Tweet thread on the excellent paper in nature genetics on GWAS of intelligence. First off, congratulations to the authors not just on the paper, but also this extensive FAQ here : thessgac.org/faqs. Paper is here. nature.com/articles/s4158…
A reminder for people who did not see this paper - the authors combined a number of cohorts (@uk_biobank - UK and @23andme - mainly US - along with others) for a ~1.1 million person of European descent GWAS on "years in education"
Years in education feels like a crummy statistic for something like intelligence, but it is something you can get for many cohorts, and ... very often sample size beats phenotyping accuracy . Years in education basically goes into two main classes "university/college or not"
(by actually having a numerical scale the authors can also regress out other factors to remove non genetic variance, eg, sex which the authors do)
Broadly - I believe the results - they did a really good job - and the "headline" numbers of >1,000 loci, 11% variance explained and association with neurological associated genes look robust
Some other take homes / concerns for me. I *still* have nagging doubts about the confounding of educational provision/style and region, even with lots of PCs, with regional genetics - true in all countries, definitely true in both UK and US.
As many - most - variants show reproducibility between cohorts, it can't be the case that this is the main driver to this result, but ... untangling the details is going to be really fiddly (see below!)
(I note the Bristol crew - @nic_timpson, @mendel_random, @carolinerelton etc understand this at a more mathematical / statistical level and I am sure will weight into this debate at some point)
I also note that there is hetreogeneity in the betas overall - ie, the cohorts don't precisely align and their effect size estimates are overdispersed, without being able to pin-point any particular set (I suspect we need some method development here)
There are two obvious explanations for this. One is subtle geographic confounding (see above) or - far more interesting - differences in the educational environment with genetics.
The next comment is that although 11% is impressive statistically, the authors - and commentators - point out that it's not good enough to predict an individual. That said, the *extremes* of the PRS do look (by eye) quite a shift in odds in whether someone goes to uni or not
In other fields people are thinking about using the extremes of the PRS in an actionable way; I don't think this is at all warranted here, but I just want to point out the difference in framing a robust PRS score
Finally I really like the association they do with other endophenotypes - self reported maths ability, other things.
I think this is the way to go - despite the statement about sample size beating phenotype accuracy, when one wants to understand *why* and *what* one has discovered, one needs to drill into the actually underlying mechanism of these genetics effects
It's going to be particularly interesting when (a) the @uk_biobank imaging comes on line and (b) when/if things like the Scottish or Danish detailed school records can be linked anonymously to genetics to allow for a higher dimensional view (multiple exams!) of "intelligence"
Two things I strongly agree with the authors from their FAQ (a) this does not have policy implications - it really is 'more research needed' and I caution anyone doing anything with things we don't understand well and haven't had different groups work at it, the science maturing
(b) This is not predictive of any individual, and in particular is not applicable outside of Europeans due to the assumptions needed to bring to the GWAS about 'mixing' population, the PRS wont simply translate to other allele frequency distributions (ie, ethnic groups)
