No. I'm not sure Lello et al have released the polygenic score in question.
> Would be nice if you could predict the height of your kids at a young age.
You can already do that with Galton's midpoint regression method (indeed, he invented the entire method of regression based on predicting children's heights from their parents!). What you could do is predict which sibling will be taller, though.
PGD has been in use to a very limited extent, testing on just 1 or a bare handful of specific genetic diseases, along with basic checks for gross abnormalities like microscope examination. Use of full SNP arrays is still very new and no one has ever used them to select based on complex traits like height.
For tall, muscular people with certain superficial features? Probably. For intelligence and deeper features, all while controlling for undesirable adverse effects (like early cancers, epilepsy, and so forth) hell no. What we don’t know about the brain alone is staggering, nevermind how tweaking it with ensembles of genes will impact the whole human.
So relax, we have plenty of dystopia without designer babies.
No code in paper, correct. Very easy to implement via R or python. One of the authors (Gustavo de los Campos) has alot of code that could probably be coped/pasted with little editing from his website.
Algorithm is actually painfully simple. Just relying on massive data (500K individuals)
Coordinate descent is simple but the resource requirements may complicate your implementation running over n=500k with tens of thousands of markers each; as it was described to me, they were using a workstation with like 128GB RAM... (Plus of course you need access to the UKBB dataset in the first place.)
That's to replicate the study; presumably if the authors released their trained model it would be much less computationally expensive to run it on your own data?
If by trained model you mean PGS, yes, it would be trivial. While the fitting of the optimal weights may be computationally intensive, it's still just a simple little linear model: look up the SNP value, add or subtract a small coefficient for each of the 30k or whatever 'activated' SNPs, and get the net total.
My sister turned out 6'3, my parents would have given a lot to predict that height some years before (as they feared she might end up even taller). Same is probably true for very small people that give their kids growth hormones. And If you are looking for a sport your child is good at it might also help. Sure you can construct very adversial uses of such a technology here but that does not mean that the tech itself is bad.
https://dna.land/ (New York Genome Center/Columbia University) has a height-model based upon genome data which you can just plug your 23andme data into.
(You can also give them your actual height so they can improve their data set)
I didn't read the article, but isn't this like predicting country of birth? I.e. how well is the prediction if you take someone from the US and let them grow up in Africa?
Like many traits, height is a mix of genetic and environmental factors. Based on things like twin studies, we know that 60-80% of the variation in height is due to genetic factors [1]. This paper shows that you can predict height from genetic information with an accuracy that is in line with that heritability estimate. It doesn't mean that it will be 100% accurate all the time, since environmental factors (e.g. nutrition) will account the remaining 20-40%.
[1] https://www.scientificamerican.com/article/how-much-of-human... "The short answer to this question is that about 60 to 80 percent of the difference in height between individuals is determined by genetic factors, whereas 20 to 40 percent can be attributed to environmental effects, mainly nutrition."
Good post/paper in general. Predicting traits in humans has been more challenging than most geneticist want to admit. The good news is that most high heritable traits will become "solved" as we enter a new order of magnitude of data availability (i.e 5 million people instead of 500k).
Much work left to do though, and there are interesting methods being developed in this space
Genes are a strong predictor of height (in certain populations), AND it has gone up tremendously in places like Japan. Both are true.
To speak of the influence of genes also assumes a certain environment. If we look at western populations for example, the heritability of height in adulthood is something like 0.8. But this is predicated on that western environment where ~nobody has nutritional deficiencies. A uniform, high-quality environment means that there is little besides genes that can influence height. If something happened and a large portion of the population had such deficiencies, the influence of genes would be diminished.
In his book Behave[1], biologist Robert Sapolosky explains the issue as follows:
> Heritability scores are relevant only to the environments in which the traits have been studied. The more environments you study a trait in, the lower the heritability is likely to be.
He goes on for several pages, describing the ways the surrounding environment can change an organism's gene expression, and I think this quote best summarizes his point:
> Here’s a rule of thumb for recognizing gene/environment interactions, translated into English: You are studying the behavioral effects of a gene in two environments. Someone asks, “What are the effects of the gene on some behavior?” You answer, “It depends on the environment.” Then they ask, “What are the effects of environment on this behavior?” And you answer, “It depends on the version of the gene.” “It depends” = a gene/environment interaction.
