Response to Euny Hong’s critique of 23andMe

Screenshot 2016-08-26 19.23.09

Update: In light of further comments I may have been wrong about Hong’s recent admixture! See the comments below (also, further discussion with Spencer Wells offline). I don’t have total clarity on what’s going on, because I’m sure my friends weren’t lying…but they were also early adopters, and the methods may have changed. And, I do think 23andMe has the talent and methods to resolve Korean ancestry, so it’s a matter of investment, not data.

All that being said, all individuals should pull down the raw data and do a reanalaysis.

End update

Quartz has an article up, 23andMe has a problem when it comes to ancestry reports for people of color, which I want to comment on at length. Though literally taken the title is not something I’d disagree with too much, the tone and details I have serious issues with.

First, some disclosure. Hong talked to me on the phone for an hour about this story. Mostly we talked about her Korean ancestry results. More on that later. Second, I consulted for 2.5 years for Family Tree DNA, am friends with Spencer Wells (who is quoted), and am on friendly terms (I’d like to think!) with Joanna Mountain, and quite respect many of the scientists at 23andMe (e.g., Kaisa Bryc and Ivan Juric off the top of my head).

I will go through the article point by point. First:

I doubt that most 23andMe users realize how paltry the company’s data is for non-Caucasians. For example: The data set that 23andMe used to generate my report has 76 Koreans in it, according to Dr. Joanna Mountain, the company’s senior director of research. 76 Koreans. It is estimated there are at least 7 million Koreans living outside of the Korean peninsula—including 1.7 million in the US—among a worldwide population of 83 million.

Seventy-six Koreans seemed small to me, but what do I know? I’m just a journalist. So I spoke to geneticist Spencer Wells, founder and former director of National Geographic’s Genographic Project (arguably a 23andMe competitor), which he ran from 2005-2015. “[76] is a really low number,” he concurred.

The small sample sizes seem really, really problematic if you are a lay person, or a journalist. The issue is that with genotype technology that looks for common polymorphisms you really don’t get that much more information from 1,000 individuals than you do from 100. All things equal, more sample size is better, but the gap between 10 and 100 is much much greater than 100 and 1,000 or 100 and 10,000. You can see this in the robustness of results for model-based clustering conditional on different sample sizes. For a homogeneous population like the peoples of the Korean peninsula, who seem relatively panmictic, a bigger sample size would have only marginal effect on the overall outcomes using these methods (also, it might matter if you were looking at low-frequency alleles from whole genome sequencing).

Before I talked to Hong I checked in with a friend who was half north Korean (in that her father’s family was from the northern half of the peninsula and migrated south) and half central Korean (i.e., her mother’s family was from around Seoul). Just like her husband, whose family was from Busan in the far south, her results came back as 99% Korean. Some genetic research has been done on Koreans, and there just isn’t that much structure. The Koreans have a composite origin if you go far back enough, but they’ve been intermarrying with each other a long time.

Next:

Also, astonishingly, the report shows that I am 13.4% Japanese and 14% Chinese—and only 61.6% Korean. I was looking forward to watching my parents freak out. My sister texted me, “Oh [Dad will] probably blame Mom.”

To my disappointment, my parents did not freak out, nor did they get into an amusing argument about which of their ancestors was the ho. Because they simply did not believe the data. And, for once, they were right.

The public relies on journalists for the truth. Sometimes the truth can be slippery. But sometimes it is clear. Most of conversation between Hong and myself was about her Korean ancestry. As I said to her, I asked a handful of my Korean friends about their 23andMe results before we spoke. From that I told Hong I was 99% sure that she had recent non-Korean ancestry. 23andMe’s results are really robust. I tried to emphasize that over and over. Hong can believe what she wants, but it is obvious that she almost certainly has non-Korean ancestry relatively recently in the past.

Because 23andMe uses chromosome painting, you can see she has very long segments of inferred Chinese and Japanese ancestry. This non-Korean ancestry is probably from within the last three generations because ancestry tract lengths indicate that recombination hasn’t broken apart the associations across the chromosomes (there are 20-40 recombination events across the genome per generation).

Next:

I asked Wells whether my percentage breakdowns of Korean, Chinese, and Japanese meant anything. “Yes,” he said, “but I think it is misleading to go to a decimal place or even to go out two digits.” Wells said that another problem with the data is that “Most of those [samples] are from the US. They’re not terribly useful for studies of indigenous composition—which is effectively what this analysis is trying to do.”

I had a long text conversation with Spencer on this after the article came out. I can see where he’s coming from. And 23andMe does have a shortfall of indigenous and non-European samples. But as I said, I asked around to Korean friends who had used 23andMe before and the population is pretty homogeneous, and the friends’ results I cited above were representative. I have also worked with and seen samples from Family Tree DNA, and it’s the same story. There might be undersampled populations from Korea, but I’d bet against it. Koreans are relatively homogeneous, with a position between Japanese and North Chinese. Where you would expect them to be.

Spencer is correct about the decimal places issue. They give people a false impression of precision. I do know that scientists within DTC companies struggle against it. But scientists don’t always win these arguments.

Next:

I also interviewed Harvard geneticist Robert Green, who made the important point that private companies have different methods and standards from those of an academic lab. “There is a difference between analysis you can do with hundreds of [genetic] markers at a research level, and the kind of analysis that even the best companies can do, which is more an approximation,” he said.

