Listen on
Overview
Everyone says AI is transforming drug discovery, and yet a new drug still takes about a decade and billions of dollars. Both statements are true at once, and the reason is specific. Large language models are transforming the work around drug discovery, the coding, automation, research, and regulatory writing that all revolve around words. They barely touch the hardest part of the problem, which is predicting how a molecule actually behaves in the body. That prediction is physics, not language, and it is where the real work still has to happen.
Woody Sherman is Founder and Chief Innovation Officer at PsiThera, a biotechnology company designing oral small molecules for immunology and inflammatory diseases, starting with the TNF superfamily. His career runs from physical chemistry at MIT, through more than a decade at Schrödinger building computational tools the industry still relies on, to founding Silicon Therapeutics, where his team took a small molecule from concept to clinic in roughly three years on a target most people had written off. He scaled that platform at Roivant and now leads the science at PsiThera. He has published more than 100 peer-reviewed papers across molecular dynamics, quantum mechanics, free energy simulations, and machine learning for drug design.
In this episode, host Ross Katz and Woody work through why language models stall at the molecule, what physical AI means as a category built on quantum-mechanical embeddings rather than language tokens, and why molecular representation (SMILES strings, 2D graphs, and PsiThera’s quantum-mechanical Psiformer) is the whole game. They dig into the STING program at Silicon Therapeutics, where the team engineered a small molecule to bind to itself through a quantum-mechanical interaction that had not been designed for before, why the field’s excitement over fast co-folding models like Boltz-2 needs a dose of skepticism, and what it takes to build a database and a team culture where chemists, biologists, and data scientists can understand each other.
Key Takeaways
Large language models transform the words around drug discovery, not the molecules
Language models are genuinely strong at coding, automation, research, and regulatory writing, the parts of drug discovery that revolve around text, and that work is getting faster. Predicting how a molecule behaves is a different problem, because a molecule is physics, and words cannot describe how it binds, moves, or clears the body. Sherman’s view is that more data will not make language models fluent in molecules; the field needs a different kind of model. He also names the harder truth underneath the hype: the low-hanging fruit is largely gone, so remaining targets are harder and the biology less understood, which no tooling fixes on its own.
Physical AI represents molecules as quantum-mechanical objects, not language tokens
A language model’s tokens are words and word-pieces. A physical AI model’s tokens are orbital-level, quantum-mechanical representations of atoms and molecules. Sherman’s argument is direct: quantum mechanics is the fundamental nature of reality, so a model that describes molecules that way can be both more accurate and better at extrapolating into chemistry no one has made yet. That is also what separates physical AI from a plain physics engine, which is the next point.
Molecular representation is the whole game
Most machine learning in drug discovery has treated a molecule as a SMILES string or a 2D graph. Both are convenient and both are lossy: they throw away the 3D shape, the dynamics, and the quantum-mechanical properties that decide how a drug works. PsiThera’s Psiformer learns an embedding directly from quantum-mechanical calculations rather than a human-designed fingerprint, so the representation carries the physics instead of an abstraction of it. A model can only reason about what its representation preserves.
Physics-based simulation is accurate but never learns; physical AI learns from it and scales
A physics-based free energy simulation is rigorous, but it is just as good on day 1,000 as on day one, because it runs the same equations of motion every time rather than learning. It also costs roughly a day per molecule on a GPU, so it cannot screen millions. Physical AI keeps the physics but uses it differently: it trains on the outputs of those simulations, and on wet-lab data as it arrives, which lets it run millions of predictions per second and, over time, exceed simulation accuracy by training on the endpoints that actually matter. Before any wet-lab data exists, a team can even fine-tune a fast model against physics-based simulation data to get it started.
Concept to clinic in three years: define the target, then build only the tools it needs
At Silicon Therapeutics, Sherman’s team took a small molecule STING agonist from concept to clinic in roughly three years, a target the field had largely written off. It is hard because it requires activating a pathway rather than blocking one, locking a flexible protein into a productive shape, and escaping the poor drug-like properties of the natural cyclic dinucleotides. Their answer was a much smaller molecule engineered to dimerize on itself through a quantum-mechanical pi-pi stacking interaction, which Sherman believes had not been done before and which required a new quantum-mechanical method built on the fly. His philosophy: do not build a hammer and go looking for nails. Figure out the house you are trying to build, then build only the tools that house requires. A lean team of about ten computational and ten wet-lab people did the rest.
Treat benchmark claims skeptically, especially outside the training distribution
The widely shared claim that Boltz-2 delivers near-FEP-quality binding affinity at roughly 1,000x the speed did not hold up over the six months that followed. It works where you already have data around the target, but drug discovery usually operates in the novelty regime, where co-folding models have little to say. Benchmarks are frequently contaminated by data leakage: as Pat Walters has documented, even a clean date cutoff can let a “new” result be a lookup of something published years earlier. The practical move is to check whether your system resembles anything in the training data (PsiThera and collaborators published a tool for exactly this on the OpenFold portal) and to fine-tune on local or simulated data when it does not.
The database has to understand chemists, biologists, and data scientists alike
Sherman keeps returning to a point most tooling misses: computers do not make drugs, teams do. His teams capture simulation outputs, wet-lab results, and human chemist judgment (the votes and written comments experienced drug hunters leave on molecules) as structured data that generative models can learn from. Making that work took a shared vocabulary across computational and experimental staff and a relational database designed from the start to hold the molecule as the central entity, with targets, assays, and human decisions connected to it. An experienced medicinal chemist as the North Star, not the model, is what keeps the whole effort pointed at a real drug.
Related: CorrDyn helps biotech and life sciences organizations build the data infrastructure and machine learning systems this kind of work depends on, including the evaluation discipline (benchmarking, validation, and honest uncertainty) that separates a model that is genuinely useful from one that only looks the part.
Full Transcript
Jason: Welcome back to Data in Biotech, the podcast where we dig into how data and computation are reshaping the life sciences.
