Why AI Drug Discovery Is So Hard

What we want to do is decode biology. We believe there is an enormous amount of waste that goes into drug discovery today. If you look at how much money it actually takes, on average, to bring a new drug to market, it has gone up exponentially over time. And we believe a lot of that is due to our lack of deep understanding about how biology works.

A lot of the hypotheses we have are very linear, single-dimensional in nature. There's one gene or one protein that causes one disease, and perturbing that will fix the problem. And biology just doesn't work that way.

So the approach we've taken is to say: this way of performing single-dimensional, mechanistic modeling of disease is part of the problem. What we need to do is a better job of understanding biology. And the way to do that is to collect an awful lot of data about biology and chemistry, pair that with deep computation, and then actually test those predictions. Don't just make a model and take it out to the clinic — test those predictions in a continuous cycle.

If that really goes, we think there's a world in which you could get drugs for new diseases, for particular patients with particular diseases, a lot faster — exponentially faster than the current drug discovery and development pipeline. And we think the root of that is a collection of data, the building of models, and the cycling of that back through experiment in order to decode biology.

Just about everything, as it turns out. Particularly mathematicians and computer scientists who come into biology — me among them — will come in and say, "Oh man, this is easy. I'm just going to take this data, throw a machine learning model at it, and it's going to be good."

You all didn't know what you were doing. I remember very clearly the first rotation project I did in my PhD, where the postdoc I was working with gave me some datasets and said, "Okay, now figure out which compound is going to interact with these proteins" — to predict drugs. I'd just come off engineering's machine learning class. I was like, all right, no problem, I got this.

Three months later, all of my correlation plots looked like somebody had just sandpapered my screen. There was no correlation whatsoever. And here I am thinking, I failed the test. I'm not going to get to do a PhD. They're going to kick me out. I walked into the room to talk about it with the postdoc and my advisor, showed them the results, and they said: "And thus you're enlightened."

So the challenge is really manifold. Even at the smallest scale — individual molecules, individual targets — there's an awful lot we don't understand about how compounds interact with particular proteins, how the proteins might change in different contexts, and so on. You zoom out: every drug interacts with not just one target, but many, maybe hundreds of different targets in the body, at different concentrations, in different tissues. We have a fairly poor understanding of how all of those things add up.

As you keep layering on more levels of complexity, what you realize is that the simple reductive models we have at the base — "this drug hits this target, this target causes this disease" — are fundamentally wrong, and you have to zoom out from that. Biology is a system that has evolved over billions of years to work. It didn't evolve with a design principle that says, "this has to be simple and modellable in the easiest way you'd like." It's more like a billion years of somebody saying, "yeah, that's good enough, that'll do." And that's the system we're trying to work with.

If you've got The Sims at 1 and the weather at 10, I'd say trying to do small-molecule drug discovery is somewhere out at like 50 or 100.

But it's a really interesting example. Even over the last 10, 20 years, weather prediction has gotten dramatically more accurate. It was sort of unthinkable that you could pull up your phone and have a reasonably accurate hour-by-hour forecast, or that the 10-day forecast was anything more than pure guesswork. And now they're actually pretty decent.

Part of that is that we have some understanding of the physics going into these systems, and that's gotten better. A big part of it is that our access to data has gotten massively better. If you look at what's happening in weather, there's a lot more satellite data collection — you can see what's happening with cloud systems, where they're moving, get measurements of the wind and temperatures all across the globe. That data feeds into computational models. At the same time, we have wide networks of weather sensors — temperature, wind, precipitation — that can be fed back in to optimize those models.

I think those closed loops — that combination of modeling and really high-dimensional, high-throughput data collection — are things that have been absent in biology. It's a great example, because it's very well aligned with how we see the direction of moving biology forward.

That's actually not incorrect. What I'd say is that it's very broad — that's a process you could use to describe almost any pharma, where you say, "there's a disease, there are systems I can set up, experiments I can do, I can decide which programs might be interesting for a portfolio, and I'll advance them that way." So the interesting part is actually in the details.

Let me establish this by contrast. If you're at a typical target-driven pharma company, you might say, "Look, there's this awesome paper that just came out in Cell a couple of years ago, and they think this gene may be linked to this particular disease. Let's set up an experimental system around that gene, screen a bunch of compounds, have some chemists work really hard for a couple of years optimizing that compound, put it through some mice," and so on. Even as I say it, it sounds like a long, slow process.

