Latest / Razib Khan's Unsupervised Learning / Alex Young: IQ, disease and statistical genomics
Transcript
- 0:00This podcast is brought to you by the Albany Public Library
- 0:03Main Branch and the
- 0:04SPEAKER_01: generosity of listeners like you.
- 0:09SPEAKER_00: What is a podcast?
- 0:16SPEAKER_00: God, Daddy, these people talk as much as you do.
- 0:26SPEAKER_01: Razeeb Khan's Unsupervised Learning.
- 0:29SPEAKER_01: Hey, everybody.
- 0:29SPEAKER_01: This is Razeeb with the Unsupervised Learning Podcast.
- 0:31SPEAKER_01: And I'm here today with my friend Alexander Young.
- 0:34SPEAKER_01: And Alex is someone I've had on the podcast before,
- 0:38SPEAKER_01: actually two and a half years ago,
- 0:39SPEAKER_01: I think with James Lee.
- 0:41SPEAKER_01: You guys can check that podcast on the links.
- 0:43I also actually recorded a podcast with him on the Insight
- 0:46many,
- 0:47many years ago.
- 0:47SPEAKER_01: But unfortunately, the Insight archives are gone.
- 0:50SPEAKER_00: I think there's another one on Unsupervised Learning as well.
- 0:54SPEAKER_01: Oh, okay.
- 0:55SPEAKER_00: The cognitive genomics one.
- 0:56SPEAKER_01: Yes.
- 0:57SPEAKER_01: So that was a, yeah, yeah, yeah.
- 0:58SPEAKER_01: So yeah, there's many times.
- 1:02SPEAKER_01: So yeah,
- 1:03we're here though to talk about cognitive genomics and
- 1:06everything that's going on
- 1:07SPEAKER_01: with that.
- 1:08SPEAKER_01: We have a new paper.
- 1:08SPEAKER_01: You guys have a new paper that's out and we will talk about that.
- 1:12SPEAKER_01: But first,
- 1:13tell us where we are in this field in terms of what do we
- 1:18understand about
- 1:19SPEAKER_01: heritability?
- 1:20SPEAKER_01: What has GWAS told us?
- 1:22SPEAKER_01: Where are we with family-based studies, et cetera, et cetera?
- 1:26SPEAKER_01: And then,
- 1:27SPEAKER_01: yeah,
- 1:27SPEAKER_01: let's talk about the science first,
- 1:28SPEAKER_01: and then we can talk a little bit about the political controversies,
- 1:31SPEAKER_01: and then we'll go into the parasite-related paper.
- 1:35And actually,
- 1:35I should just have you speak to your affiliations real quickly.
- 1:39SPEAKER_01: I can tell them, but why don't you tell us?
- 1:42SPEAKER_00: Sure, yeah.
- 1:43So I'm currently an assistant professor in the human genetics
- 1:46department at UCLA.
- 1:49Before that,
- 1:50I did my PhD at Oxford,
- 1:53also a postdoc joint between Oxford and DECODE.
- 1:57SPEAKER_00: And I got into family-based studies,
- 2:01SPEAKER_00: mostly because of my time at DECODE,
- 2:03SPEAKER_00: where there was a lot of family data,
- 2:05SPEAKER_00: which was very different from what was available in
- 2:09generally in the field and that's actually proved to be a very
- 2:12important pathway
- 2:14for the future of the field both in terms of getting a better
- 2:18understanding of
- 2:18heritability what GWAS is telling us what polygenic predictions
- 2:23genetic predictions
- 2:24of traits are telling us and also in embryo selection IVF which
- 2:31I've been working
- 2:32SPEAKER_00: in with a company Herocyte
- 2:35SPEAKER_00: So I don't know if you want to,
- 2:36SPEAKER_00: should I go into some detail about GWAS or what should I talk about?
- 2:42SPEAKER_00: That's a big, that's a broad topic.
- 2:43SPEAKER_01: Yeah, yeah, yeah.
- 2:44So,
- 2:45you know,
- 2:46GWAS is a buzzword that I think a lot of the listeners have
- 2:48heard.
- 2:48Like,
- 2:48tell us what a GWAS is and tell us how it's useful in
- 2:53understanding heritability.
- 2:54SPEAKER_00: Yeah, so GWAS stands for Genome-Wide Association Study.
- 2:59And this became the key study design in human genetics since
- 3:06about 2007 to 2010.
- 3:08SPEAKER_00: And it's really dominated the field since then.
- 3:13SPEAKER_00: And it's a very simple study design.
- 3:16All that you're doing really is looking for correlations
- 3:19between genetic variants
- 3:21SPEAKER_00: and diseases or traits like height,
- 3:25And it's actually simpler than the studies that went before
- 3:28these linkage studies
- 3:28SPEAKER_00: that were based on pedigrees and families,
- 3:30SPEAKER_00: which were kind of cool.
- 3:32SPEAKER_00: But the GWAS design,
- 3:34SPEAKER_00: which is much more just a simple regression of phenotype onto genotype,
- 3:39SPEAKER_00: that ended up being much more powerful and simpler,
- 3:44easier to gather large amounts of data in these big
- 3:47meta-analyses as well.
- 3:49SPEAKER_00: And that's...
- 3:51SPEAKER_00: kind of revolutionized the field.
- 3:53SPEAKER_00: There were a lot of critics of GWAS early on,
- 3:56especially people that thought it wasn't going to tell us much
- 3:59about biology.
- 4:00SPEAKER_00: And they were partly right, partly wrong, I would say.