Ross Katz: I’m Ross Katz, principal and data science lead at CorrDyn. Here’s a contradiction a lot of you have probably felt. Everyone says AI is transforming drug discovery, and yet drugs still take a decade to develop and cost billions of dollars. Both things can be true at once. Yesterday, I caught up with Woody Sherman, who’s a computational chemist who has spent over a decade at Schrödinger building the tools the industry still uses, founded Silicon Therapeutics, where his team took a molecule from concept to clinic in about three years on a target most people had written off, scaled that platform across an entire portfolio at Roivant, and now leads the science team at PsiThera, designing oral drugs for immunology targets that today only injectable biologicals can hit. In this conversation, we get into why language models are brilliant at code and paperwork but hit a wall when it comes to molecules. What Woody means by physical AI and why he thinks drug discovery needs a different kind of model than the one that’s powering Claude or ChatGPT. How his team represents molecules as quantum mechanical objects instead of text strings and why that matters. And the very human problem of getting chemists, biologists, and data engineers to speak the same language and why the database has to understand all of them. If you’ve ever wondered where AI genuinely helps in drug discovery versus where it’s oversold, this one’s for you. And with that, here’s our conversation with Woody Sherman. Woody Sherman, welcome to Data in Biotech.
Woody Sherman: Great, thank you.
Ross Katz: Well Woody, your career has spanned physical chemistry at MIT, a decade-plus at Schrödinger, founding Silicon Therapeutics, scaling platforms at Roivant, and now at PsiThera. Can you walk us through the thread or threads that connect all of those chapters and what problem or problems you kept running into that pulled you from one stage to the next?
Woody Sherman: Yeah, absolutely. Thanks for the opportunity to be here. I think the underlying thread has been how do atoms and molecules influence biological systems. And in undergraduate, I started by running a bunch of quantum mechanical calculations for some research on organic polymers, looking at nonlinear optical properties, which was super interesting and deep in the weeds of quantum mechanics. But I saw a presenter talk about molecular dynamics simulations and how proteins move and how that can be used to explain biological function and mechanism. And that got me hooked. And so I went to grad school in physical chemistry to continue those types of simulations. But while I was there, my advisor ended up moving into the Computer Science and Artificial Intelligence Lab, CSAIL. And so I’ve been at that interface of more traditional physics-based simulations, which are super powerful but complicated and slow and don’t really learn as they evolve, and the AI world, which at the time, decades ago wasn’t nearly what it is now, but one could see that there was promise and it’s all starting to come together now and it’s always been this, the core thread has been how do molecules, how do you design molecules to modulate biology and other properties, basically molecular systems and predicting properties of molecular systems using computational simulations and tying that all together in the real world through experimental validation and iterative design-make-test cycle. Yeah, so at PsiThera, we’re a fully integrated therapeutics company and so we have a wet lab with biology, chemistry, structural biophysics, etc., right here, but also a very strong computational platform that we’ve built over the years and continue to develop, and we’re focused on oral medicines for challenging immunology diseases. We’re focused primarily on validated targets where there’s antibody solutions, so injections, and we’re focusing on oral medicines. So think injection in a pill, taking well-known biology but that’s been hard to modulate with traditional oral molecules and using this computationally heavy simulation approach integrated with the wet lab to design drugs for these challenging targets.
Ross Katz: Yeah. So, in your background, you took a small molecule as I understand it from concept to clinic in roughly three years at Silicon Therapeutics, targeting STING, which is known to be an extremely difficult target. Can you talk to us a little bit about the engineering or computational scaffolding that made that timeline possible?
Woody Sherman: Yeah, so that was a great program because it did really combine many aspects of the computational and wet lab capabilities and all the accumulated learnings over the years. So STING, we were looking for a STING agonist, which is something that activates a pathway. So a lot of drugs that people think about are antagonists that block a pathway, which it ends up that there’s a lot of ways that you can mess up a pathway. It’s a lot harder to activate the pathway. And so that was one of the key challenges and it was a motion-based thing. STING is a flexible protein that undergoes a conformational change in this activation process and so we had to not just bind, but bind and lock it into a conformation that was productive for signaling. And the third bit that was particularly hard there is that the endogenous molecules that bind to STING are cyclic dinucleotides. And these are big, ugly molecules that don’t really have drug-like properties and people historically that have worked on STING have started with these cyclic dinucleotides and they could find molecules that work fine in a cell but they didn’t work as drugs. They didn’t have drug-like properties, they got cleared out of the body before they could do anything to the target. And so we took a very novel approach. We found a small molecule that was much smaller than these cyclic dinucleotides that had drug-like properties and we got it to dimerize on itself, bind to itself, through a quantum mechanical pi-pi stacking interaction, which I don’t think has been done before. And so in order to do that, we had to run a lot of quantum mechanical simulations, designing molecules that enhanced this intramolecular interaction. We had to run a lot of molecular dynamics simulations understanding how the protein moves in the context of these molecules binding. A lot of structural biology and biophysics, a lot of integration with the lab, the iterative design-make-test cycle, and building an organization where all the people were bought into this idea. Computation can be super powerful when used appropriately, but you can’t make drugs on computers. You can come up with ideas, you can triage vast chemical space, hypothesis generation is super powerful. But designing that design-make-test cycle still requires making and testing in the lab. And so building that company from scratch with the right people that were experts in drug discovery who were fully bought into the computational platform approach was super important as well. And so there’s a lot of technical and scientific aspects in the platform that made it possible. There’s also the human psychology and behavioral team dynamics that came together really nicely and, yeah, we were able to crack this nut that I think is still unprecedented for designing a small molecule human STING agonist.
Ross Katz: Yeah, and as I understand it the CDNs that you were replacing were also only able to be administered using an injection and then the small molecule approach you turned it into a pill so it’s sort of analogous to the work you’re doing at PsiThera today, is that right? So the combination of the wet lab and the dry lab, the computational approaches and the wet lab, and then also the serendipity of having the right people and the right team together. Can you maybe speak a little bit about what uniquely worked about that team or that approach that you take into your work at PsiThera?
