Daphne Koller - The Future of AI in Biology and Drug Discovery

Hi Ravid. It's a pleasure to be here. Thank you for inviting me.

I think I agree with a lot of elements of the bitter lesson. We have absolutely seen that if you have a domain where there is a lot of data and where your task is defined as pattern recognition, pattern matching, you can find those patterns. The machine can find those patterns in the data like nothing else has ever done, including the human brain, and extract incredible capabilities to generate data, generate responses, generate insights. I think that has been a transformative technology for the field and it is absolutely derived from the bitter lesson, but I don't think that means that structure is irrelevant, and I also don't believe it means that causality is irrelevant because the other element of probabilistic graphical models is that they do provide, specifically the Bayesian network and its various progeny, a formal framework on causality. So let me talk a little bit about each of those separately. Starting from structure, I think structure matters in a few different places. One is in domains where no matter how we think about the world, data will never be quite as much as we would like. In the domain that I'm in, which is biology, the complexity and richness shows that we're going to need a tremendous amount of data that is not available to us today. Now, I'm not saying that there's never going to be a world where we have incredibly sophisticated measurement devices, but we're not there yet. And when we don't have that amount of data, then I think you need to combine some of the technologies from the bitter lesson in places where sufficient data exists, but also leverage structure in some ways to create a more compact representation. That's number one.

I don't think pattern recognition and function approximation are wrong. I just think they're not the only thing that we need. So I think the right future lies in taking those incredible capabilities and incorporating them into models that also include elements of causality and some higher level structure. And if you talk to the people who are the greats in this space, like Yoshua Bengio and Yann LeCun, they will tell you that structure and causality are the next frontier.

So that's a great question. First of all, let me say that the tools that have been developed, the automatic coding agents, co-work, all of those technologies are incredibly valuable and we use them all the time in our work, including agentic technologies for helping design experiments, monitor the experiments, interpret the data from the experiments, and so on. It's absolutely an incredibly powerful capability. I would also say that there are certain sub-problems within the space of biology where by a combination of sufficient data collection and the fact that these problems are sufficiently close to physics, we have been able to create models that are reasonably predictive. Protein folding is a great example of that. Tremendous amount of effort on data collection. 40 years of people collecting sequence to structure data and a domain where the prediction problem is very self-contained. You have a sequence, a protein. You want to predict that singular event of how does the protein fold or arguably how does a ligand bind to the protein. And it's a problem where the fundamental underlying principles are basically those of physics. The other problems in drug discovery, however, are not about the chemistry or the way in which a protein folds. It's about the way in which human disease comes to be and the intervention points that allow us to stop or ideally reverse the disease process. That is a question that has to do with measuring not individual proteins, but the multiple interconnected layers of biology that go from DNA to RNA to protein to cellular organelles to cells that talk to other cells within organs, circulating blood, whatever compounds, molecules that circulate in the blood and are degraded at different rates and go to different places based on signals that we have never measured and do not understand. And so the fundamental problem in making a drug today isn't when you get to the singular well-defined place of okay, I need to make a molecule that does X. The singular problem in drug discovery is figuring out what that X is. What is the X that I need to intervene in in order to revert disease? And the fundamental even bigger problem is that the places where that question has the biggest unmet need is in diseases where the only place we can really measure them is in human beings because there are certain processes that are relatively conserved. Cell metabolism, for example, pretty well conserved and you can measure it in yeast and other systems. Cardiovascular function arguably you can measure it in mammals and what you measure translates to humans. Neurological diseases do not exist in other lower order species. Even non-human primates do not get ALS, do not get schizophrenia, do not get Alzheimer's disease. So in order to measure those processes, you have to measure them in humans, which makes the whole measurement problem far more difficult, far more expensive, and frankly far more ethically fraught.

