There are several mechanisms that could reduce fitness. Perhaps most important would be co-evolution of adjacent exons, which would render unlinked exons more likely to be incompatible. This is, for example, clearly an important dynamic in HLA evolution.
Also, it seems from the paper from @davecarlson, the magnitude of effect is low (10% to 2%), even if the p-value is low.
Sure, when there’s positive selection for maintenance of divergent alleles. But do most alleles in most genes within a population diverge enough for coevolution to have much significance? Perhaps if there is a single, widely spearated residues in the protein that interact.
I think the easiest way to test for this would be to compare the results of SNP-based measurements of recombination hotspots (where some selection has presumably occurred) with results from a non-population genetics approach that minimizes the opportunity for selection to occur.
For example, Brick et al. 2012 used CHIP-seq on a variety of mouse crosses from different genetic backgrounds (including homozygous PRDM9 knockouts) to directly identify double-stranded breaks. I would expect that this set up minimizes the effects of prior selection (with the obvious exception that completely nonviable games are not seen) compared to SNP-based methods.
They found that the wild-type PRDM9 backgrounds have hotspots that are directed away from functional elements (specifically transcript start sites) compared to mice without functional PRDM9:
Despite the clear critical role of PRDM9 in the initiation of genetic recombination meiotic DSBs are present in Prdm9 knockout mice9. Remarkably, we found that initiation of recombination in these mice is not random or uniform, but is still clustered in hotspots. The vast majority (99%) of these hotspots do not overlap the hotspots detected in any of the wild type strains, nevertheless, 94% of Prdm9 −/− hotspots still overlap H3K4me3 marks. Most of these marks (92%) are present in wild type testis and almost half are not specific to germline cells (Supplementary Fig. 5d). As H3K4me3 marks outnumber DSB hotspots, they do not appear to be sufficient for hotspot formation (Supplementary Fig. 6). H3K4me3 is a general mark of active transcription primarily associated with gene promoters16, enhancers17,18 and possibly other functional genomic elements. We found that in the absence of PRDM9 and PRDM9-introduced H3K4me3 marks, recombination hotspots are re-routed to these alternative H3K4me3 sites (Fig. 2a). Almost half (44%) of recombination hotspots in the Prdm9 −/− mice localize to the promoters of annotated genes compared to just 3% in wild type.
Here is a cartoon illustrating their proposed model of how PRDM9 directs recombination away from functional elements:
Proposed role of the PRDM9 protein. Left: In cells containing a functional copy of PRDM9 (Wild type) the DSB formation machinery (scissors) is directed to preferred DSB sites / PRDM9 binding sites. Right: In the absence of PRDM9, the DSB formation machinery opportunistically makes breaks at PRDM9-independent H3K4me3 marks such as those at promoters and enhancers. This results in inefficient DSB repair and meiotic arrest.
Unfortunately, this paper talks mostly about recombination at promoters. They don’t discuss the relative frequency of DSBs within exons in wild type and PRDM9 knockouts, so without analyzing the data myself, I’m not sure we can draw any strong conclusions regarding whether or not recombination avoids genes specifically.
It would seem to require at least one non-silent site per exon. How likely is that without significant divergence, on the order of that typically seen between closely related species? 20% of human and chimp proteins are identical, and most of the rest have a single difference.