November 10, 2020

It all adds up: mathematical simulations in biopharma research

By Conor McKechnie and Dodi Axelson

It all adds up: mathematical simulations in biopharma research

Computer simulation helps us design better cars before we make a physical model. What if we could use this technology to do the same for purifying new biotherapeutics? These podcast guests think we can and will by bridging math and biochemistry.

DODI: It is all going to add up, Conor. We're getting straight down to business on Discovery Matters today.

CONOR: Okay, let's just hit the road. Let's go, no wasting of time.

DODI: Right on. So, you know our colleague Gunnar?

CONOR: I do. Gunnar is something of a legend around the business.

GUNNAR MALMQUIST: I'm a principal scientist at Cytiva.

DODI: But do you know Gunnar’s backstory?

CONOR: Oh, no.

DODI: He got interested in science from an early age.

GUNNAR: I had a very early connection with chemistry. You could buy these kinds of chemistry toolboxes in the toy stores. And they were much more fun at that time than they are today. Because there were things included that that wouldn't be allowed today.

CONOR: So, let's be honest, and this is coming from someone who did his own experiments. Well, never mind. Gunnar was, like blowing things up. Is that right?

DODI: That's right. Absolutely. All fingers intact, though. Thank goodness. And it was Gunnar who explained to me the idea of mechanistic modeling.

CONOR: Okay, so using computer simulations to decrease the number of experiments you need to do during the development of a process. Essentially, doing kind of in computer modeling, to be more efficient ahead of trying something in the real world.

DODI: Well, yes. Not unlike a recent article that we read in Scientific American that stated there's a 50:50 chance that we are all living in a computer simulation.

CONOR: Yeah, yeah. And the guy that wrote that program, and that simulation, right now, I'm not happy with you. This is bad coding.

DODI: Bad coding, indeed. But let's talk about mechanistic modeling on this episode of Discovery Matters.

GUNNAR: The good thing about modeling is that it allows you to understand an event like a chromatography separation. And it allows you to simulate it. So, you can do many, many experiments in the computer. The thing with mechanistic modeling, as opposed to statistical empirical modeling, is that the equations you use are actually based on fundamental physical chemical processes that take place during chromatography. The binding of a protein to the ligand is described by an isotherm, and the transport of proteins between the particles and inside the particles is described by equations. That means that you can actually interpret the outcome of these models in chromatography terms. And that allows you to do process development, troubleshooting, looking at potential deviation in manufacturing in a much more efficient way.

CONOR: And in big picture terms, what that means is that you can get better medicines faster to the market, is that right?

GUNNAR: Absolutely. I think that has been a very, very strong driver for the biopharma companies picking up on mechanistic modeling in the recent years. It allows you to shortcut some of the labors tasks that you have during process development and process characterization.

CONOR: If it's that clear that you don't use materials, and you can do all these experiments on your computer and not in the lab, why isn't everyone just doing it? Why aren't we running simulations before we do anything?

DODI: Exactly what I asked Gunnar, and he says it comes down to that threshold. The fact is, there's a pretty steep learning curve to get to the point where you feel familiar and confident in doing this.

GUNNAR: The equations involved are rather complex, it's partial differential equations. It is currently described in the language that may not be the typical chromatography language. The good news, though, is that things are starting to change here. And that is why I think it's starting to pick up. In the 1990s this was very much an academic exercise. You can build a complete thesis on mechanistic modeling today. There is software that allows you to do this without being a software engineer. You don't have to program, and things are moving in the right direction. The threshold is still rather high to get over. But the more people who get into the game, the more this is discussed at scientific meetings, I think that threshold will go down in the future.

DODI: Somebody else who is actively trying to lower that threshold is Tobias.

TOBIAS HAHN: I'm a cofounder of the company GoSilico. I'm a mathematician by training.

DODI: We're gonna get to the threshold that he's trying to lower in just a little bit. But first, as always, a backstory. In 2012, Tobias stumbled into bioengineering as one does. By accident. He was working on fluid dynamics in the mathematical department at the Karlsruhe Institute of Technology. Basically, Tobias got bored, trying to figure out a problem at work. Have you ever experienced that, Conor?