Boy, so I think there are two halves to this question. There is the half that has to do with drug discovery and how do we retain appropriate ethical oversight on the data collection that we need for drug discovery, and then there is the other half which is how do we make sure that researchers that have never interacted with human subjects research know what is the minimum set of ethics. I mean honestly for the latter half I wish I had a better answer for you. We have been running experiments on human beings at scale in the tech industry for quite a while. I mean some of the experiments that are run by social networks under the guise of AB testing. We know that those have been very problematic in some cases and the recent spate of lawsuits about did you encourage my children to commit suicide or have an eating disorder are the results of algorithmic refinements which are basically driven from experiments in human beings and no one has exercised ethical oversight over those experiments. I mean, if we were doing experiments like that in a controlled setting within an academic environment and asking whether certain types of intervention encourage teenage girls who are not even adults to have an eating disorder, there's no way that would pass an IRB. But yet we do it because it's somehow outside the sphere of a regulated environment which frankly I think has led us to a lot of the issues that we see in how certain technologies are deployed and the impact that they have. So I'm going to put that aside because that is above my pay grade. That is something that governments should be thinking about how to create the appropriate oversight on experiments in human beings at the scale of billions. I think for the purposes of what we do, it's more of a question of how do we create the right resources with the appropriate ethical safeguards to give researchers the appropriate data that they need to measure diseases that are either intrinsically or predominantly human at sufficient scale so that we can learn what the right intervention points are. And this is a place where I think some really valuable steps have been taken. The announcement yesterday on the official launch and enrichment of the All of Us corpus here in the United States I thought was a huge step forward. We make tremendous use of national biobanks. UK Biobank has been our go-to resource for human data. And we and many other researchers have extracted unbelievable amounts of insight from that resource. It's not only about medical issues but also about issues of socioeconomic intervention effects and factors and education and so on, which have been hugely valuable to understanding how to make the world a better place for people. I really applaud All of Us for following suit and learning from that and creating a resource that has some limitations relative to UK Biobank specifically in the richness of the phenotyping but also some huge advantages specifically the breadth and diversity of the cohort because UK Biobank is largely Eurocentric and All of Us was by design intended to capture the much greater diversity of the US population. So I think creating additional resources like that, ones that are ideally even larger, maybe not. I mean, you have to kind of always balance the depth and the breadth. And so how do we create an appropriately consented, appropriately anonymized, appropriately protected cohort, but one that unlocks the amazing amount of value that exists by studying human beings and extracting the kind of insights that you are never going to extract from any other system. Not from animal models and not from single cells because you can study reductionist pieces of disease in single cells but if you're studying things that are complex multicellular emergent properties of a complex system, we don't have a simile of a human being that we can study in a dish and I don't know when we will have such a thing.

That's a great question and I wish we might have started with that because I think it would have allowed me to give a better answer to one of the earlier questions. So, drug discovery has largely three phases and each of them is absolutely critical. The first is what I call disease to mechanism. So it's taking a complex, multifaceted, often ill-defined disease such as ALS or Alzheimer's disease or heart failure. Often times these are not one disease. There are multiple biologies that converge to create what we as humans have declared to be the name of the disease. So this is not a biological definition. It's a symptomatic definition. And you need to extract from that an individual biological mechanism which is where you need to make an intervention. Like for example, I want to stop this protein binding to this other protein. That is an example of a mechanism. So going from a disease to a mechanism where you'd like to intervene is step one. Step two is okay, I want to stop the interaction between this protein and that protein. I need to make a molecule that does that. So there's different modalities. There's small molecules and biologics and RNA drugs and whatever. You need to pick the modality and you need to make sure that it actually does the thing that you set out to do. And that furthermore, it does it within the construct of a molecule that you can actually deliver to humans. It doesn't clump up and turn into a gel. It goes into the right place in your body. It doesn't degrade within 20 minutes. And it doesn't bind to all sorts of other things that you don't want it to bind to that might actually have really detrimental effects. So that's what I call mechanism to drug. And then the third piece is drug to patient. So okay, so I have a molecule, I have a disease, which are the right patients, when do I deliver the drug? How do I measure whether the drug is working? Like how do I prove the point that the drug is actually having the desired effect? And each of those phases is really important. Each of those, by the way, can hugely benefit from AI technologies, but they are quite different. Now when I go back to some of the questions that you asked earlier, a lot of the big successes that we've seen in AI deployed towards biology sits in the middle bucket which is I want to design a molecule. People like that, they gravitate to that particular question because it's self-contained. It's well defined and you can achieve or not achieve success within a very limited amount of time say 18 months. You can look at the molecule and say I set out to perturb this particular binding property for example like the binding of this protein to that protein. Have I successfully done that? And is the molecule a good molecule? That's great.

