Thank you for the invite.
Yeah, it was great for the conversation. Maybe just to start at a high level, obviously AI, the integration of AI in healthcare, very topical, a big topic to wrap your arms around as an investor. But, but maybe to start, you know, where are the current capabilities in AI machine learning at Recursion in the space broadly? And, you know, any specific areas you hope to see growing in the next few years?
Yeah, I'm happy to start and then jump in. So, I think one of the really important things is AI is not more than a tool. It's a fantastic and powerful tool, but it is just a tool, and so if you think about what it allows you to do, it allows you to take very large data sets and make sense of them. It allows you to have much faster loops of learning to be able to validate, train, test and make predictions in the future, and so what we actually do across the company is try and think about why a clinical trial might fail and then create a predictive system to do that, and usually AI is best able to make those predictions because it's able to take a multi-parameter environment and think about all of the different possibilities of this environment.
What we do is we had two legacy organizations with legacy Recursion and legacy Exscientia focus on slightly different aspects of AI. On the legacy Recursion, it was much more large data sets, genomics, imaging, transcriptomics, real-world patient data. On the legacy Exscientia, it was more sparse data problems where you are trying to solve very particular chemistry aspects. But both of them have really been able to come together pretty seamlessly. And do exactly what I was talking about, figure out why, for this particular drug and this particular patient population, the clinical trial is likely to fail, and can we predict that earlier and design a better drug or design a better program for it?
Yeah, and you know, one of the things that I've heard over the years, Alex, is, you know, when are we gonna be able to point to a molecule or a drug that is AI? And I don't know that that's the best way to be thinking about it. I mean, as Ben mentioned, you know, we're looking at tools, and what we're gonna see over the years is these AI tools have an influence and flavor the way drugs are designed, the way, you know, biology is discovered, by the way, clinical trials are run. So I think there's hopefully over the years there'll be a sort of an evolving understanding of what an AI drug, you know, looks like.
Obviously, you know, Recursion is perhaps a little more advanced in using that tools, whether it be via chemistry, patient selection, biology. We, this is where the company is, you know, heavily leveraged, but we expect, you know, a lot of companies will also, you know, start to use these tools, and hopefully, you know, we'll be able to stay in the lead.
Okay. Yeah, and to your point, David, you know, one question we get from investors is, you know, is there a ChatGPT moment on the horizon for AI in healthcare, or is it sort of a gradual shift in the way that R&D is run? And hopefully cumulatively over consecutive years, you get an increase of quality and quantity of drugs that are, you know, helping patients. It, how would you sort of think about this? Or, yeah.
Yeah, it's a really great question, and I think going back to the two different core uses, we do use a lot of foundation models internally. In fact, there was a competition recently. There's 22 different predictive models that are tested, and it's open to people all over the world, and we won 17 out of the 22 categories using a single foundation model. This was all in ADMET predictions and trying to do that. The five categories that we didn't were all on single-use models for that exact prediction. Of course, next year, hopefully we'll win all 22, but I think that there is an incredible power behind foundation models. The problem comes down to when you're talking about a drug. It's incredibly specific.
And so you'll be able to use foundation models to go from what seems infinite down to a definable source of problems that you wanna solve, but actually getting to should this hydrogen atom be here and how will this specific protein react to it inside of a larger biological environment? I think we're still going to be for the near term focusing on specific models or specific incorporation of models and experimentation. So I don't think there's a single moment where the foundation models break and take over. But I think what we will see is drugs repeatedly being having some aspect of AI incorporated into their design, clinical trials getting smarter, patient selection getting smarter, novel biological insights because those are all of the things that we're seeing internally right now.
And part of why I think, you know, five years ago, large pharma did not talk about their AI very much. Biotech didn't talk about it at all, except for the few of us who were working on it. And I think now, the latest numbers I've seen, three quarters of biotech incorporate AI into their discovery and design in some ways. And all of the large pharma have internal programs. We're just getting better results. So I think from an industry perspective, there has been a lot of adoption. Still in early days, but you're seeing it come through.
Okay. And to your point on foundation models, as sort of the bread and butter for the industry today, do you see that being the case, moving forward? I think, you know, on the LLM side, we've seen DeepSeek sort of validate that maybe you don't need to have the biggest compute to, you know, derive an interesting model. Biology's a little bit different than word processing in terms of the multi-modality of the data. Do you see foundation models and scale on the compute side as being, you know, sort of required for the industry for the next few years?