The deeper challenge, in my opinion, is that when you do that, you collect data in the course of these experiments — but you collect very narrow slices, very narrow verticals. You have this one target that you looked at, all of your compounds tested against that, and then little islands of related molecules around them.

The way we're going about things is to turn that picture on its head. What if we could collect all of that data up front, and then look inside the data to identify the most promising areas — and to do so without reference to somebody's favorite Nature paper? To take a broader look at what we're seeing in the known scientific literature, in patient genetic and omic data, and so on.

To give you an example: the way we build what we call our map of biology is to take human cells in culture — typically primary cells, sometimes cell lines. We grow them in plates about the size of your hand, with over 1,500 wells, or little spaces for cells. Then we perturb each of those wells with different agents. When I say perturb, that might mean we knock out a particular gene, or put a particular chemical or drug starting point on there at a particular concentration. We might put proteins, antibodies, toxins, cytokines on them.

And then, literally, we look at them. We take those cells and put them under a microscope. The remarkable thing about biology is that form follows function — when cells are in a different state, doing something different, they actually look different under the microscope too.

So we take pictures of those perturbed cells, and with the deep learning image-processing models we've built in-house, we relate those different states to each other. The really remarkable thing is that what comes out is what we call a biologically meaningful space — meaning directions and magnitudes in the space mean something.

Imagine a ball, or a sphere. Cells that were unperturbed sit in the center. Maybe you knock out a particular receptor — that sits over here at x equals 1. As you take wells where you've dosed an antagonist for that receptor, something that blocks it, you'll see the features coming out of the machine learning sequentially move closer and closer to that knockout, until you've inhibited it. And if you dose a receptor agonist — something that activates it rather than inhibiting it — you'll see those embeddings move in the opposite direction.

So when we say we're building a map of biology, in one sense it's very literal. We're literally taking biology and trying to find out where these different perturbations, these different states, live in space, and how we can get from one to another. What's really neat is that once you can do that, you can relate all of these perturbations to each other. When we screen a particular compound on our platform, it's not like that earlier case where you get the result for one target. You can see its relationships against all of the genetic knockouts you've done, in one go. You also get its similarity to all of the compounds you've ever screened, in one go. So every time we run a new experiment, we're increasing the number of relationships — the amount of understanding — at a super-linear pace.

The discovery strategy becomes totally inverted. Rather than saying, "I have one interesting gene I found in a paper, let me go build a whole experimental platform around that," you can say, "let me look at the totality of what I've collected." At this point in Recursion's history, that's about 3 trillion relationships, I think. And you ask: what are the strongest relationships? Which ones look real and not just noise? Which ones look novel — not things people have well understood and really probed? And which ones look therapeutically relevant — like they could actually make a difference in a disease for particular patients? And then you can choose to go after those.

One piece of language we use is that we're broadening the funnel at the top to narrow it at the bottom. Rather than having a bunch of things and sequentially winnowing them down — 50%, 50%, 50% — we want to look at an enormously large number of hypotheses up front, with real data, not just guesswork, and then immediately subselect to just the ones we think are going to go.

That's exactly the goal. It really comes from a standpoint of acknowledging that we don't understand everything in biology. There are pieces we know, but we really don't know everything.

So can I ask a quick follow-on there? It sounds like you have multiple levers to advance the power of your approach. One lever might be to expand the number of cell lines you can test on. Another might be to expand the chemical space you can feed into your wells, to observe their effects. A third might be the power of your machine learning models, and so on. So let's say you have half a dozen to a dozen levers. Is there one or two you expect will drive the biggest breakthrough? And what are you looking for as a signal that the approach is working, getting better, getting more powerful — is it the type of insight you just described, or something else?

It's a couple of things. Seeing examples like the CDK12/RBM39 case I've shared are certainly pieces of interest and validation for us, and we've done a few things like that in other contexts. Seeing these progress through the pipeline, seeing some of them really have success, is very interesting to us.

Like a number of other companies, COVID was an opportunity for us to really accelerate that process of discovery, and we got some interesting signal there.

