* Lots can go wrong during meiosis; many gametes have severe genetic abnormalities (e.g. aneuploidy). Due to these abnormalities, upwards of 20% of conceptions result in spontaneous abortion (often before the mother even knows she’s pregnant).
* Normal stem cell lines already have a proclivity to undergo cancer-like selection when being cultured, acquiring growth-enhancing mutations.
* Haploid cell lines are genetically unstable and have a proclivity to revert back to a diploid state (or quickly acquire chromosomal abnormalities in the process of trying to revert to a diploid state).
* Haploid cell lines that manage to remain haploid can select for cancerous mutations much more quickly than their diploid counterparts.
This doesn’t even touch on issues with polygenic scores, or that optimizing for a genome with the most alleles contributing to a PGS for a given trait has no guarantee of actually maximizing that trait, since many alleles are epistatic—their effects are dependent on the presence or absence of a specific set of other alleles, which is not captured by the simple linear mixed models used to compute a PGS. Analogously, throwing together all the most commonly used spices across all 3 Michelin star restaurants would result in a horrible dish, even though each spice on its own would contribute highly to a “polyspice dish score.”
Even if we had some magic way of solving all the aforementioned problems, we still have no way of knowing that optimizing for a certain set of alleles that definitively results in some desirable traits would not also inadvertently result in some other horrible traits, which we would have no way of knowing from population statistics because they’ve been selected out of the population.
I really like the food analogy here. This is a perfect explanation for why embryo selection using this method would absolutely fail. Polygenic scores are touted as this magic bullet for understanding the contribution of genetics to a phenotype but are utterly imperfect models for what actually occurs due to pleiotropy and epigenetics (among other things). In my opinion the best way for embryo selection is to just have a lot of offspring. Let natural selection do it's job. It gave us us after all.
> Polygenic scores are touted as this magic bullet for understanding the contribution of genetics to a phenotype but are utterly imperfect models for what actually occurs due to pleiotropy and epigenetics
Agreed! To be a bit more charitable to polygenic scores, they can be mediocre descriptive models—given a real individual’s genotype, they can sometimes decently predict their phenotype (if you consider explaining ~10% of phenotypic variance “decent”).
They are unequivocally horrible generative models—you will not obtain a synthetic genotype that would result in a desired phenotype by sampling from a PGS model, i.e. by randomly maximizing the number of high-scoring PGS alleles in a synthetic genome, as proposed by the article.
I like the contrast between descriptive vs generative here. Really puts this into context. If you want to understand the contribution of any particular allele PGS is reasonable. But using PGS to generate a phenotype violates so many assumptions.
Author here (sorry for the late response, I didn't see this got posted to HN).
These are all valid concerns (especially about spontaneous diploidization), but I think they can be overcome with sufficiently rigorous quality control (probably whole genome sequencing, not just genotyping).
The issues with epistasis (i.e. will polygenic scores continue to be valid at extreme values) are also worth noting. I think that improved models (besides the linear ones in current use) are likely to solve this though.
I’m not a biologist, but there’s one thing about this proposal that seems very odd to me: If the goal were to produce many haploid cells, select for some phenotypic trait, and repeat, then this would make sense. When one is trying to evolve some trait, growing a large number of organisms and letting them (or helping them) select for that trait makes sense. But in this proposal, the goal seems to be to find a input to a function on a computer that gets a good score. What’s the point of using actual cells?
I am also not a biologist, but I apparently think I know some of this? Hopefully if I'm wrong enough, a real biologist can correct me. :)
I think we're still very far from arbitrary genetic editing and whole genome design and synthesis. I don't know how much of this is "technically feasible but way too expensive" vs "we have promising research directions on some of the challenges, but definitely don't know how to do this effectively at all", but it all adds up to that not being an option today.
The best we actually have available today is embryo selection. We don't know exactly what phenotype each embryo will express, but we've got enough statistical clues that we can make fairly good guesses most of the time. So, you make 100 embryos, sequence their DNA, then choose whichever embryo is given the best score by your statistical analysis, and use that one.
Embryo selection is far from perfect, but it lets you avoid any genetic conditions that we know how to detect, and it lets you get an embryo with more of the preferred variants of genes statistically correlated with whatever you want to select on, and it's absurdly cheaper and more efficient than alternatives like "Let each embryo grow for 30 years, then evaluate the results of the genome".
With iterated embryo selection, the idea is that instead of one large batch of 100, you do a small batch of 10, choose the best from that group, and use that as source material to make another batch of 10. This acts as a ratchet, letting you get more selection with fewer embryos.
