Adeno-associated viruses (AAVs) are the DNA-delivery vehicle of choice for many gene-therapy companies, but they’re not perfect. Many past attempts to engineer improved AAVs have failed, because the protein shell of the virus—the capsid—is like a Rubik’s Cube. Improving one feature often throws others out of line.
A new study suggests that machine learning could help solve that molecular puzzle. And a new start-up, Dyno Therapeutics, has been quietly raising money to test the concept on an industrial scale.
A team led by George Church at Harvard Medical School created all possible single codon mutations—substitutions, deletions, and insertions—in the capsid of an AAV variant called AAV2. The team then tested how the mutations altered the AAVs’ immunogenicity, thermostability, ability to multiply in cells, and distribution to different tissues in mice.
With all this data in hand, the team set out to introduce multiple mutations in AAVs to improve their delivery to the liver. A collection of viruses designed with this method performed better than viruses with random mutations. But making multiple changes broke most of the AAVs and made them nonfunctional (Science 2019, DOI: 10.1126/science.aaw2900).
“What they’ve done here is truly a remarkable tour de force,” says Luk Vandenberghe, director of the Grousbeck Gene Therapy Center at Massachusetts Eye and Ear. The study highlights the potential of machine learning for AAV design, he adds, though it falls short of actually designing a significantly improved AAV that is ready for prime time in clinical testing.
Eric Kelsic, who as a postdoc worked on the project in Church’s lab, is Dyno’s CEO. He says the project gives “a snapshot of what’s important, and it’s a starting point for engineering capsids in a principled way.” Dyno is now applying machine learning to design better capsids.