I’m basing this off of selection, not editing. I haven’t looked into the genetics stuff very much, because the bottleneck is biotech, not polygenic scores.
Would look forward to your rebuttal! I just hope you’ll respond to the strongest arguments, not the weakest. In particular, if you want to argue against the potential effectiveness of selection methods, I think you’d want to either argue that PGSes aren’t picking up causal variants at all (I mean, that there’s a large amount of correlation that isn’t causation); or that the causality would top out / have strongly diminishing returns on the trait. Selection methods would capture approximately all of the causal stuff that the PGS is picking up, even if it’s not even due to SNPs but rather rarer SNVs. (However, this would not apply to population stratification or something; then I’d think you’d want to argue that this is much / most of what PGSes are picking up, and there’d be already-made counterarguments to this that you should respond to in order to be convincing.)
I’ll repeat that I’m not very learned about genetics, so if you want to convince even me in particular, the best way is to respond to the strongest case, which I can’t present. But ok:
First I’ll say that an empirical set of facts I’d quite like to have for many traits (disease, mental disease, IQ, personality) would be validation of tails. E.g. if you look at the bottom 1% on a disease PRS, what’s the probability of disease? Similarly for IQ.
or beyond those?
I rarely make claims about going much beyond natural results; generally I think it’s pretty plausible there’s some meaningful thing we could feasibly do that’s like +6 -- +8 SDs on intelligence, but I’m much less confident about the +8 SD claim, and not super confident of the +6 SD. Like, I think the default expectation ought to be that we can meaningfully get to +6 SDs; this seems like the straightforward conclusion. (I’m just restating the intuition / impression.)
Why do you expect that only a very small fraction of natural SNP differences are needed to get the extremes of natural results
Assuming linearity, the math is fairly straightforward. In the simplest model, with 10,000 fair +1/-1 coins (representing all the variance in a trait, so some coins are environmental), an SD is 50 coins and the average is 5,000. So there’s 100 SDs of variance available. Obviously this is mostly meaningless in terms of the trait, as linearity would not remotely hold, but my point is that the issue isn’t the math of additive selection. See here for more (e.g. about if the coins are biased https://tsvibt.blogspot.com/2022/08/the-power-of-selection.html#7-the-limits-of-selection ).
IQ seems to have thousands of small contributions from different regions. 10% of the variance is therefore in the ballpark of 10 trait SDs. Again, I’m not saying you can get to 250 IQ; what I’m saying is that the math of selection and variance isn’t the problem. Lee et al. state “In the WLS, the MTAG score predicts 9.7% of the variance in cognitive performance[...]”; this was in 2018, I would bet we can do substantially better now.
Lee, James J., Robbee Wedow, Aysu Okbay, Edward Kong, Omeed Maghzian, Meghan Zacher, Tuan Anh Nguyen-Viet, et al. ‘Gene Discovery and Polygenic Prediction from a 1.1-Million-Person GWAS of Educational Attainment’. Nature Genetics 50, no. 8 (August 2018): 1112–21. https://doi.org/10.1038/s41588-018-0147-3.
Why do you expect that effects would be linear?
TBC I certainly don’t expect the effects to be literally linear even in the typical human range; it’s more that expect them to be fairly linear. Like if the answer is that a trait-mean couple that selects their child’s genome to have a (carefully, accurately as best anyone can) predicted IQ of mean 170 actually tends to have a child with mean IQ 155, I’d shrug and be like “huh, that’s weird and surprising, let’s investigate and make sure to communicate this fact to parents”; and I think this possibility is nontrivially strategically relevant; and it means we should accurately describe this plausible outcome, in order to not overhype etc.; but I wouldn’t be totally shocked. If the child tends to have a mean IQ of 125, I would be shocked, yeah. (The 170 vs. 155 thing would be hard to notice for a while because testing intelligence at that range is barely feasible, but just saying for illustration.)
Also, certainly I would expect strong nonlinearities at the extreme tails at some point. I’d certainly strongly advice against, maybe even condemn, pushing noticeably outside the regime of adaptedness.
Why do I expect the effects to be fairly linear in the human envelop, to the point where increasing a bunch of causal variants increases the trait?
Some of my impressions come from here: https://arxiv.org/pdf/1408.3421
“On the genetic architecture of intelligence and other quantitative traits”, Stephen D.H. Hsu, 2014
(TBC, I’m not saying these sources present the strongest arguments; I’m just saying where my impressions historically come from.)
