Latest / Razib Khan's Unsupervised Learning / Davide Piffer: how Europeans became white
Transcript
- 0:00This podcast is brought to you by the Albany Public Library
- 0:03Main Branch and the
- 0:04SPEAKER_02: generosity of listeners like you.
- 0:09SPEAKER_02: What is a podcast?
- 0:16SPEAKER_02: God, Daddy, these people talk as much as you do.
- 0:26SPEAKER_02: Razeeb Khan's Unsupervised Learning.
- 0:29SPEAKER_00: Hey everybody,
- 0:29SPEAKER_00: this is Razeeb Khan with the Unsupervised Learning Podcast,
- 0:33SPEAKER_00: and today I am here with David Pifer,
- 0:37SPEAKER_00: who is,
- 0:38SPEAKER_00: some of you,
- 0:38SPEAKER_00: actually a lot of you probably know him from his substack,
- 0:43SPEAKER_00: pifer.substack.com,
- 0:45SPEAKER_00: and he also has a website,
- 0:46SPEAKER_00: David Pifer.
- 0:48SPEAKER_00: And the reason I'm talking to you,
- 0:50SPEAKER_00: David,
- 0:50SPEAKER_00: is because,
- 0:52SPEAKER_00: you know,
- 0:52SPEAKER_00: you've been putting a lot of posts up.
- 0:54SPEAKER_00: I want to talk to you about your posts on ancient DNA and pigmentation.
- 0:57SPEAKER_00: I thought it was super interesting.
- 0:58SPEAKER_00: You guys should check them out,
- 1:02SPEAKER_00: you know,
- 1:03SPEAKER_00: after this podcast,
- 1:04SPEAKER_00: after listening to the whole podcast,
- 1:05SPEAKER_00: obviously.
- 1:06SPEAKER_00: But before we go there, before we talk about ancient DNA and, you know,
- 1:12SPEAKER_00: what you are finding, what you're looking at there.
- 1:15I guess the main thing is can you talk about what you're
- 1:20interested in and what a
- 1:22polygenic score is or a polygenic index is and how that works
- 1:26first just in terms
- 1:27of the mechanics before we talk about ancient DNA and then what
- 1:30results you found.
- 1:32So I've always been interested in studying genetic variation
- 1:37between populations.
- 1:39This interest of mine started when I was a high school student
- 1:46and I read Cavalli Sporza's book.
- 1:50I think it was just being published,
- 1:53the Italian version,
- 1:54History and Geography of Human Genes.
- 1:57SPEAKER_01: And I read the Italian version,
- 1:59SPEAKER_01: of course,
- 2:00SPEAKER_01: and I remember looking at the maps and thinking,
- 2:06SPEAKER_01: wow,
- 2:06SPEAKER_01: this is really cool.
- 2:09SPEAKER_01: They had maps of Italy, France, and general all over the world.
- 2:14SPEAKER_01: But this was very primitive.
- 2:18SPEAKER_01: It was cool,
- 2:19but it was ahead of this time that it was primitive because it
- 2:23was based on just a
- 2:26SPEAKER_01: few genes.
- 2:28SPEAKER_01: Now we have much better data and also much bigger data sets.
- 2:35And we can also,
- 2:38on top of that,
- 2:39we also have this Polygenic course,
- 2:44so not just our ancestry.
- 2:46SPEAKER_01: genome-wide ancestry,
- 2:49SPEAKER_01: which is what Cavallis,
- 2:51SPEAKER_01: Fords,
- 2:52SPEAKER_01: and a lot of geneticists have been using for many years.
- 2:57But in the last 15 years,
- 3:02scientists have started building polygenic pores,
- 3:05which
- 3:07are a sort of genetic score that is built using your
- 3:15genome-wide association studies.
- 3:17SPEAKER_01: So like genome-wide association studies,
- 3:20SPEAKER_01: they study a phenotype,
- 3:21SPEAKER_01: a tree with large samples of people
- 3:26SPEAKER_01: then they correlate the allele counts,
- 3:33SPEAKER_01: the number of alleles within each individual to the values of a chain.
- 3:42SPEAKER_01: So, for example, for height, we find
- 3:48SPEAKER_01: Some alleys are correlated to taller starches.
- 3:53We assign a positive effect,
- 3:56depending on the correlation,
- 3:58the effect size we find in the GWAS.
- 4:00SPEAKER_01: And then we sum all these variants together to get polygenic score.
