Invest Like the Best with Patrick O'Shaughnessy
· 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.
· Patrick O'Shaughnessy, Alex Karnal
[00:00:00] S0: That's the top of the funnel. [00:00:00] S1: Exactly. And so now, superstar, brilliant, amazing scientists who go down their life journey to become a scientist that wants to fight against disease, wants to do work that matters – are no longer going to spend two-thirds of their career at the side of a bench pipetting all day long. They're going to have all that time back to point their energy and their enthusiasm, their optimism towards trying to figure out, how do I use the systems and machines that I have and point them in the right direction at the codes that we should be cracking? That's really enabled when you have the combination of brilliant human scientists, force-multiplied by the agentic power of these systems that are being built. And the more those agentic systems can be fully automated, where all of that lab experimentation is happening in real time and without the need for human interaction, I think the quicker and quicker it can scale, and I think the quicker and quicker we can achieve that level of scientific superintelligence.
· Patrick O'Shaughnessy
[00:00:00] S0: What do you make of this whole peptides plus citizen pharmacology side of things? I think they just came out and said, there's going to be 14 something peptides that are allowed to be compounded. And it seems to be like deregulation of sorts happening in this space. People say peptides, I think not really knowing what that even means. So maybe define it. I'm just curious what you think of this. It really plays to your earlier point about these generations – these younger generations especially – seem to be a new attitude towards all of this stuff, which is like, I want to own this. And part of owning it is a willingness to take risks that are above and beyond what is mandated or regulated that I'm allowed to take. This subculture is fascinating to me. You read about BPC-157, I think that's the number. All these interesting different things that are being tried. Do you have a take on the impact that this will have?