6 Comments
User's avatar
Realist's avatar

Emil, thanks so very much for the analysis of genomic predictions. The complexity is disconcerting but not discouraging. While the challenges are daunting, I believe that genomic enhancement of positive traits will happen. AI will be an excellent tool in this endeavor.

Peter Frost's avatar

It depends on what you're trying to find out. If you simply want to know the direction and intensity of natural selection on a trait (like height or cognitive ability), you can make do with a lot fewer genomes, since the adaptive value of the trait should be the same for all individuals within the same natural and cultural environment.

Of course, an environment may contain micro-environments with their own specific selection pressures. But that's the general rule.

Emil O. W. Kirkegaard's avatar

This article is concerned with building predictive models between individuals, not populations, or merely establishing selection coefficients.

Stig's avatar

What if you only want to rank order individuals, for example in an embryo selection scenario. How would that affect the needed sample size?

Emil O. W. Kirkegaard's avatar

The effectiveness of embryo selection depends on the predictive validity of polygenic scores within families. Since for some traits, the between family genetic models have less validity within families, this requires either some family design (siblings or parent-child data), or some previously unknown methodological advance. Since these designs have less variation to work with, they probably require larger samples yet again.

User's avatar
Comment deleted
Mar 25, 2025
Comment deleted
Emil O. W. Kirkegaard's avatar

Those kinds of simulations have already been done. See e.g.

https://www.cell.com/cell/fulltext/S0092-8674(19)31210-3?dgcid=raven_jbs_etoc_email

Typically, the within family decline is ~20% for IQ predictors suitably optimized. Mind you that r² metrics will not work well here, you need the slope.