Woody Sherman: Yeah, well, team is so important when you’re doing anything, but the harder the problem is and the more challenges, I think the more the right people and the right team dynamics is essential. I’ve always felt that way, we knew that from the beginning. It was a unique moment in time in early 2016 where this computational approach and AI-driven drug discovery was still quite nascent. And so there wasn’t a big pool of people that were out there that were interested and that would believe and leave a big pharma company or a more traditional setting to join. And so there was a self-selection process that took place in that the people that were interested in this concept were the forward-thinking folks from the lab that could see the value of computation. But also, we didn’t want people that were blindly diving in head first without knowing how deep the pool is. And so it requires the right kind of balance of experienced people that have that right skepticism and understand that the computers alone are not going to do everything. And so some of the folks came out of the founder’s labs out of Harvard Medical School, and that was super valuable to have those connections. Many of the computational folks came out of my network and then a lot of talking at conferences, getting the word out, publishing papers, and letting the science really attract the right people to the company.
Ross Katz: Yeah, that makes sense. And obviously three years from concept to clinic is very fast. I’m interested in both what were the main bottlenecks that made it so that it didn’t move even faster, whether it be compute or synthesis or wet lab assays or human judgment or just the thought processes that needed to work themselves out, and also what were the bottlenecks that you didn’t experience that you felt allowed you to go from concept to clinic that quickly?
Woody Sherman: Right, so I think one of the lynchpins for most of the folks getting into drug discovery from more of the technology or AI front is that you do need an experienced card-carrying medicinal chemist who’s made drugs before. There’s a lot of things that we can predict pretty well on the computer, but we can’t predict everything and biology is enormously complex and human experience is still essential and there’s a lot of things about not just what’s the right molecule now, but what’s the right molecule that can go to the next step, that can go to the next phase, that can be manufactured and produced and all of these other pieces that go into making a drug. And so Lijun Sun, who was one of the founders of Silicon Therapeutics and at Harvard Medical School, has done some phenomenal work in industry and in academia and having somebody like that really as the founding medicinal chemistry drug hunter set the right foundation for the experimental work, the wet lab portion, making sure we had the rigor in the lab and the experience to execute on a drug discovery campaign beyond just screening molecules and seeing if they bind, but thinking about all those other properties that are required to make a drug. So it’s essential to have your North Star. When you start a project to know what are you aiming for? You’re not just wandering around in the dark doing cool calculations and making molecules and seeing if any of ’em looks like a drug. You really have to know what your objectives are and drive the project toward those objectives. And so that was one of the key things is that we knew at the beginning what we wanted to do, what we were looking for, and building the right assays so that we could be measuring what matters as we go through these design-make-test cycles. The computation was nascent when we were building it but we brought in the right people and we built the tools that we needed on the fly to solve our problems. And so rather than building some big, complex computational infrastructure and then looking for a problem, building a hammer and trying to look for things that are nails, we wanted to build a house and we had to figure out what was the landscape, what was the environment where we were trying to go, what were the aspects of the house, the architecting of the house and then build the appropriate tools to solve those problems. And so for example, that quantum mechanical, that dimerization, that pi-pi stacking interaction between the molecule binding to itself in the context of the protein binding site is unprecedented and there were no tools out there to do that and we wouldn’t have anticipated that before getting into the project that that was the right approach and so we developed a quantum mechanical method on the fly that worked exceptionally well. It was only needed for a few months of the project but it was really a critical part of the project where we could enumerate all of the substitutions on this molecule to make this pi-pi stacking interaction as strong as possible and then building out the other tooling once you have this dimer molecule in a protein. None of the other commercial tools out there work for that ’cause it’s never been done before and there’s no reasons for somebody to develop a tool for that. So we had to build out the molecular dynamics infrastructure and the free energy calculations and the machine learning AI capabilities all around that particular problem. And so we built the tooling, first of all a strong infrastructure as is always essential, having the right compute resources. We built our own GPU cluster because this was before the cloud was as accessible as it is now. And so we had to build the right infrastructure hardware, we had a great HPC team, all of those pieces had to come together. And then build, as we encountered problems as the project progressed and we identified the issues, we had to build out those capabilities. This was in the very early days of AI. We had some phenomenal young AI researchers who had just come out of grad school and postdocs and as we collected data on the program we built the AI tools necessary to address those problems. And then that accumulation has then built the foundation of the platform that we continue to use and develop, but at the time it was really build what was needed to solve the specific problems. As a startup you have to be laser focused on the problems that matter. We can’t be just thinking about here’s all the cool things that maybe useful in the future. We were a small team, a lean team, lean but mighty team of about 10 computational people and 10 in a wet lab folks and doing whatever it takes to solve the problems that arose on that given day of the project.
Ross Katz: Yeah, that’s great. And I love both sides of the coin that you gave there. There’s sort of the having that highest possible quality medicinal chemist who serves as the human in the loop, or the leader of the humans in the loop to guide all of the computational work in the right direction and make sure that you’re seeing all of the things that the computational work can’t see and then obviously having the crack team of computational people who can build the tools you need on the fly. So both together is what enables the level of success that you were able to achieve in the time you were able to achieve it. So that’s great. Previously when we were talking you described this contradiction that I think has come up on this podcast and a lot of our listeners might feel, which is that AI is supposedly transforming drug discovery, yet drugs still take a decade to produce and cost billions of dollars. In your view why are those two statements both true at the same time?
