The New Eugenics Companies

Oops, I meant to say generational health and embryo selection

If you are just returning from a nice vacation on Mars, you may not have heard that there is a wave of startup companies doing various kinds of genetic testing on people, their children, and especially their embryos. Ross Douthat just interviewed Noor Siddiqui, who has a startup called Orchid. There was a splashy roll out of a company called Nucleus, with Robert Plomin as a consultant. Then there is Herasight, led by Alex Young and Tobias Wolfram, with no less than Jonathan Anomaly (an open eugenics advocate) as a team member.

The companies, or at least their websites, are somewhat different in outlook. Orchid focuses on whole-genome sequencing, arguing that it allows them to do traditional testing for major genetic disorders better than had previously been possible. Nucleus is gauzy, with sort of a MAHA self-improvement feel. It is difficult to tell from the website exactly what they are doing, but we are led to believe that involves being “healthier” in a “generational” way. Herasight is more outwardly data-driven, with a sciency-looking fixed-width courier font on the website.

There is obviously a great deal to say about this development, not much of it positive. I will limit myself to a few topics.

They are hedging across different kinds of genetic testing

Siddiqui is particularly clear about this in the Douthat interview. People have tested embryos for major genetic disorders for a long time. Traditionally this was done in utero, but embryos fertilized in vitro are also tested for known, almost entirely monogenic, disorders. People have abortions and discard embryos if they are shown to carry major disorders. This is all relatively uncontroversial, and if the whole genome sequencing they are doing at Orchid allows it to be done better, fine.

But then there are several new levels that are much more complicated biologically and ethically. First are complex medical diseases like Type II diabetes or hypertension. These diseases are substantially heritable in twins, at least moderately SNP-heritable in unrelated individuals, and can be predicted reasonably well (but not decisively) with polygenic scores. From an ethical point of view, they have the advantage of being more or less consensually bad: one can make a reasonable argument that the world would be better off if no one had diabetes. If most of us were offered a guarantee of being free from complex chronic disease, we would take it. It’s not quite that simple, but good enough for now.

The next level is psychiatric disease. Psychiatric syndromes are quite heritable in twins, but substantially less so in SNPs; for many diseases the SNP heritablity is under .1. Polygenic scores perform very poorly in psychiatry, and for most syndromes the within-family performance (the only thing that counts when selecting embryos) is unknown. It is hard to get enough sibling pairs who differ in their depression diagnosis to get a good estimate. And while the most severe psychiatric diseases like schizophrenia share with chronic physical disease that they are pretty much universally negative, for much of psychiatry it isn’t so simple. Sure, it would be nice if no one ever got depressed, but what would that actually mean? People are going to be sad, and the most modern models of psychiatric disorders view most of them as extremes on continuous dimensions of normal behavior. I don’t think we want some kind of genetic version of Soma.

Finally, there is variation in normal (in the sense that it applies to everyone) complex traits, like personality and the eugenic elephant in the room, intelligence. These traits share with psychiatric disease that SNP heritabilities aren’t very high and polygenic scores perform poorly. Intelligence-related traits are also the ones where genetic effects decline the most rapidly within families as opposed to between them. They lack any of the consensual ethics that makes selection against disease at least somewhat plausible. From the beginning, it has been received tech-bro wisdom that of course we want to create people who are as intelligent as possible, but how do they know this? We don’t want people who are as honest as possible, or as loving, or as rule-abiding, or norm-shattering? This has always been one of the bottom-line problems with eugenics, one that I don’t need to pursue here. Francis Galton was the original tech bro.

Anyway, the problem with the new eugenics companies is that they are on the safest ground for the monogenic conditions that are not, if we are honest, what these companies are all about, and on the shakiest ground for the complex human traits which are really the reason people care so much about all of this. That’s why Nucleus is so vague about what exactly it is up to in making people “generationally healthier”. It is why Herasight professes to be entirely about disease prevention, but includes intelligence in its widget that users are encouraged to “play” with.

Pleiotropy is a Huge Problem

Pleiotropy is one of several terms from classical genetics that is applied to modern, complex human genetics without much thought, but which has a very different meaning now than it used to have. Pleiotropy refers to genes that have an effect on more than one phenotype. If you have a gene that is related to segmentation patterns in beetles, and then that gene also turns out to cause differences in coloration, that is pleiotropy. But in human behavioral polygenic scores, we generally have no knowledge whatsoever about which genes do what or when. A polygenic score, by design, is acausal. It is a correlation accumulator, designed to predict some outcome by any means necessary. So what we mean when we say that a polygenic score for depression is pleiotropic, we mean that something that is (by design) correlated with depression is also correlated with a million other things. There are no behavioral polygenic scores for which this isn’t true. Pleiotropy is just a general way of saying we have no idea what is going on with PGS on a causal level.