- 4:03SPEAKER_00: So the original impetus for GWAS was to find
- 4:07Genetic variants that were strongly associated with some
- 4:09disease like diabetes.
- 4:11SPEAKER_00: And we found some of those early on.
- 4:14SPEAKER_00: Variants associated with body mass index, diabetes, height, whatever.
- 4:19SPEAKER_00: And from those strongly associated variants in the population,
- 4:23SPEAKER_00: we would be able to figure out something about the traits biology.
- 4:27SPEAKER_00: And that was a moderately successful research program.
- 4:31SPEAKER_00: But...
- 4:32SPEAKER_01: Well, actually, let me just interrupt real quick.
- 4:35SPEAKER_01: So when you talk about traits biology,
- 4:36SPEAKER_01: just to give the listeners an example,
- 4:39SPEAKER_01: so you do GWAS on height,
- 4:41SPEAKER_01: and there's something called Gene Ontology Go,
- 4:44SPEAKER_01: the categories of what are the genes doing functionally.
- 4:47SPEAKER_01: It turns out that a lot of the height-related genes are in...
- 4:52SPEAKER_01: Involved in bone morphogenesis.
- 4:54SPEAKER_01: Okay,
- 4:54SPEAKER_01: this is,
- 4:55SPEAKER_01: as you know,
- 4:55and this is exactly what you would expect if the GWAS is
- 4:59actually zooming in on
- 5:01SPEAKER_01: real biological pathways of relevance to the characteristic,
- 5:04SPEAKER_01: right?
- 5:05SPEAKER_01: And you see something similar with,
- 5:06SPEAKER_01: you know,
- 5:07SPEAKER_01: cognitive genomic stuff that you're interested in,
- 5:09SPEAKER_01: correct?
- 5:10SPEAKER_00: Yeah, although the cognitive genomic stuff is a bit tricky.
- 5:15SPEAKER_00: I mean,
- 5:16I think these brain-related or cognitive or educational traits
- 5:21tend to be what we
- 5:22SPEAKER_00: call highly polygenic.
- 5:25SPEAKER_00: So they're affected by many, many genes across the genome.
- 5:28And there's not really any one particular gene that really
- 5:31stands out strongly from
- 5:33SPEAKER_00: these GWAS studies.
- 5:35SPEAKER_00: Now,
- 5:36SPEAKER_00: sort of skipping ahead,
- 5:37SPEAKER_00: I guess,
- 5:37to a different design where they look at rare variants that
- 5:41have a damaging effect
- 5:43on a protein that sometimes that cause the protein to
- 5:47completely lose its function.
- 5:49They've more directly identified genes associated with
- 5:52cognitive ability than the
- 5:53SPEAKER_00: typical GWAS studies.
- 5:55SPEAKER_00: So
- 5:56SPEAKER_00: Yeah.
- 5:56SPEAKER_00: So in terms of discovering biology,
- 5:58GWAS has been moderately successful and it varies from trait to
- 6:01trait and disease
- 6:02SPEAKER_00: to disease.
- 6:03And now we also have these related study designs that look at
- 6:08rare variants that
- 6:09SPEAKER_00: have a known functional impact on particular proteins.
- 6:13SPEAKER_00: And we can use that to sort of directly identify,
- 6:15SPEAKER_00: oh,
- 6:15SPEAKER_00: if this gene is broken,
- 6:17SPEAKER_00: then it messes up this trait or this disease or something.
- 6:21SPEAKER_00: And pharmaceutical companies are very interested in that because you can
- 6:25SPEAKER_00: you can identify a drug target from that.
- 6:28So that's another kind of study design that sort of evolved out
- 6:30of this GWAS paradigm.
- 6:33SPEAKER_01: Yeah, and so, you know, there was this whole issue.
- 6:35SPEAKER_01: So GWAS, you're talking about GWAS.
- 6:38SPEAKER_01: Relate to what heritability is and how heritability is different,
- 6:41because I think the two are confounded or conflated in a lot of
- 6:43people's minds.
- 6:44SPEAKER_00: Yeah, yeah.
- 6:45SPEAKER_00: People do confuse this.
- 6:49I was helping Scott Alexander with a post he was doing about
- 6:53missing heritability
- 6:54SPEAKER_00: recently,
- 6:54SPEAKER_00: and I kind of realized that
- 6:56SPEAKER_00: And even for smart people who are outsiders in this field,
- 7:00it kind of gets confusing because there are all these different
- 7:03study designs and
- 7:04different ways of measuring heritability and how much of a
- 7:07trait we can predict.
- 7:08And they're kind of related to each other,
- 7:10but they're not exactly the same.
- 7:11SPEAKER_00: So it can get a bit confusing for outsiders to understand.
- 7:16SPEAKER_00: And so with a GWAS,
- 7:18we're not really trying to estimate heritability with a GWAS or
- 7:22with the GWAS data.
- 7:24SPEAKER_00: And typically, we only have data on what are called common variants.
- 7:29SPEAKER_00: So these are variants that say the minor version of the variant.
- 7:34SPEAKER_00: So you have an A or a T.
- 7:35SPEAKER_00: The less common one is present at a frequency of,
- 7:39SPEAKER_00: say,
- 7:39SPEAKER_00: at least 1% or at least 0.1%,
- 7:42SPEAKER_00: which is actually only a minority of the variants in the genome.
- 7:47But from that GWAS data,
- 7:48you can estimate something people call SNP heritability.
