Dyno sketches tens of thousands of viable next-gen AAV capsids with the help of machine learning — paper
Since its launch in May 2020, Dyno Therapeutics has touted its platform’s potential to discover viable variations of current-gen AAVs with noticeable limitations in terms of patient safety and efficacy. Now, according to a new study, Dyno has used its neural network to outline tens of thousands of variants that could add weight to its mission to build a better capsid.
In a study meant to determine how many viable variants of the AAV2 capsid it could design with the aid of machine learning, Dyno sketched out more than 100,000 viruses that could be used to carry gene therapies, according to a new paper published in Nature Biotechnology.
Using a neural network to design selectively mutated sites on a 28-amino acid chain, researchers at Dyno identified 110,689 viable variants of the AAV2 capsid — a success rate of more than half of all variations the biotech’s machine learning platform came up with as part of the experiment.
How did Dyno’s tech accomplish that feat with limited instructional data? According to co-founder Sam Sinai, Dyno’s team worked smarter and not harder with the data sets available, selectively inputting — and sometimes omitting — data to create a better prediction on less information.
“We looked at how different ways at looking at the same data — or even ignoring data that we had — can help certain machine learning models in their ability to model the space that we are trying to go into,” Sinai told Endpoints News. “This is a huge advantage, removing the burden of expensive experiments from the laboratory to the computer.”
The result was a rich variety of AAV variants that could offer the needed diversity to tackle Dyno’s signature challenge — working around patients’ natural immunity to specific AAV serotypes due to prior exposure. Paired with other research into how to design a capsid to better target specific tissues, Sinai argued his team is piecing together a potential road map for the future of AAV-based gene therapies.
The experiment has a side benefit, Sinai said, as one of the biggest experiments ever undertaken to dramatically rework the shape of a protein. The high rates of success in finding viable variations is just the cherry on top.
“The study itself is one of the largest designs of any proteins to date with machine learning in terms of bandwidth and in terms of how much change we have induced in this protein,” he said. “In that sense it’s very exciting. When we started this study in 2017, we didn’t know we could change the protein as much as we did. One of the exciting results of this study is that we could.”
With the high success in mapping AAV2, Sinai said Dyno is also focusing on additional serotypes, including AAV9 — the tech behind Novartis’ Zolgensma. The team’s computational power should work the same way across all wild serotypes, Sinai said, which could churn out millions of unique viable variants down the road.
So far, some big-name players in pharma are taking a bet on Dyno’s growing potential, with the Roche/Genentech group just recently pledging up to $1.8 billion to the team’s hunt for a better capsid. As part of the deal signed in October, Dyno went to work with the Spark team at Roche building better models of prototype AAV vectors and looking to overcome some of the barriers that have kept the therapy’s potential corralled to a limited set of organs.
The Roche deal came on the heels of similar deals with Sarepta and Novartis, both of which looked to harness the computational power behind Dyno’s designs.