4 ms·

> Allow embryo selection for 6'+ sons, that's where the real money will be! Yes, but height is influenced by hundreds of genes, each with a minuscule influence

> Allow embryo selection for 6'+ sons, that's where the real money will be! Yes, but height is influenced by hundreds of genes, each with a minuscule influence, which will make it all but impossible to create a model that predicts height accurately. Additionally, since all the embryos you can select from are from the same two parents, models created from population samples are going to be a lot less useful. So I guess you won't be able to select for tall sons, at best you might be able to pick the larger of two embryos with slightly higher than random chance.

pjscott 7y ago
All but impossible? We got a model that predicts height startlingly well, usually to within a few centimeters, back in 2017: https://infoproc.blogspot.com/2017/09/accurate-genomic-prediction-of-human.html https://infoproc.blogspot.com/2017/09/accurate-genomic-predi...
gwern 7y ago
> Yes, but height is influenced by hundreds of genes, each with a minuscule influence, which will make it all but impossible to create a model that predicts height accurately. Not at all. The current PGSes for height already predict ~40% variance (or r=0.63): see https://www.biorxiv.org/content/early/2017/09/19/190124 https://www.biorxiv.org/content/early/2017/09/19/190124 or https://www.biorxiv.org/content/early/2018/07/24/375337 https://www.biorxiv.org/content/early/2018/07/24/375337 > Additionally, since all the embryos you can select from are from the same two parents, models created from population samples are going to be a lot less useful. No? It's all genes. The genes within a family are the same as across the population. The PGS is reduced by half* simply because you are comparing half-related embryos (siblings), so they will share many of the same genes and there are no differences to select on, but that's the nature of it. You could do selection on unrelated embryos if you really wanted to, but few people seem to want to... * between-sibling PGS reduction is sometimes more depending on other factors like 'nature of nurture'; this is something that affects EDU more than IQ, but other traits like BMI or height appear to not be affected > So I guess you won't be able to select for tall sons, at best you might be able to pick the larger of two embryos with slightly higher than random chance. It'd actually be 65% with the current height PGSes (upperbounded at perhaps 71% as WGS can recover the full heritability: https://www.biorxiv.org/content/10.1101/588020v1 https://www.biorxiv.org/content/10.1101/588020v1 ).
- DuskStar 7y ago
  > > So I guess you won't be able to select for tall sons, at best you might be able to pick the larger of two embryos with slightly higher than random chance. > It'd actually be 65% with the current height PGSes (upperbounded at perhaps 71% as WGS can recover the full heritability: https://www.biorxiv.org/content/10.1101/588020v1 https://www.biorxiv.org/content/10.1101/588020v1 ). Wouldn't knowing 65% of the variance let you pick the best of two more than 75% of the time? (my intuition would be 50 + 65/2 %, but I'm probably wrong there)
  - gwern 7y ago
    No. I don't know of any simple formula relating variance with the order statistics of selecting the max, such as in the worst-case scenario of picking from only 2. (In any case, the variance I used is 40%, then halved for relatedness.) I do have an open StatsExchange question on this: https://stats.stackexchange.com/questions/427292/probability-of-selecting-maximum-in-bivariate-correlated-order-statistics https://stats.stackexchange.com/questions/427292/probability... But to calculate the probability, I just Monte Carlo it.
    - DuskStar 7y ago
      Thanks for the correction, and the simulation code!

pjscott 7y ago

All but impossible? We got a model that predicts height startlingly well, usually to within a few centimeters, back in 2017: https://infoproc.blogspot.com/2017/09/accurate-genomic-prediction-of-human.html https://infoproc.blogspot.com/2017/09/accurate-genomic-predi...

gwern 7y ago

> Yes, but height is influenced by hundreds of genes, each with a minuscule influence, which will make it all but impossible to create a model that predicts height accurately. Not at all. The current PGSes for height already predict ~40% variance (or r=0.63): see https://www.biorxiv.org/content/early/2017/09/19/190124 https://www.biorxiv.org/content/early/2017/09/19/190124 or https://www.biorxiv.org/content/early/2018/07/24/375337 https://www.biorxiv.org/content/early/2018/07/24/375337 > Additionally, since all the embryos you can select from are from the same two parents, models created from population samples are going to be a lot less useful. No? It's all genes. The genes within a family are the same as across the population. The PGS is reduced by half* simply because you are comparing half-related embryos (siblings), so they will share many of the same genes and there are no differences to select on, but that's the nature of it. You could do selection on unrelated embryos if you really wanted to, but few people seem to want to... * between-sibling PGS reduction is sometimes more depending on other factors like 'nature of nurture'; this is something that affects EDU more than IQ, but other traits like BMI or height appear to not be affected > So I guess you won't be able to select for tall sons, at best you might be able to pick the larger of two embryos with slightly higher than random chance. It'd actually be 65% with the current height PGSes (upperbounded at perhaps 71% as WGS can recover the full heritability: https://www.biorxiv.org/content/10.1101/588020v1 https://www.biorxiv.org/content/10.1101/588020v1 ).

DuskStar 7y ago
> > So I guess you won't be able to select for tall sons, at best you might be able to pick the larger of two embryos with slightly higher than random chance. > It'd actually be 65% with the current height PGSes (upperbounded at perhaps 71% as WGS can recover the full heritability: https://www.biorxiv.org/content/10.1101/588020v1 https://www.biorxiv.org/content/10.1101/588020v1 ). Wouldn't knowing 65% of the variance let you pick the best of two more than 75% of the time? (my intuition would be 50 + 65/2 %, but I'm probably wrong there)
- gwern 7y ago
  No. I don't know of any simple formula relating variance with the order statistics of selecting the max, such as in the worst-case scenario of picking from only 2. (In any case, the variance I used is 40%, then halved for relatedness.) I do have an open StatsExchange question on this: https://stats.stackexchange.com/questions/427292/probability-of-selecting-maximum-in-bivariate-correlated-order-statistics https://stats.stackexchange.com/questions/427292/probability... But to calculate the probability, I just Monte Carlo it.
  - DuskStar 7y ago
    Thanks for the correction, and the simulation code!