Green is a medical geneticist who does great work. But I’ll be generous and assume he’s taken totally out of context here, because what he says makes no sense. The genotyping platforms do have error rates (no-calls, mistypings, etc.) on the order of 1%. But they’re using hundreds of thousands of SNPs. This error rate doesn’t matter too much for what 23andMe is doing in relation to ancestry. And with population structure inference these errors usually don’t cause a major issue if they aren’t systematic.

Then there’s this:

A few of the geneticists I interviewed for this article (but not Green or Wells) outright accused 23andMe of commercially driven ethnic bias. For example, no distinction is made between northern and southern Chinese, who have very different traits. This was a serious allegation, so I put the question directly before 23andMe’s Mountain. “As a scientist, I find that insulting,” she said in a phone interview.

I brought up the issue with the Chinese to Hong, and I apologize to Mountain here if it came off as offensive, because I certainly didn’t mean it that way. My point, which I’ve brought up for years both in public, and when I have consulted for DTC companies, is that South and East Asians are huge groups, and it’s incongruous that they aren’t differentiated as much as the Europeans. These tests basically tell you are South Asian, or Chinese, or Korean, or Japanese. In the case of Koreans and Japanese there isn’t that much structure within these groups, but that is not the case with the Han Chinese. There is an decent amount of structure, but last I checked 23andMe has a catchall Han Chinese group. Why? I’ll get to that later. (It’s not because they don’t have the data.)

Though I disagree with the tone and the emphasis, a simple inspection by Hong has shed light on something that has been glaringly obvious in the genetic genealogy community: there is laser-like focus on differentiating very close Northern European groups, such as Irish and English, and not so much emphasis on differentiating diverse populations such as South Asians. This was one thing I did talk to Hong about at length. I don’t think it’s crass racism, and I think that I made that clear to her, but I’m not happy with the situation either (23andMe representatives know I’m not happy, and have talked to me about it at ASHG).

The final sections involve Hong reviewing the disparities in sample representation. As I said above, some of this overdone. But, it is a little ridiculous that there are only a few hundred African population samples in their data. Granted, it turns out that between-population genetic distance in Africa is actually not as much as you’d think based on aggregate variation (the within population variation is what makes all the news). I think Hong is correct that 23andMe should have made more effort on sample collection these past few years…but I’m not CEO of 23andMe, and Joanna Mountain and her scientists don’t call all the shots. I think Hong’s piece leaves Mountain and the researchers holding the bag for something that really isn’t their doing (perhaps it is, but I’m really skeptical of that).

Finally:

Could the company be doing a better job with collecting ethnographic data? “Absolutely they could,” Wells said, “but it’s not their raison d’être.” Which, of course, is pharma and health research. Fair enough—it’s their money. But how about a disclaimer attached to the ancestry part of the report? Like, “for entertainment purposes only?” Because data based on 76 Koreans (or any other ethnic group) is definitely not worth potentially causing family discord or a blood feud. I don’t know whether the company understands the realities of deadly global ethnic tensions and the potential damage created by people’s trust in these reports.

I think Spencer has highlighted the major dynamic here: 23andMe is pivoting towards biomedical research. It has a database of north of a million, mostly European-origin individuals. The real money now comes from leveraging the database to collect information on health, and combining it with the genotypes they already have. On the margin, getting greater population diversity is probably not a major avenue by which they could gain higher valuations. And getting from one million to ten million genotypes is nothing without increasing their database of phenotypes.

The real story here is not one of racism. It’s one of capitalism. Most of 23andMe’s customers are white European in ancestry, and a disproportionate number of those are Northern European. Is it a surprise that their tools breakdown Northern European ancestry so finely? That’s their customer base.

Second, many Asians I’ve talked to are relatively uninterested in fine-grained breakdowns in their ancestry. For several years I worked with an engineer from Fujian, and his Family Tree DNA results showed that he was shifted toward the southern end of the north-south Chinese cline. He didn’t care at all, because he was from Fujian, so of course he knew this. Many Asians seem to have this attitude where the ancestry results are viewed as confirmatory. Hong’s case, where there was a surprise, is exceptional.

If 23andMe wanted to they could easily breakdown Asians into further subcomponents. I think there are two reasons they don’t want to aside from the firm’s recent focus on health and pharma. First, they don’t have that many Asian customers. Second, their Asian customers might actually get a bit irritated!

Ultimately, Hong can think whatever she wants to about her 23andMe results. But the data are out there. It’s pretty obvious that unless there was a sample mix-up, she has recent Chinese and Japanese ancestry (she could put the raw results in the public domain and have people cross-check with other methods, like PCA; I’m pretty sure they would confirm the 23andMe results).

On a last nerdy note: the data generated by DTC companies is great. Their Illumina SNP-chips are really good, with 99% or so correct-call rates. Hong referred to data in the piece when she really meant results. The thing is that results are basically generated through a sieve of methods geared toward human digestibility. 23andMe and other DTC companies differ because of different methods and parameters in those methods, that are determined by what humans want out of these techniques. But the data, that’s pretty straightforward and robust.

If you are interested in a more philosophical take, Joe Pickrell’s What is ancestry?

Addendum: My conversation with Hong was very wide-ranging. We talked about EDAR, random mating populations, and local ancestry deconvolution. Well, perhaps not in those words. It’s a little saddening to me that ultimately what came out of all that is a piece which tries to paint 23andMe as prejudiced against minorities. The only prejudice they exhibit as a firm is against smaller market share.

Posted in Uncategorized

Comments are closed.