I want to give more specifics about my main concerns with this pre-print on polygenic scores (PGS) in carriers of rare high-risk variants (e.g. pathogenic mutations in BRCA1). /THREAD
I (and many others) have been enthusiastic about the idea for years, but if we are to move from promise to reality we need to be careful.
1) This figure summarizes the main result in the paper: PGS are associated with risk among carriers of pathogenic variants—in fact the gradient of risk described by PGS is larger in carriers than non-carriers. HOWEVER… 
This result appears to be a consequence of the model the authors fit, a logistic regression with main effects for carrier status (M) and PGS: log(odds) = b1 M + b2 PGS. (I am not 100% sure about this—I had trouble understanding exactly what was done.)
This model assumes much of what was to be proved. In particular, it assumes the relative change in odds associated with a change in PGS is the same in carriers and non-carriers.
At a minimum, we should check whether this assumption is realistic, e.g. by fitting separate PGS effects in carriers and non-carriers.
It seems the authors did this (interaction p=0.53)—although given the small numbers (111 BRCA1/2 carriers combined in the UK Biobank) there is very little power to detect meaningful differences in PGS effects between carriers and controls.
At a maximum, we should also be checking whether the log-linear odds assumption (which leads to that big spike in absolute risk in the tail) holds in carriers the same way it seems to be a good fit in the general population. ncbi.nlm.nih.gov/pubmed/28481708
The authors check this assumption by fitting polynomial models for the PGS—but it seems they have done so in the full UK Biobank sample, but not in carriers specifically.
To really understand the association between PGS and risk in carriers we will need very large sample sizes. The most informative people are carriers with high PGS scores: a small fraction of a small fraction of the population.
Luckily, at least for breast cancer, there has already been a lot of work assembling and and analyzing such large samples.
CIMBA found that a breast cancer PGS was associated with risk in 15,252 BRCA1 carriers—but that the PGS odds ratio was maybe a smidge smaller in carriers than the general population. ncbi.nlm.nih.gov/pubmed/28376175 @KKuchenbaecker @antonis02
More work to come—partly enabled by the CONFLUENCE project. Watch this space. dceg.cancer.gov/research/cance…
2) The authors report a 3.5-fold increase in odds of breast cancer for BRCA1/2 carriers in a case-control sample. This is strikingly smaller than widely-quoted previously published results, showing odds ratios of ~11.5. Why? ncbi.nlm.nih.gov/pubmed/26014596
This is partly due to the fact that the authors define carriers as women with pathogenic or “likely pathogenic” variants—probably this was necessary to boost the number of carriers, but it limits interpretability.
It may also be due to the authors' sample—both cases and controls came "from the Color Genomics commercial testing laboratory,” but no details are given about who these women are.
I suspect that these women came to Color for testing because of their personal or family history of breast cancer. Indeed, the proportion of controls with a family history of breast cancer is 43%—way higher than the proportion in the general population.
So this case-control study is not a sample from the general population, and this limits interpretability and generalizability. #designmatters.
Bottom line: PGS have the potential to inform risk-management decisions among carriers of rare high-risk variants, but we’ll need large studies and careful analyses to get honest and accurate PGS risk estimates. /END
