4 ms·

Though thousands of alleles contribute to intelligence, we would only need to edit a few hundred to get extremely large increases in IQ, especially if you perfo

by wycs 8y ago

Though thousands of alleles contribute to intelligence, we would only need to edit a few hundred to get extremely large increases in IQ, especially if you perform a round of embryo selection (or even iterated embryo selection) before editing. Editing hundreds of alleles may sound incredible to you now, but I don’t think it will sound so incredible for long. Even ignoring GM, embryo selection can get us pretty far with current technology: https://www.gwern.net/Embryo-selection https://www.gwern.net/Embryo-selection

krastanov 8y ago
Just checking: you do know that there is a difference between change of "5 standard deviations" and change of "5 IQ points". 5 standard deviations is equal to "1 in 3.5 million", which makes your claim fairly farfetched with the tech of today or of the future decade or two, especially given how IQ is much more function of nurture than of nature.
- wycs 8y ago
  I am aware of this. Genetic engineering implies far more power than Gwern mentions for one round of embryo selection. 5 standard deviations may sound a lot for you, but in animal breeding we have changed the distribution of various traits by far more than 5 standard deviations. Into the dozens.
skosuri 8y ago
What you are saying is just not supported by evidence. For example, polygenic analysis of the latest meta-analysis from >250K individuals seems to say that the top several hundred variants explain in aggregrate only 5.2% of the variation in intelligence. Source: https://www.nature.com/articles/s41588-018-0152-6 https://www.nature.com/articles/s41588-018-0152-6 (Supplement section 2.3.3 if you don't have access).
- carbocation 8y ago
  In addition, even with that much or more explanatory power, we generally do not think we have necessarily identified causal variants in most cases. We identify groups of associated variants that signify an interesting locus. There isn’t currently a generic framework that lets us decide which ones are causal.
  - sanxiyn 8y ago
    You need causal variants to do editing, but tagging variants are sufficient for selecting.
- wycs 8y ago
  There is no human with all the positive variants; it is the the distribution of these variants that explains 5.2%. GM would allow you to create such a human. That paper does not imply what you think it implies: that editing all these variants in an individual embryo would have a small effect. An individual with the top 200 IQ-enhancing variants edited in (and an otherwise normal genome) would be far, far smarter than than than you seem to think, even though the variants edited only explain 5.2% of the differences within the population.
  - skosuri 8y ago
    Ahh I see, you are just making stuff up.
    - wycs 8y ago
      No. I think the claim "one can raise IQ several standard deviations by editing <200 alleles" is likely true:https://arxiv.org/pdf/1408.3421.pdf https://arxiv.org/pdf/1408.3421.pdf
      - skosuri 8y ago
        Yeah, know his work, and I don't think he'd agree with you: http://www.genetics.org/content/210/2/477 http://www.genetics.org/content/210/2/477
        kanzure 8y ago
        Your method of modifying human height is naive because it assumes that your genetic modification technique to increase height would need to modify or even know about height alleles. edit: We last talked October 2015, hi.
        skosuri 8y ago
        Hello :-). My method???
  - 8y ago
- gwern 8y ago
  You are misunderstanding what 'variance explained' means. It means 'variance', as in, average differences between individuals; nothing to do with the average effect size or possible maxima or anything like that. It refers only to predicting the net effect after, being distributed at population frequencies, most of those hundred variants have canceled each other out. In editing, the population frequencies & distribution don't matter, only the summed effect size of all the variants flipped. Confusing one for the other is like confusing waves on a beach with a tsunami: https://www.gwern.net/Embryo-selection#limits-to-iterated-selection-the-paradox-of-polygenicity https://www.gwern.net/Embryo-selection#limits-to-iterated-se... Since they have an average effect size of ~0.05-0.1 IQ points or so, a few hundred of those would in fact be many SDs. You can download the PGS, grab the top few hundred or thousand betas, and sum them if you don't believe me. (I show this with Benyamin et al 2014 in that section above, but no reason one couldn't redo it with a better PGS; the results won't change much.) And yes, Hsu really does believe this, he is quite explicit in that 2014 paper that that is how it would work (in addition to countless places on his blog & Twitter) and I've discussed it with him in the past. (As other comments point out, the real problem with this logic is that the hits are usually not causal but tag the causal variant, even if you had some magic CRISPR which could make hundreds of arbitrary edits. That's soluble, but not easily. It is, however, not a problem for selection procedures.)
blueblimp 8y ago
I'm bullish on selection approaches generally, but I think editing is not as promising at the moment. Editing via double-strand breaks is just way too unreliable for use in embryos, and base editing is too limited to make use of GWAS info. This could change if the tech improves, but it would need to improve A LOT.

