It all started with a wooden microscope. Even at age 5, the gift sparked a lifelong curiosity that Karen Akinsanya has carried throughout her career — over 30 years defining the intersection of science, technology, and biopharma. Currently President of R&D, Therapeutics at Schrödinger, she is responsible for the discovery and development of therapeutics as well as business development and collaborations.
Karen started her career at Ferring Pharmaceuticals, where she led on discovery and development of a new treatment for prostate cancer. Much of her career was at Merck, where she wore many hats — in clinical pharmacology, discovery, preclinical and early development before moving into business development.
Karen joined me to reflect on the evolution of computational drug discovery and the responsibilities of setting realistic expectations as we enter a new age of AI-driven R&D. — Sajith Wickramasekara
* Editor’s note: The conversation has been edited for length and clarity.
Pursuing big technological shifts
Sajith Wickramasekara: Your career shows this remarkable range and curiosity for every facet of the R&D process and business. I’m curious about some of the biggest industry shifts you’ve seen — was there a tipping point when you knew computational science was the next big disruption for medicine?
Karen Akinsanya: 25 years ago, when I was still at Ferring Pharmaceuticals, and we were trying to pick targets, we'd interview academics and say: “What are you working on?”
Around that time, I picked up Nature and read this paper about a rare genetic disorder, where they'd figured out the mutation. It occurred to me that if you could do that at scale, we'd be able to do a much better job at picking targets.
I tell that story because I think my whole career has been about: how do we do a better job of making medicines for patients?
The story about joining Schrödinger was very similar. I had spent well over a decade at Merck trying to discover and develop medicines. I actually spent time in clinical pharmacology, where you’re one of those people that gets to dose a patient with a drug for the very first time. It's absolutely exhilarating.
Sajith: And scary?
Karen: Very scary. But quite often, we were killing molecules within weeks, if not months, of having done a lot of work to get it into the patient.
And I became very focused on the idea that if we could only test more ideas, test more hypotheses, and be much more efficient about coming up with the molecule, then we’d have a better chance of a molecule surviving that initial test.
At Merck, we decided that we were going to survey the landscape to see if there was a different way of doing what we were doing; I got put in charge of this “lab of the future” project.
One morning, another pharma had published the molecules that we were working on. They had a patent with almost all the same molecules, and we were like, “Oh my gosh, we've invested millions here! How are we going to recover from this?”
And we came across Schrödinger, who apparently had a platform that could identify molecules very quickly and solve for very complex target product profiles. And everyone didn't believe it. They said: “There are a lot of companies out there saying they can do all sorts of things.”
But this pilot was hugely successful. Not only did they find very unique molecules, but they found multiple series of them within a couple of months. So it occurred to me — something's happening here.
This approach to computation (which are physics-based methods) just accelerated our ability to solve a problem. What if we could do that more systematically?
And that's the story of how I ended up leaving Merck and joining Schrödinger.
Sajith: So the results were so compelling, that you were like: “I have to go.”
Karen: I had to go. I had to be able to build a pipeline, ask more questions, and test more hypotheses.
How computation is setting the new standard of speed
Sajith: So computational platforms are clearly redefining timelines. In the case of SGR-1505, a target for the treatment of non-Hodgkin’s B-cell lymphomas, it took just 10 months to discover a development candidate — which one of the fastest timelines for Schrödinger to date, and fast for the industry as a whole.
Schrödinger’s physics-powered, ML-enabled platform sorted through 8.2 billion potential compounds, identifying 78 that were synthesized and filtered through preclinical experiments to select the most promising candidate. First, congrats on progress and impact. How did you achieve this, and what does it signal for the future of drug discovery?
Karen: Before the 78 molecules we synthesized, which is everything we did to come up with a clinical candidate which is now in phase I, it was actually the 10th molecule that we made that had good enough properties to go in vivo and validate the pharmacology.
I want to be very clear about this, because I think there are high expectations with computational methods, whether it's AI/ML, or physics-based methods combined with ML.
Drug discovery is hard for a reason. You're not just trying to solve for potency, selectivity, solubility, or brain penetration. Sorting through 8.2 billion compounds was really our attempt to find a molecule that didn't just solve one of those things.
When I joined Merck, you'd screen say, 2 million compounds. The chances you're going to find an individual molecule that has the target product profile are very low. Or you’d have to keep doing successive screens and successive chemistry to find that molecule. And so the 8.2 billion allowed us to find, very rapidly, molecules that had in vivo pharmacology results that essentially compelled us to solve, not just for the potency and selectivity issue, but for a drug that has really excellent properties in humans. 78 compounds got us essentially to the development candidate and also an in vivo package.