- 4:07It's a weighted sum of the frequencies and the effect size,
- 4:12in other words.
- 4:13And this index has predictive accuracy,
- 4:22which varies a lot depending on the phenotype.
- 4:26SPEAKER_01: But for height, it reached about 50%.
- 4:32And 4AQ,
- 4:35depending,
- 4:36of course,
- 4:36there is this era site,
- 4:38which is,
- 4:38I think,
- 4:40what was that,
- 4:4125%.
- 4:41SPEAKER_01: And then there is educational attainment, reached 15%, 20%.
- 4:45SPEAKER_01: And then we have other traits that...
- 4:52SPEAKER_01: at much lower value.
- 4:53So I always try to focus on the best studied phenotypes where
- 5:00we have larger sample
- 5:01SPEAKER_01: sizes and a lot of GWAS significant hits for my studies.
- 5:11Yeah,
- 5:12so,
- 5:12you know,
- 5:12you're talking about,
- 5:13like,
- 5:13allele counts and all these things.
- 5:15I mean,
- 5:17obviously,
- 5:17this is new technology,
- 5:20you know,
- 5:21methods that could not be done.
- 5:22SPEAKER_00: You're talking about L.L.
- 5:23SPEAKER_00: Cavalli's history and geography of genes,
- 5:25SPEAKER_00: which was published,
- 5:27SPEAKER_00: well,
- 5:28SPEAKER_00: I think 1995,
- 5:28SPEAKER_00: 96,
- 5:28SPEAKER_00: something like that.
- 5:29SPEAKER_00: It was a while ago now, right?
- 5:30SPEAKER_00: And you said there weren't that many genes.
- 5:32SPEAKER_00: And, you know, I would like the listener to know, you know,
- 5:35SPEAKER_00: There weren't that many genes.
- 5:35SPEAKER_00: There was fewer than 100 genes, really, in most of those analyses.
- 5:40SPEAKER_00: Basically, 100 genetic positions, 100 variants.
- 5:44SPEAKER_00: And when you're doing analyses of variants,
- 5:46SPEAKER_00: when we're doing analyses of variants now,
- 5:48SPEAKER_00: genome-wide SNP arrays will have 500,000 positions.
- 5:52SPEAKER_00: Right.
- 5:52SPEAKER_00: So we're in a whole different world.
- 5:53Whole genome sequences obviously have three billion positions
- 5:56of which about like,
- 5:57SPEAKER_00: you know,
- 5:57four and a half million or so in the average individual will be
- 6:00a polymorphism.
- 6:02SPEAKER_00: And so,
- 6:02SPEAKER_00: you know,
- 6:03SPEAKER_00: when you're talking about polygenic scores,
- 6:05the only reason it's possible really is because of a change in
- 6:09the amount of data
- 6:10SPEAKER_00: in terms of scale of the data we have.
- 6:11SPEAKER_00: Right.
- 6:12SPEAKER_00: Like in some ways.
- 6:13I think people can think about maybe polygenic scoring as a
- 6:16little bit like
- 6:16SPEAKER_00: regression,
- 6:17like expanding it out with all of these different variables
- 6:19that you're weighting,
- 6:20SPEAKER_00: and you create a composite output index.
- 6:23SPEAKER_00: So it's really not that complicated.
- 6:28It's just that we have the data now,
- 6:30and so you just choose to look at the data.
- 6:33SPEAKER_00: Now, it's one thing to look at data
- 6:37SPEAKER_00: On height,
- 6:38SPEAKER_00: you can measure people,
- 6:40SPEAKER_00: they're alive,
- 6:41SPEAKER_00: and even if they're dead,
- 6:42SPEAKER_00: if you have good skeletons,
- 6:44SPEAKER_00: you can actually measure them,
- 6:45SPEAKER_00: too.
- 6:47SPEAKER_00: And then,
- 6:48SPEAKER_00: you know,
- 6:48SPEAKER_00: height in contemporary populations,
- 6:50SPEAKER_00: obviously you have good quality data,
- 6:51SPEAKER_00: because you have the SNP array,
- 6:53SPEAKER_00: or you have whole genome sequences.
- 6:55SPEAKER_00: So what you're doing, though, is looking at polygenic scores.
- 6:58SPEAKER_00: So we're going to talk about pigmentation.
- 7:01I think you've done educational attainment,
- 7:03a bunch of other quantitative traits.
- 7:05SPEAKER_00: These are traits you cannot measure in ancient populations.