That may be fine for some purposes, but it has pretty important implications if you are thinking about selecting humans on the basis of the PGS. Eugenics-adjacent people have a way of talking about selection as though it was easy-peasy. Pick your trait, encourage people who are high on that trait to reproduce, and voila. But it isn’t that easy, even on the farm. Modern forms of the breeder’s equation, the mathematical representation of how phenotypes respond to selection, are multivariate. Animal breeders know that selecting for one trait also winds up selecting for other traits, often undesirable or even deadly traits, once again even under controlled conditions on the farm.

Imagine a world in which a very effective polygenic score for depression is developed, and it really takes off. Everyone starts reproducing that way, and selecting embryos to try to protect their kids from depression. What would happen? We have no effing idea what would happen. That means embryo selection companies face a dilemma: they can rely on the fact that polygenic scores for behavior don’t work well enough to make a difference one way or another, or they are conducting unregulated experiments on human babies, never mind the human genome at large.

Consumer understanding of genetic information is a big problem

Quick: Suppose you are at the 80th percentile of a polygenic score for diabetes that has demonstrated an AUC (area under the curve) of .9 for dichotomously-defined diabetes in a British population. The rate of diabetes is somewhat higher in your location than it is in England. What is your risk for diabetes? Is it something you should worry about?

It would take me an hour to figure it out, and even then the practical considerations that follow— Should I change my diet? Have my kids tested?— would remain murky. How do we expect someone with no knowledge of statistics, no understanding of effect size, no notion of where polygenic scores come from, no idea of the difference between absolute and relative risk, to think about a problem like this? The eugenics companies know this, and are using people’s ignorance to line their pockets. Sure, at some level everyone has a right to whatever knowledge is out there about their own genome, but dissemination of that knowledge should be based on something more than a toxic combination of libertarianism, greed, and ignorance.

Regulation and reproductive freedom

Exactly what should be done about this is a very difficult problem that is outside my area of expertise, so I will be brief. As a matter of reproductive freedom, I don’t really see how you can tell people what they are allowed to test for in their own genomes or the genomes or their kids and embryos. Presumably it will still be allowed to test for Down’s syndrome, so how can you outlaw computing a polygenic score for extraversion? It isn’t that hard to do it on your own for free if you really want to.

So one way or the other the regulation has to come on the side of the companies pushing the testing. Suppose some lab developed a new drug for depression. There is a whole regulatory apparatus built around determining that the drug is effective and safe before it is marketed to consumers. Maybe I am missing something, but I don’t really see why this logic shouldn’t apply to genetic testing. On the one hand there is the possibility of fraud, if the tests don’t really work; on the other hand there is the possibility of real damage, if selecting for one trait turns out to create risks for others. How are those scientific consultants going to feel on the day it turns out that embryos selected for IQ are at increased risk for autism? How sure are they that it won’t happen?

The alternate model for regulating drug-like things is the supplement industry, which seems to me to be the way these companies are operating now. Supplement companies tell you, vaguely, that their product “supports the immune system,” or some bullshit like that. It sounds a lot like Nucleus telling you that you can “build generational health.” The implicit assumption behind supplements is a lot like complex trait genomics: Hey, our product might not help very much, but as far as we know it won’t do any harm either, so why not give it a try?” I assume that the system determining whether a given product is a regulated drug or an unregulated supplement is very complicated, driven by money, and probably broken. We should try to do better in the early days of genetic testing.

Where is ssgac?

The Social Science Genetic Association Consortium has been at the forefront of scientific applications of polygenic scores, almost entirely in a positive way. Only a few of them (and I sometimes use ssgac to refer to social scientists using genomics, rather than people I know to be official members of the organization) have turned up on the boards of the new companies. I wrote a few weeks ago about how social science genomics was getting squeezed by the increasingly tiny effect sizes for human behavioral heritabilities and effect sizes. This spring I wrote a post about it. As scientists have gotten better and better at controlling for non-genetic confounds of polygenic scores, the effect sizes for predicting behavioral phenotypes have been going down instead of up. In the paper I discussed in that post, they finally reached zero. As a practical matter, I have tried to use polygenic scores in my applied research, and been frustrated by the fact that they don’t predict anything well enough to be useful. A lot of the time they point in the wrong direction. I have referred to it as an “effect size crisis.”

Now ssgac is getting squeezed from the opposite direction. ssgac researchers wrote dozens of well-intended FAQs attached to the big behavioral GWAS, assuring the readers and the participants that the science wasn’t appropriate for applied commerical use, and wasn’t intended to do that. Now we have all these companies springing up doing exactly that, some of them led by people who were authors of those FAQs. The response from the ssgac establishment has been deafening silence. I don’t like it when people call me out to comment on something that they think will be a gotcha for me, so I won’t name names, but I really do think we need to hear from the responsible scientists who have been defending the use of genomics in social science, and polygenic scores in particular. There was certainly plenty of noise when the educational attainment PGS accounted for 15% of the variance in EA3, with dotted lines pointing forever upward, so where are we now? Is genomics a weak but occasionally useful scientific tool for understanding social and behavioral phenomena, or is it a “game-changer” that is about to revolutionize social science? (Plomin and Murray). Are there legitimate uses for it in the clinic, beyond identifying major genetic disorders?

The FAQ has gotten a lot more complicated.