by wycs 8y ago

krastanov 8y ago

Just checking: you do know that there is a difference between change of "5 standard deviations" and change of "5 IQ points". 5 standard deviations is equal to "1 in 3.5 million", which makes your claim fairly farfetched with the tech of today or of the future decade or two, especially given how IQ is much more function of nurture than of nature.

wycs 8y ago
I am aware of this. Genetic engineering implies far more power than Gwern mentions for one round of embryo selection. 5 standard deviations may sound a lot for you, but in animal breeding we have changed the distribution of various traits by far more than 5 standard deviations. Into the dozens.

skosuri 8y ago

What you are saying is just not supported by evidence. For example, polygenic analysis of the latest meta-analysis from >250K individuals seems to say that the top several hundred variants explain in aggregrate only 5.2% of the variation in intelligence. Source: https://www.nature.com/articles/s41588-018-0152-6 https://www.nature.com/articles/s41588-018-0152-6 (Supplement section 2.3.3 if you don't have access).

carbocation 8y ago
In addition, even with that much or more explanatory power, we generally do not think we have necessarily identified causal variants in most cases. We identify groups of associated variants that signify an interesting locus. There isn’t currently a generic framework that lets us decide which ones are causal.
- sanxiyn 8y ago
  You need causal variants to do editing, but tagging variants are sufficient for selecting.
wycs 8y ago
There is no human with all the positive variants; it is the the distribution of these variants that explains 5.2%. GM would allow you to create such a human. That paper does not imply what you think it implies: that editing all these variants in an individual embryo would have a small effect. An individual with the top 200 IQ-enhancing variants edited in (and an otherwise normal genome) would be far, far smarter than than than you seem to think, even though the variants edited only explain 5.2% of the differences within the population.
- skosuri 8y ago
  Ahh I see, you are just making stuff up.
  - wycs 8y ago
    No. I think the claim "one can raise IQ several standard deviations by editing <200 alleles" is likely true:https://arxiv.org/pdf/1408.3421.pdf https://arxiv.org/pdf/1408.3421.pdf
    - skosuri 8y ago
      Yeah, know his work, and I don't think he'd agree with you: http://www.genetics.org/content/210/2/477 http://www.genetics.org/content/210/2/477
      - kanzure 8y ago
        Your method of modifying human height is naive because it assumes that your genetic modification technique to increase height would need to modify or even know about height alleles. edit: We last talked October 2015, hi.
        skosuri 8y ago
        Hello :-). My method???
- 8y ago
gwern 8y ago
You are misunderstanding what 'variance explained' means. It means 'variance', as in, average differences between individuals; nothing to do with the average effect size or possible maxima or anything like that. It refers only to predicting the net effect after, being distributed at population frequencies, most of those hundred variants have canceled each other out. In editing, the population frequencies & distribution don't matter, only the summed effect size of all the variants flipped. Confusing one for the other is like confusing waves on a beach with a tsunami: https://www.gwern.net/Embryo-selection#limits-to-iterated-selection-the-paradox-of-polygenicity https://www.gwern.net/Embryo-selection#limits-to-iterated-se... Since they have an average effect size of ~0.05-0.1 IQ points or so, a few hundred of those would in fact be many SDs. You can download the PGS, grab the top few hundred or thousand betas, and sum them if you don't believe me. (I show this with Benyamin et al 2014 in that section above, but no reason one couldn't redo it with a better PGS; the results won't change much.) And yes, Hsu really does believe this, he is quite explicit in that 2014 paper that that is how it would work (in addition to countless places on his blog & Twitter) and I've discussed it with him in the past. (As other comments point out, the real problem with this logic is that the hits are usually not causal but tag the causal variant, even if you had some magic CRISPR which could make hundreds of arbitrary edits. That's soluble, but not easily. It is, however, not a problem for selection procedures.)

blueblimp 8y ago

I'm bullish on selection approaches generally, but I think editing is not as promising at the moment. Editing via double-strand breaks is just way too unreliable for use in embryos, and base editing is too limited to make use of GWAS info. This could change if the tech improves, but it would need to improve A LOT.