Completed on 5 Dec 2017 by Mikhail V Matz . Sourced from https://www.biorxiv.org/content/early/2017/12/02/222307.
Login to endorse this review.
Author response is in blue.
A few thoughts :
My greatest concern for GWAS is that many individuals are related, which (as far as I could see from the methods) was not accounted for in GWAS analysis. As a result the analysis would have fished out predominantly markers that best discriminate between distinct relatedness clusters (as these clusters are also highly different in torpor onset, since heritability is close to 1). Although the true torpor-driving SNPs would probably be among those, my intuition is that the confidence would be greatly inflated since all related individuals are considered as independent samples. Is there a way to perform GWAS while controlling for relatedness?
If we account for relatedness, the next concern will be the small sample size (for GWAS), which is expected to lead to many spurious associations (considering that the amount of markers profiled ). The qq plot is not terribly convincing in this case. It would be best to use randomized data to generate a null distribution of pvalues. Randomizing would be trivial if all individuals we unrelated (just shuffle phenotypes), but to be honest I am not sure how to perform shuffling in this case - but it feels like there must be a way.
Third, eQTLs: please show that your top-GWAS SNPs are more likely to be eQTLs for hybernation-associated expression than a random choice of SNPs.
Fourth: I would like to discourage the authors (and everybody else) from weaving extensive "just so stories" about genes that seem to "make sense". It is a slippery slope - you start telling what *you* think should be going on rather than objectively summarizing the data. That said, this paper is not too bad in this respect, I've seen much worse.
"genetic architecture" in the title is misplaced (as has been pointed out by a few tweeps already), it would be more appropriate to say "Natural genetic variation underlying onset ..." or something like that.
Make sure all the labels in the figures are described in the legend, for example what are the "States" in Fig 6.
Very minor: I would use points instead of population names in Fig 2 B - the names overlap too much, looks funky.
Wow, thank you for the thoughtful and detailed critique! To address your biggest concern, we did account for genetic relatedness in our GWAS (see lines 553-556 of the manuscript), because yes, most of the individuals were related. Basically, we accounted for genetic relatedness by including it as the random effect in the null linear mixed model. When we’ve looked at the results, we see the SNP genotypes segregating by differences in hibernation onset within families, not as clusters discriminating between the families.
And yes, we recognize that we have a small sample size. However, as you have suggested, when we apply a simple phenotype shuffle (not even accounting for relatedness), we see that the the resulting distribution of observed p-values largely follows the expected null distribution, which is in contrast to our results in Fig 3B. We can include this in the manuscript if it helps. As an aside, we hope readers will appreciate that quantitative genetic and genetic mapping studies are largely non-existent in the field of hibernation. So this was a risk and being exploratory, we depended upon squirrels that had been previously telemetered and collected for use in other experimental studies. We hope our results will help motivate and justify the use of more extensive genetic mapping studies (with larger sample sizes) to better understand this phenotype.