December 1, 2015 (Vol. 35, No. 21)

Bioinformatics Engine Mines Gene/Disease Data, Uncovers Causal Associations

A computer system developed by Deep Genomics can mimic how cells read DNA to sustain life. It can also show what happens within cells when DNA is altered. These capabilities, extended and enhanced by means of machine learning, allow the system to do more than suggest correlations between genetic variations and specific diseases. The system can establish causal associations.

Deep Genomics uses a form of machine learning called deep learning, a relatively new form of informatics. Deep learning, the company insists, offers a scalable, flexible, and unified computational approach for pattern discovery. The company has reported that deep learning can outperform other modeling approaches.

For example, the company indicated that it used deep-learning algorithms to derive a computational model that takes as input DNA sequences and applies general rules to predict splicing in human tissues. The model led to novel insights into the genetics of autism, cancers, and spinal muscular atrophy.

“A lot of companies talk about big data, says Brendan Frey, Ph.D., CEO and lead scientist at Deep Genomics. “Our system is different because it is based on accurate models of how biology works.”

Dr. Frey notes that there are two approaches in medical informatics. One is based on genome-wide association studies, which ignore biological complexity. The other approach tries to account for everything that occurs in the cell. It preserves context but risks becoming mired in complexity.

“Our technology combines the best of those two approaches,” Dr. Frey asserts. “We use machine learning to look for relationships throughout the many layers of biological processes whose disruption leads to disease.” This deep-learning approach is discovering both new and well-known disease mechanisms and is accurately predicting their consequences in patients. “We provide a rich, complete mapping that not only identifies mutations but shows how those mutations cause problems by changing what’s going on within cells,” Dr. Frey continues.

Other methods provide a simple score for each mutation. In contrast, the Deep Genomics system provides a cascade of consequences in a way that depends on the genetic context.

“Let’s say that two people have the same mutation,” Dr. Frey suggests. “It may be pathological in one context, and benign in the other.” In a situation such as this, the Deep Genomics system can present the ramifications of each mutation—perhaps a specific protein binding site is fated to be destroyed. “Our engine takes into account the patient’s genetic context and explains why the mutation is a problem.”

Deep Genomics is working on an integrated computational system designed to learn, predict, and interpret how genetic variation, whether naturally occurring or therapeutically induced, alters crucial cellular processes. [iStock/alanphillips]

Initial Focus: Splicing Variants

Deep Genomics’ first product, Spidex™, accounts for how genetic variants alter splicing. Research during the past decade has .shown that human cells undergo more splicing than cells of other species, and that the brain has more splicing than other organs. “Splicing is involved in cancers, cystic fibrosis, spinal muscular atrophy, autism, and other serious conditions,” notes Dr. Frey. “So we chose it as the first component to build in our Deep Genomics engine.” Spidex, he insists, has the power to transform medicine.

“When we compare our engine to the most popular machine-learning system on the market today, our engine finds 64 times more pathogenic variants than the other system,” Dr. Frey informs. “This kind of sensitivity is important for pharmaceutical companies. Recent results show that if they screen drug targets using genetic information, they will have much higher success rates and can treat patients more effectively.”

Deep Genomics has six engine modules in development and a plan for more than a dozen. “They can be used independently,” notes Dr. Frey. “But their greatest value is when they are integrated to form a comprehensive informatics engine.”

Genesis of the Idea

Deep Genomics was spun out of Dr. Frey’s research at the University of Toronto. “In the 1990s, I was part of a group of machine-learning researchers focused on what later became known as deep-learning techniques,” Dr. Frey recalls. “Members of the group included Geoff Hinton, Yann LeCun, and Yoshua Bengio. This area of artificial intelligence trains computers to understand data in multiple, complex layers.” Deep learning is now used by Google and Facebook, as well as biotechnology startup companies such as AtomWise.

A medical issue within his own family caused Dr. Frey to focus his research on the human genome and to discover deep representations that explain how the genome works. “If you want to understand how mutations cause disease, the computer needs to be able to tell you what goes on inside cells,” he says. “We published papers in Nature, Science, and Cell. This was a transformative technology, but it wasn’t catching on. So, we formed a company last year to bring the technology to the clinic and the R&D lab.”


Deep Genomics currently is raising seed capital to validate its technology clinically, to triple the size of its R&D team, and to develop additional components of its informatics engine.

Dr. Frey is talking with venture capital and angel groups to raise about $3 million in additional funding. But, he stresses, this is about more than raising money: “It’s important to us to have partners, clients, and investors that share our goals.

“Our focus is primarily upon validating the technology clinically. We want to transform different aspects of medicine, so we’re very interested in working with partners who will provide feedback and work closely with us as we develop subsequent components. We have partners to help with clinical validation. Trials are still being designed.”

Because the engine is broadly applicable both to diagnostics and drug discovery, Deep Genomics is working with diagnostics companies and hospitals to use this engine to find mutations that don’t merely correlate to disease, but that actually cause disease. Importantly, these mutations needn’t be in the literature to be identified, “so our system provides unusual value,” adds Dr. Frey.

Unlike drug developers, Deep Genomics’ analytics engine doesn’t need regulatory approval. “Our value is in providing explanations,” concludes Dr. Frey. “We can generate new information for pharmaceutical companies that lets them make connections they couldn’t make otherwise. We enrich their network of information.”   

Deep Genomics

Location: MaRS Discovery District, 101 College Street, Suite 320, Toronto, Ontario, Canada M5G 1L7
Phone: (647) 255-1339
Principal: Brendan Frey, Ph.D., CEO and Lead Scientist
Number of Employees: 8
Focus: Deep Genomics is developing a computer technology that identifies how specific genetic variations change cellular behaviors and provides causal information for pharmaceutical development, precision medicine, and diagnostics.
Previous articleHuman Longevity Acquires Cypher Genomics
Next articleAlmirall Expands in Dermatology, Buying Poli Group for $386M+