Recommendations for how to parameterize ReLERNN

[andreaswallberg]

Dear @jradrion @andrewkern et al,

I am exploring the possibility of using ReLERNN to infer recombination rates in a non-model arthropod with a large genome and high levels of nucleotide diversity (1-1.5%). Nothing is known currently about the recombination in this species and we have no ground-truth evidence to fall back on to verify results. Our assembly is relatively fragmented and our sample size is just below 70 diploids. The decay of LD seems relatively rapid in our data (phased with BEAGLE 4.0), similar to what is seen in many other arthopods.

The ReLERNN paper speaks much about the of relationship between mutation and recombination rates, and both the mutation rate and parameter "--upperRhoThetaRatio" seems to be key to successful inference.

I tried setting --upperRhoThetaRatio to 35 as in the paper and used a mutation rate typical for arthropods, and while all steps in ReLERNN worked on my machine with a powerful GPU, the inferred recombination rates came out very flat, with dips around contig brakes along scaffolds or genes with reduced levels of variation, suggesting training and parameterization has not worked well.

ReLERNN is new to me and I am not sure how to move forward.

Can you give some hints as for what parameters to tweak?

A higher or lower "--upperRhoThetaRatio"?

Removing variants with low minor allele frequencies?