Invest Like the Best with Patrick O'Shaughnessy
· Patrick O'Shaughnessy
[00:00:00] S0: I have two questions here. The process itself is beautiful and has driven so much discovery and progress, not just in medicine, but scientific method-driven progress everywhere. But I'm curious about two dimensions of shortening that feedback loop or increasing the velocity of new discoveries and progress. One is AI. And the second is what I'll call citizen pharmacology, which is people. This is becoming really prevalent in what I'll call the peptide community. People love talking about peptides these days. GLP-1 is a peptide. Collagen's a peptide. There's well-known things that are peptides, but there's this whole class of things that we can talk about. But it seems like if you take to Reddit, for example, you will find groups of people organizing to test new things on themselves outside of the normal RCT, FDA-approved process. If you shorten it, you probably increase the danger, but you decrease the time to learning or something. So you've got both of these things as interesting aspects to me of what will drive forward progress in drug discovery. Riff on each, maybe starting with AI. What is real already? What do you think will become real by virtue of the increasing skill level or reasoning level of the models?
· Alex Karnal
[00:00:00] S1: Absolutely. So I think what you're basically describing on the AI side is our quest to try to get to a level of scientific superintelligence. You can imagine that from an AI perspective, if scientists are armed with an agentic infrastructure that has a scientific intelligence that's at an Einstein level or multiple of Einstein level, the force multiplication that can occur between a superstar scientist armed with that agentic resource means that they will do science faster and with higher probability for lower cost and have greater impact. So that's important conceptual framing. And so then practically, how does that manifest and why does it matter? Well, you can imagine any human scientist on their own, tinkering in the way that I just described, scientific method and that wheel of science spinning and spinning and spinning. Us as human scientists are always going to be constrained by what we can retain in our human minds and constrained by our ability to use our human hands to go run those experiments. There are super talented, incredibly smart people that can retain volumes and volumes of scientific papers. Can't retain all of them. AI can. And if humans are running experiments, not AI driven robotics, there's just a certain throughput that is going to be limited by human hands, that would be unlimited if you had an agentic, robotically driven system. At the simplest level, AI coming into the discovery of medicines is one of the biggest unlocks that we will see in the history of medicine, because it means that we're no longer constrained by our minds. We're no longer constrained by our hands. We can generate not just a few hypotheses for what the answer can be from the total sample set of knowledge that exists. We can identify all N number of hypotheses that could exist and should exist and then figure out what is the greatest and best next experiment to go run, run that, and run those in real time, 24 hours a day, seven days a week. The sheer ability to be more comprehensive and complete in our hypothesis set and then the sheer ability to run experimentation at a scale that's unprecedented, I think practically means we are going to be able to speed breakthroughs from timelines that would typically take three to five years to timelines that can take less than a couple of years.
· Alex Karnal, Patrick O'Shaughnessy
The other important element when we think about AI and its impact on drug development is not just the potential of what could be from that conceptual framing, but where are we today? I think it's really exciting. Today, I would say we're in the early innings of being able to discover novel targets. And I would say that we have not yet, in a systematic way, been able to discover a target that didn't exist yet. But what we can do is for targets that we know exist, but that we have really struggled to be able to develop medicines against those targets – we're starting to see that code be broken by AI, and AI being a big force multiplier to enabling that to happen. What we have today are companies that can go from what would historically been running these massive screens that are incredibly time consuming, to doing all of that screening work in silico, and can go from model to molecule in what would have taken a couple years' time, in a month. [00:00:00] S0: That's already true today? [00:00:00] S1: Already true today. There are companies like Lila Sciences and Nabla, these are companies that you can basically ask the agentic system to develop you a molecule against a certain target, and it can crack that code and give us something that we can work with and start running experiments in a month's time. [00:00:00] S0: And is it your impression that this is just going to happen? There's no major risk left that AI will not have major impact on this entire, both target and therapeutic discovery process? Are we now on a curve that just feels inevitable to you?
· Patrick O'Shaughnessy, Alex Karnal
[00:00:00] S0: And is it your impression that this is just going to happen? There's no major risk left that AI will not have major impact on this entire, both target and therapeutic discovery process? Are we now on a curve that just feels inevitable to you? [00:00:00] S1: Yeah, to me we're on that curve. [00:00:00] S0: When did that flip? How recently?
· Alex Karnal
[00:00:00] S1: Very recently. For most of my career, there's been some version of machine learning and the hopes of AI really driving drug discovery and the development of novel medicines against those discoveries. And I always worried that in order for that paradigm to be true, it required data that didn't actually exist in any easy place. If all you do is train AI models based on the research that's available, you unfortunately have a paradigm of a lot of garbage-in, garbage-out. I think one of the things that people don't appreciate when I start talking to them about what data exists is that unfortunately, a lot of the recorded literature is actually incorrect. And there's been tons of studies that show if you go try to replicate the experiments that are in the literature, you don't even get the same results. Definitionally, if that is true, and these AI companies don't have a source of creating novel data, what are you going to do? How are you going to train? We've seen just how powerful LLMs have been and how impactful they are across so many different areas of life because they've been trained on trillions and trillions of novel tokens. So data ends up being a big part of what drives to advanced levels of intelligence, it drives an intelligence revolution. Well, the same has to be true in science. And so the AI companies that I believe are going to be most set up for success are the companies that have incredible AI talent, have access to significant amounts of capital, and can combine both of them with a novel way by which they can generate science tokens that don't exist in the public domain. In fact, the generation of those tokens becomes a big driver of their moat. Not only the driver of the moat, but actually what enables them to be able to get to an increasing and increasing level of scientific intelligence that starts to separate from that which can be achieved with the LLMs, without having the benefit of those tokens. And we're starting to see that. And so we've got one company that we're really close to today that's actually starting to show the bending of that curve and helping us to see that the idea of getting to scientific superintelligence is one that is likely to follow a pretty deterministic set of scaling laws.
· Patrick O'Shaughnessy, Alex Karnal
[00:00:00] S0: I think it was OpenAI that came out with this announcement, maybe working with Ginkgo or somebody, where they've created a closed-loop automated lab and discovery process of not just models, but physical world testing as well. If you close your eyes and think five, seven, ten years from now – what's your best guess, knowing what you know, at what the discovery process will look like at that point? [00:00:00] S1: I think it's going to be completely automated. I was up at Lila Sciences the other day and you go into their labs and what you see is all of these robotic arms and you see petri dishes that are moving around in an automated fashion, moving from one spot to the next spot. To me, I think what that means is the scale of data that's going to be required to be able to train these models, the trillions of tokens that are required – has got to move beyond the constraints of human hands because it would just cost so much money to do it in a human capital-driven manner. It has to be done robotically and for that to be done, it has to have an agentic engine behind it and needs to produce data at a scale that's kind of unprecedented. What I think the future of discovery looks like is agentic systems that can drive robots to do most of the experimentation that we're doing today with human hands. And I think that's a big unlock for the ability to crack the codes. [00:00:00] S0: That's the top of the funnel.