You're absolutely right. So to me 22% is the upper bound of things that we understand. It's not the lower bound. Many of those drugs were discovered by serendipity. Many of those drugs are actually addressing the symptoms. They're not addressing the root cause of the disease. They're like pain relief drugs that don't actually address the root cause of the pain. So yes, the number of drugs where we actually are intervening what are called disease modifying drugs is much much smaller than the 22% that we talked about. So yes, the vast majority of diseases we simply do not understand. And so if you're asking me what is the core problem that we need to solve in the drug discovery space, it is the problem of biological understanding. It is not the problem of making a molecule.

No, you look you're absolutely right that we don't we as humans do not need to understand but we do need to be able to predict at least one of the consequences of the disease. And then you're right often times sometimes a drug surprises us and it does better than we expect. Sometimes by the way it surprises us by doing things that are not so good that we also didn't expect and so the ability to predict would actually be useful across many things but in order for you to go through the let's just try it and it works, the only experimental system that we have that allows us to do the right experiment is called a human being. So I mean I'm hoping you're not advocating for putting a whole bunch of random stuff in a human being and seeing what happens. Although the Chinese peptide craze, I guess, is a way of doing that, but no one's actually measuring what that actually does. So let's hope that is not a good scale. So, do you need to be able to predict everything with accuracy? No, I don't think that's a I mean, I think that's an aspirational goal that we're nowhere close to. But you need when you're coming and saying I want to find a drug against ALS, putting a whole bunch of random compounds in ALS patients and hoping for the best is not a strategy.

I mean, I think that experiments like the GLP-1s that have been done at the scale that they have been done are an incredible source of insight and pharma companies have taken some of those observational learnings if you will and turned them into rigorous clinical trials and sometimes they work out and sometimes they don't because there's a lot of confounders in human observational data. So for example, a lot of times we see that things work really well in a particular group of patients that are treated with a drug versus ones that are not treated with that drug because the people who are treated with that drug tend to be of higher socioeconomic status have access to better health care and therefore they do better because they have access to a better drug not because the drug is working but because they are from a privileged population. So just coming back to our earlier discussion of association versus causation, there's a lot of these examples that have not panned out because there was a confounder that wasn't appreciated at the time that the observational data was collected, which is why clinical trials, randomized clinical trials are and should remain the gold standard for understanding whether a drug works. Having said that, I'm not I'm absolutely not dismissing the potential of the observational findings in GLP-1s. What we've learned from GLP-1s is that absolutely many things are downstream of metabolic syndrome and these drugs not only address metabolic syndrome but many of the downstream consequences. I think that was predictable and it's panned out which is amazing. We've also seen that even in patients who do not have metabolic syndrome and who do not experience weight loss, we see a diminishment in other things due to I would say parallel pathways that don't go through weight loss that go through things like inflammation goes down in these drugs before you even see the weight loss. And we're scientists, not we specifically but scientists are trying to understand how there is an effect on inflammation that doesn't go through weight loss. And then to your point, yes, this acts on the brain and acting on the brain interestingly suggests that there is these common mechanisms of addiction that leads to addiction to food in a similar way that you have addiction to smoking, drinking, gambling and so on. I think that remains to be proven in rigorous clinical trials but it's a very exciting mechanism if that turns out to be the case. So coming back to the meta level point we should be learning from human data as much as we can anytime we can because human data is the thing that keeps surprising us and giving us insights on the thing that we care about the most which is how to cure humans. And so the better we get at collecting, processing and making fit for purpose for machine learning human acquired data the faster we will learn about human biology and human disease.

So the R&D function of Insitro is half wet half dry. So half of our people basically make data, do experiments that make data and then take the insights that emerge from the data in order to design the next round of experiments. And some of those experiments are at scale data generation that are intended to map human biology or in this case cell biology at frankly enormous scale. We've generated tens of billions of cells so far and continue to generate more. We're, you know, growing. I think we've doubled capacity three months ago. We've doubled capacity again this month. We intend to double capacity yet again next year. So I mean in terms of generating cellular data and then...

