Schrodinger, the New York-based computational drug discovery shop, has emerged the past couple years as a key behind-the-scenes enabler of progress at a couple of rising startups – Nimbus Therapeutics and Morphic Therapeutic.
Bill Gates has now invested in Schrodinger for four consecutive funding rounds. Schrodinger announced on Friday it has raised its latest large sum — $85 million — to further extend its approach across the pharmaceutical industry. The Bill and Melinda Gates Foundation Trust and WuXi AppTec’s Corporate Venture Fund co-led the round, and was joined by Deerfield Management, Baron, Qiming Venture Partners, and GV.
Gates, in the statement, said Schrodinger: “has demonstrated that precise molecular design can significantly accelerate drug discovery and lead to unexpected solutions that stand to benefit patients.”
I met with Schrodinger CEO Ramy Farid and chief biomedical scientist Karen Akinsanya, a Merck veteran, on a visit to their New York offices in early December. You can listen to my conversation with Akinsanya on The Long Run podcast. With Farid, I asked more about the nitty gritty of what Schrodinger does with computational drug discovery.
Below is a condensed and edited version of the conversation. It covers a lot of useful scientific concepts, and insight into Farid’s thinking about business strategy. Given the big $400 million investment in Cambridge, Mass.-based Relay Therapeutics, another one of the emerging computational drug discovery companies, this is a timely relevant conversation for many people in drug discovery. Warning: This was a 45-minute conversation. It’s long.
Luke Timmerman: How did Schrodinger get started? It wasn’t for drug discovery, was it?
Ramy Farid: Schrodinger started [in 1990] as a pure software company, focused on quantum mechanics. The focus was initially interacting with academics. The company was funded by SBIR grants from NIH. The vision back then, from Rich Friesner (Columbia University) and Bill Goddard (Caltech), was ‘if we can get the physics right and really understand molecular interactions in a deep and rigorous way, accurately using physics-based approaches, we could solve problems that were not possible with traditional methods. At the time, people were using QSAR (Quantitative Structure Activity Relationships) or machine learning. As you know, machine learning and AI have been around a really long time. Chem informatics, it’s sometimes called. Back when the company was founded it was very difficult to imagine what computer power would be like in 20 years. Even though you might know what the actual formula is, you can’t imagine it.
LT: You can graph it out, but you can’t see how the advances will be applied.
RF: Exactly. It turned out back then that computers were so slow, you couldn’t even model all the atoms in the molecule. You had to eliminate the hydrogens. Two-thirds of the atoms in the molecule. You had to make all these approximations, because computers were so slow. It was a really difficult time. There was a lot of hype back then, in the 1980s, and early 1990s, which said that computers really can transform the way molecules are designed. Both drugs, and materials. It took A LOT longer than people thought, to get to the point where you could actually run a calculation on a computer and predict a property of a molecule reliably.
LT: While all this is going on, Moore’s Law is progressing, right? We were gaining all kinds of new applications that were suddenly possible, because of that. But not this [drugs designed on computers]. Not that vision, not right away, anyway.
RF: No. It really took two things. It wasn’t just computing power. It turns out that something so simple – the binding affinity of a molecule – a lot of people still think of that as “You can just use pattern recognition for that.” The medchem literature is enormous. There’s so much information. A lot of chemists look at a molecule and say ‘I’ve done this project before, and know that if I add methyl groups here, and move this nitrogen around, or add a hydrogen bond donor over here, and make it more greasy, I usually improve the affinity of the molecule.’
LT: Isn’t this some of the “art” of medchem. Trial and error.
RF: Exactly. Trial and error. But here’s what happened. During that time, two things were really happening. One is that people forgot their freshman chemistry classes. They forgot the concept of entropy and the concept of free energy. And they forgot the concept of enthalpy-entropy compensation. These are very fundamental principles. They also were exposed to a lot of computational chemists claiming they had solved this problem when in fact, they hadn’t. They were limited by the science and by compute power.
