,
12 tweets,
3 min read
Read on Twitter
Re-upping my thread on the utility of small-effect GWAS findings in light of discussion at #AACRMolEpi Workshop (ahem @travisgerke). And adding a few comments about polygenic risk scores (PRS). /thread
While the potential utility of PRS in any particular context is up for debate, a blanket statement that “GWAS findings are not useful for prediction" is unwarranted.
Often the “modest” c-statistics for PRS (from 0.6 to 0.7) are cited as evidence that PRS are not useful. But there is no universal c-statistic threshold that determines utility. “How good is good enough” will depend on the context, what you want to do with the PRS.
(C-statistic = area under the ROC Curve, aka AUROC aka ROC.)
If you want to use PRS as a diagnostic, then yeah, you need really high sensitivity and specificity (c-stat>>0.7) to achieve a high positive predictive value.
But it you want to use PRS to define risk strata for a stratified screening program, for example, then a smaller c-statistic may suffice.
Here is a hypothetical example, where different screening strategies are recommended to women at different levels of polygenic risk. The width of the bars is proportional to the fraction of women in each risk category, based on a PRS with a c-statistic ~0.65. 
This strategy may or may not be better than current one-size-fits-all screening recommendations (e.g. biennial screening starting at age 50)—that will depend on concrete benefits and risks of screening in different risk strata, and how one defines “better.
So whether a breast cancer PRS is useful for stratified screening is partly an empirical question, partly a modeling question, and partly a question of values (e.g. relative costs of overdiagnoses and preventable breast cancer deaths).
People can make different modeling assumptions and value outcomes differently and come to different conclusions about utility. But we can’t simply say, “because the c-statistic is under 0.8, the PRS is crap.”
On a quest to convince @travisgerke that GWAS should not be consigned to the dustbin of history.
And FWIW, the article Travis cited saying GWAS are not useful for prediction (PMID:29716745) went on to say that (a) as sample sizes get bigger and stat methods get better, PRS will get better, and (b) utility depends on context (e.g. diagnostic vs stratification). /fin
