This story is over 5 years old.

This Startup Says AI Can Predict the Effects of Gene Editing

Can a Canadian startup with some big names attached set itself apart in the competitive world of bioinformatics?
Image: Flickr/John Goode

Groundbreaking gene editing techniques like CRISPR/Cas9 have recently experienced a sharp uptick in interest from researchers, and their fair share of scrutiny, too. But in a world where bespoke genetic mutations are not unheard of, how can you be sure that a little DNA tweak will actually work?

Deep Genomics is a Canadian startup that wants to employ deep learning techniques—when a computer "teaches" itself how to recognize categories of things using a large set of training data, like a whole genome or a section—to predict with some accuracy what effect a given gene edit, or mutation, will have.


The company, which launched on Wednesday, was founded by Brendan Frey, head of the Probabilistic and Statistical Inference Group at the University of Toronto. Frey was also once a student of Geoffrey Hinton, the Canadian computer scientist now largely recognized as the founding father of modern deep learning research. Yann LeCun, Facebook's director of AI research and former postdoctoral research associate in Hinton's lab, is also on board as a company advisor.

"We can use our system to determine the efficacy of therapies, whether it's a drug, or a CRISPR/Cas-9 gene editing system, whatever it is, our technology allows us to predict the effects of those modifications," Frey said. "That's a very difficult thing to do computationally, most of the approaches are experimental."

However, Deep Genomics's participation in gene editing requires a world where gene editing is common—and we're certainly not there yet. Gene editing remains a controversial practice, and the US National Institutes of Health flatly stated it will not fund research on ethical grounds.

"I'm not getting the sense that everyone's running to use the software and implement it"

For now, Deep Genomics has set its sights on the slightly less futuristic word of genetic testing; for example, the system might come in handy when a mutation in the genes of a patient suffering from a particular disease doesn't exactly line up with what's in the literature. "Our technique can be used to filter and search over those mutations and rapidly provide information that really helps the diagnostician figure out what's going on," Frey said.


Deep Genomics is the commercial application of work that Frey and his colleagues have been doing for a number of years. A 2010 paper describing their approach made the cover of Nature in that year. Last year, he led a study that used the computational approach that Deep Genomics employs to identify genes that make their carriers more likely to develop conditions like autism and cancer. The results of that study were published in December in Science.

But it didn't catch on in the research world, Frey said, an outcome that disappointed him. The reason, Frey told the Globe and Mail, is likely that not enough research groups are equipped to deal with genomics and AI together like Deep Genomics does. Indeed, the 2015 Science paper has only been cited 22 times.

"These bioinformatic softwares, when they're used and they're successful, tend to be highly cited," said David Smith, a professor of microbial genome evolution at Western university, said. "I'm not getting the sense that everyone's running to use the software and implement it."

Watch more from Motherboard: The Singularity of Ray Kurzweil

"They seem very ambitious and wanting to get this stuff out there, and into the zeitgeist," Smith added. "And it makes you wonder if this is all smoke and no fire. I'm not sure it's that, but it's hard to tell. The true test will be if people cite the software. Time will tell, but it doesn't appear that it's getting cited that much more than other softwares that are appearing every day."

Deep Genomics will be entering a hugely promising and also highly competitive market in commercial genetics. 23andme, a controversial genetic testing company that recently made the move to drug research and development, is another company designing algorithms to "read" the human genome.

The system also has a key inherent limitation, Frey said. While gene mutations can potentially cause disease in many different ways, Deep Genomics' approach is limited to detecting only mutations that affect the process of splicing.

Deep Genomics will also make its database of genetic variants and their predicted effects on the splicing mechanism, SPIDEX, free for non-commercial purposes.

Whether Deep Genomics will push ahead of its competition remains to be seen, but with the gene-reading industry gaining steam, it seems more certain than ever that we'll need computers to help us understand what makes us human.