So what was the situation? You had chemists, who are humans, who usually have an overinflated view of their ability to do pattern recognition. Then you have this unbelievably complicated phenomenon. A molecule binding to a protein. The protein is flexible. The ligand is flexible. Waters are flexible. There’s entropy. There’s enthalpy. Those two things are compensating for each other.
LT: Then you get into things like dynamic equilibrium.
RF: There are dynamic systems. We can get into that later.
But now you’ve got a situation where people are feeling like they can use their intuition and pattern recognition to design molecules with certain properties and it’s just not working. It’s really unreliable. This is what Schrodinger focused on. We asked, maybe we can get the physics right. If we can model entropy. If we can model enthalpy. Maybe we can model dynamics of a system. When I joined the company in 2002, I’m not going to claim anyone told me it’s going to happen in X number of years. But I thought – and everybody thought – we sort of had this sense of “we’re almost there, it’s a couple years away.” Again, it turned out to be WAY further out, because it was a really hard problem.
That’s where Bill Gates came in. Bill said, “Alright, this is going to be a solved problem, but it’s not going to take a few years. This is really hard. It’s going to take a lot of computers, a lot of scientists working on the physics, and a lot of people just plugging away. Don’t worry about breaking even, don’t worry about the short-term revenue. Just focus on the science.” When we started doing that, a lot of public companies had to start abandoning [computational medicinal chemistry.] No VCs are going to tolerate 10 years of just developing an atomic force field without knowing what it’s ever going to be used for.
LT: I started covering biotech in 2001 at The Seattle Times. I remember people talking about “in silico drug discovery.” That was the phrase then. It didn’t happen.
RF: That’s right. But what’s worse is there was a lot of hype. Then the natural hype curve kicked in. We went way down. There was severe skepticism. We suffered from that. We contributed to it, too. Not knowingly, but we did. We didn’t know. We thought we had something, but remember, we were a software company. You’ve got software engineers, and the software was saying “this is so predictive, what’s wrong with those chemists? Why aren’t they using it to design molecules?” The reason is when you’re doing retrospective validation, it’s very easy to fool yourself. You think it works better than it really does. Until you really try it in the real world. And until you try it in the real world, you don’t know what the real problems are.
For example, we were told this many times: Potency is not the only problem. You have to worry about selectivity and solubility. Or whether it’s permeable. We had the core of something really exciting with our technology, but we didn’t have the context. We didn’t have the respect of the people in the trenches who were really trying to develop drugs.
They were right. That’s why we started Nimbus [in 2009]. That’s why we developed a biotech organization within [Schrodinger]. It was so that we knew which problems to solve, and so we could really test the programs prospectively.
LT: That sounds like a classic tech startup with a solution in search of a problem. You think you know what you’re doing, but you don’t.
RF: That’s a really good way to put it. But that’s great.
LT: That’s often how things get started.
RF: Exactly. If we had started a biotech company 20 years earlier – and by the way, we had VCs pressuring us to do that – if we had done that to create a billion-dollar molecule, we would have taken our eye off the ball. We would have said “let’s not worry about the physics and the little details, let’s just get the drug. We’re burning money like crazy, these experiments are expensive, let’s stop doing research.” That didn’t happen. The fact that David Shaw and Bill Gates stuck with the company – actually they are responsible for a major advance in the field. We got the physics right, and now we can apply it to real problems.
LT: Before we get to the Nimbus story, and I know that’s really important, what’s the business model here in the beginning? You’re a software company, so is it annual licenses to pharma companies, and that kept the lights on and people employed?
RF: That’s exactly what it was. In retrospect, it was maybe a bit naïve. We said “Without actually doing drug discovery ourselves, we’re going to go to these pharma companies and bang our hands on the desk and say: You guys are idiots, look how stupid you’re being in developing drugs. You should be using this technology like this.” That doesn’t work very well. You have to have the credibility, and you have to have prospective validation. You have to know what you’re talking about.