CONOR: I don't think boredom is necessarily a recurring factor in my work life at the moment. But not figuring out problems is pretty regular.

TOBIAS: It was too simple somehow.

DODI: So, he decided he needed to push himself even further and find something else, something almost impossible to figure out. And by chance, Professor Hubbuch moved to Karlsruhe, so he switched universities.

TOBIAS: He worked on simulating chromatography at his previous research group. And because of licensing issues, he could not bring the code with him. So, he asked in the math department is there somebody that could help out writing a small simulator for chromatography, and that ended up on my desk because I had free resources and wanted to work on something else for a change. And after two weeks, this first simulator was done. He was really happy with that and acquired more funding to deepen the simulation work and also the research in that area.

DODI: So, from this Tobias along with three cofounders started the company GoSilico in 2016.

TOBIAS: We simulate processes in biopharma, especially the purification process of drug candidates mostly.

CONOR: Okay, so let me get this right. A professor started working at the Karlsruhe Institute of Technology. And this professor has this code, but he couldn't transport it. And so Tobias just simulated the code.

DODI: Exactly.

CONOR: Wow. Okay, I've got it. But how is it that simulations help science with the actual process of purifying proteins? Well, that's got to be physical, right?

TOBIAS: The simulations in the end help us to perform virtual experiments. One of the main applications is to do ‘what if’ analysis, and try out scenarios that are hard to perform in the lab at all. For example, in biopharma, we have those quality by design (QbD) guidelines that define that we should understand the process and also all the parameters and how those process parameters influence the quality of the drug. And there are some scenarios that can happen in production, which we cannot just test all scenarios; there are too many parameters. So, the dimensionality is enormous if you wanted to change very simple things like flow rates, pHs, temperatures. And those ‘what if’ analyses can be done, very simply and quickly when they have simulations.

DODI: So now we're coming back to that threshold that Tobias is trying to lower.

TOBIAS: The new challenge for me is now to rep this technology in a way that it becomes usable, and that it unfolds all its power to create tools that create an understanding if a model is good or bad, and how trustworthy those results are. This is actually tricky. But that's the same problem for the statistics people. You can have arbitrarily bad statistical predictions. The same with our models. If the input data is bad, then the model will also be not predictive to create an understanding and for the user, and ideally quantify the predictive power. This is something that I like working on at the moment.

CONOR: Okay, so this sounds so convincing. The next side of the question has got to be why would we conduct any real experiments anymore? If we can do kind of everything in a simulated environment first, why would we do anything in the real world?

DODI: It's like you were in the room, Conor, because I asked that, too. So Tobias says this is a typical question that you would only really get in biopharma.

TOBIAS: You wouldn't get this in chemical drug development anymore. Process simulations are an essential part of process development plan design for chemical drugs for two or three decades. And it's the same with other industries. Nobody would develop a new airplane without simulating it first. The same with cars. It could go on piping, even groundwater flows. Everything Is simulated first; all these things are simulated today before a prototype is even built. So, the question is very specific to biopharma. The reason is that it's a new technology, for biopharma, and the adoption rate is rather slow. But on the other hand, having a model in place would show that the process understanding is already there. I think we will get there over time.

CONOR: So, the question is, will we ever get full-on total sci fi 100% simulation, and I can just lie in my bed plugged in for the rest of my life?

TOBIAS: We'd love to go full simulations, but I perfectly understand the hesitance. And so the regulators must also be educated. It's a new field for our industry.

CONOR: And so what are some of the examples that Tobias and InSilico are working on?

DODI: He described something that he and GoSilico are working on called virtual piloting.

TOBIAS: You can go from very small scales, and all the experiments are done on a robotic liquid handling station directly to a production plant. And we have shown that this is actually working, we showed that we can use something very simple and generate a model from those very early experiments that are normally done as a very first step in process development. And then use that information that we have predicted how a chromatography run could look in the lab scale, and then also predict how would this process look like in 12 000 L fermenter scale, and it's matched perfectly. And that's a feature that our simulations can bring that we can scale up and scale down effortlessly.