At the beginning of March, April of 2020, COVID was this hugely exploding phenomenon — a lot of interest in identifying whether there were drugs that could be repurposed to treat it. So we initiated two screening programs. One looked at infection with SARS-CoV-2, the causative agent, to see if there were drugs that could help in the acute phase of infection. The other looked at a model of severe COVID — the cytokine storm that results.

The really cool thing that came out of that: we did those screens, published the results, others went ahead and did clinical trials on them. And when we look retrospectively at the large trials that others did, we got eight out of nine of the results right. And we sort of know the reason why that last one didn't go. That gave us real belief that this platform —

Obviously this is hypothesis — I haven't gone and done the exact experiments to follow it up. But as I mentioned, the way we built that screen, we had cells perturbed by a cocktail of cytokines, modeling the cytokines you see in severe COVID cases. Dexamethasone is a steroid that's going to have anti-inflammatory effects. It was probably acting on a different cell type to suppress the release of those cytokines. So dexamethasone would have acted upstream, but we basically short-circuited the way it would have gotten in there. That's not something that would have been modeled in the particular experimental system we used.

That's exactly right. Of course, the human system is more complicated than this. Let's assume a two-component system, where you have cells in the blood vessels or the lung, and you have immune cells. Those immune cells are the secretors of the cytokines that then have systemic effects on the rest of the body. The way we modeled this was by taking just the cytokine mixture and putting it directly on cells. It's likely that dexamethasone was not acting there, blocking the response of those cells to cytokines — it's likely it was acting upstream, preventing cytokine release in the first place.

I guess that could also be an example of the type of paradigm you're trying to avoid — the narrow slice of biology, the experimental constructs of the traditional approach you described earlier. Would you say that's just an artifact of an accelerated experimental process in the context of COVID, where a very specific experiment was designed in short order? Or, I guess what I'm wondering is: how do you systematically avoid that type of miss in your broader endeavors, beyond COVID?

That's a great question, and it gets to some of those levers you mentioned. Absolutely — at that time, April 2020, we were seeing what was going on around the world and trying to figure out the most reasonable translational model we could develop under the constraints we had.

In general, there are always going to be trade-offs with models. You have to decide which parts of the system you're going to model and which you aren't. At the end of the day, a culture of human cells in a plate is not the same thing as an entire organism. So there are things we do along those levers to reduce the occurrence of that. We can model multiple cell types. We can look at multiple readouts, whether imaging-based or transcriptomic-based. And of course things end up going through downstream tests in animal models.

One thing we've optimized our discovery strategy toward is not worrying quite as much about false negatives — where there's something that in principle would be good, but you don't get it. Because the principle is: how do we do this at scale? How can we find enough really good things that, if you've got 100 things that work but you missed one that would have worked, that's still okay? You've still come out way ahead.

Or even that you find, given the value of new medications, one or two that actually work — even if you were to miss 50 or 100 — it would probably still be a worthwhile endeavor.

That's exactly right. It was something that was mind-blowing to me as a PhD student, and it's still true today. I remember my advisor telling me: if you have a 98% error rate on your model, then what you have is a revolutionary invention, because you're twice as good as the rest of the world. Now, depending on where we take the slice, that number changes. But the message is still true. Even if you're wrong 80% of the time, given 90% failure rates depending on where you look in the pipeline, that is a revolutionary change in getting drugs to patients.

Imran, can I build on that? Just to think about the systematic problem of drug discovery in terms of cost. My understanding is that most of that cost is driven by the drugs people think will work, that go into the clinic, incur the costs of an expensive trial — financial costs, but also human cost — and then ultimately fail. In terms of the biggest hurdles for Recursion to overcome, to reduce that type of failure at that stage of the process — what are those hurdles, and how optimistic are you that Recursion is going to crack that problem?

I'm quite optimistic. Well, I certainly hope so.

It's money, and it's time — and that's the one resource that's non-negotiable for all of us.

Ultimately, the reason things fail in the clinic most of the time is not that your chemistry is no good. It's not usually that you have a molecule that gets metabolized and excreted too fast, or doesn't interact with the target you thought it would. We're fairly good at those things — metabolism and excretion failures get filtered out in phase one, if not earlier. By the time something gets to those later phases, you have chemical matter. The problem is that you didn't understand well enough the biology it was going to interact with in order to produce an effect. That's really why we've focused our strategy internally on the biology aspect.