No, this only says that the about 20000 SNPs (mutations) that they can accurately measure have a limited effect on variance. What the millions of other SNPs do is not known. For example there are estimates for many many thousands of SNPs that influece your height by 0.1 mm. The single variant effect is tiny but altogether it adds up a lot.
Oh, and lets not forget that again this is only point mutations (SNPs). But there are also genomic insertions, deletions, larger structural variants and poly variant interactions, all of which we do not really have much idea about or even the tech available to accurately measure.
This looks pretty good: digging into the supplement it appears they controlled for age, which is how the controlled for things like improved nutrition, elimination of smoking while pregnant etc
The models are predicted off of "~20k activated SNPs" - 20,000 separate mutations, most of which have really small effects on height. "Designer babies" are pretty far off (IVF is awesome and great, but not trivial, so it would be risky to do this just to make sure a baby will be...probably somewhere near average height). Engineering the right mutations with CRISPR/Cas9-based gene editing would be astronomically expensive (and is realistically impossible); filtering embryos for the right genetic background would also be incredibly expensive and probably impossible, as you'd be relying on the natural, stochastic arrangement of those 20,000 mutations to be in sync with the vision you have for your baby.
Having small mutations is irrelevant; it could be dozens of large effects or thousands of small effects, it doesn't matter, you get a normal distribution anyway. The question is how much variance there is and how much of it you can predict.
Or let me put it this way: how much do siblings (ie embryos) vary in height? It's by quite a bit, often several inches. Most of this is due to genetics. And this PGS is able to predict half of the genetic contribution. So...
(If you're curious, my best estimate is that embryo selection could boost height by about an inch on average.)
> filtering embryos for the right genetic background would also be incredibly expensive and probably impossible
It would cost about $2000 for the biopsies and SNP arrays of the usual ~5 embryos, and can be done either now or in the next few years and could have been done years ago if any real effort was put into it.
You would only be able to select from those 5 backgrounds, however, and the predicted heights for those embryos will be drawn from a normal distribution; you can only "design" from what you've sampled.
On the other hand, if you sampled millions of embryos, you could find the rare few predicted to grow to be 6'5" person -- but this would be very expensive and basically impossible without synthetic embryos.
> You would only be able to select from those 5 backgrounds, however, and the predicted heights for those embryos will be drawn from a normal distribution; you can only "design" from what you've sampled.
Yes, and this is cheap and does lead to gains. (Specifically: around 1 inch.)
> but this would be very expensive and basically impossible without synthetic embryos.
First, this is not remotely what you said. Second, it's not true: you only need more eggs and they don't have to be 'synthetic' whatever that means, and there's a lot of work on inducing egg development from stem cells so in another 10 years it may well be possible for parents to do massive selection like that. Third, you don't need millions of embryos if you just want a very tall person, as the advantage is cumulative over generations (which is the critical insight behind Iterated Embryo Selection: it's much more efficient to take a few hundred embryos through multiple generations of selection than it is to try to brute force a single selective step). That's omitting any gains from CRISPR or genome synthesis or other methods not yet thought of. Fourth, why would any parent want that in the first place as that's into the realm of potential healthy problems (even if we're assuming only male embryos) and beyond the useful level of height advantages, especially when they could be instead spending that count of embryos to maximize a weighted sum of all health and other complex traits? (Remember, just because everyone talks about doing embryo selection on a single trait at a time doesn't make that remotely a good idea; there are big gains to selecting on many traits simultaneously.)
Even embryo selection and genomic prediction alone gets you pretty far if you’re willing to discard a lot of embryos (is it possible to read the DNA out of sperm and egg cells non-destructively? I don’t know much about the technology involved). Just fertilize a bunch of embryos and discard all of them except the top few (selected on e.g. some weighted combination of height, intelligence, and whatever else you can predict). Then your kids will be overwhelming likely to be taller and smarter (and whatever else you selected for) than if you’d just picked embryos at random.
> is it possible to read the DNA out of sperm and egg cells non-destructively?
No. Who knows what the future will bring, though? There's already some neat new CRISPR-based methods for detecting DNA and viruses etc which were published the other day.
47 comments
[ 2.8 ms ] story [ 101 ms ] threadWould be nice if you could predict the height of your kids at a young age.
Yes.
> and find the SNPs and calculate on your own?
No. I'm not sure Lello et al have released the polygenic score in question.
> Would be nice if you could predict the height of your kids at a young age.