- 7:53And that's the amount of trait variation you can explain with a
- 7:57linear model of the
- 7:59SPEAKER_00: genotypes that you have available for your GWAS.
- 8:02SPEAKER_00: And that can put an upper bound on...
- 8:06how much of the trait we can predict using that kind of data
- 8:10that we have for GWAS
- 8:11SPEAKER_00: with what we call a polygenic score.
- 8:14SPEAKER_00: So a polygenic score takes the GWAS results across the genome.
- 8:20SPEAKER_00: And some of these methods are quite sophisticated,
- 8:23SPEAKER_00: but at a basic level,
- 8:25all it's doing is taking the associations from a GWAS and using
- 8:29them to add up the
- 8:30SPEAKER_00: effects,
- 8:31the estimated effects of all these different variants across
- 8:33the genome in a simple
- 8:34SPEAKER_00: kind of linear model.
- 8:36SPEAKER_00: And that is another measure of explained variation.
- 8:41SPEAKER_00: So you could have,
- 8:43SPEAKER_00: say,
- 8:43SPEAKER_00: for something like height,
- 8:46SPEAKER_00: the SNP heritability might be 45% when we estimate it from GWAS data.
- 8:52SPEAKER_00: Then we have a polygenic score,
- 8:55and they're actually very good for height now because we have
- 8:57sample sizes around 5
- 8:59SPEAKER_00: million that are available for height.
- 9:02So the polygenic score might mean we're able to actually
- 9:05predict around 40% of the
- 9:08SPEAKER_00: variation in height.
- 9:12SPEAKER_00: And then this is in contrast to more traditional... Let me just be clear.
- 9:20SPEAKER_01: This is population-wide variation.
- 9:22SPEAKER_00: Yes, yes.
- 9:22SPEAKER_00: This is a variation in the population you're explaining.
- 9:25SPEAKER_00: And actually,
- 9:26SPEAKER_00: it's a little complicated because it's not necessarily...
- 9:29The GWAS-type estimates of heritability and what we can explain
- 9:33with apologetic
- 9:34SPEAKER_00: score,
- 9:35SPEAKER_00: they're not...
- 9:37SPEAKER_00: necessarily just capturing the causal effect of genetics.
- 9:41SPEAKER_00: They can also be capturing gene environment correlations or confounding.
- 9:46It gets a bit complicated,
- 9:47but assortative mating can also inflate these estimates.
- 9:51So they're not as well-controlled estimates of heritability as
- 9:59the classical
- 10:00SPEAKER_00: estimates of heritability that mostly come from twin studies.
- 10:04SPEAKER_00: So twin studies,
- 10:05SPEAKER_00: the classic twin study design,
- 10:07compares the correlations between identical monozygotic twins
- 10:12and non-identical
- 10:13SPEAKER_00: dizygotic twins.
- 10:15And the key assumption there is that any greater correlation
- 10:19between identical
- 10:20SPEAKER_00: twins and non-identical twins is due to genetics,
- 10:23SPEAKER_00: because they both are raised in the same kind of family environment.
- 10:26And it seems like a reasonable assumption that,
- 10:30for example,
- 10:30with height...
- 10:31SPEAKER_00: those estimates usually come out at like 80%.
- 10:34So this is a bigger number than we get from the GWAS data,
- 10:39which is say like 45%.
- 10:40SPEAKER_00: Now there are various complexities in comparing those numbers,
- 10:43SPEAKER_00: but this gap has been labeled the missing heritability problem.
- 10:48And that's something we could maybe go into a bit more detail
- 10:52on if you want.
- 10:53SPEAKER_01: Yeah,
- 10:54SPEAKER_01: so,
- 10:54SPEAKER_01: you know,
- 10:55SPEAKER_01: missing heritability just means that,
- 10:56SPEAKER_01: you know,
- 10:56SPEAKER_01: you have these twin studies that show up.
- 10:59SPEAKER_01: I mean,
- 10:59SPEAKER_01: I think it's like,
- 11:01SPEAKER_01: you could say beyond twin studies,
- 11:02SPEAKER_01: just like classical design of these sorts of,
- 11:04SPEAKER_01: you know,
- 11:05SPEAKER_01: inference of heritabilities,
- 11:06SPEAKER_01: because it's not just like in human,
- 11:07SPEAKER_01: you know,
- 11:07SPEAKER_01: constructs.
- 11:08Like heritabilities obviously exist in animals and stuff and
- 11:10agricultural genetics.
- 11:11SPEAKER_01: And then what happened initially,
- 11:13SPEAKER_01: I think,
- 11:13SPEAKER_01: with missing heritability,
- 11:14SPEAKER_01: and I'm talking about,
- 11:16SPEAKER_01: say,
- 11:16SPEAKER_01: 2010,
- 11:17SPEAKER_01: around that period,
- 11:18SPEAKER_01: is you had these genotyping arrays with hundreds of thousands.
- 11:21SPEAKER_01: Actually, at that point, sometimes it was tens of thousands of markers.
- 11:25And they were not picking up the heritability and the hits on
- 11:30the GWAS,
- 11:30right?
- 11:31SPEAKER_01: And so statistically, you knew that it was supposedly there.
- 11:35SPEAKER_01: But you weren't figuring it,
- 11:37SPEAKER_01: you weren't,
- 11:39SPEAKER_01: you know,
- 11:39SPEAKER_01: pinning down the causal variance in terms of like a molecular,
- 11:42SPEAKER_01: genetic,
- 11:44SPEAKER_01: you know,
- 11:44SPEAKER_01: mechanistic fashion,
- 11:45SPEAKER_01: right?