Breeding programs in non-humans work well, so there’s plenty of variance for those traits, no huge nonlinear walls that you hit, etc. This is far from dispositive; intelligence could plausibly be different from traits like egg production or weight, maybe it’s important that you’re checking along the way, etc.
Linear PGSes work well for many traits (height in humans; various traits for cows I think; even IQ, up to 10%).
SNP heritability estimates are substantial; the number in my head for IQ is >.3 of the variance. Though my impression is also that these are maybe controversial? I dunno.
It’s far from obvious to me that there’s been much selection for IQ in the past couple thousand years. But if there were a case that the selection has been strong, that would shift me.
There’s a theoretical argument, which IDK if it should hold much weight, but I like it: since DNA segments get shuffled around a lot, there’s selection pressure for things to work reasonably well with other things. E.g. DNA segments that have really bad effects when combined with some other segments would be selected against; and DNA segments that improve mechanisms for repairing/smoothing-out/compensating bad epistases between other segments will be selected for. In general this smooths things, which makes the landscape more linear. (I understand that specific epistases are much rarer than single variants, and therefore relatively invisible to selection; but I think my point stands somewhat, though this could be clarified and maybe basically disproven with good quantitative analysis.)
I just haven’t seen evidence of this nonlinear wall that’s right between 140 and 160, or whatever the claim is. It’s just people saying “maybe there’s a U-shaped curve of something” or “maybe there’s a high fan-in latent with a cutoff after the latent in the final IQ sum” which makes sense, but AFAIK is basically just speculation. It also isn’t super compelling speculation if we’re talking about a super polygenic trait in a super-complex organ where I’d expect there to be lots and lots of ways to tune and fix and just upregulate stuff. Like, my actual guess would be that there’s a whole spectrum of functional forms, from linear (substantial, according to h estimates!) through small ORs of ANDs, through large ORs of highly sensitive ANDs and other nonlinear forms; and these are all mixed together; and this does imply something; but it doesn’t imply that you can’t have quite large effects on the trait with germline engineering.
My impression is that the upper tail of IQ does get a little weird, and maybe g stops existing as much / the distribution of different tests stops being as one-dimensional? But IIUC (not sure, heard this from a psychometrician) there’s no observed threshold in the effect of IQ on other traits, despite people looking, though it’s quite hard to measure past 150ish. And e.g. this random paper claims to find quite substantial SNP heritability in a cohort with estimated rarity >4 SDs, though it’s not a huge cohort (1238 in the selected cohort) and I didn’t study it so maybe it’s very flawed / meaningless, IDK. https://www.nature.com/articles/mp2017121 In other words, to the small extent that we can look at the extreme tails, I at least haven’t heard of big results saying “aha! actually the genetics of IQ on the tails is quite different than near the mean!”.
(Also there’s sibling studies that IIUC say we are indeed picking up causality, though that’s not directly relevant to linearity.)
near-evolutionarily-optimal range. That has not happened with intelligence,
What makes you think this? As I said, it’s not clear to me that there’s been much selection pressure for intelligence in the past few thousand years.
Also, the “evolutionary optimum” can change. E.g. calories are not much of a problem in the developed world, but that’s recent.
Also, there’s always an influx of de novo mutations, and evolution has limited selection power. I’m not clear on the math here exactly, and I think kman has suggested that mutational load isn’t the main source of IQ-associated SNPs, but it demonstrates that it’s far from ironclad logic to infer from evolutionary pressure on a trait that the trait should be near optimum in linear variants. The brain is one of the organs with the most diverse gene expression profile (I mean, more genes are expressed in the brain than in most other tissues); and IIRC most genes are expressed in the brain (not confident of this, maybe it’s more like 1⁄3 or 1⁄2. But anyway, there’s a lot of genes potentially relevant to brain function, so there’s a lot of surface area for mutational load to drag things down a bit.
genes are not simply choosing a level of intelligence.
I don’t know what you mean by this. Are you talking about pleiotropy? Between what and what? I mean of course genes do lots of things, but IIUC so far as we’ve observed, the correlations between most measured traits are pretty small (and usually positive between traits most people would judge desirable, e.g. lower risk of mental illness and higher intelligence).
