23andMe Pursues Health Research in the Shadow of the FDA

(CORRECTION 4/17/14: An earlier version of this article stated that 23andMe customers could only consent out of genetic health research by closing their accounts. This is incorrect. While customers' data can always be used for internal research related to product development, 23andMe's health research has a robust consent process on an opt-in basis.)

By Aaron Krol 

March 24, 2014 | There’s a magic number that Anne Wojcicki sometimes invokes when she’s out on the media circuit. As co-founder and CEO of 23andMe, and a leading evangelist for genomic medicine, it’s not unusual to see Wojcicki in a TV studio talking up the benefits of personal genetic testing. From WSJ Live to CBS This Morning, Wojcicki has used her magic number to persuade viewers that buying 23andMe’s Personal Genome Service (PGS) is an investment in a new vision for medicine.

The number is one million, and it’s explained in a post Wojcicki wrote for the 23andMe blog in December 2012, announcing a drop in the price of the PGS from $299 to $99:

One million customers can be the tipping point that moves medicine into the molecular era…  A genetic data resource of this magnitude has enormous potential to address unanswered questions related to the contributions of genes, the environment and your health… [This is] an ambitious plan that could transform medicine for generations to come.

You start to get a sense of Wojcicki’s rhetorical style. Still, the chain of thought is easy to follow: 23andMe wants to help you counter any dangers lurking in your genome by explaining what your DNA says about your future health. This service requires a lot of research into which genetic variants are involved in which diseases. 23andMe is happy to do that research itself, but it needs lots of data; in fact, it needs one million customers, with genotypes and personal health information, to reach a tipping point where discoveries will come faster and faster. In pursuit of that goal, the company will offer a steep discount on its service, just to get your data in the hands of its researchers.

23andMe thought this benchmark of one million customers could be reached by early 2014. And for eleven months the process hummed along smoothly, with record numbers of Americans shelling out $99 and a vial of saliva for a peek at their genes and a chance to help make Wojcicki’s healthcare transformation a reality.

Then, on November 22, 2013, the FDA sent a letter to 23andMe’s address in Mountain View, California, writing, “23andMe must immediately discontinue marketing the PGS.”

Fateful Choices 

In 2006, Joyce Tung was working as a post-doc at Stanford, researching the genetic basis of skin pigmentation, when she saw a flyer advertising jobs at a new startup called 23andMe. “I was on my way to do a Southern blot, which is a terrible, terrible experiment,” she remembers. “So I was probably particularly open to suggestion at the time. I thought, this is exactly what people ought to be doing with genetics.”

It was the year that personal genetic testing seemed to make it into the air. Falling costs of DNA analysis were fueling the discovery of single-point mutations linked to a wide range of diseases, including Alzheimer’s, celiac disease, and a variety of cancers. 23andMe was just one of several companies – Navigenics and deCODE genetics among them – who had decided the time was ripe to turn genetic health testing into a commercial service.

Tung started working as a contractor for the budding company, and became a full-time employee the following year. Along the way, she picked up the spirit of scientific populism that has helped brand 23andMe as a sort of rebel health reformer. “We were really excited about making genetics available to people,” she says of her early days at the company. “Science should just be more accessible to people, and I think we were all excited about taking something that had previously been locked up in stuffy old journals, and bringing it to the rest of the world.”

 

23andMe’s Mountain View headquarters in 2011. Reprinted with permission of the artist 

In 2007, 23andMe was not yet doing in-house research. Tung’s first priority as a member of the scientific staff was building the reports that customers would see: their ancestry information, a smattering of miscellaneous traits, and most importantly, their disease risks.

This is slippery territory, because our knowledge of the genetic contributors to disease is so incomplete. The company had to decide, with no clear precedent, which conditions to include in the PGS reports. This involved combing through the scientific literature for studies that reported gene-disease associations. “We have a pretty robust process for determining which studies we thought were the most valid,” says Tung. “That was a collaboration between the content team and the research team.”

One early decision in particular may have left a large regulatory footprint. When 23andMe’s health reports show a customer’s risk for a given disease, they don’t simply note an association between gene and phenotype as a journal article would. Instead, they specify a percent chance of getting that disease over a lifetime. To generate these risk figures, the scientific team built a set of algorithms that consider the prevalence of each disease, an estimate of how each genetic variant skews the odds, and some personal information supplied in the customer’s profile, most importantly ethnic background. (Tung, for instance, who learned from her own PGS that she is 100% Chinese, would have different background disease risks than her husband, who has Native American, African, and European ancestry.)