Woody Sherman: So, before we get into the technical aspects of what AI is good for and what it’s not good for, one of the things that is important is that a lot of the low-hanging fruits in drug discovery have been picked already. And so targets that we’re going after are getting more challenging, the diseases that we’re going after are more complex, the biology that we’re trying to unlock has a lot of unknowns. And so even if the tooling is getting better and better, the problems that we’re going after are getting harder and harder, and so the whole regulatory system and a lot of those other pieces keep getting more and more expensive and bogged down and there’s a lot of things that computers just aren’t going to solve that make the cost continue to go up. But on whether AI is transforming drug discovery or not, they’re both true at the same time. There’s certainly things that these large language models are enhancing undoubtedly across the ecosystem. So automation, orchestration, research, reading, writing, anything that has to do with words and perhaps most importantly is coding. And the language models are phenomenal with coding. People that never had any coding abilities and practice or training can ask questions and get code that solves specific problems and that’s amazing and we use that heavily and I think everybody is using that heavily now. And so there are aspects of drug discovery where that helps. The aspects like regulatory filings, a lot of the paperwork, all those things language models are great. Where language models don’t have much in the way of strength is dealing with molecules. And drug discovery is all about molecules. How do molecules interact, small molecules, large molecules, targets, biology, DNA, RNA, cells, all those things you could use all the words in the world and you’re not going to be able to describe those processes. And so that’s I think where the gap is that folks are really focused on the automation and streamlining things and all of the language-based tasks that are performed exceptionally well including generative chemistry type stuff where you can explore chemical space, you can come up with lots of ideas, but then changing all of those ideas or prioritizing all of those ideas based on accurate models, actual accurate predictions is still a major gap. And I think that’s the area that folks need to fill. Many people are hoping that just you keep throwing more data at it and these language models will figure it out. I disagree with that approach. I think we actually need a new kind of AI for drug discovery, in the pre-clinical discovery stage where you’re making molecules, and that AI is physical AI. It’s building world models for the nanoscale. And we’re working on that, there’s a very small number of emerging companies that are working on that. I suspect some of these larger pure tech LLM type companies are going to be moving in that direction. We saw Anthropic, for example, just hired John Jumper, who came out of DeepMind, who won the Nobel Prize a couple of years ago. That approach has very little to do with language models. There’s some language stuff in there in evolutionary sequence information etc., but he’s really an atom guy. He came from D.E. Shaw Research, we were in the same building together when I was at Schrödinger. He understands atoms, molecules, molecular simulation, how things move and how to tie that into AI. And so I think that we are in a very exciting new space where folks have seen the value that comes out of these language models. Many are just going all in with the language models, but I think there’s a emerging group of people and companies who are saying we need a new kind of AI and that’s what people are hearing more and more about, physical AI and world models and what we’re talking about specifically is that nanoscale world model dealing with atoms and molecules.
Ross Katz: Yeah, so maybe you’ve just defined the term, but since we’ve gone there, physical AI and specifically nanoscale physical AI is the paradigm that you’ve said can close the gap between what we want AI to do inside of drug discovery and what it’s doing today. Can you maybe give us a richer definition of what nanoscale physical AI constitutes and what makes it different than traditional physics-based simulations that might be out there. I think you’ve already really distinguished it nicely from language models.
Woody Sherman: Yeah, so the first distinction from more traditional language-based AI is that the representations or the embeddings of a language-based model are words, these tokens, these words, letters, pieces of words. What we’re talking about in the molecular physical AI space are tokens, representations, embeddings that have to do with atoms and molecules. So getting away from words and really thinking about quantum mechanical descriptions of molecules where the tokens are essentially orbital type of objects. And so we are describing molecules as this combination of a bunch of orbitals and quantum mechanics is the underlying principle that defines everything if there’s one thing that I think we can all agree on is that quantum mechanics is the fundamental nature of reality. And so if we can explain things that way, we’re going to be able to make better predictions, we’re going to be able to both be more accurate and be able to extrapolate more. And so now where does that differ from the traditional physics-based simulations? Physics-based simulations rely ultimately on a quantum mechanical understanding of the system. There’s a lot of approximations often that have to go into place to model a protein, which is too big to model with pure quantum mechanics, and so you either model part of it with quantum mechanics and then make approximations in other place or build simplified models. But when you’re running these models, the models that you’re running really are based on physics and they’re just as good on day one as they are on day 100 or day 1000 of the project. They’re not really learning. They’re just running these Newton equations of motion or some other formula-based approach. So we’re using equations of motion and mapping out how molecules move over time. And while that is a rigorous and accurate way to do things, it’s also computationally very expensive. The physics-based free energy simulations that we run cost about a day per molecule on a GPU. And knowing the cost of GPUs etc., we can’t be looking at millions of molecules a day using these rigorous physics-based simulation approaches. And furthermore, as our programs evolve and we collect data, we should be using those data. And so the idea of physical AI is to use the principles of physics and physics-based simulation, but not to directly simulate the system, but to extract embeddings that describe the system and then you could work more in the latent space to do a lot more, both generative chemistry as well as predictive modeling based on data that either has been already generated previously, either in the public literature or in-house, or data that’s arising over the course of the program.
Ross Katz: Yeah, that was a really rich and interesting description. So thank you very much. What comes to mind for me is, does this mean that a lot of the art of constructing these models is figuring out how to construct the embeddings that describe the quantum mechanics and how to featurize that space in a way that is meaningful for drug discovery?
Woody Sherman: Yeah, so that’s been a lot of the historic challenge. That’s exactly right. Folks have generated fingerprints of all different flavors and graph representations etc., trying to simplify molecules from these complex three-dimensional dynamic objects to a machine-interpretable language, which tends to be overly simplified. And so that’s been one of the biggest gaps in the field so far is that folks that have been applying machine learning approaches to drug discovery typically treat molecules as a linear string, a fingerprint, or sometimes now more emerging a 2D graph. But those both still have big limitations when we know the reality is molecules are these three-dimensional quantum mechanical objects. And so what we’ve done most recently within PsiLabs, which is our AI lab within PsiThera, is to build out what we call the Psiformer, which is a true quantum mechanical embedding of molecules. And so it’s a learned embedding based on these orbital-like representations of molecules. We train on quantum mechanical calculations. And so it’s no longer like a human-tuned or human derived extraction of things that we think are related to the molecule, it’s just representing the underlying quantum mechanical object through these orbital-like embeddings. And then that’s used, so we train on that, and then we can use that to apply to fine-tune on various different endpoints. And I think that’s the direction the world is going to go. We started with these linear SMILES strings and fingerprints. They were not great but they were useful. And as folks have moved more into the 2D graph representations, that’s closer to what molecules look like. They’re definitely more like graphs. But that doesn’t encapsulate the 3D, it doesn’t encapsulate the dynamics, and it doesn’t encapsulate any of the quantum mechanical properties of molecules. And so we’re going straight to that endpoint. And I think that’s what’s going to really open up this nanoscale physical AI revolution in drug discovery. That plus other types of approaches, the co-folding approaches that were pioneered by AlphaFold, for example, that we’re continuing to push and democratize through the OpenFold consortium, lots of other things going on, but I think really those are the key pieces is understanding proteins as proteins, as three-dimensional objects, and other biomolecules as that, and understanding drugs and small molecules as these quantum mechanical entities, which you can use then to train on endpoints of interest based on public or internal data.