Downvoting because it seems like you’ve barely read anything I wrote and also don’t know anything about genetics or intelligence, and are now posting AI slop, but I will upvote a thoughtful post making an argument using information and logic that address why people think it might work.
I’m basing this off of selection, not editing. I haven’t looked into the genetics stuff very much, because the bottleneck is biotech, not polygenic scores.
Would look forward to your rebuttal! I just hope you’ll respond to the strongest arguments, not the weakest. In particular, if you want to argue against the potential effectiveness of selection methods, I think you’d want to either argue that PGSes aren’t picking up causal variants at all (I mean, that there’s a large amount of correlation that isn’t causation); or that the causality would top out / have strongly diminishing returns on the trait. Selection methods would capture approximately all of the causal stuff that the PGS is picking up, even if it’s not even due to SNPs but rather rarer SNVs. (However, this would not apply to population stratification or something; then I’d think you’d want to argue that this is much / most of what PGSes are picking up, and there’d be already-made counterarguments to this that you should respond to in order to be convincing.)
I’ll repeat that I’m not very learned about genetics, so if you want to convince even me in particular, the best way is to respond to the strongest case, which I can’t present. But ok:
First I’ll say that an empirical set of facts I’d quite like to have for many traits (disease, mental disease, IQ, personality) would be validation of tails. E.g. if you look at the bottom 1% on a disease PRS, what’s the probability of disease? Similarly for IQ.
I rarely make claims about going much beyond natural results; generally I think it’s pretty plausible there’s some meaningful thing we could feasibly do that’s like +6 -- +8 SDs on intelligence, but I’m much less confident about the +8 SD claim, and not super confident of the +6 SD. Like, I think the default expectation ought to be that we can meaningfully get to +6 SDs; this seems like the straightforward conclusion. (I’m just restating the intuition / impression.)
Assuming linearity, the math is fairly straightforward. In the simplest model, with 10,000 fair +1/-1 coins (representing all the variance in a trait, so some coins are environmental), an SD is 50 coins and the average is 5,000. So there’s 100 SDs of variance available. Obviously this is mostly meaningless in terms of the trait, as linearity would not remotely hold, but my point is that the issue isn’t the math of additive selection. See here for more (e.g. about if the coins are biased https://tsvibt.blogspot.com/2022/08/the-power-of-selection.html#7-the-limits-of-selection ).
IQ seems to have thousands of small contributions from different regions. 10% of the variance is therefore in the ballpark of 10 trait SDs. Again, I’m not saying you can get to 250 IQ; what I’m saying is that the math of selection and variance isn’t the problem. Lee et al. state “In the WLS, the MTAG score predicts 9.7% of the variance in cognitive performance[...]”; this was in 2018, I would bet we can do substantially better now.
Lee, James J., Robbee Wedow, Aysu Okbay, Edward Kong, Omeed Maghzian, Meghan Zacher, Tuan Anh Nguyen-Viet, et al. ‘Gene Discovery and Polygenic Prediction from a 1.1-Million-Person GWAS of Educational Attainment’. Nature Genetics 50, no. 8 (August 2018): 1112–21. https://doi.org/10.1038/s41588-018-0147-3.
TBC I certainly don’t expect the effects to be literally linear even in the typical human range; it’s more that expect them to be fairly linear. Like if the answer is that a trait-mean couple that selects their child’s genome to have a (carefully, accurately as best anyone can) predicted IQ of mean 170 actually tends to have a child with mean IQ 155, I’d shrug and be like “huh, that’s weird and surprising, let’s investigate and make sure to communicate this fact to parents”; and I think this possibility is nontrivially strategically relevant; and it means we should accurately describe this plausible outcome, in order to not overhype etc.; but I wouldn’t be totally shocked. If the child tends to have a mean IQ of 125, I would be shocked, yeah. (The 170 vs. 155 thing would be hard to notice for a while because testing intelligence at that range is barely feasible, but just saying for illustration.)
Also, certainly I would expect strong nonlinearities at the extreme tails at some point. I’d certainly strongly advice against, maybe even condemn, pushing noticeably outside the regime of adaptedness.
Why do I expect the effects to be fairly linear in the human envelop, to the point where increasing a bunch of causal variants increases the trait?
Some of my impressions come from here: https://arxiv.org/pdf/1408.3421 “On the genetic architecture of intelligence and other quantitative traits”, Stephen D.H. Hsu, 2014
and from https://gwern.net/embryo-selection
(TBC, I’m not saying these sources present the strongest arguments; I’m just saying where my impressions historically come from.)