“Those algorithms were derived when we had just one science team,” says Tung. “We weren’t that big, and we didn’t divide the group into research and not research. We just had a group of scientists who worked together on those algorithms.”

The algorithms are significant, because one of the FDA’s biggest concerns with direct-to-consumer genetic testing has been users overreacting to genetic risks, especially if they get incorrect information. The psychological difference between seeing that you carry a genetic variant involved in heart disease, and seeing that 23andMe estimates a “30% chance” of developing heart disease, may be profound. In its November letter, the FDA references the PGS report on BRCA mutations, which indicate high risks of breast cancer, warning that a 23andMe customer with a false positive BRCA result might rush out to get unnecessary “prophylactic surgery” – i.e., a mastectomy.

If you think that sounds a little hysterical, you’re not alone. Misha Angrist, of the Duke Institute for Genome Sciences and Policy, was just one of many observers who considered this scenario “borderline absurd.” Ronald Bailey, author of Liberation Biology: The Moral and Scientific Case for the Biotech Revolution, wrote that “what the test results would actually lead patients to do is get another test and talk with their physicians.” But the FDA believes its job is to take this scenario seriously.

At issue are the twin questions of analytical and clinical validation: whether 23andMe’s microarrays correctly genotype its customers, and whether its interpretations of those genotypes – the risk percentages – are well-supported by evidence and clearly explained. The former is fairly easy to prove; the latter, much more difficult.

The Other Shoe 

To understand the stakes here, we have to get a bit into the weeds of the FDA’s medical device classifications.

By 2013, 23andMe was the only company still offering customers health information based on their genotypes, and for four years, it had been wrangling with the FDA over something called a 510(k). This is a petition to treat a new product, like 23andMe’s PGS, as a class II medical device, rather than a more strictly regulated class III device, the default for novel products. A 510(k) is a popular route for new in vitro tests, which rarely pose the “potential unreasonable risk of illness or injury” that defines class III.

23andMe’s PGS is essentially an in vitro test – spit sample in, answers out – so this would seem like the company’s best bet for clearance. Unlike class III devices, which go through the arduous process of premarket approval, tests in class II only need to meet some mutually agreed-upon standards for labeling and quality controls, to assure the FDA that the test is accurate and understood by users. In short, analytical validation is crucial for both classes of test, but in class III, clinical validation is much more prominent.

That was the conversation the two sides were having between July of 2009 and sometime in early 2013. The FDA letter recounts “more than 14 face-to-face and teleconference meetings, hundreds of email exchanges, and dozens of written communications” in this time, through which the agency considered marketing pathways, what disclaimers the PGS should carry, and whether at least certain disease risks reported in the PGS could appropriately fall under class II regulations.

Sometime in early 2013, however, it appears that 23andMe decided to blow off the 510(k) process – perhaps gambling that it could get away with not being classed as a medical device at all. In late 2012, the FDA sent the company requests for additional information, which were never addressed. In March 2013, the FDA informed 23andMe that its 510(k) was considered withdrawn. Finally, in May, 23andMe stopped replying to the FDA’s communications. This has been the most shocking revelation of the FDA letter, what Matthew Herper of Forbes called “the single dumbest regulatory strategy I have seen in 13 years of covering the Food and Drug Administration.”

Seen in this light, it’s almost surprising that the FDA waited until November to declare that the PGS fell by default under class III, did not have premarket approval, and could not be sold.

After 14 days of deliberation and intense media attention, 23andMe reached a painful compromise: it could continue selling the PGS, but all health information was stripped from the reports. Only the raw genotypes and ancestry information remained. This has been a serious blow to the company’s ability to attract customers, as Wojcicki acknowledged in an appearance at the SXSW Festival in Austin last week, saying that the FDA action “has slowed up the number of people signing up.”

She also revealed how many customers have bought into the PGS so far: 650,000. That’s a lot of genotypes, but it’s not the magic one million.

Before the FDA action, it didn’t much matter whether one million customers was really the tipping point for research in genetic medicine. It was useful buzz, a way to inspire people on the fence about personal genetics to join in. Now, with signups stalled short of the finish line, one has to ask: will two-thirds of a million do?

“We Are Looking to Impact People’s Health” 

Tung, who comes across as mellow but outgoing and has a habit of saying gosh or shoot when she’s thinking over a question, is brimming with enthusiasm for her work. Today, she’s 23andMe’s Director of Research, coordinating the projects that drill through customers’ genotypes for new findings on heredity. When Wojcicki talks about the discovery power of one million customers, Tung’s work is what she wants you to think of.

