Common patterns in human population simulation

[hammer]

(Posted by @eric-czech)

This is a summary of patterns I've seen in papers that simulate human genotypes primarily for the purpose of validating association models and kinship estimation methods.

The 3 most common approaches are:

1. Resample from existing datasets to a test dataset of any size
   • e.g. hapgen2 starts from a reference panel of haplotypes and simulates cross-over events with markov chains
   • The primary advantage of this is that it preserves realistic LD patterns. Most simulation methods concerning association modeling and/or kinship estimation don't really care about this since it's correlation between samples that they're always trying to control for (independence of SNPs is often assumed)
2. Simulate phenotypes or exposures using real datasets, or possibly subsets of them (no "upsampling")
   • This appears to be something of a gold standard for association modeling or mendelian randomization methods
   • The GAW conference is dedicated to producing datasets like this and the datasets from it make for common validation targets
3. Create fully synthetic datasets
   • This is generally where validations through simulation start, and they frequently follow a similar pattern:
     • First, the Balding-Nichols model is used to define allele frequencies (AFs) for subpopulations by:
       • Drawing an ancestral AF for each SNP
       • Defining a Beta distribution for each SNP within a subpopulation using the ancestral AF above and a provided Fst value (so the user can control the extent to which the populations differentiate, on average)
     • Second, admixture is added to the simulation using the Pritchard-Stephens-Donnelly model by:
       • Drawing subpopulation mixture proportions from a Dirichlet distribution
       • Genotypes for founders are then defined by a Binomial distribution with the sole probability parameter for each SNP equal to the average of the Beta draws (in [0, 1]) from the subpopulation distributions for that same SNP, weighted by the Dirichlet mixing proportion
     • Third, pairs of founders are matched up and progeny for them are simulated based on standard Mendelian transmission rules
       • These forward-in-time steps are then often carried out for pedigrees of some preconfigured size (e.g. 3 generations as in SAIGE)

To summarize, the purpose of the synthetic simulations is only to:

• produce configurable numbers of populations
• enable admixture amongst founders
• reflect kinship in relatively small pedigrees

There is an exception or two, but essentially all of these synthetic methods simulate genotypes directly rather than appealing to sequences with mutations. I'd love to hear how this contrasts with the work you often do @jeromekelleher, if you're up for it. I'm guessing that genotype simulation became more common in this domain because it's simpler and easier to scale. Couldn't say for sure though.

Also, I wanted to emphasize how often realistic LD is discounted in simulations for human GWAS pipelines in favor of focusing on the more difficult problems with ancestry and relatedness inference. Hopefully this will help us in defining what useful test datasets should look like.

References

I'm basing this on commonalities in these papers:

• SAIGE (2018)
• BOLT-LMM (2015)
• ggmix (2020)
• FaST-LMM (2011)
• PC-Relate (2016)
• PhenotypeSimulator (2018)
• GWA publications (specifically 14 and 20)

[hammer]

While browsing GitHub aimlessly tonight I happened to notice that Stephanie Gogarten has been working recently on simulate_phenotypes which is in R but may be useful for us. The vignette has an example use case.

[hammer]

• Automatic Inference of Demographic Parameters Using Generative Adversarial Networks (2020): https://github.com/mathiesonlab/pg-gan is an interesting approach to simulating genotypes under an evolutionary model.