Ross Katz: Yeah, that makes a lot of sense. So when we were talking previously you made this analogy of aerospace where generative AI can produce novel aircraft designs, but the engineers still rely on physical models to determine whether a design can actually fly. How does that analogy map onto drug discovery and where do you think it breaks down?
Woody Sherman: Yeah, so it’s very similar in a lot of ways. With airplanes for example, they come up with a lot of ideas and the physical models tend to be based on a lot of these traditional physics models of materials, of strength and aerodynamics etc., and then they make pieces and they still put things in wind tunnels and measure the stuff that’s still overly complex for a pure physical model to understand. And we’re in the same type of regime with biology. I would say that biology is substantially more complex than aerodynamics, and so there’s a lot more moving parts, there’s a lot more underlying formulas, there’s a lot more complexity, there’s a lot more interactions between many more objects in the system. But conceptually it’s that same type of idea. And if the airplane is still of that level where we can’t just build it all on a computer, drug discovery is even more so that. And it requires that iterative design-make-test cycle. So you can iterate through a lot of designs with a airplane computationally and ultimately you have to do some testing in the lab. And the same thing on the drug discovery side. We can iterate through a lot of ideas computationally. And the question is, how do we go from these 10 to the 60 possible drug-like molecules that people talk about to the next five molecules that we’re going to make in the program. And so it’s a bigger space to explore, the interactions are more complex to model with quantum mechanics, the biological systems are more complex. Human biology is the most complex thing that we know in the universe and there’s just so many facets of it that you need that experimental testing. But if you can pull these pieces together, like the airplane you should be able to make a lot better products a lot faster, a lot more efficiently.
Ross Katz: That’s interesting and as I was hearing you talk, it occurred to me that when we’re designing an airplane, we’re designing an airplane for one specific purpose, which is to fly. And what you were describing earlier with your work on STING was that you weren’t trying to create a platform to solve every biological problem. You had a focus on a particular biological problem and you created the tools and processes that you needed to solve that problem. And so when you narrow the scope of the problem you’re trying to solve, maybe it is more like designing an airplane because you know what flying looks like for the particular molecule that you’re trying to design, but when it’s open-ended and you’re trying to design a method that solves all of biology, then the complexity and the scope and the high dimensionality of the space just is there to bite you.
Woody Sherman: Yeah, that’s right. Yeah, no, I think the more you can define your problem in advance the better. And at the same time, there is common tooling and the tools that we did build for STING, it’s not like we throw those away when the project is done. That becomes part of the foundation of what we’re building and then we build further on that and many of those tools continue to be used today and have been built on today. And so, but the number of problems is vast and yeah, we can’t, we don’t, we’re not at a, the field is at a stage of having a button that you press to get you to a drug. Making a drug has a lot of different challenges and depending on the disease of interest, and what are the indications you’re going after and what are the landscape of both competition and patients, it introduces different types of challenges. And so we published a paper a few months ago on the grand challenges for predictive modeling in small molecule drug discovery. Each aspect of drug discovery are going to have these grand challenges. Antibodies and biologicals and cell therapy and gene therapy, so depending on your approach there’s these different types of problems that arise and most importantly is understanding what are these problems that you’re trying to solve and then either just starting drug discovery and addressing those as they arise, which is what we started with and now where we are now is trying to build that complete toolkit to solve those problems. So a priori you could come into ’em. And it’s not an infinite number of problems. We know what’s needed to make a drug. You need to engage with the target, you need to be selective against potential off-target liabilities, you need to be soluble and permeable, you need to be in the body for the right amount of time and not be metabolized and cleared. So if we can have predictors for these various different endpoints that we’re going after, then you can start to think about a true generative framework that can explore chemical space for the disease of interest. But what we first need is to have not just the automation and the generative tooling, but to have those predictive models under the hood so we can drive toward our objective.
Ross Katz: Yeah, that makes a lot of sense. One thing that you mentioned as you were describing nanoscale physical AI was this idea that traditional physics-based simulations like the kind you helped build and commercialize when you were at Schrödinger are accurate but slow and critically they start from scratch each time. So they’re not learning from the data that you’re feeding to them. So I’m interested in both how and what changes when you build a physical AI system that learns from physics-based data.
Woody Sherman: So I think the things that change are first of all that there’s a lot of upfront training that needs to take place. With the traditional physics-based simulations the work goes into building the model in a different way. It’s building the equations, the equations of motion and the parameters associated with that, and once you have that you run it. With the physical AI world you first have to think about what is the embedding, which is in a sense the same thinking as how are we going to run the simulation, what are the equations of motion etc. And then the training is the big piece that comes in which is, what are the data that we need in order to train these models, which in our space ends up being some combination of synthetic data, in silico generated data, and physical data. And I think we’re moving more and more into this in silico data generation space where with those right embeddings, again the right description of the molecule, that is really quantum mechanics. And it’s one of the few areas of the world where you could actually get more accurate information from a simulation than an experiment in many cases. When you get down to that quantum mechanical level, that is the description of the molecule and it’s as good as it gets. And it’s not just a single number, it’s a full three-dimensional electron distribution.
Ross Katz: Yeah, as I understand it that’s because the measurement tools that we have at our disposal, the level of error in those measurement tools is such that the quantum mechanics are as good or close to as good as actually measuring it in the lab.