The basic pattern of these projects is a constant. 23andMe creates customer surveys asking about traits that the research team believes may have a genetic basis. The company casts a wide net, exploring everything from Parkinson’s disease to motion sickness, from risk factors for diabetes to risk factors for smelling asparagus in pee. Tung estimates that there are around 50 of these surveys now in circulation, along with hundreds of one-off questions.

 

The cycle of personal genetics, as shown in an early 23andMe blog post. Image credit: 23andMe 

The company also stores troves of genetic data on each of its customers, the fruits of its genotyping service. In its latest version, the custom 23andMe microarray chip captures just over 600,000 single nucleotide polymorphisms (SNPs) from each user’s genome. (That’s actually a substantial decrease from the previous chip, but 23andMe is convinced it has zeroed in on more informative SNPs now, and that many of the SNPs it eliminated can be inferred from the data it has already.) Compare the SNPs to the survey answers, and trends may begin to emerge.

So which traits is 23andMe most interested in? Cancer or back hair? Given the FDA’s crackdown on health results, I half expect Tung to waffle a bit on this question, but she doesn’t hesitate.

“We are looking to make discoveries that impact people’s health,” she says firmly. “So we do care a lot about the more medically related conditions. And we also care about people’s response to different treatments.”

Treatment responses are an area 23andMe is well-positioned to study, thanks to its information on customers’ medication histories. “At this point, we have some customers who have been with us for over six years,” Tung points out. “It’s useful to look at [their drug responses] over time. It’s great to know when they started taking the medication, and if and when they started to see results or side effects.”

This kind of longitudinal data is considered the gold standard for research in genetic health. The ease with which Tung’s team can collect this information is a major advantage of working with an involved user community like 23andMe’s customer base. Where other studies would have to go to great lengths to assemble and genotype a cohort of medication users, 23andMe just polls its customers.

When I ask Tung if she has any favorite findings, she brings up a paper her team published in Nature Genetics last October, describing regions of the genome associated with allergies.

“That was pretty cool,” she says. “It’s exciting to make discoveries in a field where there’s not a ton of, say, government funding, because allergies don’t kill the number of people that cancer or cardiovascular disease do. But they really impact people’s lives, and a ton of people have them. So trying to understand how people develop allergies, what makes them susceptible, and how we might be able to treat them, is something I really appreciate.”

It’s a well-chosen example, demonstrating a real unmet need that 23andMe can fill, by studying conditions of pressing concern to ordinary people but less interest to traditional funding bodies. The paper also appeared in a top-tier journal, which, despite Tung’s earlier prod at “stuffy old journals,” offers 23andMe some much-needed vindication of its methods. Survey results are not always accepted as reliable measures of clinical phenotypes, although Tung insists that “as long as you’re asking the right questions, they’re as good as a medical record.”

Now that 23andMe is barred from reporting its discoveries to customers directly, publishing in scientific journals may be more important than ever. “We’ve always published, and I think we intend to continue,” says Tung. “Our goal is to advance biomedical research, and part of that is sharing our findings with the scientific community.”

 “Certain kinds of findings are [also] made available via our blog, to our customers and other fans,” she adds. That raises a perplexing question about the FDA action against 23andMe. The company can still post its findings in public forums – in fact, there’s probably no legal action the FDA could take to stop this if it wanted to. And the agency has signaled it’s not interested in blocking companies from offering people their raw genetic data. So 23andMe users could theoretically just keep abreast of new findings from the company, and check their own genotypes against any variants flagged in the research.

Most won’t bother, of course. It’s a huge hassle to delve through the relevant literature, and search your genotype for variants one by one. But it’s peculiar to think that our current national consensus on personal genetics is that you can have your genotype – or even your whole genome, if you can find someone to sequence it – and you can have all the information you want on genetic disease risks, but you can’t have both in the same place.

Research and Commerce 

It’s not clear how much the research team will have to adapt its methods to the new regulatory realities. On the one hand, it will certainly have a harder time collecting new data. And there’s plenty of opportunity for the content and marketing teams to influence the course of research, if they think it could help draw customers in. “The company’s not that big,” says Tung. “We all eat lunch together every day, so we keep the lines of communication open. “

On the other hand, the company’s research philosophy hasn’t been forgotten. Asked if 23andMe will now focus more on the ancestry information it can legally share with customers, Tung says, “The mission of the company remains on health. Obviously we’re interested in ancestry as well, but it’s still our goal to advance biomedical research. I don’t think that’s changed.”

