In 2012, Moti Shniberg sold his face recognition startup to Facebook and started looking for a new challenge. “We wanted to take our expertise and do something good,” he says.
Then he met the head of a medical genetics center, who explained the difficulty of diagnosing rare genetic disorders in children. Specialists sometimes use the shape and appearance of a child’s face as a clue because some conditions, such as Down syndrome, give a child’s face a distinctive appearance. For many other diseases, however, the signs are more subtle, and the cases very rare.
That got Shniberg’s mind racing. “Right away we knew we could really help,” he says.
In 2014, Shniberg’s new startup, FDNA, launched an app called Face2Gene. It was built around a machine learning algorithm like the one he previously used to recognize individuals. Only FDNA’s algorithm analyzes a face to suggest genetic disorders a person might have.
Face2Gene is now used by thousands of geneticists worldwide. Its core algorithm can recognize about 300 disorders with high accuracy from a patient’s face. That’s a boon to geneticists and families searching for a diagnosis—but the face algorithm still can’t see most genetic conditions. For the rarest, FDNA doesn’t have the seven or more photos from different patients needed to train its algorithm to detect the disorder.
“We’re able to now work on disorders that the system didn’t learn or wasn’t trained on.”
Peter Krawitz, chief scientific officer, FDNA
Last month, scientists from FDNA and several international institutions published results from a new algorithm called GestaltMatcher they say can distinguish about 1,000 conditions—a roughly threefold increase from FDNA’s original algorithm. It is now available in the Face2Gene app.
A tricky case from 2017 helped prove out the new approach. Two unrelated families in Norway and Germany each sought help from local doctors for a son with growth problems, tremors, and an unusually triangular face. For both boys, tests for known genetic conditions turned up blank. Doctors in each country independently used gene sequencing to discover each boy had a previously undescribed mutation in a gene called LEMD2.
The two teams got connected by a site called GeneMatcher that helps researchers find others puzzling over cases involving the same gene. The boys’ similar symptoms strongly suggested their shared rare mutation was to blame, but the researchers looked for additional evidence. They got it from a combination of conventional biology research and an experimental algorithm from researchers at the University of Bonn who collaborate with FDNA.
Lab studies suggested that the boys’ mutation had similar effects on their cells to progeria, a fatal genetic disorder whose patients also have distinctively triangular faces. It is caused by mutations in genes with similar functions to LEMD2.
The experimental algorithm, a prototype inspired by that powering Face2Gene, backed up those findings. It didn’t attempt to identify the specific disorder of a person in a photo. Instead, it calculated how similar a face was to those of other patients. It reported that the Norwegian and German boys had very similar faces, even though they had different ethnic backgrounds. Their faces were similar to children with progeria but distinct from it and other known disorders. “It helped to validate our impression that this is something new,” says Felix Marbach, a physician at the University of Heidelberg who worked on the project while at University of Cologne. The pan-European team of researchers published their discovery of the disorder in 2019. Identifying the gene didn’t open up new treatment options, Marbach says, but could lead to subsequent research that does.
The project showed that it was possible to use a face algorithm to help identify conditions where data was scarce, or missing altogether. “That was kind of the first time that it worked,” says Peter Krawitz, FDNA’s chief scientific officer and head of a genomics institute at University of Bonn, Germany. “We’re able to now work on disorders that the system didn’t learn or wasn’t trained on.”