Woody Sherman: That’s right, exactly. And when you start to measure a molecule and you do it through, if you want to get the structure through crystallography, for example, you end up having a frozen representation of that. With simulations we could have dynamic representations of those and we can see, we can actually extract a lot more information. We can extract essentially the whole wave function for the molecule and with the wave function you can describe all properties of the molecule. And so the quantum mechanics of the data generation component ends up being potentially a very large investment that requires a lot of expertise. But then once you have that representation and that understanding of molecules, now you can start training as data come in. And so generating data for the project that’s going to make those models more accurate. So where the old physics-based simulations, once you built that initial model, that’s what you got throughout the course of the project and throughout the course of all the other programs and maybe you could tune some kind of settings in how you set things up and how you deal with your protein protonation states and little things like that, but you’re not truly learning as the data comes in. In this new paradigm, the models can’t make predictions without some data to train on. But once you start having those data, the models get more and more accurate and not just the accuracy which can supersede the physics-based simulation accuracy because you’re actually training directly on the true endpoints that matter, but you also get scale. And so rather than being able to simulate one molecule per day per GPU with the traditional physics-based simulations, we can now do millions per second. And that’s a game changer.
Ross Katz: Right, so even if in principle the oracle or the surrogate model that you’re using to evaluate the candidates that you have available is less accurate than the simulation would be if you ran the one per day, when you’re running millions and you’re able to explore space that much faster, you’re able to move toward your solution and know which data to generate next that’s going to get you the information you need to get closer. Am I thinking about that right?
Woody Sherman: Yeah. As you accumulate data we’re finding that the AI models do get to be more accurate than the physics-based simulations because you’re training directly on the endpoints that matter. And we have to remind ourselves that our physics-based simulations still have a lot of approximations that go into them. So we’re not actually running quantum mechanical simulations of the entire system that matters. We’re making approximations. These one-day simulations still have a lot of approximations. We’re using these Newtonian force fields to describe molecules, which are accurate and useful, but they’re not quantum mechanics. They’re approximations to quantum mechanics. And so until we have a quantum computer, for example, that can simulate a protein-sized system and all the associated dynamics and interactions of that, the physics-based models have intrinsic limitations in the types of interactions that are not modeled well. When we go back to the STING story at Silicon Therapeutics, we ran a lot of those traditional molecular dynamics simulations, but we still had to drop down to pure quantum mechanics to understand that pi-pi stacking interaction, which is something that just doesn’t get represented with the simplified force fields.
Ross Katz: That’s very interesting. Six months ago Recursion and MIT released Boltz-2 under an MIT license with the headline claim of near FEP quality binding affinity at roughly 1000x the speed. If that claim holds, then it sort of pressures the case for needing physics-based simulation in the loop at all. So I’m interested in, from your perspective, where would you trust a model like that in your workflow today and where does it break or would you expect it to break when incorporating it in?
Woody Sherman: Yeah, so I can first of all say that that proposal doesn’t hold and that’s what the field has been learning over the course of the last six months. It was exciting news, it was a great press release, but it’s just not the reality. It holds in cases where you already have enough data around the target and the drugs that are of interest, but typically when you embark on a new program that’s an interesting program, you don’t have those data. And so these models, at this point the co-folding models in their ability to predict affinity are nowhere close to the physics-based free energy methods. They just don’t know anything about the target of interest. And so in a lot of these benchmarks there’s a lot of data leakage and people think that they’re benchmarking in an honest way and I think people truly are trying to benchmark in an honest way. But Pat Walters just published a really nice article about a week ago that shows that even when folks do date cutoffs and say that we’re going to train on everything before 2025 and then we’re going to test on everything after 2025, you look and you’ll find that after 2025 somebody published something that they had already been published 10 years before and maybe it was a slightly different condition and slight, but ultimately you predict this thing in the future but it’s still a lookup table of something that’s been done before. Whereas in drug discovery, we’re often really trying to do something novel that nobody’s done before. And so Boltz and these other co-folding type of approaches for predicting binding and affinity, have a big gap in terms of when they’re in regimes of high similarity they do well, but in the novelty space which is mostly what we care about in drug discovery, they need data. And they need local data to become accurate. And that’s where this concept of fine-tuning comes into play. And so even if you have a good foundation model that’s a good starting place, you still ultimately have to tune that around something. And experimental data is quite expensive. So what we’re finding and other people are starting to find as well, is that you could actually start to do that fine-tuning around the physics-based simulations. And so if you start with the assumption that the physics-based simulations are accurate and they’re a representation of reality, before you even have wet lab data, you could start training on the physical simulation, the free energy calculation, and that starts to build your AI model. And then you could get something like Boltz-2 to recapitulate the physics-based simulations and do it a lot faster. And then as the program runs and you start to accumulate more data from the wet lab, then those physical models, those AI models continue to get better and better. And so there is some truth in that Boltz statement that it is a thousand times faster, it is easy to run and anybody can run it. And the caveat is you have to know whether the molecules you’re looking at are somewhat close to the training data. And we’ve just published something on the OpenFold consortium portal that allows folks to go directly to the portal and say, here’s the system I care about, is there anything that’s like this in the training data? And that’s the first step to take. And if the answer is yes, then you might have a chance of getting something that works right out of the box. And if the answer is no, then you have to think carefully about how you’re using the AI tools and whether they’re going to be useful at all or if they are going to be useful how to make them useful by building out the right data set. And then in some cases, that might only take a few dozen data points or it might just be possible to do with synthetic data, from the computer.
Ross Katz: Yeah, I love the guidance that you gave there and it’s also just a reminder that all of the announcements that come out of the modeling community and the technology community in general need to be validated before you take them as truth and bring them into your company as such. You’ve talked about how you use physics-based simulations to enter your modeling processes and as I understand it these physics-based simulations generate enormous amounts of data, not just binding free energies but also conformational ensembles and solvation maps, etc. I’m interested in how you think about using all of the data that are coming out of these simulations to support the training and validation of your models.
