Thread by @venkmurthy

New paper out today in @JAMACardio that originated from a question @RaviShah_MD had several years ago:

what are the relative contributions of genetics, fitness and physical activity to the development of obesity?

jamanetwork.com/journals/jamac…

We struggled for long time how best to capture impact of genetics. Fortunately, recent polygenic risk scores (PRS) offered an answer.

We used @skathire @amitvkhera genomewide PRS as most comprehensive measure of genetic risk of obesity to date:

sciencedirect.com/science/articl…

We decided to examine CARDIA cohort (Coronary Artery Risk Development in Young Adults), sponsored by @nih_nhlbi ~30 yrs ago to study black & white men & women age 18-30, followed serially for 30 yrs (though we only had 25 yr data for this analysis)

cardia.dopm.uab.edu

So this was a rare opportunity to apply genetics in a racially diverse cohort.

We started by asking the question, how well does knowing a person's age, sex, race and how they answered a question asking whether either of their parents were overweight or obese do for predicting a person's BMI over 25 yrs?

This was our "base model" and didn't include *baseline* BMI today (as is common in the genetic literature).

You can see the variance explained in the top row of this grid. Each column represents a different year of follow-up.

For both whites and blacks, this simple information only explains about 5-10% of the variation in BMI in the population over time.

Adding PRS bumped that up to 11-14% of variance in BMI we can account for.

Fitness, by comparison, explained ~18-25% of the variance over time

Activity (based on a questionnaire) was a bit lower.

Putting all of these data together only got use to 22-27% though.

However, most epidemiologists would say for *clinical prediction*, we should account for baseline BMI at year 0.

When we do that, the picture changes dramatically.

Now the base model + year 0 BMI account for between ~50% and 80% variation of subsequent BMI and PRS, fitness, and physical activity offer only marginal gains.

You can go further and ask, if we measure BMI serially, as most people would have at their PCPs office, what happens then?

The gains of PRS, fitness and activity measures are even smaller.

So how accurate are predictions for BMI 25 years later made with just age, sex, race, family history and PRS?

Here are Bland-Altman plots showing the typical errors are massive (std deviation of 11.4 to 14.5 BMI units) for whites (A) and blacks (D)

Adding in baseline BMI instead of PRS helps a lot. Whites in (B) and blacks in (E)

Serially monitoring BMI does even better. Whites in (C) and blacks in (F).

So what is conclusion? Measuring BMI, cheap & easy has much more information content for understanding how someone's BMI is going to change over time than PRS, fitness, activity.

1 measure at age 18-30 is *very* good over 25 yrs.

Refining it with serial measures is even better.

Need to thank coauthors especially Rui Xia, @HFpEF, Abi Baldridge, Myriam Fornage for enabling this work, the participants of CARDIA study and the CARDIA investigators for collecting this data.

@HFpEF Was this helpful?

@HFpEF A couple of bonus tweets...

A while back I generated this simulated data of how well a PRS would do and caught a lot of flak for it:

https://twitter.com/venkmurthy/status/1120714201142583296?s=20

Turns out the simulation was not quite spot on but was rather close. Here is the real data. The top row are for whites and bottom row for blacks.

Notice that the figures on the left which use age, sex, family history of overweight/obesity and PRS really are not very tilted.

Meaning that the range of predicted BMI really doesn't change much regardless of what the *real* BMI is.

Meanwhile, the right hand column which uses baseline BMI to predict BMI 25 years later is not all that bad with a cloud that appropriately slopes upwards.