[gwern]

'Genetic engineering' routinely moves traits by 5 standard deviations. Countless breeding programs have accomplished that or much more, and even under the narrowest definition of 'genetic engineering', something like editing in chestnut blight resistance will increase the trait by many SDs (going from ~0% chestnut trees surviving to a large fraction of them surviving). That is, after all, how evolution works.

[bayesian_horse]

In animals or plants, for simple traits, yes.

Complicated, polygenetic traits with virtually unlimited interactions between genes and environmental factors? Much more difficult.

Add in the serious consequences of mistakes. The limited supply of subjects. The difficulty of grading the success of any modification.

For intelligence, we don't even know what kinds of changes we would want to make, even if we could make them reliably.

[gwern]

> In animals or plants, for simple traits, yes.

No, complex traits too. Come on. Dog intelligence and personalities? The domesticated foxes? Any of the many mouse/rat selective breeding experiments like Tryon? Personality, intelligence, these sorts of traits are all highly heritable and about the only thing everyone can agree on that heritability means is that you can select on it. 'virtually unlimited interactions between genes and environmental factors'? Give me a break!

> Add in the serious consequences of mistakes. The limited supply of subjects. The difficulty of grading the success of any modification.

??? None of that is a real problem. You are wildly gesturing at hypotheticals and making stuff up. Being 'polygenetic' is not a problem, it's a blessing.

> For intelligence, we don't even know what kinds of changes we would want to make, even if we could make them reliably.

Yes, we do. The PGSes have identified hundreds of variants at genome-wide significance, and there are thousands at high posterior probability, with nontrivial cumulative predictive power in the general population.

[bayesian_horse]

You don't understand the science behind animal breeding, or quantitative genetics and I assume that is why you think you can do the same to Humans.

The number of subjects absolutely does matter. Selecting mice for intelligence is all nice and well, but you have to breed thousands, and "discard" most of them. Both in traditional breeding as well as with genetic engineering, you have to have sufficient number of trials to conclude anything.

With Human subjects, given ethical and practical constraints (lifespan), that's going to be a lot tougher.

You can reliably quantify the intelligence of a mouse at a few weeks of age and breed it soon thereafter (though you would probably use that "IQ" to score its parents). You can't reliably quantify health and intelligence of a Human within years, especially if you are aiming for "multiple standard deviations" above average. At the very least the practical generational span is 20 years, and usually much longer if the subjects have any say in the matter.

Genome wide association studies... another thing people get terribly confused about. The mutations they screen for are virtually never causal, but merely markers associated with actually causal sequences. There may be occasional better-researched mutations here and there. But overall the individual contributions of any such marker are already extremely low and don't sum up to a significant portion of the suspected heritability. And heritability of intelligence in Humans is another problem where people don't seem to understand what the research is actually saying. Whenever people say things like "IQ in Humans has a heritability of 40-80%", they are at best summarizing in a misleading fashion if not outright lying.

To my knowledge there are virtually no candidate genes/mutations proposed for genetic improvement of intelligence in Humans. Especially not with the confidence you'd need to even attempt this. Even without ethical considerations.

[gwern]

> The number of subjects absolutely does matter. Selecting mice for intelligence is all nice and well, but you have to breed thousands, and "discard" most of them. Both in traditional breeding as well as with genetic engineering, you have to have sufficient number of trials to conclude anything.

The number of 'subjects' is elastic. If applied to IVF, that's everyone who uses IVF. In embryo selection, you are selecting out of available embryos, eg 1 out of 5. With other techniques like ovarian biopsy or gametogenesis, you are potentially selecting 1 out of 100+. Hence, plenty of subjects.

> You can reliably quantify the intelligence of a mouse at a few weeks of age and breed it soon thereafter (though you would probably use that "IQ" to score its parents). You can't reliably quantify health and intelligence of a Human within years, especially if you are aiming for "multiple standard deviations" above average. At the very least the practical generational span is 20 years, and usually much longer if the subjects have any say in the matter.

All irrelevant, since no one is suggesting phenotyping. That's the point of the polygenic scores.

> The mutations they screen for are virtually never causal, but merely markers associated with actually causal sequences.

Also irrelevant. When you are doing prediction for selection, all you need to do is predict the one with the highest scores. There is zero need to identify even a single causal variant. GWASes are so predictive because the SNPs are in LD with the causal variants, and the LD does not change abruptly in a single selection step.

However, since you think that none have been identified, you should go read the appendix to Lee et al 2018, among other IQ GWASes, where they do estimates, and turn up scores of SNPs which have causal probabilities >90%.