Woody Sherman: So first and foremost, before you even start running these simulations, my recommendation and our approach is to build the right relational database. And so you got to start with something that’s going to be able to capture the simulations, the full simulations as well as all the associated metadata so that you can work with it later. And so it’s a lot easier to work with these data if you have the right database structure to start with. And so that was something that was important that we did in the early days of Silicon Therapeutics and that’s continued through Roivant to where we are today. So we’ve been able to continue to accumulate data, capture these data etc. And then there’s various ways that you could use those, in the context of a project. Some of it’s just learnings, what works doesn’t work, where do you spend the human time digging into things? A lot of it comes through in visualization, complementing human intuition with physical data from the simulations and so we’ve created an interface that allows humans to seamlessly interact with not just these free energy endpoints like you said as a number, but to drill into those and say what’s actually happening in the system? How is the protein moving? How is the small molecule moving? What are the interactions that are maintained or broken etc. And a lot of that visualization can build insights that can maybe not produce a final prediction, but can help build out new hypotheses that, here’s a particular interaction that is not being maintained well. How can we enhance that? And then that leads to the next round of predictions. And then we’re starting to capture data on the human choices, which is really interesting. Human decision, as we mentioned at the beginning, these experienced medicinal chemists have a ton of intuition that they’ve built out over decades of seeing programs and it’s hard to put those into a formula. But we just have humans voting on molecules that they like or don’t like, just like the reinforcement that took place with the language model space. And then the generative models learn from that and they can start to create molecules that move in the direction. So that human feedback reinforcement learning is a super useful component. And so we capture all of the information not just about our in silico simulations and not just the wet lab data but also the human interactions and the human decisions and not just their votes, but comments and other things that now the language models are quite good at processing and we’ve been capturing this data for years even before the language models were something that and we knew that one way or another we’d be able to get through these data and use it usefully and the language models have now made that a lot more useful. And then as we keep accumulating all this trajectory data which we have, we’ll probably be able to start using that for augmenting these co-folding models. And so most co-folding models have been trained on static structures from the protein data bank. There’s been some early work from Boltz and a few others that have started to augment with molecular dynamics data. I think that’s shown some promise, although it’s still very early, but I think that’s an exciting direction to go is data augmentation with molecular simulations to capture more of the complete ensemble of structures. And so it’s the right data storage, the right data capture so that when you are ready to build those larger models the data’s there and ready to use.
Ross Katz: Yeah, I love the idea of having natural language comments on the molecules themselves that you can then feed into the language models to condition the generation still further. That’s a great use case. Are there any unique failure modes that you’ve hit when training models using simulation data?
Woody Sherman: Yeah, absolutely. First of all these simulations are still limited both in terms of the force field representation, the accuracy of the model, as well as the sampling. When we talk about running these simulations for a day on a GPU, sometimes that’s insufficient. It depends on the system, it depends on the motion, it depends on the time scale. We’re trying to simulate biological time scales and there’s certain events in biology that happen on the nanosecond or microsecond time scale, which is roughly what we can simulate, and there’s certain motions that happen on the millisecond time scale or even longer time scales which we aren’t going to run a single calculation for a month or a year just to get that one data point. And so there are limitations in terms of just the accuracy and the sampling of these underlying simulations that are getting used. I think there’s still a lot of value in capturing what is there, but there’s definitely gaps and then this quantum mechanical nature of molecules, being able to deal with polarizability. That’s a big aspect that’s missing from most force fields. Most force fields are trained on quantum mechanics for a molecule in isolation, which is good, it’s a good baseline, first-order approximation, but as molecules get close to each other the electron clouds interact and that polarizes the molecules and that changes their behavior. And so in that case again back to that STING dimerization, that was a polarization effect, couldn’t be captured with traditional methods. We had to rely on the quantum mechanics. And so we’re in a field where there’s rigorous quantum mechanics but only on very small systems. There’s more approximate molecular simulations on big systems that get you toward the biology but miss out on the subtle interactions and there’s a lot of very specific interactions in drug discovery that are governed by these more subtle quantum mechanical features.
Ross Katz: No, I love that. It’s really interesting. Another thing you mentioned earlier was this idea that you’re capturing all of these different kinds of data that maybe your typical off-the-shelf system like an ELN or a LIMS might not be designed to capture. So I’m interested in PsiVision and what you built in order to enable you to house and interact with the different types of data that you’re now capturing when you do drug discovery in this way.
Woody Sherman: Yeah, so super important, first and foremost, starting with the right relational database, the right schema, the understanding of what those data may look like, what does a quantum mechanical calculation produce? What does a molecular dynamics simulation produce? What do these machine learning models need in order to build models and for inference? And that requires all the different types of representations that we might see of molecules. We certainly still deal with SMILES strings, like we mentioned at the beginning that are not great for building machine learning models but they are a very efficient representation of the molecules themselves that can be used for simple querying and looking for redundancy and it is your database key. This is the molecule, all the data is associated with that. So you have to build things from the context of molecules are the central entities that matter and a lot of stuff derives from there, but not everything derives from those molecules. You also have your target protein, you have the biological assay and all of those pieces have to be combined. And so what we’ve found is to have phenomenal engineers who know how to build out the databases and build out the systems that can scale. To have data scientists who understand the data itself and can build the right structure for those data to live. And then to have the right people on the side of the lab. And that means biophysicists, structural biologists, chemists, and biologists who all sit down and work with the data scientists and work with the engineers to figure out how to make all these pieces connect. I don’t think there’s anybody out there who knows all of this stuff and can keep it all in their head. And so this collaboration is super important and it’s not just to say, how do we want to pull things into the database, but how are people going to leverage that and make decisions? And how is that going to, what are people going to want to pull from the database in the context of the human interactions, in the context of the AI ML models and in the context of just learning and presentations and understanding how to extract the maximum insights and information from the data. So it’s a process and it takes iterations and figuring it out requires a lot of conversations. A lot of it is, we had the benefit of doing it in real time in a real project with some really phenomenal engineers and data scientists and folks in the lab that were all willing to put in the time and effort to understand each other. We built out a vocabulary, basically a dictionary that translates and helps people understand from all those different disciplines. A lot of times there’s the same concept that’s represented in different words and so your database has to understand the ontologies as well, which language models are starting to get better at and that’s super helpful, but still the database structure itself has to be flexible and robust with all of the human-specific information in there to make it useful.