- 11:46SPEAKER_01: Mm-hmm.
- 11:47SPEAKER_00: Yeah, yeah.
- 11:48SPEAKER_00: There are different sort of notions of missing heritability.
- 11:51SPEAKER_00: One is like,
- 11:52SPEAKER_00: I mean,
- 11:53SPEAKER_00: I think I've heard Eric Terkheimer kind of propose this,
- 11:55that we should only really consider heritability explained if
- 11:58we have like an exact
- 12:00SPEAKER_00: causal mechanism for each variant.
- 12:03SPEAKER_00: We know the exact causal variant.
- 12:04SPEAKER_00: We know its exact biological mechanism.
- 12:06SPEAKER_00: I would say that is not a widely accepted definition.
- 12:10And then there are more sort of statistical notions of like how
- 12:13much variation we
- 12:14SPEAKER_00: can explain,
- 12:15SPEAKER_00: even if we don't necessarily know
- 12:17which causal variants are explaining it or through what
- 12:20biological mechanisms
- 12:21they're acting and that's sort of more where the debate is now
- 12:25um but yeah this
- 12:27debate is sort of it does relate to how well we can predict
- 12:31traits genetically um
- 12:33but they're not exactly the same thing and i've seen a lot of
- 12:35people get confused
- 12:36SPEAKER_00: some people intentionally i think so confusion about this but um
- 12:41SPEAKER_00: If we can only predict 2% of the variation in a trait,
- 12:46SPEAKER_00: but we actually know that from the GWAS data,
- 12:50SPEAKER_00: if we add a bigger sample or something,
- 12:52SPEAKER_00: maybe that could go up to 20%.
- 12:53SPEAKER_00: It doesn't mean that only 2% is the heritability.
- 13:00SPEAKER_00: Yeah,
- 13:00you just have to be careful about these different notions like
- 13:03what we can predict
- 13:04SPEAKER_00: versus what we could in theory explain.
- 13:07And then the maximum possible that we might be able to explain
- 13:11if we could account
- 13:13SPEAKER_00: for all genetic variants,
- 13:14SPEAKER_00: not just the ones that we have for typical GOS data.
- 13:19SPEAKER_01: Yeah,
- 13:19SPEAKER_01: I mean,
- 13:20SPEAKER_01: at this point,
- 13:21when it comes to missing heritability and these sorts of things
- 13:23and understanding
- 13:23SPEAKER_01: causality,
- 13:25SPEAKER_01: you're getting into deeper philosophical waters,
- 13:27SPEAKER_01: right?
- 13:28SPEAKER_01: But I would say that in terms of explanatory power,
- 13:33SPEAKER_01: for the regular person,
- 13:34if you can get a medical test and that gives you a certain
- 13:38polygenic risk score
- 13:40SPEAKER_01: from certain variants,
- 13:42SPEAKER_01: I think for them,
- 13:43SPEAKER_01: that's causality.
- 13:44SPEAKER_00: Yeah,
- 13:45SPEAKER_00: I mean,
- 13:45the polygenic score performance is a bit more concrete and
- 13:50directly relevant to
- 13:51SPEAKER_00: real-world applications.
- 13:55There are promising results now looking at how we can use
- 13:58polygenic risk scores for
- 14:01prostate cancer or breast cancer to improve our decisions about
- 14:05when to bring
- 14:05SPEAKER_00: people in for an MRI or some sort of other screening test.
- 14:10So we can measure the performance of these genetic predictors
- 14:14quite concretely and
- 14:15reason about whether they're going to be worthwhile for various
- 14:19applications.
- 14:20SPEAKER_00: The missing heritability debate is more abstract and maybe academic.
- 14:25SPEAKER_00: I think it's important because it does get at this point.
- 14:29more fundamental question of like nature versus nurture to what
- 14:32degree especially
- 14:33when we talk about things like socioeconomic outcomes
- 14:36intelligence education then
- 14:39this missing heritability debate like it does get to
- 14:43fundamental questions about
- 14:45inequality in society and things like that and then the the the
- 14:49relevant parameters
- 14:50are maybe less how well other you know we'll talk more about
- 14:53the applications of
- 14:54SPEAKER_00: genetic prediction of intelligence but
- 14:57at a more high-level societal level than knowing to what degree
- 15:02are the differences
- 15:04SPEAKER_00: in school outcomes or intelligence,
- 15:05SPEAKER_00: are they genetic?
- 15:06SPEAKER_00: And that is important.
- 15:07SPEAKER_00: And there's still an ongoing debate,
- 15:08SPEAKER_00: like,
- 15:09SPEAKER_00: are twin studies overestimating heritability for these traits?
- 15:13SPEAKER_00: I think there is evidence,
- 15:14SPEAKER_00: actually,
- 15:15SPEAKER_00: that they are overestimating heritability for educational attainment.
- 15:19SPEAKER_00: We've seen that by comparing.
- 15:21SPEAKER_00: So there are more robust methods for estimating heritability.
- 15:26One uses relatedness variation between siblings,
- 15:29which is basically random.
- 15:30SPEAKER_00: It's due to random segregation of genetic material during meiosis,
- 15:34like whether you inherit one or the other copy of each allele
- 15:37from your mother and
- 15:38SPEAKER_00: your father.
- 15:38SPEAKER_00: So this varies between siblings.
- 15:40SPEAKER_00: So siblings can vary between 60% related and 40% related, roughly.