- 7:09SPEAKER_00: You cannot measure a Mesolithic European's educational attainment.
- 7:13SPEAKER_00: They've had an educational attainment.
- 7:16SPEAKER_00: But you also usually cannot measure – I mean –
- 7:21SPEAKER_00: You can kind of see some pigmentation.
- 7:23SPEAKER_00: I do have to say it's pretty clear that a lot of the old stuff,
- 7:27SPEAKER_00: when they said,
- 7:28SPEAKER_00: oh,
- 7:28SPEAKER_00: like this person has light hair,
- 7:31SPEAKER_00: when it's like a 5,000-year-old,
- 7:33SPEAKER_00: you know,
- 7:34SPEAKER_00: decomposed or desiccated individual,
- 7:37SPEAKER_00: you're not actually getting a lot of times the correct hair.
- 7:40SPEAKER_00: I think it's pretty clear, actually.
- 7:41SPEAKER_01: So when it comes to pigmentation... Like the skin, it's not per se...
- 7:47SPEAKER_01: Yes.
- 7:48SPEAKER_00: Yeah, the skin is also discolored.
- 7:50SPEAKER_00: So you can't get phenotypic measures on that.
- 7:52SPEAKER_00: And then also there's the other thing that I want to bring up,
- 7:56SPEAKER_00: which is the quality of the genomes.
- 7:59So,
- 8:00you know,
- 8:00obviously you have a couple of issues here,
- 8:03but I mean,
- 8:04the primary issue would be
- 8:05SPEAKER_00: Well,
- 8:05SPEAKER_00: I mean,
- 8:05SPEAKER_00: there's two issues that I guess I want to bring up,
- 8:07SPEAKER_00: like contamination and,
- 8:10SPEAKER_00: you know,
- 8:10SPEAKER_00: obviously the degradation of the DNA.
- 8:13And so you get,
- 8:13like,
- 8:13a lot less markers back for a polygenic spore,
- 8:18which is super important.
- 8:19SPEAKER_00: And then in terms of contamination, you can get the wrong result.
- 8:22SPEAKER_00: Now,
- 8:22SPEAKER_00: I will say,
- 8:23SPEAKER_00: when we're talking about pigmentation,
- 8:24SPEAKER_00: I will say there's a paper,
- 8:25SPEAKER_00: and I tweeted about this,
- 8:28SPEAKER_00: paper of Andronovo samples from,
- 8:30SPEAKER_00: like,
- 8:31SPEAKER_00: at the Altai in 2009.
- 8:34SPEAKER_00: And that paper is used to show how pale the early Andronovo are.
- 8:42SPEAKER_00: When later samples came out, they didn't really match that.
- 8:46SPEAKER_00: If you look at the frequencies...
- 8:48of blue eyes and blonde hair,
- 8:51they're suspiciously close to modern Russian frequencies.
- 8:54SPEAKER_00: So I'm just going to say,
- 8:55SPEAKER_00: I think that that was contamination,
- 8:58and it's actually pretty well known that a lot of the older
- 9:01ancient DNA literature,
- 9:03SPEAKER_00: including a lot of Svante Pabllo's work,
- 9:06SPEAKER_00: I mean,
- 9:06SPEAKER_00: some of it is explicitly known as contamination,
- 9:08SPEAKER_00: but...
- 9:09SPEAKER_00: Yeah,
- 9:11SPEAKER_01: for example,
- 9:11SPEAKER_01: you get a lot of contamination as well.
- 9:14SPEAKER_00: Yeah, yeah, yeah.
- 9:16SPEAKER_00: So these are issues.
- 9:17SPEAKER_00: How do you deal with them?
- 9:19SPEAKER_00: How do you think about these in terms of your analyses?
- 9:21SPEAKER_00: Because they're not trivial.
- 9:24So the data quality in terms of the DNA degradation affects the
- 9:33depth of coverage.
- 9:35SPEAKER_01: So how many sequences are retrieved.
- 9:39So the coverage is the percentage of the genome that is
- 9:44actually sequenced.
- 9:47And the depth of coverage is the average number of times each
- 9:52base of the DNA is sequenced.
- 9:55so what the dna data comes with depth of coverage information
- 10:02which gives us a
- 10:04reliable measure of dna degradation so i always add this
- 10:13measure to the regression
- 10:15uh so when i run the regressions with the polygeny score as the
- 10:20dependent variable
- 10:21SPEAKER_01: and then
- 10:23SPEAKER_01: date, which is years before present, a lot of geographical variables.
