Today, a big project from my PhD was published
. It was the result of a multi-year collaboration between Harvard, Google AI, and Dyno Therapeutics (the gene therapy company I work at). In short, we used machine learning to re-design the capsid protein of a benign virus (AAV) that is useful for gene therapy. In the picture below I highlight the part we modified (a 28-amino-acid segment) in pink. You can see it on the full capsid (made of 60 subunits), a hexameric (6 bit) assembly with higher detail, and the single protein. As you may see it is complex bit of the protein interacting with many other parts of the capsid to assemble it into a functional vector.
To do this, we trained multiple machine learning models on datasets of a few mutations (changes to the segment), and then asked them to design sequences with many mutations. We found that the models can produce functional and novel AAVs with a high amount of diversity with high efficiency. We modified the virus segment that we targeted farther than any two natural “serotypes” are apart. For the more technically-oriented, here is a “tweetorial”.
Why would you want to do this? you may ask. Well, currently a lot of people are not eligible to receive gene therapies because they have pre-existing immunity to naturally occurring AAVs that are used for treatment. Diversifying these capsids away from what the body has seen can help them avoid the immune system of potential patients, and safely deliver the treatment to the right part of the body. It was truly an exciting adventure to be participating at the boundary of what is possible in protein engineering. This study further encouraged us that we can use ML to solve problems in gene therapy, which is why Dyno was born.