I remember when I joined the company [in 2002]. We said “We’re going to transform the industry. We’re just going to convince people this is the right way to do it.” I like that we did that. It’s the good kind of arrogance. It’s the Steve Jobs kind of arrogance. We never used the phrase “the customer is always right” at Schrodinger. We said “let’s just get the physics right and somehow good things will happen.” That might sound a little naïve, but I think it served us well. We stayed focused in a good kind of arrogant way.
LT: This is where naivete can be an advantage. Because if you really knew how difficult the thing was that you were attempting, you’d never attempt it in the first place.
RF: We’re not afraid of hard problems. We instilled that culture. Don’t worry if it will take five years. Don’t be afraid. That’s what it took. It also took patient capital.
LT: You’re a chemist. Trained at Caltech. You were on faculty at Rutgers, right?
LT: So you had a chemistry background, but not necessarily the software, is that right?
RF: I actually did have software experience. Do you know my brother Hany Farid? [a computer science professor.]
RF: He founded a field called digital forensics. He’s in the news a lot. He’s at Berkeley now, was at Dartmouth. He actually taught me a lot. I’m not a professional software engineer, but when I was at Rutgers, I developed protein design software. I was not only a chemist, but in my lab at Rutgers, I was involved in protein design.
LT: From scratch?
RF: From scratch. We were actually designing what we called “de novo proteins.” To study electron transfer.
LT: David Baker is known for this stuff now at the University of Washington.
RF: Yes. I developed software in my lab, and my brother helped me, as did one employee at Schrodinger who my brother knew before I did. Peter Shenkin was a colleague of my brother’s. They had developed side-chain prediction software. I had used that. I was not a professional engineer, but at least had some knowledge of engineering. And large molecules.
LT: In 2002, what was your mandate from the board or from investors?
RF: The company had focused mostly until then on small molecules. There wasn’t a lot of expertise in the company on protein structures. It turns out that around that time, Rich Friesner (Columbia professor and Schrodinger co-founder), had developed homology modeling software. But it was in an academic setting. I was brought in to understand the market, understand what people are doing with homology modeling, and productize it. I was brought in as a product manager. It was for what we call Prime now. The company had 30 people or so. It was a real opportunity for me to get my hands in a lot of stuff. Anything I saw that looked like a problem, I got involved. This was a place that allowed that.
LT: Was it always here in New York, close to Columbia and Rich Friesner?
RF: It was in Jersey City, right before I joined. It started there. Then moved to this building, where David Shaw had his whole hedge fund. That’s why we were here. We’ve been here ever since.
LT: OK, these early years, you’re getting in on the ground floor as a product manager. It’s still about software and services.
RF: It was just software.
LT: How does this thing begin to morph through the Great Recession years, and start coming out now with real drug discovery programs in this environment with SO much going on in biopharma?
RF: The technology just kept getting better and better. The science started to work. It became more and more predictive. We started to gain a reputation of just being a science company. A lot of pharma companies started to invest in Schrodinger. I think that’s the way they looked at it. It was a matter of “Alright, maybe the software isn’t working that great, but we know they are on the right path.”
LT: Was this a matter of layering in some of the other factors in discovery? Things like the entropy and enthalpy like you were talking about earlier? Running more calculations, covering more possibilities? Or did you have to blow things up and start over?
RF: It was more the latter. We had to return back to first principles. Here’s the thing. It all builds on itself.
For example, to get a free energy model to work, you have to have fast molecular dynamics. And you have to have an accurate atomic force field. And you have to have the ability to predict the protonation state of a molecule. And you have to have the ability to predict the protein structure. All of these things we worked on in parallel. It’s kind of like we were developing all the pieces, and then you have to put it together. You can sort of tell you’re getting the pieces right, because you can test the pieces against experiment. Then you can put it all together and test the whole thing as a big solution. That’s what we did. A little progress here and there. The computers would get faster. The pieces got more accurate. They (pieces) got faster as computers got faster, the algorithms got faster. There’s a lot of software that goes in here, it’s not just about computers getting faster. You have to keep making the software more efficient. We have a lot of engineers who think about optimizing code. The software just started to work. It started to get predictive. We keep improving the correlation between the calculation and the experiment.
