Schrödinger, Inc. (SDGR) Earnings Call Transcript & Summary

Good afternoon. Welcome again to the Morgan Stanley 18th Annual Global Healthcare Conference. I'm one of the biotech analysts here. My name is David Lebowitz. Before we get started, I have to go through the requisite disclosures. Please note that this webcast is for Morgan Stanley's clients and appropriate Morgan Stanley employees only. This webcast is not for members of the press. If you are a member of the press, please disconnect and reach out separately. For important disclosures, please see the Morgan Stanley research disclosure website at www.morganstanley.com/researchdisclosures. If you have any questions, please reach out to your Morgan Stanley sales representative. And with that, I'm happy to begin the presentation by welcoming from Schrödinger, CEO and President, Ramy Farid; Executive VP and CFO, Joel Lebowitz, unrelated; and of course, the Chief Biomedical Scientist and Head of Discovery, R&D, Karen Akinsanya. Did I pronounce that correct?

It's...

I did my best. I guess before we get started, if you could give a top-level overview of the company. This is clearly a unique business in that it straddles the technology world and the biotechnology world. And I guess bring us up to speed of how this unique company came together.

Sure. Absolutely. Happy to. So at the moment, we have around 500 people in the company. About half the company is focused on developing, and has been for the last 30 years, since the company was founded, a computational platform based on physics-based methods for predicting the properties of molecule with very high accuracy. We started as a software company, selling the software to predominantly pharma companies, biotech companies, academic institutions and government labs all over the world. And as the technology continued to evolve and become more accurate, we -- and more transformative, we found that in order to really validate the technology, to demonstrate to the pharma industry the potential power of it, one of the most effective ways to do that is to start using the software ourselves to do drug discovery. So about 11 years ago or so, we cofounded with Atlas Ventures Nimbus Therapeutics. That was very successful. Still doing great, Nimbus Therapeutics is, but it was very successful. I think it proved that the technology was working. That led to the formation of a number of other biotech companies for which we have an equity stake in. And now we have in the company, out of the 500-or-so people, a little bit over 80 people that are focused on doing drug discovery, about 25 to 30 drug discovery programs. Most of them are in collaboration with these companies that we've either cofounded or have an equity stake in. A number of them have already gone public. A number have compounds in the clinic. And then more recently, we started to work on our own internal drug discovery projects. And one of the most exciting things about this, it may seem unusual to have a software business integrated with a drug discovery business, but what we realized really early on is the extraordinary synergies between these businesses. Of course, I already implied one of them, which is that the success of the drug discovery business was obviously helping to validate the platform, and that was leading to growth in the software business. But it goes the other way, too. We have close relationships with all the pharma companies and lots of biotech companies. And a lot of the know-how that comes from those interactions gets fed back into the software, which, of course, benefits the whole community, but also, of course, benefits the drug discovery business. So we continue to realize those synergies, and we continue to invest heavily in the computational platform. That's a really differentiating -- differentiated part of the company. And in some sense, we have identified these 2 synergistic ways of monetizing the platform.

I guess if you could try to distill what the software is and how does it work. Certainly, people tend to think of it in the terms of the life sciences and helping to develop drugs, but it really isn't limited to that. It's agnostic to what you're trying to structure. I guess if you could run us through that.

Absolutely. Yes, I can explain that. Thanks. So what does physics-based mean? So that's just a word. We have to sort of describe it. So what this means is that these are methods that at a very, very basic level, capture the physics of molecular interactions in a rigorous way. So we're really using these rigorous methods to compute what's really -- what is the free energy of an interaction. What does free energy mean? It means how stable a particular state is. So now that can be abstracted to many different things. So when a molecule binds to a protein, what you're really measuring and determining is the free energy of that binding. When a molecule dissolves in water, you're describing the free energy of that process. When a organic light-emitting diode emits light, there's a free energy associated with that. Every molecular process, every molecular property has a free energy that describes this property. And if you have a method that uses these first principles to describe these properties, it's, of course, agnostic to the system. So what does that mean? On the drug discovery side, it means that the method applies to discovery of small molecule therapeutics, biologics, peptides, macrocycles, protein-protein interactions. And of course, as I just alluded to, it can be used to predict the properties of polymers, for example, that coat airplane wings. It can be used to predict the properties of organic light-emitting diodes and other materials. We do supplement these physics-based methods with machine learning in some very specific applications. But to be clear, these are totally distinct from machine learning or what we often refer to as AI which are, by definition, highly localized -- can only produce highly localized models and are very specific to the problem being solved. The advantage, of course, of physics-based methods is they can be applied to really any system -- molecular system.