- 10:29SPEAKER_01: And then I also add coverage to take into account the quality.
- 10:36I should also say that recently I also found another metric
- 10:41because the genomes I'm
- 10:42SPEAKER_01: using are actually imputed.
- 10:44SPEAKER_01: So I've actually imputed genomes using Glimpse, which
- 10:49SPEAKER_00: Really quick, really quick.
- 10:51SPEAKER_00: I do want to tell my listeners what imputation is, right?
- 10:55Yeah,
- 10:55the imputation infers the missing data based on the nearby
- 11:03genetic variants.
- 11:04SPEAKER_01: So there are algorithms that are very computationally passive,
- 11:09but they enable you to assign a probability to each allele at
- 11:17the genomic positions
- 11:18SPEAKER_01: based on the nearby SNPs or genetic variants.
- 11:24SPEAKER_01: So,
- 11:25SPEAKER_01: instead of saying,
- 11:26SPEAKER_01: oh,
- 11:27SPEAKER_01: this GWAS SNP is missing,
- 11:30SPEAKER_01: I'm not going to count it for the Polygenius course,
- 11:34SPEAKER_01: I used the imputed variant.
- 11:39SPEAKER_01: So, of course, each imputed variant has a different probability.
- 11:46SPEAKER_01: And in my more recent paper on the Industrial Revolution,
- 11:51I also used the genome imputation probabilities to actually
- 11:59account for this
- 12:00SPEAKER_01: imputation error or variance.
- 12:03SPEAKER_01: And I added this to the regression.
- 12:06SPEAKER_01: It's strongly correlated to coverage, but it also adds some secrets.
- 12:12To answer your other question about the contamination,
- 12:16I don't directly take that into account.
- 12:21SPEAKER_01: I check the papers that come with the genomes,
- 12:27and usually they exclude the samples that have to mask
- 12:30contamination in the
- 12:32SPEAKER_01: downstream analysis.
- 12:35SPEAKER_01: Yeah.
- 12:37SPEAKER_00: Well, okay, so let's...
- 12:39SPEAKER_00: Let's take it as a given.
- 12:40SPEAKER_00: So actually,
- 12:41SPEAKER_00: I do have to just,
- 12:42SPEAKER_00: with the imputation,
- 12:43one thing that I guess I just want you to address real quickly
- 12:46is you have
- 12:47SPEAKER_00: Mesolithic Europeans,
- 12:48SPEAKER_00: for example.
- 12:49SPEAKER_00: There's no population that's genetically like them.
- 12:53SPEAKER_00: And so in terms of imputing,
- 12:55SPEAKER_00: because you have to have the reference panels that are high quality,
- 12:59SPEAKER_02: right?
- 13:00SPEAKER_00: Yeah.
- 13:01SPEAKER_00: And so those are presumably modern samples.
- 13:03So you're imputing the variation of modern samples into,
- 13:06say,
- 13:06a Mesolithic European.
- 13:08SPEAKER_00: I mean, how do you address that question?
- 13:12SPEAKER_01: Yeah, I actually used the 1kg, 1000 genome samples.
- 13:19SPEAKER_01: And I think,
- 13:21SPEAKER_01: of course,
- 13:22the imputation accuracy will draw the further the ancient
- 13:28samples are from the
- 13:30SPEAKER_01: modern samples.
- 13:32SPEAKER_01: So there will be more in this, the more ancient data.
- 13:38SPEAKER_01: Yeah,
- 13:41SPEAKER_01: I'm not an imputation expert,
- 13:43SPEAKER_01: but I use Glynx Jew,
- 13:44SPEAKER_01: and they recommend using 1000 genomes for ancient genomes imputation.
- 13:52SPEAKER_01: But of course, with mesolithic samples, you get noisier data.
- 14:00SPEAKER_01: But not necessarily,
- 14:02SPEAKER_01: it's not going to necessarily be biased in,
- 14:06SPEAKER_01: create a bias in the polygenic scores.
- 14:10SPEAKER_00: Well,
- 14:11SPEAKER_00: I guess what I would say is,
- 14:12SPEAKER_00: you know,
- 14:12SPEAKER_00: doing the pairwise FST,
- 14:13SPEAKER_00: looking at genetic differences,
- 14:14SPEAKER_00: Mesolithic Europeans are genetically distinct.