LT: There was no one “aha” moment?
RF: Definitely not. Of course, at the time, it felt like things were going so slowly. Now it’s cool looking back, we can say it felt like it was overnight. But it was a real, real struggle. It took an incredible amount of patience, and this belief that it would work.
LT: Then how did Nimbus come about? You and Bruce Booth [Atlas Venture partner] got together …?
RF: It’s a really neat story. As we started to get a reputation for having technology that was working and that could impact projects, we started getting inquiries from academics who said “I can’t go run a $1 million HTS (high-throughput screen). I can’t do what pharma does.” Pharma of course can throw a lot of money at these things. Academics can’t.
They came to us and said “Will you help us with virtual screening? Can you help us with optimization [of small molecules] using some of this technology?” We said, “Sure, we want to gain experience. We’ll do it for free.” But of course, if it succeeds, you’ll pay us. But there were no upfronts. We were hungry for the experience. One of the groups we did it for, we made great progress with. That person starting pitching to VCs to start a newco around what we had done. One of those VCs was Atlas. Bruce Booth had just become a partner, or was about to. It was early in his career at Atlas. He liked the biology behind this company, but he thought, “Maybe we should do it ourselves.” Let’s do the experiment. Let’s build a company around the idea that we don’t have to do drug discovery around trial and error. Let’s use physics. Let’s use these predictive models. That’s how Nimbus came about. But it took years. There were a lot of things we had to struggle with. One was the domain of applicability. The technology then was pretty limited. We took about a year and a half just to identify targets that the technology worked on.
LT: What was the starting point here? Did you have good crystallography images to start with? Did you have a series of images you could use of the protein?
RF: Crystal structures were one of them.
LT: Through a series of images, a series of snapshots? I’ve heard Praveen Tipirneni of Morphic tell the story of how crucial the series of images is for them. Those came from Tim Springer, and they came later.
RF: Exactly. That was the key. We wouldn’t have been involved in the company without the images. That’s the key to a physics model. You need to have an accurate, high-res structure. Nimbus set out to do that. It took a long time to identify targets that were interesting from the biology, and had the – you call them “images.” I like that, but it’s really the structure, the coordinates of the atoms of a protein.
LT: This is a tangent, but it reminds me a bit of a time I went to visit Ariad Pharmaceuticals years ago. They’d show me an image of Novartis’ imatinib (Gleevec) on a computer screen. And show how that drug binds in its pocket. And how their drug [ponatinib] was a slight modification off of that backbone. It was supposed to treat some of the patients who were Gleevec resistant. It ended up being unsafe… and I think it was pulled from the market for a while.
RF: I’m glad you brought that up. That image you were looking at? That’s half the problem. That’s how chemists design molecules. They look at that image. They look at how the molecule bound to the protein. In a bound state. There aren’t any water molecules there. They aren’t considering what the molecule looks like in solution, when it’s apart. It turns out, binding affinity, one of the most important properties of a drug, isn’t only correlated to what the molecule looks like when it’s bound to the protein. It’s also based on what it looks like when they’re apart.
That’s what I mean by physics. A human can’t do that. What you were looking at is only half the problem. You can’t see the other half, the things that separate (the molecule from the protein). You can’t see the water molecules. That’s why you need physics, why you need molecular simulation, fast molecular dynamics. It’s why you need an atomic force field. To model the entire system.
LT: You’re looking at a snapshot. And life doesn’t work that way. There’s continuous motion.
RF: There’s continuous motion. Free energy of binding is the energy of the bound state, minus the energy of the unbound state. That’s physics. You can’t get around that. If you only look at one part of it, you’re missing the other part.
Solubility is the same thing. Taking a solid and going into a solution. Most people look at whether something is going to be soluble by looking into whether the molecule is going to like water. You also have to think about whether it likes itself. Whether it wants to be in the crystal. How soluble something is, is how unhappy it is in a crystal and how happy it is in solution. If you only look at half the problem, you’re missing the other half. Same exact thing.