We've recently made acquisitions and investments to bolster ourselves on the chemistry and chemistry-machine-learning side, because those are essential. But we think the transformative piece is really the deeper understanding of biology, on which we've pushed inside-out — starting with in vitro modeling. New areas we're focusing on are deeper connectivity to things like patient- and population-scale datasets and genetics, in order to drive tighter connectivity with what patient populations will be responsive: for whom is this a good drug, not just is it a good drug, period?

So bringing that question back down to those levers, those axes on which Recursion can continue to develop and build its strengths — where do you think the greatest impact is going to come from, in the acceleration of progress?

What do you mean by the greatest impact?

Let's say there's a spectrum of capabilities you guys have today. It'd be great to touch on the recent acquisitions too, because those have added a few more tools to the utility belt. I guess what I'm wondering is: what's the biggest hurdle to progress in the field? Is it one specific aspect across that spectrum of capabilities? Is it compute, or the power of the models? Or the quality of the imaging data? Or the size of the chemical design space? Or going from cell lines — one type of cell in a culture — to something like an organoid, a more sophisticated model? Where do you see the greatest potential progress across that landscape? It could be other things I haven't mentioned.

Musty was asking earlier what's hard about biology — seems like he should have asked the other Imran, because you gave a pretty good summary right there. The answer is that all of those things are hard.

We've made our investments in very particular areas. We think assessing high-dimensional biology in this way is going to unlock what we can do, and that, at least for us, given where we are, next investing in chemistry is the right move to accelerate the progress of that chemical matter. If we're talking about the field as a whole, I think all of those things are potentially true.

From the machine learning perspective — and this is more of the personal reason I came to Recursion — I think the biggest drag on doing effective machine learning in biology and chemistry has been not just access to high-quality, high-dimensional data, but the ability to generate more of it. What you find is a lot of high-powered computational teams who don't have the ability to make more data, and high-powered experimental teams who don't necessarily have the ability to model it well. I think building both the technical capability and the culture to bring those two things together is really transformative.

I'll answer the meta question first, and the answer is no, not really. It's interesting work and it's noble work. I didn't get into it because I thought it was easy — I got into it because I thought it was a good thing to do. And I'm also sanguine about it: not everything we try is going to work. But I have an underlying conviction that there are things in this space that will work.

There are a number of things that have to happen here. And we're also looking at where the goalposts are — this is a fairly slow feedback system. If you make a change at the beginning of a discovery pipeline, just the current lag in how long it takes to go from discovery to development means you've got probably a decade lag from any change to actually seeing it manifest on the market, or a little faster in trials. So this is something where some patience is required, and it's a little difficult sometimes to see exactly where that lands.

Coming back to the non-meta question — that's a challenging one. For the reasons the race to claim an AI-discovered or AI-designed drug is so hot, it's hard to tell exactly what that means. Different people mean different things by it. That said, I don't really have any doubt in my mind that we will see — depending on how you think about it — that either we already have seen, or we will see, drugs in which AI and machine learning methods were essential components to designing them and getting them into the clinic. In the future, not even necessarily the far future — potentially it's already happened.

There's this joke in the field: as soon as something starts working, you stop calling it AI, you call it something else. Under that standard, well, you're never going to have an AI-designed drug, because AI is just the stuff that doesn't work. But I think the underlying methods themselves are sound. I think the underlying ways of thinking are sound. There are necessary transformations in the way we work and deal with the complexity of biology and the data involved, and I think the impact there is going to be undoubtable.

Just to add to that — I think there was one drug discovered to have a potential impact during COVID, which was baricitinib, by a British company, BenevolentAI. I think they published something on this; it was actually used in patients in India, and there was a research trial to show some benefit. So, as you say, it's a really great point around what it actually means to be an AI-discovered drug in patients. Is it one where a connection for an already-approved drug has been identified? Or one where the clinical trial has been accelerated using analytics in some way? Or one where the fundamental interaction between a small molecule and a protein is actually originated by an AI model?

You could say that AI is in the eye of the beholder. Different people would call different aspects of that AI, and I think that's what makes the discourse so difficult. From my perspective, the interesting question is a little less the label you choose to associate with any particular thing, and more: can we zoom out? Is there a broader trend of these methods, these ways of looking at, interpreting, and acquiring data, making impact on how we do this business?

What are you most excited about, say, over the next five- to ten-year horizon in this space?

