Thread by @ewanbirney

In general the response I think to the announcement of a polygenic-risk-score informed embryo selection has been right - one where the science is wrong, the clinical harm/benefit therefore also wrong, and one where ethical/societal considerations have to be folded in. However...

There are some people who say "but even if this is wrong now, it might not in the future" (true) and also "if genetics works, then this should work" often with some handwaving towards farm animal genetics/breeding/selection. In this twitter thread I'd like to tackle this.

(Context: I am a geneticist/genomicist. My two favourite organisms to study humans and Japanese rice paddy fish. I'm on the experiments/practical data science side, but have a pretty good understanding of the theory/stats side, partly because I've coded it myself/in my group)

So - the first point, saying no now doesn't mean no forever is true, but does show the need for good regulation, and there are at least 3 strands.

1. *can* one actually do something useful (science) 2. does the utility substantially outweigh harms (clinical) and 3. *should* we allow people to do this, if 1 and 2 are true (ethics and societal). 3 "trumps" 1 and 2, and 3 has to be decided by a societal process.

But lets go into 1 - science - can one actually do something useful now in theory? - if you have (say) 20 embryos, take one cell out, sequence each genome, and score not only for rare disease (which one does now for CFTR, other diseases) but also common via PRS?

One would then, as one does for pre-implantation genetic diagnosis, then implant a set of "good" embryos back into the mother (assumming she provided the eggs; the example that triggered this it was an egg donor with a surrogate mother - another layer of ethics).

First off, it's important to stress that the PRS for choosing between embryos is constrained by the parents (at least in the simple additive model from PRS); the parents already put bounds on what is feasible.

Second, PRSs are constructed to "predict" an average person presenting to a research/clinician, and in particular current PRSs accept they don't have the actual causal change in their model- it is linked to on average in the population. This is not the case in the embryo scenario

So we have an extra bit of uncertainity which is the linkage disequilibrium "trick" from the PRS chosen variants to the causal one, and if, for this parent, the causal variant is not there, no matter how many embryos one samples, the causal variant ... still wont be there.

Third we know the out-sample prediction (when we look at unknown people coming through the door, not how well the model thinks it is going to do) is worse than the model (no surprise) - its definitely not zero, but not worse.

The consequence of 1, 2 and 3 is that we can be confident that PRS for embryo selection is going to have far far more variance than PRS for out of sample in the population, itself worse than "variance explained".

In weighing up potential benefit vs potential harms, our potential benefit is clearly way down from our model readout.

This is *not* like farm animal breeding. In farm animal breeding the thing we are doing is selecting a parent (classically bull semen) on the basis of the often known and predicted score of their offspring for the trait.

Furthermore the ability to blend both genetics and observations of this genetics over multiple generations in the same environment means that we constrain the prediction problem on the parent alot.

But there's an even bigger issue - unintended consequences. This is best illustrated for me by the successful genomic selection on chicken layers (laying eggs) and broilers (for eating) which created great individual birds for both via genetics... but...

when these "Elite chickens" (this is the right term) became flocks of chickens, yields dropped. What happened? In this case, the selection for the these traits led to aggressive birds (I can't remember which one more) which was not being tracked.

Now, this is pretty extreme - multiple generations of selection on a narrowly focused trait - but it goes to the problem of these models which is that we don't know the universe of traits for a human that we care about.

There is a possibility that selecting for good attributes for one trait is going to have weird knock on impacts somewhere else.

Here rare disease is different as we know that the genetics is just bad and as the vast majority of the world doesn't have it, selecting the "correct" variant is selecting for something pretty much everyone has.

A valid question is could I *ever* see PRS for embryo selection happening sometime in the future? The closest I can imagine is the scenario you are doing PGD anyway (so the harms are accepted for rare disease genetic reasons) and a PRS with clear harms (eg, LDL cholesterol) >>

<< where the PRS was constructed from confirmed causal variants (no LD one-step-away with the embryo) and the PRS clearly did not have unintended consequences on a very very large number of traits. That latter one is a tough bar to get over; I can't imagine it this decade.

Circling back to original piece, and as I've said multiple times; I am positive about using PRSs in medicine in the right place, but this is *not* a good of PRS - for scientific reasons, for clinical reasons (I'm not an expert) and I'd also argue for ethical reasons.