Ross Katz: Yeah, I have so many follow-up questions to this, but we’re getting toward the end of our conversation so I’ll have to let you go. I can only imagine the nature of the semantic layer that you’re developing on top of that data that allows for all those different domain experts to be able to access it using their own individualized expertise and language. So PsiThera is focused on oral therapeutics for challenging immunology targets, starting with the TNF superfamily. Can you let us in on why that specific combination of modality and therapeutic area works well for your computational platform?
Woody Sherman: Absolutely. First of all why focus on this area in the first place? We feel there’s a real unmet need for patients. Immunology diseases continue to grow for various reasons. The antibodies have been very powerful, but in some sense in many cases maybe too powerful. And there’s the cost and the convenience factor. And so the value of having an oral for patients is huge. And from a drug discovery company perspective, the value inflections come earlier when you have a highly validated target like the TNF superfamily, TNF alpha, which was the target of Humira, which was the greatest selling drug of all times. That validation that there is a target that’s important in the context of human disease means that if you have a molecule that can do something to that target and maybe have better properties from the perspective of patients, you can measure that value a lot earlier in the process. You don’t have to get all the way to like a phase two clinical readout to see that you’re on the right path and creating value. As far as why this is, what’s our advantage in working on this? First of all the computational platform actually is quite broad and vast. At Silicon, we were working on oncology and immuno-oncology targets. We’ve worked historically on cardiovascular targets and even neurological targets. For us the critical factor is the relationship between the target and the disease and that mechanistic understanding of the pathway. So that’s always the gate for us. We’re not going into something where we don’t understand that relationship because then we’re just shooting in the dark and we don’t have any competitive advantage. So with the TNF superfamily and immunology targets more generally, there tends to be an understanding that this is the target in this particular pathway that’s associated with this particular disease and modulating this target in a particular way can lead to a beneficial outcome, whether that’s blocking it or activating that or doing something in between those. So it’s that target-disease relationship and that underlying mechanistic understanding as well as the de-risking on the target that makes this area so interesting for us. And then from the drug discovery perspective, the platform is well-suited for these types of systems. And so once we understand all that stuff, we can point it in the direction of immunology, oncology, neurology, cardiovascular etc. The TNF superfamily in the I and I space tends to be important. There’s a big patient need. The targets are fairly well characterized from a structural perspective but they’re also very hard and that’s another thing that we prioritize. We can make drugs for hard targets. We don’t want to just compete in a place where it’s about fast followers and me-too programs because there’s other people that do that very well. Our specialty is doing more of the novelty and hard things and the TNF superfamily falls right into that space. Many of the targets have no precedent. There’s no small molecule chemical matter. There’s great antibodies that are validated targets, we know that they’re good targets to modulate but absolutely zero small molecule chemical information, which is where the physics-based approach comes into play before we start to accumulate data and the lab-based approach where we could actually run wet lab screens.
Ross Katz: Awesome. Before I let you go, is there any advice that you would give to people doing computational chemistry or data science in the biotech space, some hard-won lessons that you have from your illustrious career?
Woody Sherman: Yeah, so I think cross-functional learning I’d say is the most important thing. I’d say have great depth in one area at least, more areas the better, but you can’t be an expert in everything. Make sure you’re an expert in something, but that you continue to communicate across all the disciplines. Drug discovery is hard and no one discipline is going to solve drug discovery. So I’d say from a computational chemistry perspective first, learn as much as you can about chemistry, the limitations and challenges of synthesis. You had asked before about some of the bottlenecks and the slowest parts of the cycle and we didn’t get to that, but it is still chemical synthesis primarily in the design-make-test cycle is the rate-limiting step and so you better understand chemistry and understand if the molecule you’re proposing is going to take a day or a week or a month or a year to synthesize. So understand chemistry, understand biology, work closely with experts, be a voracious reader, I’d say be curious, is probably the most important thing, is just be curious. And when you start to see data, dig in and try to understand what it means and why, and if you don’t have the expertise to do it, try to find somebody in that particular discipline who can help you understand that and keep accumulating that depth of knowledge.
Ross Katz: Great advice. Well Woody, it’s been a pleasure having you on the podcast. I really appreciate it.
Woody Sherman: Yep, thanks Ross. It’s been a pleasure. Yep. Take care.
Ross Katz: That’s our conversation with Woody Sherman. A few things I’m taking away from it. The first is AI isn’t one thing. The language models that are transforming coding and paperwork aren’t the ones that predict how a molecule behaves in the body. Those are different problems that need different tools. The second is that the representation of those molecules is the whole game. Whether you treat a molecule as a text string, a 2D graph, or a quantum mechanical object changes what you can predict. And also the ways that you represent the phenomena that are available to you in that data influences what the model can take out of them. And a good portion of the field is still using the simplest or easiest version that can be integrated with existing model architectures. The third is that computers don’t make drugs, teams do. Woody kept coming back to the human side. The experienced medicinal chemist as the North Star and the cross-functional trust that lets a computational prediction change what happens in the lab. And I found it really interesting to hear him talk about the team dynamics that they wanted to create and how a really great small team can accomplish a lot in a very short period of time. And the last was Woody’s description of how PsiThera is picking hard problems on purpose. They’re going after validated targets with no small molecule precedent because that’s where a physics-first approach has an edge before any data exists. And it’s interesting to think about how choosing the right kind of hard problem creates a moat for a computationally heavy industry like biotech. And then his closing advice is worth repeating as well. You go deep in one thing and stay curious across everything else. And obviously when you see data that you don’t understand, then you dig into it until you do. If you got something out of this, then please share it with someone on your team who’s trying to figure out where AI fits into their own work. And please subscribe wherever you listen and we’ll see you next time on Data in Biotech.
Jason: And that’s it for this episode of Data in Biotech. If you enjoyed the episode, please subscribe, rate, or leave a review in your podcast platform of choice. See you next time.