When you read articles about AI and finding the molecule, of course that's not the end of it. You have to get it into tox studies and pharmacology studies, and do human dose prediction, and ultimately get it approved.
Within that 78 molecules, SGR-1505 had fantastic properties, great human dose predictions, and was an amazing journey.
But I want to be clear. It's only one of about 50 journeys we've been on. Schrödinger is a 35-year-old company. 35 years ago, compute wasn't fast enough. The math wasn't working. So many things weren't working. But over that 30 year journey, we've taken many approaches to come up with development candidates. Not 8.2 billion — that’s pretty standard today, but it wasn’t a while ago.
Why you can’t develop technology in a vacuum
Sajith: Schrödinger was a software company when you joined. But today, it has a unique business model; there’s dual licensing of the software, but Schrödinger also has its own drug pipeline, with some phenomenal successes. Can you tell us about that evolution?
Karen: The hallmark of Schrödinger is that the method started to work at a time when there were very profound claims being made about computation and how it was going to change drug discovery. Unfortunately, it was a very disappointing period where computation wasn't working, and it wasn't replacing what medicinal chemists were doing.
Sajith: Was this in the ‘90s?
Karen: Exactly.
Sajith: There have been a couple of moments where biotech has proclaimed that computation is going to solve medicine.
Karen: I think we’re in one of those moments now, with AI. But at the time, there was a lot of disappointment in pharma that these methods weren't working very well.
So Schrödinger decided to do something pretty interesting, as the methods really began to work. They decided to co-found a company called Nimbus Therapeutics, where the idea was to use the platform to come up with drugs, and then sell them to pharma. This was about 15 years ago.
The first molecule that they were able to identify was an ACC inhibitor. That was one of the largest transactions at the time with Gilead.
Even at Merck, we were working on the target ACC. No one could get a selective molecule. The Nimbus molecule was the first example of a selective compound.
Sajith: Do you remember how big that transaction was with Gilead?
Karen: $400 million upfront — and that was years ago.
Sajith: And there was a much more impressive one more recently?
Karen: The second program that they partnered with Takeda was for $4 billion. So with these billion dollar molecules, what did we take away from that?
“You cannot develop technology in a vacuum. You need to pressure test these molecules against problems that pharma cares about.”
I think that's the lesson of Nimbus, which has now been followed by Morphic and Petra, and all of these companies that we've co-founded, who exclusively use our platform to come up with drugs. Those transactions are a litmus test — not only does the technology work, but the molecules are winning molecules.
Breakthroughs take a lot of patience (and cat videos?)
Sajith: You mentioned there was a point at which the methods really started to work. What made the scales tip?
Karen: Compute got fast enough. The physics started working. The math started working. Free energy perturbation calculations became almost routine and scalable. Today we're combining free energy calculations with machine learning to actually generate datasets for targets. So it was really this convergence of technology and the fundamental science, the fundamental physics behind all this atomistic modeling, and structural biology. We’re at an interesting time.
Sajith: What do you think gave the organization enough conviction to keep going? I'm sure there was lots of success along the way, but it was probably a 20-year journey before things started to really work, right? Biotech is riddled with failure in making medicines, and I think understanding how organizations persevere is really interesting.
Karen: There are several stories here, one of which is an investment story. The two investors who came from other domains were David Shaw, the father of quantitative trading, and Bill Gates. They heard about Schrödinger, and had conviction that this physical chemist Rich Friesner, who was the founder of Schrödinger, was right — if you could get the math and physics right, and combine it with computation, it was going to be transformational.
Sajith: So it took actually outsiders helping.
Karen: If it weren't for those two patient investors waiting for the physics, the math, and all of that to work — and the computation to get fast enough — I'm not sure Schrödinger would actually be here. Because VCs have a shorter time horizon.
Sajith: At a recent event, you posed this thought-provoking idea about how data generation outside of biology (like internet cat videos) can advance image and pattern recognition that later leads to breakthroughs in biology. Can you connect the dots for us?
Karen: I would say that advances are built on other advances. We wouldn't have the quality of machine learning models and image recognition, if it were not for Internet videos, Internet cats, or Internet images. And here we are, as biologists, chemists, and physicists, leveraging that advance that was really driven by the creator economy and people uploading thousands of images of cats. I would say the same thing about ChatGPT, in a way.
Of course, language and video are very different from chemistry. But I think there's some synergy in terms of the models that are being built, and the ability to build on top of those innovations that come from totally different fields.
The domain-specific models that are being built now, like with Alphafold, go back 50 years. The PDB was founded in 1971. It was 50 years of PhD students and postdocs solving structures in labs over many years. That created a very specific dataset that we can now leverage for some of the problems that we're looking to solve in terms of protein structure.
