Polyploidy Polyploidy Polyploidy

[conJUSTover]

This is an initial issue to start the discussion around the logistic of implementing polyploidy into stdpopsim. Here are a few discussion points to kick around:

engine:
The SLIM script from Booker and Schrider 2024 is probably the best place to start. Right now, that script starts with a tetraploid population at gen0, and can include selection, various dominance relationships between the 3 heterozygous states, and transitions between auto/allo via controlling whether pairing occurs at random between 4 genome, or between main/spare (p1 and p2) populations of genomes.

Outputting data:
SLIM does not support vcf files with tetraploid SNP calls, although it is trivial to output a diploid VCF where two 'individuals' contain diploid genotype calls for the same tetraploid individual. Could create a quick wrapper script to combine diploid SNPs together to create tetraploid individuals instead of 2 diploid individuals (but should also consider diffs between auto/allo here, and whether the user might want to 'phase' genotype calls into subgenomes, even if they aren't fully phased into haplotypes)
I don't have experience with the tree sequencing recording for the tetraploid model, but my assumption is that it is similar to VCF output.

Published demographic models:

1. Capsella bursa-pastoris (allo) from Blischak et al 2022, but includes HEs as a 'migration-like' parameter in dadi, so it may not transition into new SLiM modeling correctly (but that the only way so far to model HEs, so 🤷)
2. Brassica napus, including interploidy introgression from Wang et al 2023
3. Arabidopsis arenosa (auto) from Monnahan et al 2018

DFE models:
For allos, there is a basic modeling question of whether the DFE should be subgenome-specific (i.e. a diploid DFE) or model co-segregating variants between subgenomes (i.e. a tetraploid DFE). Does standard popsim support joint DFE? Does SLiM?
Diploid DFEs:

1. Arabidopsis kamchatica from Paape et al 2018

Some additional potential features to build onto here:

1. Transition from diploid to tetraploid, instead of starting with tetraploid in gen0. Could also include multiple polyploid formation events or interploidy introgression. I think mixed cytotypes (e.g. triploid that could create diploid or haploid gametes) might be too ambitious for now, but could be a cool implementation
2. Do we need different scripts for autos vs allos? For more sound burn-in for allos, we'd also need to simulate the divergence of progenitor diploids (to capture shared diversity and fixed differences between subgenomes), but burn-in for autos could be done without diploid progenitor pop
3. Add in double reduction for autos, homoeologous exchanges for allos, and homoeologous gene conversion for allos. None of these are natively supported by SLiM, so would need to be hard-coded and carefully validated. But, this is not trivial, since these steps would also require tracking centromeres and possibly revisiting how SLiM handles recombination (to allow for some chromatids that do not undergo recombination)

[jeffspence]

Very interesting! Thanks for this!

I'm not a polyploid person, but some stray thoughts:

For the divergence of the progenitor diploids -- what would be the right amount of burn-in to do? It seems like the amount of divergence between them would need to be specified somehow, and there would need to be some burn-in before those diploid populations split (to allow for incomplete lineage sorting). Are you thinking that all of this would be represented in the demographic model somehow?

Regarding the DFE -- I believe most of the machinery on the stdpopsim side is here: https://github.com/popsim-consortium/stdpopsim/blob/main/stdpopsim/dfe.py#L18 and then that gets used in the SLiM engine throughout (https://github.com/popsim-consortium/stdpopsim/blob/main/stdpopsim/slim_engine.py)

[conJUSTover]

The amount of burn-in for the diploids shouldn't be affected, so typically 10N generations or so for neutral sims (although there have been discussions about this even for non-polyploid sims). But these will need to be very long simulations, to capture the burn-in, divergence of diploid species, polyploid formation, and however many generations post-polyploidy one would want to model.

The tricky part with these allo models are that most papers ignore the polyploid aspect, and just model each subgenome as an independent diploid lineage (except for the two papers I linked above). But the newer demographic models in dadi from Blischak et al will hopefully generate some more demographic models that incorporate both subgenomes simultaneously

[dschride]

Thanks for this, Justin, and sorry for being AWOL for a while. One quick note: for his paper Will implemented a polyploid version of the Huber et al. DFE with the h-s relationship. Even if we don't necessarily want to use that DFE, the framework we used for handling dominance effects could potentially come in handy?

Regarding tree sequences, I don't see why those couldn't be used in the same way as VCF output here.

[pblischak]

Just wanted to chime in here and say that I'm glad you filed this issue @conJUSTover! I'm happy to contribute in any way that I can.

In particular, I agree with your point about the difficulty of how to handle special cases of recombination (double reduction, homoeologous exchanges/gene conversions) within SLiM itself. It's been a while since I've looked at all of this, so apologies if I'm missing something obvious, but if there are ways to add recombination maps to include centromeres or to control rates of different events across the chromosome, then maybe that will provide the more fine-grained control that would be needed to sample realistic breakpoints for recombination or HEs? Or maybe would at least provide the infrastructure for implementing the functions needed to do the appropriate sampling under different polyploid-specific recombination models

[petrelharp]

It seems like each of the various types of polyploidization constitutes a different special case as it'd need to be implemented differently; I'd start with one of them. Formation of the polyploid species is a whole additional can of worms (but a neat one). A case where everything is effectively diploid (eg allo where one of each homolog has pseudoganized or specialized) would avoid tricky things about how to specify dominance with more than 2 alleles.

[conJUSTover]

There are some aspects of SLiM5 that might complicate the polyploidy script from Booker and Schrider:

1. How are non-autosomes handled? In the extra population in the polyploid SLiM script, all chromosomes are involved in recombination/inheritance, but would not want that for, e.g. mitochondrial genomes
2. Can/Should different contigs have different inheritance patterns (disomic v tetrasomic)? Or keep that as a global parameter for the entire genome?

I think I am leaning towards having the goal of implementing an autotetraploid model initially (perhaps Arabidopsis arenosa from Monnahan et al 2018?), and then we can worry about further developments for allos or things in between. Just adding the autos will be a big lift that will (hopefully) identify most of the problematic areas, which might be a good initial goal to have, keeping in mind that we will further develop in the future.