Training model from not single ancestry samples

[juanbigotes]

Hi everyone,

Running Gnomix is pretty straight forward when using the pretrained models that you provide. However training a new model is another story. I've trying to train a model for some days and I have some questions for you:

1. You say in the demo.ipynb that you'll create a reference file by sub-setting the part of the query file that you know are single ancestry, but what about the cases when we want to use a reference file that has admixed individuals? I mean some of them clearly show a bigger ancestry component but still non of them have one single ancestry. What should I do in those cases? (Like someone that has 80% Native American ancestry and 20% European)
2. Following the demo.ipynb you guide us to use the chr22. I tried to train that example following all the steps which are (if I understand it well) a) running bcftools to get the reference_1000g.vcf file from the chr22 query file , and b) running gnomix in train mode. For the second part you use the new reference file but you use again the same query file. Why do you use both query and reference files from the same origin (chr22)?
3. For the second part (running gnomix in train mode) I decided to use a different query file (a trio) and at the end of the training I got all AFR ancestry results in all sites even tough I now they are not African samples (they are Mexican). You mentioned in an issue post that this is due an incompatibility between the query and the reference file (i.e. not the same b37 positions), but I downloaded the chr22 in that format. Could it be possible that there is something else I'm missing?
4. Training chr3 takes a very very long time (compared to training chr22) and it always ends in bash killing the process. Would you say training large chromosomes requiere very very large amount of disk or memory space?

Thanks in advance for everything!

Juan