2024 MIT Health Science Forum: Lightning Talk - DeepCure

DEREK MILLER: Hello, everyone. My name's Derek Miller. I'm the vice president of platform research here at DeepCure. And I'm excited to talk a little bit about some of the things we've been working on. So DeepCure was founded at the MIT Media Lab six years ago. And since then, we've raised over $70 million. We have over 30 employees. We've started an automated chemistry lab that can handle multi-step synthesis. And we have three drug discovery programs. One which will be starting clinical trial in 2025.

And that is our BRD4-BD2 program. We have multiple exciting indications that we're looking into. We also have STAT6, which is entering late stage discovery. We have a couple exciting scaffolds there. We have an undisclosed target, which we'll be announcing very soon. When we approach our pipeline, we have two things that we look at. One, we're looking for an unmet need in the clinical setting that has validated biology. And two, we're looking for targets that have particular medicinal chemistry challenges that we think our next generation AI platform is built to solve.

So with that in mind, immune targets became an obvious option for us to target. One, because they tend to have protein-protein interactions. These typically require larger in 3D molecules. A lot of these proteins also have highly charged surfaces with hydrophobic surrounding areas, making it incredibly difficult to build a drug-like small molecule for. Or they have very precise target selectivity issues. So precise that even, at times, it can be the difference between a non-toxic outcome or a toxic outcome.

And these types of problems, and going after these targets are very difficult. And most AI methodologies struggled here for two different reasons. One, if there's no data at all-- we've heard about data sets and models consistently-- it's very difficult to go after these targets with no data. There's no binding mode to understand. There's no starting chemical matter. If we're going after an allosteric site, there's no peptide or anything that we can use as a basis.

Second, if there is data, sometimes that can be worse. Because you get stuck in local minima trying to design around a compound or a peptide and fulfill its interactions that may be impossible to build a drug for. With that in mind, DeepCure has developed a technology to address both of those issues. And at the heart of it, we believe that if AI is going to be in drug discovery and make an impact, it should make an impact on the most challenging targets.

With that in mind, we have our pipeline laid out here. And it's composed of about three different items. We have our PocketExpander. Our PocketExpander looks at exploring novel interactions and opportunities and how we can interact with the protein. We have MolGen, which is our generative AI algorithm. It designs compounds to completely fulfill the interactions from PocketExpander. We also, in the MolGen process, co-optimize for ADME properties, ensuring that we're not going to hit a liability down the road that's immediately obvious. And then we have our Inspired Chemistry lab, which is a completely automated chemistry lab that can handle multi-step synthesis.

So the innovation starts with our PocketExpander. Our PocketExpander looks for novel interactions. And this goes beyond just hydrogen bonds. We're looking for, through quantum mechanics, through molecular dynamics, unique interaction opportunities with a protein that could deliver the phenotype that we're trying to address. And this will create multiple hypotheses in how you could interact with the protein. Most of the times, it can be in the realm of 8 to 10. And we take those hypotheses and we use our molecular generation, which is a reinforcement learning algorithm that will try to build a compound to fulfill all the novel interactions.

And these compounds are most importantly, not only synthesizable with a very high probability, it's also cost effective. Because one of the things we wanted to do with our model is ensure that it produces a synthetic path. So it tells you exactly how it thinks it should design the compound. And finally, once we've designed the compound that specifically meets the interaction needs, we can send it to our automated chemistry lab, which we call Inspired Chemistry, which has a very high throughput. It has 28 different reaction types, which allows us to explore a large diversity of molecules. And can handle up to 10 steps.

Very recently, we had a-- completely in a automated setting, we showed that we could synthesize Paxlovid. And that we have a publication on that-- or an output on that. And finally, on the partnership opportunities, we're looking for two different areas. One's in our inspired chemistry. On the platform side, when we're developing new targets, we use Inspired Chemistry almost as a customer. And so we're actually taking customers specifically for Inspired Chemistry, and working with a few to build custom synthesis for them.

Second, we're looking at collaboration, not only on some of the new challenging targets that I was defining earlier, but some of the existing programs as well. So I want to highlight our team. We have a diverse set of backgrounds, from machine learning to computer science to biology to medicinal chemistry and automated chemistry. We also have a great group of leadership as well. And yeah, thank you again. My name's Derek. And if you want to interact afterwards, I'd be happy to.