Spuriously Low Relative Recombination Rates?

[loganghavard]

Hi, I'm using Pyrho to infer recombination rates in a non-model species. I see the issues raised about absolute scaling and the advice to take this with a grain of salt; however, after looking at the range of these results, it seems I may have some issues with the relative rates as well. Pyrho is reporting very low recombination rates on the order of 10^-14 cM/Mb (after multiplying the per-generation per-base rates by 10^8). These rates suggest something else may be going on considering the mean is on the order of 10^-2 and the maximum on the order of 10. Many of these calls seem to go away when I thin SNPs within 100 bp of each other, but many are still there. Some hard filtering of the rates seems reasonable, but I want to make sure I understand why such low relative rates are being called in the first place. I would appreciate any insight you could give!

[jeffspence]

Hello! This is the second time I'm hearing of this (the other place I saw this was here: https://www.biorxiv.org/content/10.1101/2025.02.01.635998v1 top of page 8) so you are not alone. I'm not entirely sure what's going on -- it seems like pyrho is running up a hard-coded lower bound of 10^{-20} for the minimum total recombination rate between a pair of adjacent SNPs, so there's something about those sites that makes pyrho want the recombination rate to be ~0.

A few fairly random thoughts --- what windowsize are you using when running pyrho optimize? And if you compute the LD between SNPs in some of these problematic regions do they look really high (e.g., are there lots of pairs with D' (https://en.wikipedia.org/wiki/Linkage_disequilibrium#D'_Method) close to +/- 1? or does r^2 between SNPs seem generally high)? If the LD is generally high, then it could be real, or at least pyrho is interpreting a real signal (although that signal could be caused by some kind of data artifact).

Hope this helps,
Jeff

[loganghavard]

Hi, thanks for getting back to me so quickly. After raising the issue, I noticed all of the low rates on the order of 10^-14 cM/Mb were the same value, so I assumed (as I think you’re suggesting) that these rates were an artifact of the program. Setting these to 0 shifted the lowest values to a more reasonable 10^-5. I’m still not sure this is the best approach though.

I’m currently using a window size of 75 with a block penalty of 15, which I determined from maximizing the Spearman correlation coefficient at 100 kb. It may also be worth noting that I’m running it by chromosome, and those two parameters gave me correlations for each chromosome ranging from around 0.75-0.95 in the hyperparam_results. I also just noticed all of those really low calls are on the same chromosome. Do you have a suggested way of checking LD? I computed r^2 with compute_r2, but I'm not sure how to verify using this since the really low values are not in this table. It does look like LD decays with increasing rho values though. A quick check of LD with plink suggests r^2 is actually quite low for these regions (~0.3). Thank you so much for your help so far!

[jeffspence]

Interesting -- the hyperparameters seem reasonable, and the fact that all of the low rates are on the same chromosome would make me suspicious that there's something (either biological or artifactual) going on with that one chromosome (A biological example would be a segregating inversion. An example of an artifact would be if two different regions of the genome are getting mapped to the same part of the reference genome. Either would cause the appearance of a lot of long range haplotypes that would cause pyrho to infer extremely low recombination rates).

r^2 of 0.3 is actually pretty high depending on the minor allele frequencies at the two sites. compute_r2 tells you what the distribution of r^2 is expected to look like under the demography you used to compute the lookup table (i.e., it doesn't look at or know anything about your actual genotype data). It could be useful to see if the r^2 levels you're seeing are consistent with extremely low recombination rates (i.e., if the distribution of r^2 values you're seeing is close to the distribution of r^2 values expected when the recombination rate is 0).

[loganghavard]

I think you're right. For that particular chromosome, the expected LD is showing a 25th percentile of 0.24, so I guess an r^2 of 0.3 is not too surprising. I'm really curious to know what's going on in this region and if it's something biological, but for the larger question I'm trying to answer for this project, I think it may make sense to exclude this region for now since it's such a small fraction of the genome.

Thanks so much for your guidance and feedback! Pyrho seems like quite an impressive tool!