Rapid evolution of fine-scale recombination during domestication: a perspective from population genomics

Reviewer #1 (Public review):

Liu, Li, Ge, and colleagues use whole genome sequence data to estimate the recombination landscape of domesticated chickens and their wild ancestor, Red Junglefowl. They compare landscapes estimated using the deep learning method RelERNN (Adrion et al. 2020) to understand the consequences of domestication for the evolution of recombination. The authors build on previous work in tomato, maize, and other domesticated species to examine how recombination rate and patterning evolve under the demography and selection pressures of domestication. They do so by comparing estimates of local recombination rates across chromosomes and populations, asking if/how well certain sequence and chromatin-based predictors predict recombination rate, and testing for an association between recombination rate and the proportion of introgressed ancestry from Red Junglefowl.

This study provides evidence for the hypothesis that recombination evolves rapidly in domesticated lineages -- so much so that we see little hotspot sharing between breeds in the present-day! Strengths of the paper include the collection/analysis of data from several domesticated sub-populations and efforts to control for demography and structure in the inference of recombination landscapes (given the challenges of some methods under non-equilibrium demography: https://academic.oup.com/mbe/article/35/2/335/4555533). It is also reassuring to see patterns that have been thoroughly established (e.g., the negative relationship between recombination rate and chromosome size) validated.

However, I have concerns about the data and methodology.

(1) My main concern is that the demographic and recombination rate estimates inferred using ~20 whole genomes are likely quite variable and, without quantification of the uncertainty or systematic assessment of the possible biases in the methodology, it is difficult to have confidence in analyses which make use of the RelERNN landscapes.

(a) Similar studies in rye (https://academic.oup.com/mbe/article/39/6/msac131/6605708) and tomato (https://academic.oup.com/mbe/article/39/1/msab287/6379725) used data from far more individuals (916 individuals split up into populations of size 50 for rye, >75 samples for tomato) to infer recombination maps and conduct downstream analyses. Studies in human genetics make use of an even greater number! The evidence (Lines 189-196 of the main text) that the sample size is sufficient to capture fine-scale variation in recombination is weak. In particular, correlations between the true and estimated recombination rate are based on *equilibrium* demography at sample sizes of 5, 10, and 20, yet used draw the inference "20 samples per population are sufficient to reconstruct their recombination landscapes" under the *non-equilibrium* demography (inferred using SMC+).

(b) RelERNN learns the recombination landscape by using several signatures (the decay of linkage disequilibrium and, as described in https://academic.oup.com/genetics/advance-article-abstract/doi/10.1093/genetics/iyaf108/8157390, choppiness of the allele frequency spectrum) left in present-day genomes. Both signatures depend strongly on local SNP density. It does not seem the effect of SNP density on the inferred recombination rate is examined, despite the potential for correlated noise in inferred recombination rate (in SNP-sparse regions of the genome) to confound downstream inference.

(c) It is unclear if the demographic histories for chickens (Figure S6) broadly match what have been previously estimated from whole-genome data, or if a large class of demographic models are compatible with the data (i.e., confidence intervals for the demographic histories are quite large). In Figure S6, its bottlenecks are somewhat weak and affect only a couple of the groups, despite the history of domestication and the expectation that effective sizes vary more widely. The groups affected (LX and WL) are those that have the weakest correlations between recombination rate under the equilibrium and non-equilibrium demographic models.

(2) The authors test for the effects of chromatin modifications, GC content, etc using correlations between local recombination rate and the features individually. However, joint inference of the effects under a GLM (the distribution of recombination rates is probably better described by, e.g., a Gamma distribution) would permit more straightforward causal inference, given, e.g., the potential effects of chromatin marks on deleterious mutation accumulation. I recognize this likely would not change the direction or significance of the effects in question, but it is worth noting given readers who may want to learn something from the effect sizes and the nature of causes and effects is difficult to disentangle without a multivariate approach.

Overall:

Previous work on recombination landscape evolution in birds (namely, the zebra finch and long-tailed finch; Singhal & Leffler 2015) has shown that many hotspots, i.e., small stretches of the genome that experience rates of crossing over that are much higher than the genome-wide average, are conserved over tens of millions of years of evolution. Work in tomato, maize, rye, and other flowering plants with histories of domestication have shown that hotspots can be dynamic. The results of Liu, Li, Ge, and colleagues complement those analyses and will, therefore, be of interest to those working on the evolution of recombination. Additionally, the finding that minor parent ancestry is negatively associated with recombination is interesting to an otherwise general rule in evolutionary biology. Finally, it is quite exciting to see recombination maps inferred using RelERNN, and in a demography-aware fashion!