Yeah, really, really good question. And I think a core focus, one, the foundation models we use, generally need to be more multimodal. And so you're, the data needs to come together to give you clearer perspectives on it. The second part is it's the data that matters.
Mm-hmm.
Because what we've found is we've actually spoken with large pharma about taking all of their historical data and trying to make better modeling systems out of it, and what we realized is it would be more efficient to recreate all of the data from scratch rather than try and take their data and manipulate it into a format that would be usable because you need it to be annotated right. You need to be asking the right questions. You need to have it in a format that you can actually access, and you need to know what data to collect, so I think one of the biggest advantages, honestly, that we have is for more than a decade we have been experimenting in what data is important.
Mm-hmm.
Like how do we put it together? Not only do we have one of the largest, if not the largest, data repositories with our 65 petabytes, but it's really what's in that data. It's not just all indiscriminate data that's pulled together. It's actually data that is directed towards a purpose, and that is what you want to be able to feed into your foundation models. That's what you wanna be able to do. It, it doesn't even have to be in a foundation model. It could just be in a modeling system. And so I think understanding how important that data is to getting a good result is something that the industry's just learning, and I think we've got a nice advantage on.
Okay. Maybe that can kind of circle back to the integration and the merger between Recursion and Exscientia. I mean, Recursion's built out a massive data set that's proprietary to Recursion from their phenotypic screening, the high-throughput phenotypic screening. Exscientia, you know, has clinical data access. You know, how is that integration going on the data side and?
Yeah.
How does the combined data set, sort of propel the, the discovery efforts?
Yeah, it's been really fun to watch it come together, and as someone who was a believer in the strategy of bringing it together and tried to really push for it, it was a little nerve-wracking to see, okay, we think this will work, but will it actually work, and what we've seen is literally from day one post-merger, we have been able to bring those, those skills together. So I think if you think about, the legacy Recursion data set is really about solving those big data problems and finding correlations and finding novel insights and understanding the bigger picture, where legacy Exscientia was, I've got a really particular question that there is no existing data for. How do I create a learning system that's gonna take me from no data to the right answer as quickly as possible, and so those mentalities have actually come together really nicely.
Mm-hmm.
In being able to say, all right, let's think about the most practical utilization of how we can bring this together. And I think we're on a nice path there. And David, maybe you speak to a little bit to the ClinTech and everything that we're doing there.
Yeah, yeah, I mean, I could just add to that. I mean, what, when I think about, you know, legacy Recursion, I think about novel biology and traditional chemistry. When I think about legacy Exscientia, I think about novel chemistry and maybe traditional biology. The merger has really pulled together novel, novel.
Mm-hmm.
What we've tried to do is then even extend that into how we're thinking about clinical development, how we're using data, real-world evidence, AI to predict patients, predict inclusion, exclusion criteria, predict sites, so the company in total from end to end, from hit ID to the enrollment of patients is really using novel, novel approaches.
Mm-hmm. Would you say that kind of full-stack integration of the process within the Recursion OS is maybe a competitive moat that you guys have formed over a year?
I believe so. I think that there's another element that we believe is important. You know, we have set up a feedback loop for when we have clinical data to feed back into the models, which is not something that I've ever had the opportunity. I mean, most companies that I've been in, programs that I've worked on are sort of individual. If they succeed, if they fail, okay. But what we've done is now really take back both, you know, the positive data as well as the negative data, feed that into the model. The one other thing about Recursion, when we think about the operating system and the platform, it is a platform and operating system that is intended to evolve and get better, like most operating systems.
I think that is something that is also a bit unique about the company.
Yeah. You brought up a really important point. It's interesting. One of the reasons both companies started to build so many automated laboratory facilities was the data fidelity, the quality of the turnaround, because what we have found working with CROs, you get back exactly the data that you asked for. And one of the important points that David just brought up is the bad data is actually as good as the good data. Mm-hmm.
Right, and sometimes it helps you more.
Mm-hmm.
To understand what doesn't work, and so being able to take everything and have it feed directly into our system that we do internally, or having the insight of this is what I need to ask the CRO to do if it's something that we don't do internally.
Mm-hmm.
is a really important part.
Okay. Maybe that's a good segue to sort of shift to the pipeline. You guys have a pretty extensive pipeline post-merger. It's also a really fantastic time to refine, based on the data that's coming in, and really push forward the assets you have the most conviction on. Maybe walk us through sort of the updates we saw on earnings and sort of the path forward.