As the prospects of reaching one million customers start to flag, more than 23andMe’s original research is at stake. Observers have often suggested that 23andMe’s $99 kits, which are notably less expensive than similar genotyping services, are a loss leader for its real business of selling customers’ anonymized data to commercial partners. Wojcicki has said in an interview for Bloomberg TV that 23andMe makes “a small margin” on each kit it sells, but it seems unlikely that these revenues are enough to support the company’s operating expenses.

Meanwhile, 23andMe has already partnered with pharma company Genentech on a study of the cancer drug Avastin, to find users who achieve long-term remission and discover if any of the SNPs in 23andMe’s database are responsible. Academic centers and pharmaceutical companies would be keenly interested in accessing a database of a million genotypes, complete with survey responses that stand as good proxies for clinical data. 23andMe’s financial solvency may rest in large part on whether it has amassed enough genotypes to attract this kind of commercial interest.

23andMe doesn’t keep these partnerships a secret, and valuable findings might well come from them. The company also stores customers’ personal information – names, contact and payment information – separately from its other data, so any information it shares with partners is anonymous. But knowing that their own genotypes, and survey responses, are 23andMe’s real product could easily be a turnoff for customers, and the company certainly doesn’t go out of its way to advertise this aspect of the business – as a glance over its Research Consent Document will attest. 

To be fair, 23andMe customers as a whole are eager to participate in research, and genuinely want their data used for novel medical discoveries. Tung is very enthusiastic about how customers respond to the service. “Getting more information is a relatively lightweight thing for us, because people frequently log back into their accounts,” she says. She also emphasizes that the interaction between 23andMe and its customers is a two-way street. Research participants can suggest new questions or changes to the surveys, and Tung’s team tries to join the right research to the right customers. If she wants to study childbearing, for instance, she’ll target surveys to customers of childbearing age.

“We really value our customers’ time,” she adds. “The more we can bubble up to the top the things that are most relevant and interesting to them, the better.”

But customers may not be aware that 23andMe is not the only group conducting research on their genotypes. Customers concerned about commercial partners using their data do not have to consent to research participation, but those who choose this option must withdraw their data from internal 23andMe research as well.

A Regulatory Path 

My last question for Tung is whether she thinks the quality of research coming from her department could influence the FDA’s future decisions about direct-to-consumer testing.

“It might,” she says, after a moment. “It might.”

Given her employer’s reluctance to speak publicly about why the FDA came down against the PGS health reports, and what it might take to bring them back, I hadn’t hoped for a much more forthcoming answer. Approached for comment on the FDA decision, 23andMe offered me this statement: “Our goal is to work cooperatively with the FDA to provide consumers with access to their genetic data in a way that clearly demonstrates the benefit of having that information and the science that underlies the test.”  

Like 23andMe’s other public responses, it’s both cowed and defiant. Clearly the FDA has never been worried that 23andMe would fail to communicate the benefits of personal genetics. At the same time, the statement does seem to acknowledge, between the lines, that customers could misunderstand the seeming authority of their risk reports.

Now that the FDA’s jurisdiction over direct-to-consumer genetics has been settled, the question is what will satisfy the agency of a test’s reliability. If the FDA is ready to resume the 510(k) process, we may find the dispute has been largely a matter of packaging. A different way of communicating genetic risk, some sharp disclaimers about the tentative nature of genomics, and some assurance that the 23andMe SNP chip is analytically valid, might start to clear up concerns that the PGS steps over the line into unvalidated diagnoses.

Then again, if the agency holds firm on labeling the PGS a class III device, it’s difficult to see a way forward. 23andMe would be hard-pressed to deliver rigorous clinical studies to prove the accuracy of its health calls, or that customers won’t be harmed by their results – both necessary conditions for premarket approval.

Either way, Tung’s department is taking a stance of business as usual. Although new genotypes – and possibly even new survey responses – have slowed to a trickle, there’s already plenty of material to mine. It’s also clear that the focus remains squarely on health, regardless of whether customers are allowed to see the findings or not. Before the FDA action, 23andMe made concerted pushes to bring in customers with Parkinson’s disease, sarcomas, and blood cancers, and research projects on these diseases are still underway.

The Parkinson’s project in particular is an impressive effort. 23andMe has genotyped more than 10,000 people with the disease, and published new genetic risk factors. As large genetic databases like this are leveraged toward health discoveries, the company’s methods gain new support, and a health environment that takes into account genetic risk looks increasingly inevitable. It’s far from proof that 23andMe can predict your future health from a SNP chip, but it does make the attempt seem more worthwhile.

As to whether that will affect the FDA’s thinking… It might. It might.