- 15:46SPEAKER_00: And that random variation can be used to estimate heritability.
- 15:50An interesting study done by Loic Yango and Peter Vischer in
- 15:53collaboration with
- 15:55SPEAKER_00: 23andMe,
- 15:55SPEAKER_00: this preprint just came out a month or two ago,
- 15:58SPEAKER_00: so this is very cutting-edge stuff.
- 16:00SPEAKER_00: They put together a
- 16:02SPEAKER_00: a sample of about half a million sibling pairs,
- 16:05meta-analyzing 23andMe with a study that I did in DECODE and
- 16:09then some other
- 16:09SPEAKER_00: studies that they did.
- 16:11SPEAKER_00: And they found,
- 16:12I think these provide actually probably the gold standard
- 16:14heritability estimates
- 16:15SPEAKER_00: now.
- 16:15SPEAKER_00: They're sort of eclipsing twin studies.
- 16:18SPEAKER_00: And they put the heritability of height at like 70%,
- 16:21SPEAKER_00: BMI at 50%,
- 16:23SPEAKER_00: cardiometabolic disease,
- 16:25SPEAKER_00: I think was around 26% or something like that.
- 16:27SPEAKER_00: So I think, well, then you have to adjust.
- 16:31SPEAKER_00: There's an assortative mating thing as well.
- 16:33SPEAKER_00: So actually,
- 16:34SPEAKER_00: these methods and twin studies can also underestimate heritability due to
- 16:37SPEAKER_00: assortative mating.
- 16:38SPEAKER_00: So if you count for that,
- 16:39then I think their estimate for the heritability of height was
- 16:42something like 86%.
- 16:44SPEAKER_00: which is not that far off what twin studies are doing,
- 16:47but you have to be careful about the comparison with the
- 16:50assortative mating thing.
- 16:51SPEAKER_00: So it looks like maybe twin estimates are still a bit higher.
- 16:56SPEAKER_00: than we're getting from these more robust genomic methods.
- 17:00But unfortunately,
- 17:00we don't have really precise estimates from these more robust
- 17:03methods.
- 17:04SPEAKER_00: The sibling-based method,
- 17:05there's also a method I developed called relatedness
- 17:08disquilibrium regression,
- 17:09SPEAKER_00: which is like a generalization of the sib version.
- 17:12I know some people now working on applying that to things like
- 17:15school grades and
- 17:16SPEAKER_00: cognitive outcomes in kids.
- 17:18So we're hopefully going to get some estimates from that in the
- 17:21near future.
- 17:21SPEAKER_00: But 23andMe actually,
- 17:23SPEAKER_00: they used to,
- 17:25SPEAKER_00: we used to,
- 17:26SPEAKER_00: in the SSGAC,
- 17:28SPEAKER_00: used to collaborate with them on doing education GWAS.
- 17:31SPEAKER_00: We did a 3 million person GWAS of educational attainment.
- 17:34SPEAKER_00: Most of that data came from 23andMe,
- 17:36SPEAKER_00: but they had a change of heart at some point around 2020,
- 17:38SPEAKER_00: 2021.
- 17:39SPEAKER_00: I wonder why.
- 17:41SPEAKER_00: And then decided it was unethical or bad publicity or something to do
- 17:47SPEAKER_00: any GWAS of educational attainment.
- 17:49SPEAKER_00: And actually,
- 17:51SPEAKER_00: they wouldn't allow,
- 17:52SPEAKER_00: I don't think,
- 17:53the Vischer and Yengo group to study the heritability of
- 17:57education or other
- 17:58SPEAKER_00: cognitive and educational outcomes 23 may have,
- 18:00which is a shame because there's so much debate about that
- 18:03online and they have
- 18:05SPEAKER_00: some good data on it and they're not letting us look at it,
- 18:08SPEAKER_00: unfortunately.
- 18:09SPEAKER_00: Yeah.
- 18:09SPEAKER_01: Yeah,
- 18:10SPEAKER_01: I mean,
- 18:10SPEAKER_01: so I do want to talk a little bit about the political aspect,
- 18:15SPEAKER_01: but let's dig a little bit more into the science.
- 18:21SPEAKER_01: Can you explain to me why?
- 18:22SPEAKER_01: I mean, I think I know because – I mean, whatever, like –
- 18:26I'm in a different state of knowledge about this sort of stuff
- 18:29than the listener.
- 18:29SPEAKER_01: And so I am following what you're saying.
- 18:32But I just started realizing when you're talking about
- 18:35disequilibrium regression
- 18:38SPEAKER_01: and,
- 18:38SPEAKER_01: you know,
- 18:39SPEAKER_01: I'm trying to say.
- 18:40So when you say assortative mating distorts heritability
- 18:44estimates,
- 18:45tell us why,
- 18:45Professor.
- 18:47SPEAKER_00: Right.
- 18:47So assortative mating is when parents are more alike than
- 18:54random pairs for certain traits.
- 18:57SPEAKER_00: And this is actually pretty strong for education, intelligence.
- 19:02SPEAKER_00: Height is an obvious one.
- 19:03You can see that if a woman is very tall,
- 19:08she's often with an even taller guy.
- 19:10SPEAKER_00: And this
- 19:13SPEAKER_00: over many generations,
- 19:15has the effect of increasing the genetic variation in the
- 19:18population because it
- 19:19induces positive correlations between alleles that affect
- 19:22height in the father and
- 19:24SPEAKER_00: alleles that affect height in the mother.