Now I guess when you look at your competition for the software, you probably see it more as being the bench-top research itself than the actual competing softwares. How does this compare?

Yes. Yes, that's absolutely right. Really, in some sense, the competition, as you say, is doing things in a sort of traditional way, where you brute-force and make lots of molecules in the lab experimentally. And so how does it compare? So in a typical drug discovery -- let's focus on drug discovery. Typically, around 1,000 molecules are synthesized and tested in a year. That can cost -- it's around $5,000 per molecule. It can be higher. It takes a few months to make and test a molecule. On the computer, you can explore billions of molecules in that same period of time, probably on the order of thousands in a day. And it turns out, by combining physics-based methods and machine learning, you can literally look at billions of molecules at obviously a fraction of the cost. And the advantage of that is not just that you're reducing the cost, but drug discovery is one of the most challenging multiparameter optimization problems there are -- there is. And the reason is because all the properties that have to be tuned into a molecule are anti-correlated. So the more chemical space you're exploring, the more likely you are to find a molecule that optimizes all of these properties, that you find a molecule that's highly potent, that's selective, that's soluble, that's permeable, that's clear -- not clear too rapidly and so on, many, many of those properties. And so this is -- so it's obviously faster. It's obviously cheaper. But maybe the more important thing is because you can explore so much more of chemical space, we're finding that the molecules that come out of these computationally driven drug discovery projects are much higher quality, almost by definition.

When you -- I know that the physics-based approach is not quite the same as AI. And they're kind of -- and in that way, they don't directly compete with each other. What type of capability does your approach bring to these pharmaceutical and these materials players versus what maybe a Google approach with their computing would be able to do?

Yes, it's quite simple. So what is machine learning? Machine learning is -- involves a training set. You have to produce a set of, in this case, molecules that have certain property, use that as a training set and then you build a model from that. Now by definition, any prediction that you make using machine learning, any kind of machine learning, it doesn't matter how sophisticated it is, this is the case for any kind of machine learning, the only -- it's only going to work when you're interpolating. So it will only work for molecules that are highly, highly similar to the molecules that are part of the training set. This is the same thing as saying if you build an AI to detect cats in photographs, you use pictures of cats to train it. If you then feed an image of a dog, it obviously doesn't know anything about the dog or any other species, right, by definition. This is -- that's intuitive. It's the exact same principle. Now the challenge though is what does it mean to have a molecule that's similar. That actually turns out to be very complicated, not just that it looks the same. And that's a very subtle and very, very important point. You can make a 1-atom change to a molecule, 1 atom, and that molecule will now be outside of the training set. And so that's one of the limitations of machine learning. It's highly high -- can only produce highly, highly localized models. The physics-based methods can explore, in principle, all of chemical space, which is almost infinite, something estimated to be around 10^50 or 10^60 molecules. That's very, very powerful. And that's what's required to not only get into new IP, I mean, sure, that's one reason, but to solve that challenging multiparameter optimization problem, you have to extrapolate. And there is no way, by definition, sorry to keep using that word, that cannot be done with machine learning or any kind of AI.

Would you be able to run through an example of how a company might take your software and apply it to the process of developing a molecule?

Yes, absolutely. So either a company or us, of course. As we discussed, there are pharma companies that are using our software, and all the top 20 pharma companies are and many biotech companies, and of course, we are. So we have a very good understanding of this. So here's generally how it works. One of the most challenging things at the beginning of a project is to determine even whether a molecule -- whether a target -- protein target is druggable. How do you determine that? One way is, well, you work on it for 5 years and then if you end up with a drug at the end, it was druggable. But that's not a great way to do it. It's nice to be able to determine that upfront. So that's what -- the first step or close to the first step. Of course, understanding the protein structure and so on is actually the first step. But a early step is just to assess the druggability. And we have developed physics-based methods that are very effective at doing that, that can save an enormous amount of time. You know that a large majority actually, and this is sort of depressing, projects end up not producing, right? You start working on a project, and after 5 years, there's no development candidate. A lot of the reasons for that -- it's not the only reason, but a lot of the reason is because it simply wasn't druggable, for example, with a small molecule. And it took sort of trial and error to discover that. We can determine that. So that's one of the things that is done. Then we have also developed software to identify hits. So what you can do now is you can screen existing libraries virtually. You can also screen virtual libraries that are on the order of billions of molecules very rapidly using virtual screening. So our software customers and us are doing that. Once you have the structure, once you have assessed the druggability, you do a virtual screen and you identify hits. You -- from billions of molecules, you can actually identify just a few hundred molecules that you purchase or synthesize to identify sort of weak binding hits. And then our customers and us are using the software to then optimize those leads -- or hits, convert the hits into leads and then optimize the leads to development candidates. And that involves computing all of the properties that are required to make -- that are required for a drug molecule to possess: selectivity, potency, solubility, permeability, appropriate clearance and so on. And so we -- so they're using our software to not only compute those properties but to enumerate chemical space. So we've developed software that can take a lead molecule and enumerate all the possible molecules that can be made, for example, with that common core. Or you can actually take a lead molecule, maybe even from a competitor, and you can scaffold hop or core hop from that molecule, generating new IP and then enumerate around that new IP. So that's how our customers and us are using the software, all the way up to a molecule that's then declared a development candidate, ready for GLP tox studies. Then we're done, of course. We stop when there's -- when the chemistry is done.

