Novartis and Sarepta tap Dyno Therapeutics to use machine learning for gene therapy

Novartis and Sarepta Therapeutics have struck partnerships with Dyno Therapeutics, a just-launched spin-off from the lab of geneticist George Church at Harvard Medical School. Dyno will use machine learning to design custom adeno-associated viruses (AAVs) for the two drug companies to use in their experimental gene therapies. Dyno could earn more than $2 billion from the deals.

AAVs are the linchpin of modern gene therapies. Biotech companies use them as molecular delivery vehicles to shuttle therapeutic DNA into our cells. More than 100 clinical trials of AAV gene therapies are underway, and far more preclinical AAV gene therapies are in development. But existing AAVs, mostly derived from wild viruses, are imperfect. They require high doses, are expensive to make, and are poor at reaching certain organs of the body. Several groups, including Cambridge, Massachusetts–based Dyno, are hoping to change that

As a postdoc in Church’s lab, Dyno CEO Eric Kelsic developed a way to synthesize thousands of slightly altered AAVs, test them in animals, and use machine learning to assess which changes improved their properties. When Kelsic and Church presented the project for the first time at an American Society of Gene & Cell Therapy meeting in spring 2018, it quickly garnered interest from drug companies.

That June, Kelsic met with Alan Crane, a partner at the venture capital firm Polaris Partners. “Eric came to me with this spreadsheet of dozens of companies that had proactively approached him,” Crane recalls. “I have never seen such robust interest from the pharmaceutical industry in a technology.”

Church, Crane, Kelsic and three other scientists—two from Harvard and one from Lund University—cofounded Dyno in November 2018. The firm was backed with $9 million in seed financing from Polaris and the venture capital firm CRV.

This week, Dyno formally launched from stealth mode—start-up slang for a self-imposed quiet period—and announced its partnerships with Novartis and Sarepta. The deal with Novartis focuses on gene therapies for eye diseases, while the one with Sarepta is centered on muscle diseases. In both cases, Dyno will focus on making new AAVs. Novartis and Sarepta will conduct all preclinical and clinical studies on their own.

Dyno could earn $40 million from the research phase of the Sarepta collaboration. Additional financial details were not disclosed. Dyno’s partnering strategy could be so lucrative, Crane contends, that the firm may never need the large rounds of financing typical for venture-backed biotech firms. “We have been able to attract deals that really fund the company in a significant way,” he says.

The company is centered on gene therapy’s delivery challenge: getting enough copies of a therapeutic gene to its final destination, such as particular neurons of the brain, the photoreceptors of the eye, or muscles throughout the body. “If we can solve this problem, it could have a huge impact on many disease areas,” Kelsic says.

A paper published in November 2019 highlighting work from Kelsic’s postdoc offers a glimpse into Dyno’s strategy. Kelsic and his colleagues systematically substituted, replaced, or deleted codons from the AAV genome that are responsible for making the outer shell of the virus, known as the capsid. The capsid determines the AAV’s efficiency, a measure of how well it delivers a gene into cells, as well as its specificity, the profile of cell types and organs it has an affinity for.

The researchers tested how the thousands of resulting AAV variants performed in cells and mice. They then used machine learning to analyze the data and predict how changes to the AAV capsid could be combined to make a better AAV. In many cases, compiling these changes broke the AAVs and rendered them useless. But in some instances, the modifications improved the AAVs (Science 2019, DOI: 10.1126/science.aaw2900).

“I have never seen such a compelling application of AI to biology,” Crane says.

Today, Dyno is expanding this experiment on an industrial scale. In theory, the company’s machine-learning approach should allow it to discover better AAVs with each round of experiments.

In addition to optimizing the efficiency and specificity of AAVs, Dyno is working on three other AAV traits, Kelsic says. One is the packaging size—the amount of space available for the therapeutic gene. Another is ease of manufacturing, with the goal of making AAVs that are simpler to grow in large bioreactors.

The third is the elimination of existing immunity. Some people are ineligible to receive AAV-based gene therapies since they’ve already been exposed to AAVs in nature, and their immune systems will attack the gene therapy. Changing the capsid surface could create AAVs that more people can safely receive.

Importantly, Dyno’s machine-learning process allows it to optimize multiple properties at the same time, Crane says. In the start-up’s early collaborations, it will create custom AAVs tuned to meet a partner’s wish list. But Dyno will likely discover other promising AAVs along the way, which should make future optimization projects easier and could even lead to the creation of a library of novel AAVs that Dyno could pull off its shelves and quickly license.

In fact, Dyno has already begun optimizing AAVs for central nervous system and liver diseases—which are not part of the Novartis or Sarepta deals. Dyno is making these new AAVs with the intention of eventually sharing them in other partnerships. Although Dyno is not building its own therapeutic pipeline at this time, the start-up is giving itself breathing room to potentially do so in the future.

“Our philosophy is to structure these deals so that we are not giving away exclusivity in a therapeutic area,” Crane says.

Other start-ups are designing new AAVs as well. In March, Affinia Therapeutics launched with $60 million to make new AAVs, and in April it announced a partnership with Vertex Pharmaceuticals worth up to $1.6 billion.

“This is not the small-molecule or antibody world, where the key value is the drug, and delivery just adds a little bit on top,” Crane says. “In this case, it is often the gene that is being delivered that is generic, and it is the way that you deliver it that is critical.”