So, the point of finding an input to a function on a computer that gets a good score is that it's the cheapest, most efficient way we know of so far to get a human embryo whose genome scores highly on this function. There are also some other nice properties, like having the embryo be highly correlated with the genomes of the parents, which a lot of people like, and is something you can explicitly add to the scoring if you want.
Just like any selection/optimization procedure, you can get some pretty bad outcomes when you go to extremes, as for a lot of traits, your function on the computer is only a statistical model of observed correlations, not a real comprehensive model of what the genome means and does. I imagine you're probably going to get something with severe issues if you tried to grow the embryo you get from 10,000 rounds of IES.
Embryo selection is trying to use what we think we know to choose the best embryo, given only the genetics. We don't have perfect knowledge, but we do have some knowledge, and we can make some bounded use of it.
> Genotype the haploid lines and select the best ones.
I think this is the really hard part, and part which cannot be made reliable without growing enough intermediate specimens sometimes and assessing their success in the actual environment.
If we were actually any good at predicting the success of an organism just from genetic composition, we'd be able to just genetically engineer a highly successful variant, obviating the many intermediate steps.
Isn't that a problem that machine learning is good at? There are so many things that ML will never be able to do, but making a simple prediction is what they are best at.
Only if the function that they are trying to approximate is differentiable. Also, it depends on having enough training data, and on the training data actually being representative of the domain of the function.
More specifically: polygenic scores work and are additive, while epistasis and dominance are trivial (contra the furious FUDing upthread by someone determinedly ignoring the fact that, no matter how many cute restaurant analogies you use, we have lots of phenome-wide data from UKBB and elsewhere showing additive variance is larger than epistatic variance and dominance is so irrelevant it's hard to distinguish from 0%, not that that's relevant in the first place), because they use SNPs to tag blocks or stretches of the genome which carry the causal variants, the SNPs aren't themselves usually causal. So if you directly edited all of the SNPs with nontrivial posterior probability of the effect you want, you would get much smaller effects than you expect (as much as 90% smaller). But if you operate on the blocks (haplotypes), the SNPs continue to serve their existing role as markers, and you can optimize by selecting for all of the useful blocks and dropping the rest. (This is how IES would work, especially if it could run for many generations/cycles; operating on just the SNP level is only what people who haven't thought about it think it would do.) A lot of problems go away when you are doing selection, preserving all the existing causal relationships (like the extensive positive correlation among health traits which means that you won't get any net 'backfire effects'), instead of causal interventions like editing.
This doesn't make any sense to me. I tried to parse it several times (I'm a PhD computational biologist but no expert in this area)
Thinking in snps and haplotypes has been tripping up geneticists for years. Everything in genomics is highly nonlinear, not additive, and with massive amounts of long range interactions, feedback, and state.
Also organismal phenotypes are very different from molecular phenotypes- often organismal phenotypes are actually several underlying things grouped with a common label simply because they have similar features.
For legal / "ethical" reasons, I'd try this on dogs, horses, or some other domesticated mammal first. You're not trying to engineer humanity; no, you just want to reverse the degradation of pure breed subspecies.
* Lots can go wrong during meiosis; many gametes have severe genetic abnormalities (e.g. aneuploidy). Due to these abnormalities, upwards of 20% of conceptions result in spontaneous abortion (often before the mother even knows she’s pregnant).
* Normal stem cell lines already have a proclivity to undergo cancer-like selection when being cultured, acquiring growth-enhancing mutations.
* Haploid cell lines are genetically unstable and have a proclivity to revert back to a diploid state (or quickly acquire chromosomal abnormalities in the process of trying to revert to a diploid state).
* Haploid cell lines that manage to remain haploid can select for cancerous mutations much more quickly than their diploid counterparts.
This doesn’t even touch on issues with polygenic scores, or that optimizing for a genome with the most alleles contributing to a PGS for a given trait has no guarantee of actually maximizing that trait, since many alleles are epistatic—their effects are dependent on the presence or absence of a specific set of other alleles, which is not captured by the simple linear mixed models used to compute a PGS. Analogously, throwing together all the most commonly used spices across all 3 Michelin star restaurants would result in a horrible dish, even though each spice on its own would contribute highly to a “polyspice dish score.”
Even if we had some magic way of solving all the aforementioned problems, we still have no way of knowing that optimizing for a certain set of alleles that definitively results in some desirable traits would not also inadvertently result in some other horrible traits, which we would have no way of knowing from population statistics because they’ve been selected out of the population.