- 19:26SPEAKER_00: And this inflates the genetic variation in height over many generations.
- 19:32SPEAKER_00: So this is kind of a complicated thing to think about.
- 19:36SPEAKER_00: But basically,
- 19:37SPEAKER_00: in the twin estimates and the siblings,
- 19:40SPEAKER_00: so all the family based estimates,
- 19:41SPEAKER_00: they basically don't account for the inflation of genetic variants due to
- 19:47SPEAKER_00: assortative mating.
- 19:48SPEAKER_00: So they're downward biased because of that.
- 19:51SPEAKER_00: And there are ways to adjust for that,
- 19:53but it's actually not trivial to do it in a way that is
- 19:56completely without
- 19:59SPEAKER_00: assumptions or whatever.
- 20:00SPEAKER_00: So there's always certain assumptions one has to make.
- 20:01SPEAKER_00: But basically,
- 20:03SPEAKER_00: yes,
- 20:03this applies to estimates of heritability for intelligence and
- 20:07education and height
- 20:08SPEAKER_00: that the ones coming out of twin studies and sibling-based ones or
- 20:13SPEAKER_00: This RDR method that I developed,
- 20:15SPEAKER_00: they're all downward biased by assortative mating,
- 20:18SPEAKER_00: but then other estimates from other methods can be upward biased by it.
- 20:22SPEAKER_00: So,
- 20:22SPEAKER_00: you know,
- 20:23SPEAKER_00: the fact that humans don't mate randomly,
- 20:27SPEAKER_00: it makes the lives of statistical geneticists much more complicated,
- 20:30SPEAKER_00: although maybe more interesting as well,
- 20:32SPEAKER_00: I would argue.
- 20:34SPEAKER_00: And so actually,
- 20:35SPEAKER_00: you know,
- 20:35I had Greg Clark on the podcast to talk about his work on
- 20:41heritability and social
- 20:43SPEAKER_01: status.
- 20:44SPEAKER_01: And you kind of got involved in that discussion.
- 20:47Can we relitigate that a little bit in terms of like what your
- 20:50take was on that?
- 20:51SPEAKER_00: Sure.
- 20:52SPEAKER_00: So Greg Clark put together this very impressive data set looking at,
- 20:58SPEAKER_00: I think he had various outcomes like wealth,
- 21:01SPEAKER_00: literacy,
- 21:04SPEAKER_00: some sort of general measure of socioeconomic status.
- 21:07SPEAKER_00: And he looked at how that's transmitted over generations.
- 21:10SPEAKER_00: This was using sort of marriage and
- 21:12um other kinds of records he put together going back to like
- 21:16the 1600s in england
- 21:18so it was a cool data set and i mean this actually goes back to
- 21:24one of the
- 21:24foundational questions at the at the founding of quantitative
- 21:28genetics which is
- 21:29correlations between relatives so sir francis galton cousin of
- 21:34charles darwin
- 21:36He coined the term regression in relation to this phenomenon he
- 21:41observed called
- 21:41SPEAKER_00: regression to the mean or mediocrity,
- 21:44SPEAKER_00: I think he maybe called it in the original parlance,
- 21:47SPEAKER_00: which was observed for height in the offspring.
- 21:50SPEAKER_00: If you look at the average height of the parents...
- 21:54SPEAKER_00: then the height of the offspring is correlated with that,
- 21:58SPEAKER_00: but it tends to be like a little bit lower.
- 21:59So the prediction is a little bit lower than the average height
- 22:02of the parents.
- 22:04SPEAKER_00: And then R.A.
- 22:05SPEAKER_00: Fisher,
- 22:07so he's kind of the founder of quantitative genetics and one of
- 22:13the most important
- 22:13figures in both the foundations of statistics and population
- 22:19and quantitative
- 22:20SPEAKER_00: genetics,
- 22:21SPEAKER_00: And he wrote this paper in 1918,
- 22:23SPEAKER_00: which really,
- 22:24SPEAKER_00: it just has,
- 22:25I don't know if it's as equivalent to many other fields where
- 22:27there's one,
- 22:27SPEAKER_00: maybe Newton in physics or something.
- 22:29SPEAKER_00: It's almost like Newton in physics.
- 22:31He kind of like wrote this paper that basically set the
- 22:35foundations for so much of
- 22:36SPEAKER_00: what we still do today.
- 22:37SPEAKER_00: Yeah.
- 22:39SPEAKER_00: And one of these foundational questions was,
- 22:42SPEAKER_00: how can we explain the correlations that we see between relatives?
- 22:46SPEAKER_00: So parents and offspring are correlated.
- 22:49SPEAKER_00: Parents and then the offspring and the offspring's offspring,
- 22:54SPEAKER_00: their grandchildren also correlate.
- 22:56So you can compute correlations between all different kinds of
- 22:59relatives,
- 23:00including in-laws.
- 23:02And you can think about what kind of models can explain those
- 23:06correlations.
- 23:07SPEAKER_00: So that's what Greg Clark did.
- 23:09SPEAKER_00: He took this model from R.A.
- 23:11SPEAKER_00: Fisher in his 1918 paper,
- 23:13which looks at how you can explain correlations between
- 23:17relatives based on genetic
- 23:19SPEAKER_00: inheritance and assortative mating.
- 23:21Assortative mating tends to increase correlations between
- 23:23relatives,
- 23:24so...
- 23:25SPEAKER_00: especially more distant relatives.
