Ron Hause describes Shape Therapeutics' RNAfix platform, which uses AI to design guide RNAs that edit RNA by converting adenosine to guanine. The approach generates millions of data points to train deep-learning models that design better guides and tissue-specific viral vectors, with a Parkinson's disease program moving toward the clinic.
Chapters
Everyone, yeah, so I'm Ron Haas and I head up the analytics and informatics department at Shape. It's a startup here at Seattle and we have a satellite site in Boston and we're working in the field of sort of programmable RNA medicines and gene therapy enablement. And for context for folks, if you think about the 3 problems that go into actually getting some of these generative gene therapies into the clinic, It's really how to actually make the correction, some problem related to genome editing, either how to make a CRISPR guide actually edit that intended site, how to deliver this payload to the right place. So for those of you that don't know, in the clinic, when we give something like replace and fix a gene, the vast majority of viral vectors that we use to put the payload in will go into your liver and cause dramatic hepatotoxicity. So we're trying to solve that problem to make gene therapies work. And then manufacturing. So currently, if you take a drug like Zolgensma, which is used to treat spinal muscular atrophy, it's a gene replacement, it costs $2.1 million for the patient. And that's because it takes a lot of effort to create enough virus to be able to dose a patient and get it to go into the central nervous system to cure that disease. So those are 3 problems that are great problems for ML to apply, along with high-throughput data generation. To solve and 3 problems that shapes into. The first one I'm going to talk about today very briefly is how we do RNA editing as a therapy. We have a really cool platform called RNAfix that leverages an enzyme that exists in everyone's body called ADAR. Now, ADAR normally functions early in development along with brain development as a regulator of RNA functionality. And what's nice about ADAR is akin to CRISPR, You can direct it to edit a specific site in your RNA with a guide RNA. And one of the advantages to it over a CRISPR-based system is that you don't have to give a huge Cas9 protein, which allows for a lot easier delivery and a lot more of a temporal timing in how you are doing RNA regulation. And an advantage of what we can do at Shape is we can generate millions and millions of data points experimentally of different guide RNA structures that those different letters will actually cause complementarity to a given mutation, recruit ADAR, specifically make a correction from an adenosine to a guanine, and then not make aberrant corrections to nearby adenosine. So the goal here is that you wanna fix a mutation. You can do a lot with any adenosine mutation. You can hit any ATG start codon to knock down a gene, and you can make 17 different amino acid changes of the 20 amino acids. So it's really nice in that it's a little safer, simpler payload, and we can use this data to actually build generative solutions for guide RNAs that work. So we use a lot of the methods that many of you are familiar with, deep learning methods. In this case, we're showing an example of a 2-dimensional encoding with a bit diffusion that trains a model of millions of guide RNA sequences conditioned on experimental data that we have on how well they edit an intended adenosine. And then use that model through the reverse diffusion process to spit out completely new guide RNAs that we can then experimentally validate. Now there's two-dimensional and three-dimensional ways to do this. One nice thing for those of you that aren't in the RNA field is RNA is somewhat in between DNA and protein. It has secondary and tertiary structure, yet it is only comprised of 4 letters. So when the guide RNA actually hybridizes, you can actually use structural methods like hairpins, boulders, and loops to infer complementarity that you're actually getting the intended target. And what's really neat about that, at least at shape— excuse me, the animations ran through a little quickly— is beyond RNA data, we can do this for data that we generate on billions of viral sequences. So we do high-throughput mutagenesis of billions of viruses, dose them to nonhuman primates, and then we can isolate different tissues that they actually functionally transduce. So you can build a generative model for a brain-specific virus or a heart-specific virus. And we do that for promoters so we can generate a cis-regulatory architecture that finally controls a genetic payload in what cells we want to. So 3 different language problems that we work on at Shade. And an advantage is that we can then take these generations, do in vitro validation for our quote unquote reinforcement learning. Or in vivo validation where we dose these guys into mice or non-human primates to show that they achieve what we would expect, the specific functional therapeutic editing of a disease mutation. Here we're doing one in Parkinson's that we're actually trying to bring into the clinic, and that works pretty well. So with that, I would just like to say we exist. If any of you are interested in the wealth of data that we have and how to apply different ML approaches, to solve really cool problems that are sort of bottlenecks in the application of gene therapy in the clinic. We have a lot of openings. And reach out to myself, Bora, or David here, and happy to talk and chat. Thank you. Okay, great. Thanks to all the speakers for great talks. So we have a coffee break now. We'll be restarting at half past.