- 14:17SPEAKER_00: You know, they're not like any other, any modern populations.
- 14:19On the other hand,
- 14:20the pairwise FST,
- 14:22the genetic distance,
- 14:23genetic differences,
- 14:24allele-wise,
- 14:26will be lower than between,
- 14:28say,
- 14:28modern Europeans and East Asians a little bit,
- 14:31probably.
- 14:32So you could probably use that comparison to figure out how
- 14:37much a polygenic score
- 14:38SPEAKER_00: portability would degrade,
- 14:39SPEAKER_00: right?
- 14:40I mean,
- 14:40it wouldn't be totally unpredictable,
- 14:42especially when it comes to pigmentation.
- 14:43And we can talk about that later because you've done a little
- 14:46work on just
- 14:47SPEAKER_00: portability of polygenic scores,
- 14:49SPEAKER_00: which are...
- 14:50SPEAKER_00: You know,
- 14:51SPEAKER_00: there's been a lot of issues about the portability of polygenic scores,
- 14:54SPEAKER_00: but a lot of them actually happen to be African,
- 14:56non-African because of linkage disequilibrium and just the
- 14:59genetic distances are
- 15:00SPEAKER_00: greater between those populations.
- 15:01SPEAKER_00: The degradation within Eurasians is a lot less, I would say.
- 15:04SPEAKER_00: But I mean, that's kind of a minor technical thing.
- 15:07SPEAKER_00: I really want to get to, I guess, the trait.
- 15:09SPEAKER_00: So talk about high-risk Plex.
- 15:12SPEAKER_00: You know, tell the listener what it is.
- 15:14SPEAKER_00: And then talk about why you were interested in this.
- 15:17SPEAKER_01: Yeah,
- 15:17so Irisplex is actually a scoring method based on allele
- 15:26frequencies that assigns
- 15:28SPEAKER_01: the probability of each high polar based on the combination of
- 15:33SPEAKER_01: of different aliats that affect eye color and hair color.
- 15:42SPEAKER_01: I didn't use eye respect because eye respect is not public.
- 15:46SPEAKER_01: It's a proprietary algorithm.
- 15:49SPEAKER_01: I used GWAS of eye color and hair color.
- 15:55With hair color,
- 15:56it's easier because hair color is more polygenic than eye
- 16:02color.
- 16:03SPEAKER_01: And there is a good GWAS of eye color that I used that.
- 16:10SPEAKER_01: And for eye color, I used two jeans.
- 16:14SPEAKER_01: They're not perfectly...
- 16:21correlated to Irisplex,
- 16:24but they are very strongly correlated to Irisplex.
- 16:30SPEAKER_01: I use those because the Irisplex is not published.
- 16:35SPEAKER_00: Yeah, yeah.
- 16:37So,
- 16:37I mean,
- 16:38so you have skin color,
- 16:40you have skin color,
- 16:41hair color,
- 16:41and eye color.
- 16:42SPEAKER_00: I will say the eye color ones are very,
- 16:45SPEAKER_00: very,
- 16:46SPEAKER_00: you know,
- 16:47SPEAKER_00: powerful in my experience in terms of the predictors.
- 16:50SPEAKER_00: It's not like as polygenic.
- 16:51SPEAKER_00: I mean, like a lot of times you're just looking at variation.
- 16:54not necessarily dichotomous,
- 16:55but obviously blue versus non-blue,
- 16:57that sort of thing.
- 16:58Although some of the better,
- 17:01newer methods now can get spackling and various other things.
- 17:05SPEAKER_00: You get the phenotype pretty good, actually.
- 17:07SPEAKER_00: I will say it's pretty good on my kids.
- 17:10SPEAKER_00: They have variation of eye color.
- 17:13SPEAKER_00: because parents are obviously different, or not obviously, but they are.
- 17:16SPEAKER_00: And so the prediction is actually perfect.
- 17:18SPEAKER_00: So these forensic reconstruction methods are really good,
- 17:21SPEAKER_00: and they're used in criminal cases all across the world.
- 17:25So this is not just something that's used in an abstract,
- 17:29intellectual,
- 17:30scientific way.
- 17:31SPEAKER_00: They're tried and tested, and the panels are pretty robust that way.
- 17:36SPEAKER_00: And so the key issue, and I did actually...
- 17:40SPEAKER_00: The listeners will know this.
- 17:41SPEAKER_00: I did write on a paper, I think, at an Italian lab, actually.