DODI: And this is what brings virtual reality and augmented reality, you know, the situation where Conor you could stay in your bed with the goggles on all day, and all month, and all year?

CONOR: No, no, actually it doesn't sound that good. But I'm just like, you know what? I want to go places where no one has been before.

DODI: And so to science, right? Tobias actually has high hopes in the space of virtual reality and augmented reality.

TOBIAS: What I've loved to do, for example, is what is currently ongoing with Cytiva supporting vaccine developers to scale up and scale out completely at risk. They have no idea which of the 70 vaccine candidates will make the race and which one will be the one that needs to be manufactured at a billion doses scale. And they are doing this completely at risk at the moment. The scaling up and scaling out is something that we could do easily.

CONOR: Okay, so this sounds like real game-changing stuff. It starts to challenge our sense of, you know, what's real and what isn't.

TOBIAS: It would be yes, a game changer. I don't think that there are any obstacles, the models that we are using, they describe what's happening accurately, and then experiment.

DODI: And that also means that it's going to change, we who work in the field of biotech, what do we need to know, what do we need to do differently? And Tobias and I started talking about that. And Tobias said, “You know, you should go talk to your other colleague, John.”

JOHN SCIBETTA: I'm an advanced chromatography specialist at Cytiva.

DODI: John knows an awful lot about resins and chromatography. We talked about the changes that John sees coming to the industry, if modeling were to become more common. For him, it is all about the potential of communication.

JOHN: I think that mechanistic modeling is a way to create a common language so that what goes up and what comes down is easily understood and can be interpreted by the other groups. And right now, we really don't have that. And I think that in the case of mechanistic modeling, what it has allowed the lab to do is to ask counterfactual questions. And counterfactual questions are an instantiation of imagination. And that's the top tier of human consciousness or human cognitive ability.

The first would be associations, we ask, Well, what is something? And I think that's the basis of statistical analysis. You just looking at correlations between objects or things in nature. And you say, Well, if this happens, then I see something else happening. If a bird flaps its wings, then it flies. It doesn't say anything about why then we've got something like design of experiments, which is how if I change the system, do I get a different output? Those are intervention questions. And I think that's clinical trials where we have a placebo and randomized testing that's squarely in this the second rung of the ladder of human consciousness or human intelligence.

And then there's the world of imagination, which is why does something happen? And we have never been able to address that. And what occurred to me is that for the first time, I understand why things happen. And it's because you have now ultimate causes. That is the physics of interaction in the column. And that's the brilliance of mechanistic modeling. And that is the brilliance that Tobias Hahn brought to an industry that's basically been thirsting for a clean understanding of the reasons why things work, and then why they don't work.

CONOR: Wow. I mean, that is just so challenging, existentially. I'm kind of thinking that this is the beginning of the new frontier of, you know, how we think about everything we do, everything in the virtual before we dive in and get it more right in the real.

DODI: So true. Is it real? Is it virtual? How can one improve the other? And for John, it boils down to how and why. And it's really pure.

CONOR: I love it. It's so simple. It's genius. So, we've heard from Gunnar, and we've heard from Tobias about the thresholds that we need to get over. But what does John see is the reasoning behind why it's taken us so long to get there?

JOHN: Like a lot of things in science, it's the technology, it's the supporting analytics, the ability to do the math. And that's where Tobias and his group of scientists come in. They have allowed a transfer, basically an esoteric knowledge, to be democratized. So that through the GUI, the user interface, now scientists at the bench who are biochemists can start working with it. But you don't have to understand the math equations. I think it's a good idea to start reading those papers, but what the GoSilico team has done is created a very user-friendly graphical user interface that is going to allow biochemists to do modeling without having to worry about doing the math.

CONOR: So, we're bringing this to the bench, where scientists can use it in their daily work, and it doesn't have to be mathematicians locked away and using supercomputers. Yeah, what does it mean for people who are working, you know, with their hands on the tech? And are we talking about we need to change people's training? Do people have to learn new skills? What are they going to have to have? Are they going to be wearing VR glasses or retinal implants? And just wondering how modeling and simulation is going to sort of change the future of people working in labs, working in manufacturing facilities, working in design spaces, and so on.