You can already do that with Galton's midpoint regression method (indeed, he invented the entire method of regression based on predicting children's heights from their parents!). What you could do is predict which sibling will be taller, though.
[1] https://en.wikipedia.org/wiki/Gattaca
It's not direct genetic engineering yet, but we're pretty much already in Gattaca.
So relax, we have plenty of dystopia without designer babies.
Algorithm is actually painfully simple. Just relying on massive data (500K individuals)
Would it?
(You can also give them your actual height so they can improve their data set)
You upload your dna from 23andme to them and they give you a full report without government interference.
https://promethease.com/
[1] https://www.scientificamerican.com/article/how-much-of-human... "The short answer to this question is that about 60 to 80 percent of the difference in height between individuals is determined by genetic factors, whereas 20 to 40 percent can be attributed to environmental effects, mainly nutrition."
[1] https://news.ycombinator.com/item?id=16386749
Much work left to do though, and there are interesting methods being developed in this space
Would you mind posting any relevant links? Is there general agreement in the field that more data is the solution?
This seems very low, almost like saying genes don't have any practical predictive power, right ?
We've seen average height go up tremendously in places like Japan based on improved nutrition.
To speak of the influence of genes also assumes a certain environment. If we look at western populations for example, the heritability of height in adulthood is something like 0.8. But this is predicated on that western environment where ~nobody has nutritional deficiencies. A uniform, high-quality environment means that there is little besides genes that can influence height. If something happened and a large portion of the population had such deficiencies, the influence of genes would be diminished.
> Heritability scores are relevant only to the environments in which the traits have been studied. The more environments you study a trait in, the lower the heritability is likely to be.
He goes on for several pages, describing the ways the surrounding environment can change an organism's gene expression, and I think this quote best summarizes his point:
> Here’s a rule of thumb for recognizing gene/environment interactions, translated into English: You are studying the behavioral effects of a gene in two environments. Someone asks, “What are the effects of the gene on some behavior?” You answer, “It depends on the environment.” Then they ask, “What are the effects of environment on this behavior?” And you answer, “It depends on the version of the gene.” “It depends” = a gene/environment interaction.
[1]: https://www.amazon.com/Behave-Biology-Humans-Best-Worst/dp/1...
Oh, and lets not forget that again this is only point mutations (SNPs). But there are also genomic insertions, deletions, larger structural variants and poly variant interactions, all of which we do not really have much idea about or even the tech available to accurately measure.
Or let me put it this way: how much do siblings (ie embryos) vary in height? It's by quite a bit, often several inches. Most of this is due to genetics. And this PGS is able to predict half of the genetic contribution. So...
(If you're curious, my best estimate is that embryo selection could boost height by about an inch on average.)
> filtering embryos for the right genetic background would also be incredibly expensive and probably impossible
It would cost about $2000 for the biopsies and SNP arrays of the usual ~5 embryos, and can be done either now or in the next few years and could have been done years ago if any real effort was put into it.
On the other hand, if you sampled millions of embryos, you could find the rare few predicted to grow to be 6'5" person -- but this would be very expensive and basically impossible without synthetic embryos.
Yes, and this is cheap and does lead to gains. (Specifically: around 1 inch.)
> but this would be very expensive and basically impossible without synthetic embryos.
First, this is not remotely what you said. Second, it's not true: you only need more eggs and they don't have to be 'synthetic' whatever that means, and there's a lot of work on inducing egg development from stem cells so in another 10 years it may well be possible for parents to do massive selection like that. Third, you don't need millions of embryos if you just want a very tall person, as the advantage is cumulative over generations (which is the critical insight behind Iterated Embryo Selection: it's much more efficient to take a few hundred embryos through multiple generations of selection than it is to try to brute force a single selective step). That's omitting any gains from CRISPR or genome synthesis or other methods not yet thought of. Fourth, why would any parent want that in the first place as that's into the realm of potential healthy problems (even if we're assuming only male embryos) and beyond the useful level of height advantages, especially when they could be instead spending that count of embryos to maximize a weighted sum of all health and other complex traits? (Remember, just because everyone talks about doing embryo selection on a single trait at a time doesn't make that remotely a good idea; there are big gains to selecting on many traits simultaneously.)
No. Who knows what the future will bring, though? There's already some neat new CRISPR-based methods for detecting DNA and viruses etc which were published the other day.