- 23:27SPEAKER_00: So generally,
- 23:28SPEAKER_00: without assortative mating,
- 23:31SPEAKER_00: under a linear model,
- 23:33SPEAKER_00: you expect that the correlations between relatives decay linearly.
- 23:37SPEAKER_00: So siblings would be half as correlated as monozygotic twins,
- 23:43SPEAKER_00: and then cousins would be one-eighth as correlated as monozygotic twins,
- 23:48SPEAKER_00: and this just all linearly decays.
- 23:50SPEAKER_00: When you have assortative mating, it doesn't decay quite linearly.
- 23:53SPEAKER_00: You have increased correlation between more distant relatives.
- 23:57SPEAKER_00: So from this pattern of correlations between different relative classes,
- 24:01SPEAKER_00: you can fit a model based on Fisher's 1918 paper,
- 24:04SPEAKER_00: which is what Greg Clark did,
- 24:06SPEAKER_00: to try and understand,
- 24:08can we actually explain all of these correlations just with
- 24:11genetic inheritance and
- 24:13SPEAKER_00: assortative mating?
- 24:15SPEAKER_00: And
- 24:16Clark's paper made quite a remarkable observation that just
- 24:21those two parameters,
- 24:23SPEAKER_00: if you fit them,
- 24:23SPEAKER_00: can actually explain these correlations pretty well.
- 24:27SPEAKER_00: However,
- 24:28just because you can fit a model with a simple model that
- 24:31explains a set of data,
- 24:32SPEAKER_00: it doesn't actually mean that that model is true.
- 24:34SPEAKER_00: And I think that some more
- 24:39There's been actually some pretty interesting papers coming out
- 24:42of the Scandinavian
- 24:43countries where they have these national register data sets
- 24:48where they can look at
- 24:49SPEAKER_00: huge numbers of relative pairs,
- 24:52SPEAKER_00: including in-laws.
- 24:53And they see that maybe the simple model that fits Greg Clark's
- 24:56data doesn't quite
- 24:57SPEAKER_00: fit the more complicated model.
- 24:59set of relatives they can look at in these national registers
- 25:02and i think there is
- 25:03some evidence especially for even in greg clark's work i should
- 25:06say there's
- 25:07evidence that there's non-genetic transmission of wealth which
- 25:10makes sense i mean
- 25:11we know there's a mechanism for that people leave wealth to
- 25:14their offspring in
- 25:15their wills so it can't it clearly can't be purely genetically
- 25:18inherited wealth
- 25:19right so we see that in the greg clark data and then even for
- 25:22some of these more
- 25:23SPEAKER_00: educational outcomes
- 25:25We see stuff that doesn't quite fit that simple Fisherian model
- 25:28that just has
- 25:30SPEAKER_00: additive genetic inheritance and assortative mating.
- 25:33There's probably some cultural transmission from parents to
- 25:36offspring as well.
- 25:38SPEAKER_01: Yeah,
- 25:38SPEAKER_01: which is,
- 25:38SPEAKER_01: I mean,
- 25:39SPEAKER_01: cultural transmission from parents to offspring,
- 25:40SPEAKER_01: though,
- 25:40SPEAKER_01: that's like,
- 25:41SPEAKER_01: I would say that is the Western null,
- 25:44SPEAKER_01: though.
- 25:45And so I think that that is the radical aspect of Greg Clark's
- 25:49model and this way
- 25:50SPEAKER_01: of thinking where heritability in a genetic system,
- 25:54SPEAKER_01: we're talking about heritability,
- 25:55SPEAKER_01: I mean,
- 25:55SPEAKER_01: the colloquial way to think about heritability is,
- 25:58SPEAKER_01: well,
- 25:58SPEAKER_01: it's just like parent to offspring.
- 26:00SPEAKER_01: But when we're talking about it,
- 26:01SPEAKER_01: just for the listener to be clear,
- 26:03SPEAKER_01: heritability is,
- 26:04SPEAKER_01: you know,
- 26:05the proportion of the variation of the population of a trait
- 26:09that can be explained
- 26:09SPEAKER_01: by genetics,
- 26:10SPEAKER_01: specifically genetic variation,
- 26:11SPEAKER_01: you know?
- 26:12And so I think the interesting aspect of Greg's thesis is that
- 26:18over the long term,
- 26:19SPEAKER_01: and,
- 26:19SPEAKER_01: you know,
- 26:19SPEAKER_01: this thesis has been worked out in like many,
- 26:22SPEAKER_01: many books,
- 26:22SPEAKER_01: or not many,
- 26:23SPEAKER_01: but like,
- 26:23SPEAKER_01: you know,
- 26:23SPEAKER_01: multiple books now.
- 26:25SPEAKER_01: is that there is something to social status,
- 26:29SPEAKER_01: which it's,
- 26:30SPEAKER_01: you know,
- 26:30SPEAKER_01: that's a downstream variable,
- 26:32SPEAKER_01: right?
- 26:32SPEAKER_01: But there's something where social status is like a bundle of cognitive
- 26:37SPEAKER_01: characteristics,
- 26:38SPEAKER_01: perhaps,
- 26:38SPEAKER_01: behavioral.
- 26:39SPEAKER_01: Yeah, exactly.
- 26:40SPEAKER_01: Health as well.
- 26:41SPEAKER_01: Yeah, and that's quite stable.
