5 ms·

>Such predictions, at first hit-or-miss, are becoming more accurate. One test described last year can guess a person’s height to within four centimeters, on the

>Such predictions, at first hit-or-miss, are becoming more accurate. One test described last year can guess a person’s height to within four centimeters, on the basis of 20,000 distinct DNA letters in a genome. I find that comparison highly misleading, height was always a reasonably 'simple' trait with very high heritability that is (in our society) relatively uninfluenced by environment. Here's a 2009 paper showing that you can predict height reasonably by averaging the parents' height (the 'Victorian method'). https://www.nature.com/articles/ejhg20095 https://www.nature.com/articles/ejhg20095 It's one of the few traits where you can do that well, so why would you use that as an example as to why genetics-based tests are getting better? All other phenotypes/traits are much more complex. >When they built a predictor for coronary heart disease, for instance, Kathiresan’s team discovered that the people it predicted to have the very highest risk, the top 2.5 percent, had four times the average chance of developing clogged arteries. So what's the base risk? If the base risk is 0.0005% then a four times higher chance is still tiny. I think it's this paper: https://www.nature.com/articles/ng.3914 https://www.nature.com/articles/ng.3914 As the scientists correctly pointed out, they only looked at Europeans, so who knows how this fares with Asians or Africans (likely: not well). None of the odds ratios are above 1.1, most are around 1.05, not a very extreme change to me! Why this matters: In countries like Australia, unlike Germany, there is no protection by the law from insurance companies who don't understand genetics. People can and have lost life insurances due to the results and questionable interpretations of genetic tests. I can only hope politicians don't read this article.

gwern 9y ago
There's nothing misleading about it, nor does height have a uniquely high heritability You can also predict childrens' adult IQ and many many other traits to similar degrees based on mid-parent regression, because they are all genetically influenced. The point of the height GWAS example is that because it is so very obviously, even to the naked eye of the layman, genetically influenced and objective (you can't argue height doesn't exist), it serves as a simple test. If GWASes can't predict height, they probably can't predict anything else. And this was exactly how height was used back in 2009 or so when the 'missing heritability' debate was still going: if all these traits are genetic and additive, why aren't the GWASes able to find anything for height (back then)? The answer turns out to simply be that there wasn't enough data and the linear model analyses were lousy. Now there is enough data and better algorithms are being applied, so as expected, height prediction works much much better. > All other phenotypes/traits are much more complex. No, not really. That sort of polygenicity is the norm. Look at, for example, Shi et al 2016 which attempts to model the distribution of effects & thus polygenicity of 30 human complex traits: https://www.biorxiv.org/content/early/2016/01/14/035907 https://www.biorxiv.org/content/early/2016/01/14/035907 . Height doesn't exactly pop out of any of the charts. > So what's the base risk? If the base risk is 0.0005% then a four times higher chance is still tiny. ...coronary artery disease should be quite obviously a higher base rate than 0.0005%. In any case, I believe the referenced study is Inouye et al 2018, in which case the extremized prediction is ~8% vs ~30% lifetime risk: https://www.gwern.net/images/genetics/2018-inouye-cad-riskprediction.png https://www.gwern.net/images/genetics/2018-inouye-cad-riskpr... > None of the odds ratios are above 1.1, most are around 1.05, not a very extreme change to me! It's the total that matters.
- gwern 9y ago
  (Or look at Zhang et al 2017 comparing complex architectures: http://www.biorxiv.org/content/early/2017/08/11/175406 http://www.biorxiv.org/content/early/2017/08/11/175406 )

gwern 9y ago

There's nothing misleading about it, nor does height have a uniquely high heritability You can also predict childrens' adult IQ and many many other traits to similar degrees based on mid-parent regression, because they are all genetically influenced. The point of the height GWAS example is that because it is so very obviously, even to the naked eye of the layman, genetically influenced and objective (you can't argue height doesn't exist), it serves as a simple test. If GWASes can't predict height, they probably can't predict anything else. And this was exactly how height was used back in 2009 or so when the 'missing heritability' debate was still going: if all these traits are genetic and additive, why aren't the GWASes able to find anything for height (back then)? The answer turns out to simply be that there wasn't enough data and the linear model analyses were lousy. Now there is enough data and better algorithms are being applied, so as expected, height prediction works much much better. > All other phenotypes/traits are much more complex. No, not really. That sort of polygenicity is the norm. Look at, for example, Shi et al 2016 which attempts to model the distribution of effects & thus polygenicity of 30 human complex traits: https://www.biorxiv.org/content/early/2016/01/14/035907 https://www.biorxiv.org/content/early/2016/01/14/035907 . Height doesn't exactly pop out of any of the charts. > So what's the base risk? If the base risk is 0.0005% then a four times higher chance is still tiny. ...coronary artery disease should be quite obviously a higher base rate than 0.0005%. In any case, I believe the referenced study is Inouye et al 2018, in which case the extremized prediction is ~8% vs ~30% lifetime risk: https://www.gwern.net/images/genetics/2018-inouye-cad-riskprediction.png https://www.gwern.net/images/genetics/2018-inouye-cad-riskpr... > None of the odds ratios are above 1.1, most are around 1.05, not a very extreme change to me! It's the total that matters.

gwern 9y ago
(Or look at Zhang et al 2017 comparing complex architectures: http://www.biorxiv.org/content/early/2017/08/11/175406 http://www.biorxiv.org/content/early/2017/08/11/175406 )