Breeding programs in non-humans work well, so there’s plenty of variance for those traits, no huge nonlinear walls that you hit, etc. This is far from dispositive; intelligence could plausibly be different from traits like egg production or weight, maybe it’s important that you’re checking along the way, etc.
Linear PGSes work well for many traits (height in humans; various traits for cows I think; even IQ, up to 10%).
SNP heritability estimates are substantial; the number in my head for IQ is >.3 of the variance. Though my impression is also that these are maybe controversial? I dunno.
It’s far from obvious to me that there’s been much selection for IQ in the past couple thousand years. But if there were a case that the selection has been strong, that would shift me.
There’s a theoretical argument, which IDK if it should hold much weight, but I like it: since DNA segments get shuffled around a lot, there’s selection pressure for things to work reasonably well with other things. E.g. DNA segments that have really bad effects when combined with some other segments would be selected against; and DNA segments that improve mechanisms for repairing/smoothing-out/compensating bad epistases between other segments will be selected for. In general this smooths things, which makes the landscape more linear. (I understand that specific epistases are much rarer than single variants, and therefore relatively invisible to selection; but I think my point stands somewhat, though this could be clarified and maybe basically disproven with good quantitative analysis.)
I just haven’t seen evidence of this nonlinear wall that’s right between 140 and 160, or whatever the claim is. It’s just people saying “maybe there’s a U-shaped curve of something” or “maybe there’s a high fan-in latent with a cutoff after the latent in the final IQ sum” which makes sense, but AFAIK is basically just speculation. It also isn’t super compelling speculation if we’re talking about a super polygenic trait in a super-complex organ where I’d expect there to be lots and lots of ways to tune and fix and just upregulate stuff. Like, my actual guess would be that there’s a whole spectrum of functional forms, from linear (substantial, according to h estimates!) through small ORs of ANDs, through large ORs of highly sensitive ANDs and other nonlinear forms; and these are all mixed together; and this does imply something; but it doesn’t imply that you can’t have quite large effects on the trait with germline engineering.
My impression is that the upper tail of IQ does get a little weird, and maybe g stops existing as much / the distribution of different tests stops being as one-dimensional? But IIUC (not sure, heard this from a psychometrician) there’s no observed threshold in the effect of IQ on other traits, despite people looking, though it’s quite hard to measure past 150ish. And e.g. this random paper claims to find quite substantial SNP heritability in a cohort with estimated rarity >4 SDs, though it’s not a huge cohort (1238 in the selected cohort) and I didn’t study it so maybe it’s very flawed / meaningless, IDK. https://www.nature.com/articles/mp2017121 In other words, to the small extent that we can look at the extreme tails, I at least haven’t heard of big results saying “aha! actually the genetics of IQ on the tails is quite different than near the mean!”.
(Also there’s sibling studies that IIUC say we are indeed picking up causality, though that’s not directly relevant to linearity.)
What makes you think this? As I said, it’s not clear to me that there’s been much selection pressure for intelligence in the past few thousand years.
Also, the “evolutionary optimum” can change. E.g. calories are not much of a problem in the developed world, but that’s recent.
Also, there’s always an influx of de novo mutations, and evolution has limited selection power. I’m not clear on the math here exactly, and I think kman has suggested that mutational load isn’t the main source of IQ-associated SNPs, but it demonstrates that it’s far from ironclad logic to infer from evolutionary pressure on a trait that the trait should be near optimum in linear variants. The brain is one of the organs with the most diverse gene expression profile (I mean, more genes are expressed in the brain than in most other tissues); and IIRC most genes are expressed in the brain (not confident of this, maybe it’s more like 1⁄3 or 1⁄2. But anyway, there’s a lot of genes potentially relevant to brain function, so there’s a lot of surface area for mutational load to drag things down a bit.
I don’t know what you mean by this. Are you talking about pleiotropy? Between what and what? I mean of course genes do lots of things, but IIUC so far as we’ve observed, the correlations between most measured traits are pretty small (and usually positive between traits most people would judge desirable, e.g. lower risk of mental illness and higher intelligence).
Downvoting because it seems like you’ve barely read anything I wrote and also don’t know anything about genetics or intelligence, and are now posting AI slop, but I will upvote a thoughtful post making an argument using information and logic that address why people think it might work.