Compare ancestry inference performance of v4 RF models

[gtiao]

This ticket is aimed at doing the analysis for v4 population assignments (as opposed to writing the PR #494)

[gtiao]

Someone (Alicia? Konrad?) reported getting excellent results training an ancestry RF on 1KG+HGDP and applying to UKBB pan-ancestry project. Let's compare to the gnomAD RF performance and see if we can get better results.

[gtiao]

This really only applies WGS data (v3) instead of exome (v4)

[ch-kr]

we don't have labels for v4 that we can use (we have labels for v3). Mike was planning on looking at whether we should use only use known labels (HGDP/TGP) or v3 imputed labels

[mike-w-wilson]

I moved the wrong ticket here. This will happen later in analysis, not this sprint.

[klaricch]

Decided on:

• 20 PCs
• Training samples:
• hgdp/tgp samples
• v2 samples with known pops
• v3 samples with known pop for certain cohorts only (Decisions were made based on results of an analysis to determine which v3 samples/cohorts to use as training samples. This analysis consisted of computing per sample mean Euclidean distances to all samples in a given population, and per sample the mean Euclidean distances limited to only HGDP/1KG samples in each population)
• Spike in v4 samples with race/ethnicity of "Arab" of "Persian"
• min_prob parameter decided per pop using min_recall and min_precision of 0.99 ( Minimum recall (min_recall) is used to choose per ancestry group minimum RF probabilities. This min_recall cutoff is applied first, and if the chosen minimum RF probabilities cutoff results in a precision lower than min_precision, the minimum RF probabilities with the highest recall that meets min_precision is used.)

min_prob_cutoffs={'afr': 0.93, 'ami': 0.96, 'amr': 0.86, 'asj': 0.88, 'eas': 0.96, 'fin': 0.91, 'mid': 0.56, 'nfe': 0.78, 'sas': 0.96}