Rewriting Life

Does Your Genome Predict Your Face? Not Quite Yet

Genomics pioneer J. Craig Venter says if he had your genome he could pick you out of a crowd.

On Monday, the California gene-hunting company Human Longevity published a paper making the bold claim that it can identify individuals using their genomes to predict what their faces looks like.

The assertion—that your DNA can be used to create a photo-like reconstruction of you—has potentially big implications. It would allow police to pick suspects out of a lineup using a blood spot and it would mean no genome collected for research is truly private.

But a withering reaction to the face-prediction paper by scientists on social media is probably not what Human Longevity’s founder, the famed genomics expert J. Craig Venter, had in mind.

According to two experts who reviewed the paper—and one former employee—Venter can’t actually pick a person out of a crowd using a genome, and his report had difficulty finding a publisher.

“Craig Venter cannot predict faces,” Yaniv Erlich, the chief scientific officer of MyHeritage.com, a genealogy website, said bluntly on Twitter. To make the point, Erlich posted Venter’s prediction of his own face, noting that it looked more like actor Bradley Cooper than the biologist-turned-entrepreneur.

Venter’s team used genome data to predict face shape, eye and hair color, and even what your voice sounds like and melded them into images they said were accurate enough to re-identify whose genome it was.   

But skeptics say Human Longevity actually uses a person’s race and sex—easily measured from simpler DNA tests and not a new idea—to create pictures that portray average faces, not specific ones, as the company said.

Using its method, the company reported it could pick the right person from a lineup of 20 photographs about 70 percent of the time. But after discarding people of a different sex and race than the subject, accuracy dropped drastically. Used to pick a specific European man out of a lineup of 20 other European men, it was right 11 percent of the time.

“The face prediction is just predicting the average face for your race. You will always say, ‘Wow, that kind of looks like me,’” says Jason Piper, a genetics expert who worked at Human Longevity. Piper, who is listed as an author of the paper but is now employed by Apple, also took to Twitter to criticize its conclusions as unjustified. 

High stakes

The face study was meant to showboat the capabilities of Human Longevity, which has raised about $300 million to sequence one million human genomes. The business plan: create the largest DNA database on the planet and unleash the power of the genome to make health predictions.

The effort to privately amass gene information echoed Venter’s controversial role in the original Human Genome Project, when he raced public sector scientists to complete a first draft. His effort at that time to create a Bloomberg of DNA data, through his former company Celera, didn’t pan out.

With his new venture, Venter’s big name and ambition again helped him attract stars, such as machine-learning specialist Franz Och, who joined from Google to crunch data. At a “Health Nucleus” in San Diego, Venter started offering ultra-detailed medical workups costing as much as $25,000. Included are a genome readout and a personalized scientific poster depicting a customer's genes.

The face prediction project was undertaken to demonstrate the value of those offerings. After all, if Venter can print your face from your DNA, then Human Longevity’s medical workups must be worth it. “If I tell you that you are at risk of heart disease and you need to take statins, how can you be confident?” says Piper. “If [the company] can predict what the customer looks like, it’s a ground truth for confidence.”

Guessing traits

In genetics, traits like the color of your eyes or cholesterol levels are called phenotypes. These phenotypes are determined, to a greater or lesser degree, by your particular DNA, or genotype.

That’s why identifying someone's face from their DNA isn’t just theoretically possible, but likely to be possible a few years, says Mark Shriver, who works on genes-to-face prediction at the department of anthropology at Penn State. “I think it’s in our future for sure,” he says.

However, scientists still have work to do to know exactly how, say, your DNA influences the length of your nose or the width of your mouth, or the timbre of your voice. Shriver thinks Venter’s paper didn’t add much to what can already be predicted using existing DNA tests that measure your ethnic makeup and sex.

“Calling it predicting from the genome is what’s wrong,” says Shriver.  “The main message is way overstated. They just didn’t have enough people to find the genes that distinguish people. This is not the paper that is going to convince people that this is going to affect privacy or help forensics.”

Privacy question

Since it was started in 2013, Human Longevity has lost or fired several employees. Half the authors of the new paper, for instance, no longer work at the company. Och, the Google star, decamped after a short time for blood-testing startup GrailBio. Also gone are its former chief scientist and chief information officer.

A spokesperson for Human Longevity said such turnover was normal. In response to questions about its paper, the company on Tuesday released a statement noting that its data set, based on 1,061 people, was not extremely large and that “larger cohorts would be needed to make much more precise predictions and identifications.” 

In its statement, however, the company reasserted its contested claim: “If you have a genome from the public domain, researchers can generate a picture of that individual, thus identifying that person.”

Venter’s paper did not have an easy time finding a publisher. Shriver says he acted as a reviewer of the paper for Science, the leading American scientific journal, where it was twice reviewed and ultimately rejected. It was instead published on Monday in the Proceedings of the National Academy of Sciences.

In press coverage of his result, Venter also grated on some scientists by using face prediction to argue that genome data—particularly public data—holds a serious new privacy risk. Namely, if someone had access to your de-identified genome, they might be able to link it to you by generating a photograph.

“I've always said to people that your genome is more than all the other numbers in your life that identify you: your credit card number, your date of birth, your address, your Social Security number,” Venter told the San Diego Union-Tribune.

That privacy argument could be seen as self-serving, however, since Venter's point is that private databases, such as his, are more likely to offer needed protections.

Piper, the former Human Longevity employee, says he decided to speak up about the paper because he disagrees with Venter’s privacy argument. “Genomics privacy is important. But this is not how to approach it. It’s not that we can’t share. It’s how to share without fear,” says Piper. “The overall message of the paper is not something I agree with.”