Thread by @ewanbirney

I fielded a nice, thoughtful approach by a journalist last night on @skathire's PRS in obesity paper and ... this is for the other journalists entranced by the combination of obesity, genetics, and what's going on here...

First off, @skathire's lab's paper is good, but in no way is the idea (that genetics contributes to obesity) or the methodology (polygenic risk scores) is new. Sek and crew have done a really good job of bringing PRSs to obesity but ... no "breakthrough here" just, good genetics

Second; it is a deliberate modelling choice in PRS to "not know the loci" and thus "not provide any biological insight". One effectively trades statistical power in finding loci for statistical power in "making good predictions" (more on the slippery word prediction in a moment!)

One can use genetics to understand the biological components behind this. This is called "GWAS" (Genome Wide Association Study). The latest/greatest here is europepmc.org/abstract/MED/3… or if you prefer the preprint, biorxiv.org/content/biorxi….

Note in the GWAS they state: "while BMI-associated
genes are mostly enriched among genes involved in neurogenesis and more generally involved in the development of the central nervous system. These last results therefore confirm findings from Wood et al. (2014)..."

(So - the genetics of BMI points towards neurological features, and the combination of this signal and rare disease genetics in obesity points pretty firmly towards satiation control)

(Small editorial note: This HMG paper is - to be honest - as informative if not more than Sek's paper - Sek's paper focuses like a laser on the properties of the PRS. Lots of overlap in data and authors as well. Nothing wrong with Sek's paper, but don't consider it in isolation)

Third: The PRS model does not "imply 6 million causal SNPs". Rather they fit a model with 6 million SNPs to model the underlying genetics. One of the beauties of the PRS scheme is that it (brilliantly) doesn't care about how many or where the genetic loci precisely *are*

This is perhaps best understood from where PRS conceptually comes from - animal breeding pedigrees - trying to predict the milk yield for cows of a particular bull's genetics is the "classic" animal breeding problem.

Before genetics got cheap, one did this with careful pedigrees. You knew the bull's pedigree. You knew the rest of his pedigrees (cousins of different sorts) milk yield; without seeing a signal progeny of this bull you could predict the milk yield of his children well

You did this by integrating the information across the pedigree. When genetics got cheaper the animal breeders realised first that they didn't have to keep careful records of their pedigree (the genetics told them) but they could also leverage genetics to get better predictions

This is because you not only knew which was this bull's cousins, but also which bit of DNA is shared with this cousin *and* how this rough piece of DNA contributed in other bulls (similar to the stats done by players in team sports BTW)

One could built a gestalt model which - for any bull from this overall population - one could in a clever, mosaic way, build a prediction for the bull's progeny's milk yield. It made assumptions of many loci across the genome contributing to this, and the population structure

It did - and does still - work really well, and has this extremely pleasing unity to the very old theories behind genetics, in particular RA Fisher's synthesis of mendelian genetics and biometrics in the 1920s

(geneticists will now go a bit misty eyed about the foresight of this from RA Fisher - without an understanding of the mechanisms of genetics he created a theory that still works extremely well now 100 years earlier)

Back to humans - you can make the same sort of predictions of any trait in a human, in some sense building a predictive mosaic from all the other humans you have seen. This ends up being the same as fitting a linear model with some clever tricks on 6 million SNPs.

This *does not mean* all 6 million SNPs do stuff! that's not the point of this model at all! It means the process of building a "mosaic prediction of a new human from all previous humans we've observed" is mathematically the same as fitting a linear (additive) model on SNPs

(The biologists might say - as did the mendelians say 100 years ago - wait a second, surely this can't be linear - recessive/dominance at least, and then all sorts of complex things. To which the standard answer is that - all models are wrong, some models are useful >>

The linear, additive model of genetics has been remarkably robust and demonstrably useful - not least in animal breeding but also in discovering underlying human genetics - GWAS. Now - are these PRS models in humans *useful* - well that's another thread).

This "mosaic of previously seen humans" is also why the word prediction can conjure the wrong impression of our model; that somehow we know what every variant does wrt to obesity. We absolutely don't know *at all* and this is a deliberate modelling choice

For the journalists I've also been asked for my perspective on this work. It's good solid work from @skathire's team. hat's off. Personally I think his work on CAD (heart attacks) and @GENES_PK's + others (Doug Easton) on Breast Cancer PRS is more useful now

I also think people are not bringing in the cross country perspective enough. Across the OECD obesity levels differ alot with largely identical "European origin" genetics (and the genetics will be different in details but not overall theme elsewhere): oecd.org/els/health-sys…

The Netherlands is probably your stand out comparator country to US (or UK). Something systematic in the "environment" (genetics-speak for "not genetics" - so "everything else") is different in Holland to the US leading to radically lower obesity levels.

I personally think as well as understanding the individual case, unpicking this "environmental" effect is important - is it - Tulips; food labelling; food culture; school meals; Biking and town planning; The total football philosophy; which makes NL slim? Something does...