- 26:45SPEAKER_01: And there's some,
- 26:46SPEAKER_01: like,
- 26:46SPEAKER_01: you know,
- 26:46SPEAKER_01: weird,
- 26:47SPEAKER_01: weird results from the economic historical literature of,
- 26:50SPEAKER_01: like,
- 26:51SPEAKER_01: you know,
- 26:51SPEAKER_01: the most prominent,
- 26:53SPEAKER_01: like,
- 26:53SPEAKER_01: financial families,
- 26:54SPEAKER_01: like,
- 26:55SPEAKER_01: you know,
- 26:55SPEAKER_01: in finance in Italy in,
- 26:56SPEAKER_01: like,
- 26:56SPEAKER_01: the 19th and 20th centuries,
- 26:58SPEAKER_01: their roots go back to the 14th century.
- 27:00That's,
- 27:00like,
- 27:00really crazy if you think about how many centuries that is,
- 27:03right?
- 27:03All these sort of Norman surnames that are still
- 27:06overrepresented in Oxford and
- 27:07SPEAKER_00: Cambridge and England and in military generals or whatever.
- 27:12SPEAKER_01: Yeah,
- 27:12SPEAKER_01: so I think the Norman data,
- 27:14SPEAKER_01: just to recapitulate,
- 27:16SPEAKER_01: is basically you have about what – actually,
- 27:20SPEAKER_01: the proportion of surnames have increased,
- 27:21SPEAKER_01: apparently,
- 27:22SPEAKER_01: from less than 0.5% to about almost 1%.
- 27:26SPEAKER_01: So there's been some elite over-reproduction of Normans, probably.
- 27:31SPEAKER_00: All people preferentially taking.
- 27:33SPEAKER_01: They hit the elite status.
- 27:34SPEAKER_01: Yeah.
- 27:35SPEAKER_01: So,
- 27:35SPEAKER_01: but these Norman surnames are like,
- 27:37SPEAKER_01: you know,
- 27:38SPEAKER_01: you know,
- 27:38SPEAKER_01: in some fields are still like,
- 27:39SPEAKER_01: like,
- 27:39SPEAKER_01: I think like 10 X overrepresented,
- 27:41SPEAKER_01: you know?
- 27:42SPEAKER_01: Yeah.
- 27:42SPEAKER_01: A thousand years later.
- 27:43SPEAKER_01: I think that's what you're,
- 27:44SPEAKER_01: I mean,
- 27:44SPEAKER_01: that's,
- 27:44SPEAKER_01: that's like a massive like retention of elite status.
- 27:47yeah the the clark has definitely documented this remarkable
- 27:51sort of persistence
- 27:53over generations and it it fits very well this simple genetic
- 28:00inheritance plus
- 28:01assortative mating model now i i argue that i i don't think
- 28:05that model explained i
- 28:06don't think it's a true causal model and this is why we need
- 28:09the genomic data
- 28:11SPEAKER_00: ultimately like
- 28:13SPEAKER_00: This is the advantage of having real genomic data on families as well.
- 28:18SPEAKER_00: The Greg Clark kind of data is great,
- 28:20SPEAKER_00: but it doesn't give us what we'd call causal identification.
- 28:26So you're just looking at these correlations and trying to
- 28:28explain them in different models.
- 28:30SPEAKER_00: But we can't necessarily identify,
- 28:32SPEAKER_00: is it really genetic causation or is it cultural transmission?
- 28:36SPEAKER_00: I mean,
- 28:36I think his data makes a strong argument that clearly genetic
- 28:42transmission is
- 28:42SPEAKER_00: playing an important role.
- 28:44SPEAKER_00: But this is why I think we still need these very robust...
- 28:47genomic database methods using families where we have this
- 28:51randomization it's like
- 28:52an experimental design that enables us to really identify the
- 28:56causal component of
- 28:57heritability in the current generation now things can change
- 28:59over time as well as
- 29:00SPEAKER_00: another factor but um
- 29:03SPEAKER_00: I think that,
- 29:03SPEAKER_00: you know,
- 29:04SPEAKER_00: the combination of those two kinds of data,
- 29:05this sort of long historical or national register data with
- 29:09genomic data that
- 29:10SPEAKER_00: allows us to do causal identification,
- 29:12SPEAKER_00: that's going to,
- 29:13SPEAKER_00: I think in the near term,
- 29:15SPEAKER_00: we're going to have a very clear picture of the causes of,
- 29:20SPEAKER_00: you know,
- 29:20correlations between relatives and persistence of inequality
- 29:24over generations.
- 29:26I guess the question that I have,
- 29:28when you say near term,
- 29:29like,
- 29:29what does that mean?
- 29:31SPEAKER_01: Like, I mean, like 10 years, 50 years?
- 29:33SPEAKER_00: I think less than that.
- 29:34SPEAKER_00: I would say five years.
- 29:38SPEAKER_00: Well, again, it depends a bit on what we're allowed to do.
- 29:41SPEAKER_00: A lot of these data are probably out there.
- 29:46SPEAKER_01: All right.
- 29:46SPEAKER_01: You know what?
- 29:47SPEAKER_01: You know what?
- 29:47SPEAKER_01: Like, we can move in a little bit.
- 29:49I do want to,
- 29:49like,
- 29:49touch back on the science that we're talking about the hair
- 29:51side paper.
- 29:51SPEAKER_01: All right.
- 29:52SPEAKER_00: No, I'm not trying to bring it all into politics.
- 29:54SPEAKER_00: No, no, but like – That is relevant to my question.
- 29:56SPEAKER_00: No, let's put it there.
- 29:57SPEAKER_01: So the LaRica Youngo paper that you were talking about,
- 29:59SPEAKER_01: they were not allowed to look at certain things.