Thread by @venkmurthy

How good is the recently published polygenic risk score for BMI?

Explains ~9% of variance.

What does that mean?

Plot on left is prediction vs. actual BMI in simulated population with amount of info in PRS.

Right is same except using other 91% of information.

@f2harrell @ADAlthousePhD @statsepi @MaartenvSmeden @mikejohansenmd @bogdienache @VPrasadMDMPH @drjohnm

@f2harrell @ADAlthousePhD @statsepi @MaartenvSmeden @mikejohansenmd @bogdienache @VPrasadMDMPH @drjohnm Code:

@f2harrell @ADAlthousePhD @statsepi @MaartenvSmeden @mikejohansenmd @bogdienache @VPrasadMDMPH @drjohnm Shows this might give us some interesting insights, but not yet ready to predict anything useful in a clinical setting.

@f2harrell @ADAlthousePhD @statsepi @MaartenvSmeden @mikejohansenmd @bogdienache @VPrasadMDMPH @drjohnm This is why Ruth Loos says: "Any score that captures genetic predisposition will never accurately predict future obesity. … You need to account for that environmental part."

sciencemag.org/news/2019/04/n…

@f2harrell @ADAlthousePhD @statsepi @MaartenvSmeden @mikejohansenmd @bogdienache @VPrasadMDMPH @drjohnm Ruth Loos again: “If you have a score that only represents 10% of the overall obesity susceptibility, you can never accurately predict future obesity,”

statnews.com/2019/04/18/a-n…

@f2harrell @ADAlthousePhD @statsepi @MaartenvSmeden @mikejohansenmd @bogdienache @VPrasadMDMPH @drjohnm Ruth Loos again: "I am concerned that the authors may have oversold the genetic score and that their claims of prediction are too bold."

scientificamerican.com/article/a-gene…

@f2harrell @ADAlthousePhD @statsepi @MaartenvSmeden @mikejohansenmd @bogdienache @VPrasadMDMPH @drjohnm Great question from @CardiacJoshi:

https://twitter.com/CardiacJoshi/status/1120715202675335169?s=20

@f2harrell @ADAlthousePhD @statsepi @MaartenvSmeden @mikejohansenmd @bogdienache @VPrasadMDMPH @drjohnm @CardiacJoshi My answer:

https://twitter.com/venkmurthy/status/1120715388973723648?s=20

@f2harrell @ADAlthousePhD @statsepi @MaartenvSmeden @mikejohansenmd @bogdienache @VPrasadMDMPH @drjohnm @CardiacJoshi Would you use a test which has this level of accuracy in clinical care?

Was this helpful?

Excellent concern raised by @davecurtis314:

https://twitter.com/davecurtis314/status/1120719209741332480?s=20

Response: This is a toy example. it is to illustrate how little information is in the PRS and how much info is *not* in the PRS. Finding the info to make the right graph is VERY HARD.

I should also acknowledge the first function in the code was lifted directly from here:

stackoverflow.com/questions/1329…