Help with interpretation

[pjvandehaar]

Things to tell users:

• whether to trust a phenotype
  • partially via the QQ plot, especially gc-lambda-50%ile and gc-lambda-10%ile
• whether to trust a peak
  • associations: top pval; other pvals in peak (ie, vertical line vs alone); top effect; other effects in peak (ie, do they agree); MAF; width of peak; nearby associations (both in terms of distance and LD)
  • annotations: nonsynonymous vs synonymous vs intronic vs UTR vs near gene; coding gene vs lincRNA vs pseudogene; known eQTL; known TFBS
  • context: the number of variants; variant density in that region; QQ; QQ in similar MAF; number of cases/controls/samples
  • variant quality: imputation quality, read depth, allele balance, HBE, callrate? something related to recessive/dominant/additive?
  • later, do conditional analysis to judge how many signals are in a peak

Maybe label our MGI and Sardinia data with heuristics, adjust by hand, and then train a ML?

• If we went for standard image-based approach for peak interpretation:
  • would we feed in x-axis as variants or positions? If LD will be a feature, then it might as well be by variant, right?
  • input: 1000 variants on each side of peak, a row for each feature, also global info as constant-rows?
    • how to scale to 0-255?
  • output: peak confidence
  • training:
    • use known associations to train? but we haven't mapped traits, and it'll be biased if we use peaks to map traits to use for annotating peaks
    • use SardiNIA pvalues as labels for training on MGI peaks? & vice-versa? is this a thing that people do? Is it okay that one is a founder population, or does that not make a big difference? Should MAF difference be factored into label confidence?
  • method: column-wise convolution, a local-max layer (20 variants each?), one fully-connected layer (i don't understand deep learning)