JOHN: There's a quote from Tobias, that the future biochemist will have to be a digital native. I think he's right on, I think that the people who are actually taking this on who are biochemists, you could say that they're classically trained in liquid chromatography. They are quite capable of using the software and creating workable models. It's not easy, it's probably one year of coming up to speed. So yes, future biochemists who are going to be working with mechanistic modeling will have to understand some of the math. And in the case of mechanistic modeling, I think they're going to have to start looking at the equations, because then it will make sense. So, the most exciting thing about the future is that we're about to unleash 10s of thousands of eager supple young minds with mechanistic modeling to go do better work, and there's going to be a tidal wave of discovery.

CONOR: So, when we look at what we're calling the new zoo of molecules coming through from biotech, how is modeling going to be helpful and significant for the discovery of the future?

GUNNAR: I think the recent move away from standard antibodies has accentuated the need for a better way of developing processes. Taking a standard mAb, I think most people would feel pretty confident to develop that process in pretty short time. With these new types of antibody variants, there are a number of closely related product-related impurities that are so similar to the drug molecule that the separation task is much more complicated. And this is where mechanistic modeling can provide a very efficient shortcut.

DODI: And good mechanistic modeling is going to be even more relevant as we see this zoo of molecules come onto the market.

GUNNAR: If you think of a company who are developing a broad variety of molecule types, so they can have bispecific, trispecific, fragments and so on, they need to step away from the thinking that there is one platform process that fits all molecules for these types of molecules where separation of closely related species become crucial. The fact that you can describe the separation, for instance, in a cation exchange step, you can describe that with equations, that allows you to test thousands of different separation schemes within a few minutes in the computer that will allow you to find that specific solution that provides you pure enough protein at a process economically viable yield. And that tradeoff is very important for accessibility of drugs, because if you have too low yield, your production costs go up, and the supply of that drug becomes trickier. And an efficient separation process will give you an edge over your competition.

CONOR: I was going to ask for an analogy or a metaphor of some kind, because it always helps us understand what we're talking about better. But it just seems that mechanistic modeling itself is like here's an analogy. Real world.

DODI: It's so meta.

CONOR: Exactly. Yeah, it's amazing.

GUNNAR: One can look at the modeling as being an abstract description of reality. And you look at that, and you learn about the real world. And in fact, a lot of the experiments you would have done in the lab can be replaced by these types of simulations, but you can't get away from reality anyway, because you really need to purify that protein to allow you to do analytical tests on it to allow you to move to the next process unit operation. And eventually, you have to put it into patients. And that is something that computer cannot do today. And I don't see that happening in my lifetime.

CONOR: It feels like for those of you that are of that age, William Gibson's vision of the interface between reality and the virtual world is something that might not just be so crazy. And maybe things like more modern versions of those stories like Ready Player One, they're going to come to fruition. I wonder, you know, these things come out of places where you work, where you're creating huge value for very important things. And then they kind of find their way into entertainment and the gaming space and...

DODI: Or vice versa. The imagination from the entertainment world comes into science. I think it's this cycle of, you know, curiosity tested out, make it happen. And actually, I think this whole mechanistic modeling is that it's tested out in a virtual environment before you have to use all of the material and everything when you get to the bench and do your real experiments or your real process development.

CONOR: And so I make an immediate jump to massive ethical questions here. If you go full scale on this, and if you want to read about the ethical questions here, Ender's Game by Orson Scott Card, what happens when you're running in a simulation? And you don't know it's a simulation? And you make choices…

DODI: Which might be our lives now, coming right back around.

CONOR: This is some pretty gnarly simulation. I've got to say.

DODI: Yeah, what we leave you with after this episode is it could be real. It could be a simulation.

CONOR: And the question you have to ask is, Does it actually matter? Just be good.

DODI: Thank you for listening.

CONOR: Thank you. Take care.

Listen to more podcast episodes.