That's a challenging question — five to ten years is an awfully long time in this space. As I look at what's happening, I see the convergence of a few things. I realize I sound like a bit of a broken record, but I sound like this because it's real. What I see is an explosion in the flexibility and comprehensiveness of our computational modeling capacities in biology, paired with the scale and scope of the datasets we're acquiring.

Those take multiple forms. Some take the form of the datasets Recursion is generating — high-dimensional, high-quality data about biology, where you can put in whatever perturbation or state you're interested in and read it out. Some take the form of super-large-scale observational datasets. If I look back 10 years and think about the scale of genetic and genomic data that existed — there was a revolution in our ability to acquire that data with microarrays, then low-cost next-generation sequencing, then the ultra-scale stuff we're doing now. We're seeing big changes now, not just at the DNA and RNA levels, but in the ability to read out things like proteomics, lipids, post-translational modifications on proteins — that will give us a lot greater understanding of what's happening in the biological system.

That paired revolution of data acquisition — both people and assays on engineered systems to probe biology, and computational methods to model all of it — that's the thing that gets me really fired up. That's the trifecta you need to be able to observe, experiment, model, and close that loop and get better at it.

Awesome. So gather more data, find more and better ways to interact with those biological systems, and then more powerful ways to process everything you're learning.

That's right. A model is only interesting insofar as its prediction is useful. We talk about decoding biology — you could read that as understanding biology. And the way I read understanding is: I can make a prediction about what's going to happen if I make a particular change. Our ability to predict the results of an experiment is increasing, and will only increase, by application of that. Do the experiment, see what happened, see if you can predict the next one any better, and improve in that cycle.

That's very kind of you. There are maybe two different kinds of answers. In terms of habits, I am sort of aggressively a generalist. Especially in academia, there's a strong tendency to go down a very narrow rabbit hole, but my interests are extremely broad. One of the things that's brought me success and understanding is saying, "Yeah, I'm interested in the genetics of this, and I'll take the biochemistry class and understand the kinetics of how this is going — but I also want to zoom out and see, okay, the fundamentals here are actually the very same thing we're doing over there, we're just calling it a different name. And this is how we can put these pieces together."

Being willing to come out of that rabbit hole and engage with other disciplines and people in other disciplines is really critical to the space, because it's so broad that no one person is going to have all of it in their heads. Being able to understand the words the folks on the other side of the aisle are using is really valuable. One of the things I find myself doing a lot at Recursion, and elsewhere in my career, is simply playing translator — because you'll have different people nominally working on the same problem, nominally using the same words, but the words mean completely different things to each other. Being able to bring that together is really critical.

Oh, constantly. The way you say it — my computer science friends think I'm a biologist, my biology friends think I'm a computer scientist, and the chemists will definitely not accept me as one of their own. So, absolutely, constantly.

There's a grain of truth in what you're saying: there's a difference between being a generalist and being a dilettante. You don't want to be superficial. To be a good generalist, you actually have to have enough depth in a number of these areas to know what's what. But at the same time, it is possible to go so deep on something that there's more gain from learning another thing.

Not every position is a fit for that, and it takes some amount of looking. Before I came to Recursion, one of the things I explored was going back to academia. I'd kept up my publication record, I was interested in being a professor. And what I realized was that that environment really selected for — really wanted — people who are hyper-narrow in one thing, rather than that kind of generalism, bringing together a number of fields. It's changing; there are some departments and disciplines where that's more desired. But there's a strong cultural inertia — for understandable reasons. When you want to push the frontier of a small piece of knowledge, you want somebody who's very narrow. So some of it is a matter of finding the right environment for what it is you want.

I'm not sure I like that. Anyway — one thing I'll give a shout-out for, something I did that's a little unusual for folks who go into the sciences: I was an avid speech and debate person in high school. Most of my friends from those days are now lawyers, so it's sort of obvious why I'd select for that. But the experience of needing to communicate clearly and concisely about complicated topics, to a mix of audiences — some deeply versed in the technical aspects of what you're doing, some of whom may be lay people — is a skill that's really stuck with me and that I've found generalizes across all the contexts I've worked in.

I remember having relatives ask, "But you did all that debate — how come you're not taking the LSAT? Isn't that a waste?" Absolutely not a waste. Fundamental skills that have served me well through the rest of my life.