That all said, it is difficult to have certainty in the results due to the relatively limited sample size for each of the populations, the lack of control for SNP density, the uncertainty in both recombination maps and demographic histories, and the lack of a joint modelling framework to carefully tease apart effects that are reported in isolation.

Reviewer #2 (Public review):

Summary:

Liu et al. use whole genome sequencing data from several strains of chicken as well as a subspecies of the chicken wild ancestor to study the impact of domestication on the recombination landscape. They analyze these data using several machine-learning/AI based methods, using simulation to partially inform their analysis. The authors claim to find substantial deviations in the fine-scale recombination landscape between breeds, and surprising patterns between recombination and introgression/selection. However, there are substantial inconsistencies between the author's findings and the current understanding in the field, supported by indirect evidence that is hard to interpret at best.

Strengths:

The data produced by the authors of this and a previous paper is well-suited to answer the questions that they pose. The authors use simulations to support some decisions made in analyzing this data, which partially alleviates some potential questions, and could be extended to address additional concerns. Should further analysis support the claims currently made regarding hotspot turnover and introgression frequency vs. recombination rate, these findings would indeed be striking observations at odds with current understanding in the field.

Weaknesses:

I have several major concerns regarding the ability of the analyses to support the claims in this paper, summarized below.

Substantial deviations from field-standard benchmarks the estimated recombination landscape appear to have been disregarded, particularly with regard to the WL breed.
o For example, the number of detected hotspots per subspecies ranges from maybe 500 to over 100,000 based on figure 2A. While the mean is indeed comparable to estimates from other species (lines 315-317), this characterization masks that each recombination map has far too few or too many hotspots to be biologically accurate (at least without substantial corroboration from more direct analyses). As such, statements about hotspot overlap between breeds and hotspot conservation cannot be taken at face value. Authors might consider using alternative methods to detect hotspots, assessing their power to detect hotspots in each breed, and evaluating hotspot overlap between breeds with respect to random expectation.
o Furthermore, the authors consider the recombination landscape at promoters (Figure S10) and H3K4me3 sites (Figure 2C) and find that levels are slightly elevated, but the magnitude of the elevation (negligible to ~1.5x) is substantially lower than that of any other species studied to date without PRDM9. The magnitude of elevation for both comparisons is especially small for WL, which suggests that the recombination estimates for this breed are particularly noisy, and yet this breed is the focus of the introgression analysis.

Introgression and strong selection can both be thought of as changing the local Ne along the genome. Estimating recombination from patterns of LD most directly estimates rho (the population recombination rate, 4*Ne*r), and disentangling local changes in Ne from local changes in r is non-trivial. Furthermore, selective sweeps, particularly easy-to-detect hard sweeps, are often characterized by having very little genetic variation. Estimating recombination rate from patterns of LD in regions with very little variation seems particularly challenging, and could bias results such as in Figure S15. The authors do not discuss the implications of these challenges for their analyses, which seems particularly relevant for their analyses of introgression and selection with recombination, as well as comparisons between WL (which the authors report to have undergone more selection and introgression) with other breeds. Authors should quantify their ability/power to detect recombination rates and hotspots under these conditions using simulation - some of these simulations are already mentioned in the paper, but are not analyzed in this way. Also useful would be quantifying the impact of simulated bottlenecks on estimates of recombination rate.

In many analyses (e.g. hotspot and coldspot overlap, histone mark analysis), authors appear to use 1000 randomly selected regions of the same length as a control. If this characterization is accurate, authors should match the number of control regions to the number of features that they're comparing to. A more careful analysis might also select random regions from the same chromosome, match for GC content where appropriate, etc.

Authors provide very little detail about the number/locations of coldspots or selective sweeps- how many were detected in each subspecies? Does the fraction of hotspots and coldspots which overlap selective sweeps vary between species? It is unclear whether the numbers in the text (lines 356-364) represent a single breed or an analysis across breeds.