By the way, there’s another thing humans can’t do. When they’re looking at a molecule, there’s no way that in their heads, they can figure out how the molecule is packing in the solid. That’s why we’re not so good at predicting solubility. Sometimes we get it right. Because there are some patterns. But often, it’s very, very difficult for humans. That’s where physics is important.
LT: I remember talking with Nimbus in their early going, and they had some classic medchem people who looked at this and said “Yeah, we’ve heard this sort of thing before. We’ll see. We’re still indispensable.”
RF: We’ve talked about this a lot, too. Humans can be great at understanding the problem. They are great at designing molecules. We are just saying if you have an algorithm that can predict the properties, let the algorithm do it. There’s a role for humans, a very important role. This is not a replacement for humans. Humans have to do different tasks. Of course, this always happens in fields when there’s a technology transition, humans do worry about that. They anticipate, “this computer is going to take my job.” They then become defensive. It’s natural. It happens in every field. But we work very well with many chemists at companies like Nimbus and Morphic. They eventually get used to there being a better way to do things.
LT: What was the key proof of concept at Nimbus where you and Bruce and others said ‘OK, it’s actually doing what we want?’
RF: Two things. One was at a local level when we were able to accurately predict the properties of molecules we were designing for ACC [a NASH drug candidate acquired by Gilead] and for the other targets like IRAK4, TYK2 and so on. What’s really neat – and this is where you convert people in chemistry – when the computers predict something that’s not intuitive. When it’s a surprise. Then it was getting to the finish line: Actually getting to a development candidate with not that many compounds made. It was done very rapidly, on a very hard target that pharma had failed on. That was pretty validating.
LT: What kind of predictions of properties were you able to make? Could you say “If we design our molecule this way in lead optimization, we’ll see side effect XYZ?” And then you’d actually see it?
RF: I wish it was that. That would be awesome. Side effects, we can’t do. It’s because we don’t understand the physics of some side effects. The physics-based methods are really good at predicting affinity, potency and selectivity, since that’s the same thing, just on a different target. It’s good at solubility. On a number of other properties, it’s not as accurate as on the ones I’ve just described. But we’re getting pretty good models…we’re getting pretty good at permeability. Then there are other properties where we have to rely on pure machine learning. Which means it works pretty well when it’s a local model and interpolating. It’s terrible when you’re extrapolating.
We still have surprises. We haven’t solved the drug discovery problem. There’s still a lot of properties we can’t predict accurately. That’s why we have a team of close to 200 people actually trying to solve those problems. [Schrodinger’s total headcount is close to 400.]
LT: What are you shooting for from hit-to-lead? Shorter time frame, less expense, higher degree of confidence?
RF: Let’s say there was just one property we could predict and everything else was completely random.
That one property could be potency, let’s just say. Here’s what you do. The way drug discovery is done, you design a molecule that’s potent, and sometimes you start trying to fix it. You start trying to now make it soluble. You have to make it so it’s permeable. All while maintaining potency. That’s the key. What happens because of steep activity cliffs and so on, usually what happens is you try to fix one thing and it becomes a whack-a-mole problem. That’s what everyone says. Multi-parameter optimization is whack-a-mole. Fix one property, you break another one. The ability to make sure that every compound you make has at least maintained potency, solubility, selectivity, some of the other properties – it allows you to play around with some of the other properties, even when you don’t have a good model for them. That’s why you need medchem people. They try to fix those problems. But before you make it, you run it through these models to make sure that before you go spending a month making a hard compound and spending thousands of dollars in assaying it, it at least passes these rigorous models.
It’s making things a lot more reliable.
LT: You didn’t need to make a million compounds to get one drug candidate.
RF: In some sense, you did make a million compounds. You just did it on a computer. If the models are accurate enough, it’s equivalent.