Yeah. So, I think, you know, the pipeline has evolved. I think what was, you know, representative of the pipeline over the past years were version 0.5 of the operating system. Some programs we have run the experiment, we've gotten the data and we've made, you know, decisions on the data. I think, you know, the resultant pipeline is, you know, representative of a portfolio that is more, you know, representative of sort of the best of what we have, whether it be first in class, best in class. There is a, you know, sort of uniform approach that the portfolio should be able to achieve rapid proof of concept with registration plans that are clear and, you know, decision points that should come in sort of rapid, you know, succession.
I think that is representative of sort of an evolving company, representative of two companies coming together.
That's, I think that's pretty well put. I guess, you know, when you think about the best ways to validate the platform and I guess the individual assets, is there sort of a sweet spot in terms of, you know, indication or, you know, area of development that you think you can hit on that de-risking quickly?
Yeah. I mean, you know, if you think about what you know, the company really wants to achieve are really two things: demonstrate that we can bring molecules into the clinic that in a more efficient way, right?
Mm-hmm.
In a more expedited way.
Also show that the rate or sort of the batting average, so to speak, is higher than, you know, the industry average.
Mm-hmm.
I think that's what we are seeing now in our pipeline, right? We are seeing, you know, the oncology focus gives one an opportunity to achieve proof of concept quickly. You know, every molecule that I've ever worked on that translated from a therapy into a drug, you know, relatively early, and I think the pipeline, you know, represents that. The indications that we have, our approach in terms of what we are doing from a, you know, clinical trial ClinTech perspective should allow us to achieve proof of concept.
Mm-hmm.
You know, more expeditiously.
I think to your point, we've done some work on this and the industry average for success rate of drugs. I think it was like one in 12,000 that are like nominated. It's not a hard benchmark to beat.
Exactly.
but it does take time to get through.
Right. Yeah. And it's, you know, it's important that we have multiple shots on goal. You know, we're not gonna be successful with every drug, but we think that over time the pipeline will show that, you know, the success rate in the clinic coupled with our time to the clinic will hopefully, you know, be the secret sauce.
Yeah.
Of what Recursion can do.
I, I'd also say there's multiple different forms of validation, and so obviously we are very focused on getting to the clinical proof of concept validation, that we all want, and we've got a number of drugs that are getting close.
Mm-hmm.
I think in the interim, what we've also looked at is validation is the technology doing what it was expected to do. So earlier in the year, we saw a CDK7 where, you know, from what was a really, really complex target product profile to try and get all of the ADMET right.
Mm-hmm.
For a biological mechanism that should work if you can get that therapeutic window, we saw everything come out within about 5% of where we predicted. Now in human biology, right? Now playing out in the in the real world. And, and we did see, you know, that the PK/PD, the safety profile come out the way that we wanted. We did see some unexpected efficacy that was very early, but nice. I think on FAP was another, you know, recent example where there had never been a connection between MEK and FAP before. We, we uncovered it with the phenotypic platform. And so seeing, again, early, but, those responses not only on the, number of polyps, but also on the dysplasia, really looking at, I mean, cells that are going wrong show that they are going wrong.
And so when you start to see that return to a more normal looking state, that's really encouraging as well. So those are the things that are telling us we're going in the right direction. We had already showed cost and time benefits by dramatically reducing those aspects of drug discovery. Now we're seeing early clinical results that are saying, you know what, you're going in the right direction here too with things that weren't able to be achieved historically. Okay. That makes sense. And that sort of gets to my next question, which is, are there any, you know, obviously you love all your children, but like within the pipeline, are there any assets that maybe best exemplify the current capabilities of the platform?
Yeah. I mean, I think I'll give you a few. One that we think has achieved some degree of validation, which is our MEK inhibitor. That's, that is, you know, coming off of sort of the earliest versions of our pipeline. And then we have our, you know, RBM39 degrader, which has really come out of the Recursion platform, representative of sort of a later version of the operating system. And then you look at CDK7 and you're looking at a compound that was really sort of AI driven in its design. So, you know, getting back, you know, to where we started, is there going to be an aha moment that that is the drug?
But what we're seeing as we look across the pipeline, sort of each of these molecules, you know, represent very strong elements of AI-driven design, whether that's on the biology front or the chemistry front. So we think even from early to sort of mid-stage platform, you know, we have a, you know, a very nice pipeline.
