26 Jul Loneliness and the Future of GWAS
Last week, after the publication of EA3, Cecile Janssens tweeted about another GWAS she had come across.
The new low in GWAS:— Cecile Janssens (@cecilejanssens) July 20, 2018
"Further loci were identified for regular attendance at a sports club or gym (N = 6 loci), pub or social club (N = 13) or religious group (N = 18)."
The good news:
Only 3 people wasted their time. pic.twitter.com/ycWwcqXSBj
Unsurprisingly this got a lot of pushback, for example this from Stuart Ritchie:
Follow this thread for a really weird, nonsensical criticism of genome-wide association studies (apparently this GWAS is a "new low" because lonely people might get offended by scientists working out which genes are linked to higher reported loneliness? Is that the argument?). https://t.co/ERXJje9m8w— Stuart Ritchie (@StuartJRitchie) July 21, 2018
I don’t think Janssens’ criticism was nonsensical at all, in fact I think I know exactly what she meant, but here and in a bunch of other tweets I don’t think she got her finger exactly on the problem. (I probably shouldn’t characterize her position, which is spread out over a bunch of discussions with a bunch of different people. So her tweet is just my starting point, and I am sure she will jump in with her own version.)
I think Janssens suggestion is that although GWAS makes sense for some traits, for something as complex and highly psychological as loneliness it doesn’t apply:
Yes, we need to study it, couldn't agree more. But why genomics when the problems are hardly genetic? I don't see that.— Cecile Janssens (@cecilejanssens) July 21, 2018
In fact in other posts she is more enthusiastic about applications for EA3 than I would ever be:
New genetic study on education concludes that polygenic risk score might be useful for research. Yet, if fig 4a is correct, it has huge potential for screening college admissions too. Only 10% in low PGS quintile completed college vs ~50% in high PGS. https://t.co/sJrxfM7vbs pic.twitter.com/E2L9o7DMx6— Cecile Janssens (@cecilejanssens) July 23, 2018
So, I think the idea here is that GWAS makes sense for real, chunky traits like EA and heart disease and depression, for which some kind of bottom-up genetic causation seems plausible. But it doesn’t make sense for super-complex, intuitively psychological traits like loneliness.
The problem is that thanks to the first law of behavior genetics that distinction is a very difficult one to maintain. Unsurprisingly, twin studies show that loneliness is heritable. Everything is heritable. And because loneliness is heritable, there are SNPs correlated with loneliness, and if you do a big enough GWAS you are going to find them. Then you can string them together to make a PGS etc etc etc, and if you can eventually make the sample big enough it is all going to come out looking pretty much like EA3.
The possibility of a GWAS of loneliness means one of two things. Either it is a proof of concept for GWAS (and the genetics of) pretty much anything, or it is a reductio ad absurdum of the broader project that we are so now prone to take seriously when the topic is EA or depression. Again, I don’t want to put ideas in Janssens head, but my version of the argument is that it doesn’t matter that there are SNPs correlated with who winds up lonely. The fact that certain outcomes are correlated with the likelihood of certain lower level components like genes does not mean that you can explain the outcomes in terms of the components. As I once said somewhere, you aren’t going to explain plate tectonics by looking at the chemical composition of the rocks in your backyard, despite the fact that tectonic dynamics are ultimately composed of rocks.
We need to go back to twin studies and the first law of BG for a moment. Once upon a time, when we first started doing twin studies, the research focused on all the usual suspects that seemed to make sense for genetic analysis: intelligence, depression, etc. And ohmygod they are all genetic! It must be only a matter of time, people thought, before we understand the genetic mechanisms that underlie twin study heritability of educational attainment. But then something funny happened. It turned out that how much TV you watch is just about as heritable as extraversion; your marital status is just about as heritable as depression. So the basic distinction we wanted to establish, between real, deep, “biological” things that were ripe for genetic analysis, and complex psychological things that were not, broke down. Everything is heritable (First Law), in exactly the same causally vague way (Fourth Law).
In fact it was even worse than that. Multivariate behavior genetics came along, and that meant that we were no longer constrained by the traits that it had occurred to someone to put on a questionnaire. Any rotation of any combination of traits could be studied, and they were all heritable. What did this mean? It meant that the traits themselves didn’t actually have anything to do with it. There is no twin genetics of depression, there is only the observation that one way or another MZ twins wind up more similar for DZ twins for everything. Depression is just one rotated region in an infinite multidimensional space. So is TV-watching.
As a consequence, you can use twins to predict anything about people you want. Educational attainment? Sure. Marital status? You bet?.Having red hair and over size 12 feet? Works great. People who start out more similar genetically wind up more similar phenotypically. Anything beyond that about “genetic mechanims,” “biological pathways” or whatever is speculation. Show me the money.
We are now right in the middle of re-experiencing this story in molecular genetics. We have impressive GWAS for schizophrenia and EA, they are coming along for depression and personality. Sooner or later we will have them for going to the gym. It is right where twin genetics was in say, 1990. What bothered Janssen, I am speculating, is what might have (or should have) bothered a twin researcher when it turned out that marital status is just as heritable as depression. If the heritability of depression is evidence that it is a “genetic disorder,” and marital status is heritable, then…. uh oh.
The whole multivariate scenario is going down right now for GWAS. The recent Genomic SEM method developed by Tucker-Drob et al does for GWAS exactly what Martin and Eaves did for twin studies in 1991. On the one hand it’s a spectactular tool at the level of conducting studies and computing statistics; on the other hand it is the end of the time when we can think of EA or depression or anything as some kind of specially genetic trait the “genetics” of which has been discovered. There are just two complex multidimensional spaces– genomic variation and phenotypic variation– and they can be rotated in different ways to bring them into different kinds of alignment. The genotypic component isn’t causing the phenotypic component in any biologically understandable way, it’s just a statistical exercise. Sometimes statistical exercises can be very useful.
So neither of the two original possibilities– that GWAS of loneliness is meaningless, or that there is a meaningful molecular genetics of loneliness– is correct. GWAS is what it is: a way of finding SNPs that are correlated with a phenotypic outcome, and there are SNPs associated with ALL phenotypic outcomes, by the first law of BG. But this is not evidence that it makes any sense to explain loneliness in terms of the SNPs. Have you noticed that all the discussion about EA3 has involved raw prediction? How much have you heard about the biological annotation that presumably follows from those SNPs? Not much, and my contention is that the silence is because even the downtown GWAS community knows it isn’t working. We are witnessing the quiet end of the whole gene-finding project that was supposed to be the point of the exercise.
All of this puts very severe constraints on the real-world interpretability of SNP correlations. Because we have no idea about the mechanism of the SNP correlations, we have no way of knowing how general, or how modifiable, the association between the SNPs and the outcome might be. And that makes it impossible to know whether EA3 is ultimately more like some kind of relatively hard-wired genetic effect (there aren’t really any good examples, but something like lactose persistence) or just genetic background noise, as our intuitions tell us will be the case for how often you go to the gym.
The long run danger for GWAS is not that it is going to turn out to be wrong, it is that it will be swallowed by the gloomy prospect, just like twin research was.