used.

[joansaurina]

Hey @CarmenRamoser .

A new version was released this week.

Could you retry now?

Thanks!

Joan

[jcgrenier]

Sorry to open again this thread. We are still struggling with this analysis on our side at a different level. We were wondering if it would have been possible to share a list of positions that were used to do that experiment in the paper, just to at least eliminate that variable out. Thanks a lot for your help!

[joansaurina]

Hi @jcgrenier @CarmenRamoser ,

The experiments in the paper were done by a previous author, and Neural Admixture 1 (the version used in the paper) tends to perform poorly on biobank-scale datasets like UK Biobank. The current version (1.6.7) shows much better results, but we are also actively working on releasing Neural Admixture 2, which should further improve performance and scalability.

Thanks for your patience and for following up!

[RaphaelRaphael]

Hello,

I strongly suspect that the set of position used in the paper is the main driver for the reproducibility of it s figure. I do not find the specific process used to filter the genomic variations. Could you provide the filtering step you used ?

Thank you