It's a combination of it's a there's a lot of kaizen going on multiple things that come together to allow us to double capacity. So lab automation absolutely part of it. Better biochemistry so that you can get better yield from each plate of cells another part of it. The agentic technology that allows us to go faster from a concept of we would like to measure a certain thing to how quickly we have a system in which that thing can be measured in a scalable way. Another huge thing. So a lot of different contributions. Some of them from lab automation, some from agentic automation, some from better biological tools that we develop and constantly improve and enhance like we've doubled yield capacity, the yield of a single experiment by getting better readouts from ourselves by measuring them in a different way. So, it's actually amazing to see how each one contributes 20% 30% either to yield or to time or to volume and so on and it just keeps getting better and better. The other half of the company does the machine learning AI on what we do. So these are machine learning scientists, software engineers and so on. And many of the improvements also come from that side. Like for example, if we have AI methods that can for example autofocus by looking better at the cells, we again increase by 20 to 30% the number of cells that we can read. The number of cells that we need in order to derive insights because we have better embeddings of our cellular images. Again, you get an improvement from that. The ability to automate the process with the scientists of the experimental design using agentic technology. And so all of these actually come from this act actually intertwining of the kind of wet lab components and the computational components which is why the company is called Insitro because it is the seamless integration of in silico in the computer and in vitro which means in the lab.

I mean only in the small sense that we've been using things like for example alternative nucleic acids, modified nucleic acids, modified amino acids and various drug design that tends to create molecules that are less susceptible to the standard degradation processes of the body. So they're longer lived, have different efficacy profiles. I think that's amazing. I don't know enough about silicon-based biology to try and hypothesize whether that is a thing.

So we use both. There are certain cases where the problem, the objective that you're trying to achieve, is well-defined. So when coming back to our earlier taxonomy, once you've reduced the biology to a mechanism and you're trying to design something, that is where you have a very well-defined objective function that you can measure and a circumscribed search space, like the space of whatever protein sequences or a particular set of assay conditions, and you're optimizing within that space. The agents do an amazing job of, you know, we generate a bunch of experiments, you run them through together, you collect the data, you see what the results were, you learn from that. You can think of it as kind of like an agent-enabled simultaneous hill climbing. Sort of like the agentic version of what we used to call genetic algorithms, which are obviously long obsolete. But you have a bunch of experiments, you simultaneously optimize and get better and better. That's awesome, and we do that all day long. A lot of the problems that we are facing are ones that belong into the ill-defined category where you can't even tell the agent what it is that you're looking to do, other than, you know, I'm trying to make a drug for Alzheimer's disease or ALS. And we have not as of yet seen agentic technology that you can just let loose on that problem and it comes out with something meaningful. And that's where I think human taste, human judgment, human critique, human prioritization of a lot — I mean, agents are great at coming up with lots and lots of ideas. Picking which ones are actually more likely to get you to a meaningful answer — that's where a lot of taste and judgment comes in. So I think for a lot of the building block problems, you might call them, we use automated agentic technology and frankly are hoping to automate more and more of that as the integration between the agentic and the lab automation becomes tighter, we can just make that loop entirely closed. The higher-level thinking pieces, it's very much a partnership with a human.

I think absolutely it will become more efficient. I mean everything is becoming more efficient by leveraging, you know, modern-day AI, modern-day interpretation methods, agents that help you think through problems and so on. Everything will become more efficient. I don't think we are at the verge of having AI just go out and solve the problem of curing human disease. I mean, I think there are multiple challenges. The data question being among them. The slowness of the experimental cycle being another, because as we discussed, sure, you can experiment on a cellular system or a biochemical assay very effectively and do this closed-loop iterative design where the agent gets this feedback. Agents are amazing at learning. When you give them immediate quantitative feedback and they can just optimize to that, they will just bang their head against that feedback and get better and better. The systems that we work with are not the ones where you get that type of feedback. Maybe you can argue cellular systems are somewhere on that spectrum because you get feedback within a matter of weeks — not seconds, weeks — because, you know, cells need time to grow and respond to the stimulus, and no matter how many GPUs you throw at the cells they're not going to grow any faster. So even that is not your immediate feedback loop. And when your experimental system is a human being, then the feedback loop is not weeks or months, it's years. And so there's a — that's not the problems that AI is just going to crack instantaneously.