- 17:44SPEAKER_00: And you're actually based in Italy.
- 17:46SPEAKER_00: So I don't know if it's coincidence or not.
- 17:49SPEAKER_00: But,
- 17:50SPEAKER_00: you know,
- 17:52SPEAKER_00: they used,
- 17:53SPEAKER_00: you know,
- 17:53SPEAKER_00: probabilistic methods,
- 17:54Bayesian calling methods to improve the information yield that
- 17:57they get from their
- 17:58SPEAKER_00: genotypes.
- 17:59And,
- 17:59you know,
- 18:00they got pretty much,
- 18:01they get a very similar result to you,
- 18:03actually.
- 18:03SPEAKER_00: So, you know, we're talking about... You write that paper.
- 18:07SPEAKER_00: Yeah, yes, yeah.
- 18:08SPEAKER_00: Which is good.
- 18:09SPEAKER_00: They should get the similar result to you.
- 18:11SPEAKER_01: I read that actually after I published my blog and I was like, it's cool.
- 18:20SPEAKER_01: I've actually revalidated each other.
- 18:24SPEAKER_00: Yeah, yeah, yeah.
- 18:24SPEAKER_00: That's what I'm saying.
- 18:25SPEAKER_00: So it's basically a double validation.
- 18:28The piece I used where I found this other Asian genome paper
- 18:35where they were trying
- 18:35SPEAKER_01: to pair the Ipolo.
- 18:38SPEAKER_01: And then I looked up the snips they were using, and then I used those.
- 18:45SPEAKER_01: So, yeah.
- 18:46SPEAKER_01: But it was another paper.
- 18:48SPEAKER_01: It was not this Italian paper.
- 18:49SPEAKER_01: I also learned about this one, yeah.
- 18:52Is there one where they show the acceleration in the
- 18:58pigmentation of the last one?
- 19:03SPEAKER_00: Yeah.
- 19:03SPEAKER_00: So, I mean, I guess, like, talk about, like, your major finding.
- 19:06SPEAKER_00: And,
- 19:07SPEAKER_00: like,
- 19:07SPEAKER_00: then we can get into the details of,
- 19:09SPEAKER_00: like,
- 19:09SPEAKER_00: you know,
- 19:09SPEAKER_00: how,
- 19:09SPEAKER_00: like,
- 19:10SPEAKER_00: admixture and selection work and all this stuff like that.
- 19:13SPEAKER_00: You know, like, talk about it.
- 19:14SPEAKER_00: And,
- 19:15SPEAKER_00: like,
- 19:15SPEAKER_00: outline,
- 19:15SPEAKER_00: you know,
- 19:16SPEAKER_00: period by period so that the listener,
- 19:18SPEAKER_00: because,
- 19:18SPEAKER_00: you know,
- 19:18SPEAKER_00: obviously...
- 19:19Hopefully they'll look at your Substack after they're done
- 19:21listening to this,
- 19:22SPEAKER_00: but obviously they can't see anything.
- 19:24SPEAKER_01: Yeah, I wrote quite a few Substack articles on this.
- 19:31SPEAKER_01: So one is on blonde hair,
- 19:33SPEAKER_01: the other one is skin color,
- 19:35SPEAKER_01: the other one is eye color,
- 19:38SPEAKER_01: and the other one is about four lineages,
- 19:42SPEAKER_01: four races of Europe.
- 19:45SPEAKER_01: And what I do is I check
- 19:49SPEAKER_01: The time trend.
- 19:50SPEAKER_01: So it's a time series model.
- 19:52SPEAKER_01: I checked.
- 19:53SPEAKER_01: My samples span from 40,000 years ago.
- 19:58SPEAKER_01: I think a couple of centuries ago.
- 20:02SPEAKER_01: And they come from Western Eurasia.
- 20:06I don't say Europe because there is also samples from Anatolia
- 20:13and Iran,
- 20:14you know.
- 20:16so Middle East and so what I find is what I do I compute based
- 20:26on the genomes I use
- 20:28that mixture which is a software an algorithm an unsupervised
- 20:35algorithm that
- 20:36SPEAKER_01: estimates the ancestry components in your sample so
- 20:44uh something um they did the ancestry components that uh keep
- 20:50popping out kept
- 20:51popping up in different samples where i like one was like a
- 20:58more common in the step
- 21:01SPEAKER_01: uh say um um like uh pontic caspian step people so