Yeah. I'm gonna zoom it out, because I think it's really important not to feel like we're dodging your questions. It's actually a business model point.
Mm-hmm.
So if you look at what our mandate is, the people that have given us money and what they say you need to do this, it's actually to change the binary risk profile of biotech investing.
Mm-hmm.
Most of our investors are people that have come to us and say, I either don't want to or won't invest in biotech because of this binary risk profile that is throughout the industry. And most of the time the binary risk is negative, right? I mean, it's a 95% failure rate environment. And so they actually want a business model that is more diversified. And so we've got the partnership business. We've brought in nearly $500 million in cash through those partnerships. We're hitting on a lot of the discovery milestones with Sanofi and Roche and moving those things forward, and then we've got our pipeline, which is not a single agent. There's no lead compound. We just don't think about it that way internally. And part of that is because we need to be able to make disciplined data driven decisions like we did.
Hey, if we think it is a lower impact or lower probability of success, we are going to kill it.
Mm-hmm.
We don't have a lead compound that we think this is. This is too sacred to do something with, and that's a core part of why people want to invest in us. They wanna see the technology play out, but they want it to be a business model and not a binary risk.
Yeah. And I mean, the traditional development model is you have three or four assets that, you know, could make or break the company. And, you know, it took years and years to push those forward. For you guys, it's a flywheel of learning, but you also have probably too many, you know, assets or targets that you could go after yourself. So, you know, maybe that's a good segue into the partnerships. You know, you touched on Sanofi, Roche, Genentech. What are sort of the scope of these partnerships?
Yeah.
and is there additional dry powder for more down the road?
Yeah. So to answer the question, yes, definitely. I think it's interesting, Sanofi and Roche. Sanofi was legacy Exscientia, Roche was legacy Recursion. And so the Roche partnership focused more on novel target discovery, doing things in biology that had never been done before. And we delivered the first of the neuro maps last year and hopefully more to come there. Really, really exciting. The feedback we got from Roche is 90% of the targets that they are uncovering with us were not things that they had ever uncovered before.
Mm-hmm.
And this is in neurological diseases where we are desperately in need of new targets. On the Sanofi side, it's more of it had been more focused on design. So the fact that we have advanced four different programs means there are four different problems that Sanofi was struggling with or didn't know how to do better or the traditional methods had failed on, where we have now been able to solve those initial discovery things and move forward. And there's more than just four programs in that, those are just the ones that hit the milestones. So what we're trying to do now is also bring that all together, right? Like we can do more of the target ID with Sanofi. We can do more of the design work with Roche. And they're really excited about that.
We also do have Bayer and Merck KGaA, but they're just smaller partnerships.
Yeah. Yeah. Okay. That makes sense. We've got a couple of minutes left and, you know, that this is, you know, you function in the broader backdrop of everything else that's going on, and it's day to day, for the most part, but we have seen some positive, I guess, inclinations that the FDA could be receptive to using AI in their sort of regulatory process. Maybe walk us through, you know, from where you guys sit, how that relationship is evolving.
Yeah. I mean, you know, so to be clear and just to you know, clarify, you know, the FDA is has you know taken a position or is thinking about using AI for biologics. Most of what we do is small molecules. That said, our whole sort of discovery process is focused on building and using AI based models to predict.
Mm-hmm.
So it's, you know, encouraging to see that the FDA is starting to go the way that we have been going. I think it'll take some time as we, you know, generate data and, you know, engage in dialogue with the FDA. But this is good news for a company that is really spending a lot of, you know, you know, intellectual capital trying to think about what will discovery and tech and clinical development look like, not a year from now, but five, 10 years from now.
Yeah. And I'd say, I think, we're really excited to see the FDA focusing on things like animal testing, where it's a big part of the cost and time, and it's often not as predictive as you would hope for. So, I think we will continue to watch where it comes out. They also just released guidance on how they can use AI in their review process, which would be terrific as well. And so we're big supporters. Okay. Yeah. I have heard that even, you know, going back to Gottlieb, like this was something on the table internally there.
Yeah. Absolutely.
But it takes our new board member, Najat Khan, was one of the main authors of the bill that, or the guidance, that David was talking about. So, yeah. Okay. Great. Well, good to see the progress, both for you guys and for the industry. And, I think for the sake of time, we'll have to leave it there. But, please join me in thanking David and Ben for the engaging discussion. Thanks. Thanks, guys.
Thank you. Thank you.