My last question on the software side. If you think in terms of probabilities, the traditional small molecule that enters the clinic has maybe below a 15%, like a 13-or-so percent chance of getting through to the end of the -- again, through the end. How could a software like this change that probability? And I know that it's just too early to really have a true number of what -- of how it is different. What would you speculate it might do for that number?

Yes. So there are 2 sources of failure for drug discovery projects. One is biology. The target is simply -- even engaging the target, it didn't -- inhibiting it or activating, it just simply didn't have the effect that was predicted in -- from earlier studies. So that's one source of failure that we are not addressing. The other source of failure is what we are talking about before. Essentially, what it is, is not being able to solve the multiparameter optimization problem. You make a few thousand -- typically, what's made, maybe 5,000 molecules, and none of those molecules over the 3 to 4 to 5 years of making those few thousand molecules satisfy all the properties that are required for it to be a drug candidate. And the project is killed, even though the biology was fine. What we're finding is by exploring enormous amounts of chemical space with these accurate models, we've significantly increased the probability of getting to a development candidate. Now as you said, you probably need to run hundreds of projects, right, to really build up statistics where you can say the percentage is something point something with high precision. But what we're finding is roughly -- first of all, the projects that we're choosing, we're being very careful to try and reduce the biology risk by working on well-precedented targets. So that's one area that's really increasing the success probability. With regard to then the chemistry risk, we are finding that the probability of success is very, very high. I mean above 80% right now is what we're seeing of projects that -- or let's put it, failures of 20% due to not being able to solve the multiparameter optimization problem. It's very unusual for us. It's above 80%. That's a significant increase from the sort of now published industry standards.

That was very helpful. Moving on to the drug discovery side. First, we'll talk about the partnerships. You have 25 to 30 partnerships that you had talked about earlier. I guess one of the challenges with being an investor is trying to decipher it all. I guess how many clinical-stage projects are there right now? And how many do you expect there to be maybe added in over the next year? And from the P&L standpoint, how could we expect that to impact milestones, equity income, various things like that?

Karen, do you want to start?

Yes. So just to answer your question about clinical-stage programs, I think if we're right up to date, it's probably 4 programs that are in the clinic actively being pursued in Phase I through Phase IIb. And I think an interesting way to look at what's to come is that there are 9 collaboration programs in the late stage of drug discovery. That means that if you factor in getting through GLP tox, we expect a good portion of those actually to transition into the clinic over the next 2 years. And as you said, a lot of those projects are associated with milestones for getting into GLP tox or getting those patients dosed or an IND accepted. They all sort of vary a little bit. So that's, I think, what we should be looking at in terms of the trajectory of the collaboration portfolio.

And then with regard to the equity part, so far, 4 of our earlier programs, we've realized value from the equity. Two of the companies have gone public, and one has done deals where distributions from those deals have gone back to shareholders. And another one was acquired. So we are realizing the value from this equity. And as you can see from our filings, the amount of equity that we have in these companies is growing. As we -- with the first one that we set up, of course, who knew if it was going to work? But once that one worked and the next one worked and the next one, you can imagine that the amount of equity that we're able to obtain from these partnerships is increasing. You can see that. I mean we have some partnerships where it's as high as 50%. The economics around the milestones are also increasing. So as these later programs mature to the stages that Karen saw, we certainly are projecting an increase in revenue from the collaborative drug discovery business.

I guess when you're looking at these partners, which particular molecules right now do you think we should pay attention to, at least in the next 6 to 12 months?