I do think those foundational pieces around ML models coming from videos and images really propelled this whole field forward.
The big responsibility with AI
Sajith: I feel like the biopharma industry can do a better job of being part of the tech and AI zeitgeist. How do we make it so the industry is faster at adopting technology from other fields?
Karen: I think that biotechs, as we described with the Nimbus and Morphic stories, are probably the faster adopters relative to biopharma.
Sajith: You had to literally create a new company from the ground up to do it.
Karen: Correct. But I think what we're seeing now is that large pharma are recognizing the moment we're in, and are adopting it faster.
What we do as an industry is extraordinarily complex. Any complexity that we can reduce by leveraging these methods, validating them, and demonstrating the advancements that we're making (versus with more empirical methods) — that's absolutely key.
But what we must also be very careful about is not overhyping it, because then we set ourselves up for disappointment. If the expectations are too high, you can drive the whole space into this winter of discontent, where the results don't quite match the expectation.
So I think it's a huge and profound responsibility that we have to balance validation and demonstration of impact, with the excitement we all have: that computation is really here to stay, and is going to make a huge difference.
“As an industry, we have a responsibility, in the sense that we have to be realistic about what's possible.”
Breaking down silos and bringing teams together
Sajith: Schrödinger takes a very different approach to drug discovery. You obviously have computational methods incorporated from the very beginning. Traditional teams in biopharma might have a 10:1 ratio of medicinal chemists to computational chemists, but your teams are structured differently — what impact does this have on drug discovery?
Karen: When I joined the industry, biologists were outnumbered by chemists by a significant degree. Biologists obviously became more important on project teams at some point. But now our teams are really different.
At Schrödinger, we have physicists and machine learning experts. Rather than having one computational chemist supporting 5-6 programs (which is how it was when I was in the Discovery group at Merck), we have an equal number of computational chemists to medicinal chemists on our teams.
We have this incredible multidisciplinary team of clinical people and translational people, all trying to tackle the problem in parallel instead of in series, which is how we used to do things.
Sajith: Even before Schrödinger, you had a lot of different roles across the entire R&D process. You've worked in clinical pharmacology, discovery, even business development. How has this shaped your vision for how an R&D organization should work?
Karen: Yeah, well, I'm fundamentally nosy.
I was known at Merck for going into other divisions and asking what their problems were. There's a lot of interdependencies between different specialty areas in a pharmaceutical company, and I think we're breaking down those silos.
The problem, in my opinion, was the handoffs. The chemists would hand off to the biologists. The biologists would hand off to the pharmacologist, and a lot of context gets lost.
“A lot of the challenge of drug discovery is figuring out what's going to kill the program as soon as possible. The more you integrate the people who have to make those series of decisions as one group, the better decisions you end up making.”
So the idea for me was, I wanted this very broad toolbox. I wanted to understand: what problems are our clinical people facing?
Sajith: Do you think every discovery scientist should do a tour of duty in clinical pharmacology or business development?
Karen: Absolutely, I even encourage it. I think it's really important for people to spend time in other domains.
I describe it as the “T”. We all have a long arm of the “T”, that's our expertise that we trained in, but the top bar of the “T”, we have to explore as broadly as possible — so you understand what the rest of the value chain is going through.
The biggest hurdles for AI-driven R&D
Sajith: Do you see a future where every biopharma and biotech company uses AI/ML and computation to accelerate their R&D? And how quickly can we get there?
Karen: I would say the future is now.
“I don't think there is any biopharma company, whether that's a biotech or a large pharma company that isn’t using AI, ML, or computation in some way, shape, or form.”
Whether they're using it to write regulatory documents, or at the very front end to come up with the hits, I think everybody's doing it already. It's the question of how much are they using it, and how much are they connecting it across different groups?
Sajith: I’m sure you meet with a lot of biopharma companies who come to Schrödinger and maybe feel a little bit behind on their AI journey, and want to accelerate. What hurdles do they face, and what advice do you have for them?
Karen: I think the hurdles are that, especially in larger companies, they're set up a certain way. The Merck campus was massive; there was a whole building full of chemists. How do you undo that? How do you retrain people to embrace a whole different workflow? That is a challenge.
I think the other hurdle is that there’s a shortage of people with specific kinds of training, like this idea that we now have physician-scientists who also know how to write ML code.
Sajith: That's a lot for a person. You'd been in school for a long time…
Karen: That sort of multilingual type of talent is at a shortage right now. That's one of the biggest challenges, and then the other is the proof points: Does this actually work before I go too far and invest too much? I think that is a real challenge. But we're making a huge amount of progress as an industry.