If you think of the chemical space, it’s almost infinite. What that means is, by definition, you have a lot of shots on goal. This is what’s really neat. Once you’ve determined a protein is druggable, and actually has a binding site, just think about – what are the chances you’ll find a molecule out of the 10 to the 50th power of possible structures that has all the properties you want? It’s pretty high, but not if you only make a few thousand compounds. A few thousand out of there – there’s a good chance you’ll get to the end of your program and not have something good enough. But if you can explore maybe a billion, a hundred billion, or a trillion [molecules], that’s actually what this is about. It’s about exploring more chemical space, so you have a higher chance of having a molecule that has all the properties required to make a drug.
LT: So you guys have proven your concept with Nimbus, somewhere in about 2012 to 2015. Then along comes Morphic, with a great substrate of images from Tim Springer’s lab, where you said “Let’s put the computers to work on determining structures against those integrin structure images.”
RF: Tim Springer came to us and said “I’ve heard good stories about what’s happened at Nimbus, I’d like to do the same thing, but I’m going to bring something amazing to the table.” That’s what he did. And chemists who bought into this idea that there’s a new way to do things. They really embraced it. The chemists who see this as an opportunity, not a threat. That’s the type of team that was built at Morphic.
LT: It was less than three years from company formation on a white board to entering the clinic with a big partnership with AbbVie. That’s fast.
RF: It’s very nicely validating.
LT: Now you’ve got more options as a business. You’re no longer just selling software. You’re an investor in companies that are worth something now (Nimbus and Morphic). Are you going to do more venture-type investments in the future? What about building your own in-house drug company?
RF: Let me be very clear. We are going to do both things. We are unbelievably committed to our software business. We think it’s critical that we continue to make all the technology available to pharma companies, which play such an important role in curing diseases. It’s almost an obligation we have to make all that science available to everybody. For a price of course. I’m not saying we’ll give it away, but we’ll charge what it’s worth. The value it brings. We’re really committed to that part of the business. It generates revenue. It funds the science. It’s the right thing to do. But the other thing that’s cool is we learn a lot.
We learn a lot by interacting with pharma companies, learning about their problems. It helps guide what we do. We are committed to the drug discovery side of the business to. How are we going to be a good software company if we don’t use our own software?
LT: You eat your own dog food.
RF: We eat our own dog food. I think it’s really important. We see these two as critical. In the same company, totally integrated as they are. Of course, there’s an IP firewall [from pharma customers], but not a know-how firewall, as in general know-how. That’s freely exchanged within the company. It’s interesting what happened. In order to make sure we were good partners, we found ourselves needing to hire medicinal chemists. Biologists who could interact with CROs. We did a deal with Takeda where they were expecting us to run a program. We started to build up that expertise of managing CROs, of picking targets. Of driving projects from the beginning – target selection – all the way to development candidate. We had that capability, and we naturally saw that as part of having a balanced portfolio, it made sense to have some internal programs. But not any one of these things will somehow take over the company.
LT: What about drug discovery for biologics?
RF: Great question. To the extent the things we have been talking about are physics based, it doesn’t matter if you’re designing a protein, a drug, or a material. The challenge is you need structures. You need high-quality structures. There are a lot more small molecule-protein structures than there are antibody-antigen structures. That’s one thing. But there are a lot of other challenges besides binding affinity. There’s protein aggregation. There’s viscosity. Immunogenicity. We don’t have very good models for those. We’re actively working on those. We have some projects on designing biologics, but we don’t have any current drug discovery, milestone-based or risk-based biologics programs. We’re not sure the science is there yet.
LT: The protein-folding problem is still there.
RF: That’s part of it. It’s a really hard problem. Protein structure prediction of big molecules is complicated. It’s an area we’re committed to. We have a lot of software our customers use that helps in that area. We don’t think we’re quite ready yet, or have the expertise, to have drug discovery projects there. But I think that will change in the near future.
LT: The business sounds like a three-legged stool. Software, some venture investments, and some in-house drug discovery.
RF: We’re de-emphasizing the venture investments. We’re not a venture capitalist. That’s not our business. We have in our agreements the right to invest in new companies. We have done it. But we will likely not be doing that very much going forward, if at all. The reason is because we have found that was important when we were still validating the technology. It’s validated now. We’re doing deals where we’re receiving a significant amount of equity.
