By all means, let’s have a nuanced discussion of pros and cons of bringing genetic risk scores to the clinic. But “deep flaws” is laying it on a little thick, bending the stick too far in the other direction. 1/
To my mind, the most interesting and important discussion here is given a risk profile similar to that shown in the recent polygenic risk papers—and it need not be a genetic risk factor, could be BMI or circulating plasma rhubarb—what’s the clinical and public health utility? 2/
How do/should clinicians and patients interpret risk scores/screening tests? How does society, how do individuals balance risks and benefits? This is a general set of Qs, not specific to genetics, cf debates over PSA and mammography. 3/
Polygenic risk scores are already better than a lot of risk tools that are used in the clinic, so they can’t be dismissed out of hand as having no useful discriminatory ability. 4/
Secondarily, on the wonkier issues (even @FiveThirtyEight which cares about stuff like overfitting and simple versus complex modeling would put this in the footnotes): (1) AUC is not the only metric for clinical utility. 5/
(2) Stipulating for the sake of argument that the clinical utility of a 74-SNP risk model and a 6-million-SNP risk model are the same, one might still choose the 6-million-SNP model for logistical reasons. (Economies of scale, data can be used for other diseases.) 6/
(3) School of hard knocks experience and nice recent work by @nilanjan10c tells us it gets harder to improve discrimination after you’ve captured “low hanging fruit”—large samples needed to capture signal we’re confident is out there. 7/
But per (1) and (2) that does not necessarily mean the effort is not worth it. (Never mind payoffs on basic science front.) /fin