LT: Like with Takeda and Sanofi?
RF: In those, we don’t have equity.
LT: They are fee-for-service?
RF: With Sanofi, it’s more like that. With Takeda, those are our programs, and they have an option to acquire the development candidates.
LT: That’s you being a drug discovery company.
RF: Yes. And with Faxian, the newco with WuXi, we have a more meaningful equity stake. The intent is not to invest cash, it’s to invest our services. In exchange for the equity we have. We think that’s a better model.
LT: How did Nimbus and Morphic’s success affect the software side of the business, in terms of Total Addressable Market penetration?
RF: It wasn’t just those two, it was all the partners. When you really get into a drug discovery project, a real one, where you actually have to develop a drug, you learn so much. About what problems need to be solved. You get the data. It’s had a profound effect on what direction the company has gone, what problems we work on, and our ability to actually solve those problems. It’s because we had real-world testing. In real-time. With a pharma company, it’s not because a partner doesn’t want to share data freely. They can’t. They have all kinds of restrictions. It’s very difficult for them to do that. Here [Nimbus and Morphic], the teams are interacting with each other every day. We have access to the data as it’s being generated. That was really, really useful for advancing the technology. Then the world benefitted from it.
[Companies we work with] like that, actually. Nobody wants to work with a company that’s only working with just them.
LT: Why raise more capital?
RF: We see an enormous opportunity ahead of us. We have in a sense de-risked the science. Now it’s time to invest heavily. By investing now in the science, it will accelerate the rate of breakthroughs. But that’s what always happens. You get the basic science right, and now you can turn the crank, almost, and start solving really hard problems, because the underlying infrastructure is established. The other reason is we are really excited about what we think we can do with our partners. On the internal drug discovery side, it’s time to start growing that group. The same thing with materials. That’s doing well, and we want to invest ahead of revenue.
LT: We’re coming off a year where a gazillion companies got funded AI for this and that and drug discovery. 90 percent of the time, I don’t even know what they’re talking about.
RF: Neither do we.
LT: They haven’t really explained the problem in a way I can understand, like with using computers to run simulations around protein structures and small-molecule binding optimization. This I get in concept. But do you hear competitive footsteps?
RF: We don’t. We have an AI group. Everybody can do that. A lot of that code, all of it as far as we know, is open-source. Google is open sourced. Vijay Pande’s DeepChem, which we’ve integrated into our software, everyone has that. We don’t think – AI is not magic, machine learning is not magic, deep learning is not magic. Those algorithms are widely available. It’s about understanding the problem and the data.
LT: What about the quality of the underlying data?
RF: Also incredibly important. We have found that a lot of AI companies describe what they’re doing as a black box. Everyone should get very nervous when people talk about science as a black box. People should be suspicious. Everything we do is published. We understand what it is. It’s physics. Machine learning has a role, for sure, and we should be transparent about what it is. It’s not magic, and no one has a competitive advantage there.
LT: Who do you consider competitors?
RF: At the moment, there are no commercial competitors in the physics space that we see. We do – there are some academic efforts. We try to engage with them. We work with some. We have a strong SAB with lots of academics who benefit from those interactions. But we have a massive effort that is very difficult for academics to duplicate. We have a commercial environment. We have amazing and patient investors. We have professional software engineers who are here for a long time, not a short time, which you see in academia, like in the grad student or postdoc life cycle. I’m trying to be very clear that it’s important that academics work in the space.
LT: But that could be a threat in the future, if they dump solutions to half the problem in the public domain.
RF: It can be. But that goes back to the question you asked before, which was “Why raise the money?” It’s to continue to invest aggressively in the science. That’s the way to stay ahead.
LT: To maintain your lead.
RF: To maintain the huge lead we have right now. If we became complacent and said “Oh, we’re so smart, patting ourselves on the back and saying we solved the problem,” we’d wake up one day and there would be someone knocking on our door. We’re not doing that. We’re not arrogant. There’s a lot more science to be done. We’re investing really, really heavily to maintain the lead.