Dave Ricks, CEO of Eli Lilly, on GLP-1s and the business of pharma

**John** (0:00)
Dave Ricks is CEO of Eli Lilly, which is now a $700 billion company and the world's most valuable pharma company. Eli Lilly is 150 years old. They grew up as the first company to mass produce insulin in the 20th century. But today, most of the company's business is in the new GLP-1 diabetes and weight loss drugs, where they've become the market leader. Simultaneously, Eli Lilly is upending the traditional model by selling directly to their consumers over the internet with LillyDirect, rather than through the traditional middleman.

**Patrick** (0:27)
All right. Cheers. Thanks for coming.

**John** (0:31)
I'm pretty impressed that you came and you just poured your own pint.

**Dave Ricks** (0:34)
Poured my own pint.

**John** (0:35)
Major flex.

**Dave Ricks** (0:36)
Have glass will pour.

**John** (0:37)
Exactly. Actually, a good place to start. Tell us about your Nvidia announcement that you just had.

**Dave Ricks** (0:41)
Yeah. So today at the, what's it called? GTC conference they have, they unveiled that were well underway actually, should be done by the end of the year, but building a supercomputer on-prem for us, really just to run proprietary drug discovery models. We think it's the biggest biologically focused supercomputer there is. And certainly the biggest pharma is done. Yes. With B300s latest chipset. And yeah, we're only constrained by power like everyone else. But yeah, we built a bunch of tools, we'll run them on that. And scientists use it to sort of co-invent, co-develop, focus mostly on chemistry to begin with, but we'll expand from there.

**Patrick** (1:22)
And so is the idea here you have some target, you've had some challenges actually drugging it, and so you give it to one of these new chemistry models, and you ask it whether it can come up with something totally orthogonal beyond what a human might have tried.

**Dave Ricks** (1:37)
So take a really good popular example is like GLP-1. So that's a hormone peptide that we all excrete.
And it engages targets that are what we call G-protein-coupled receptors. So they're hard to drug targets on the outside of cells. And to try to mimic a big, huge protein with a very small chemical is a complicated undertaking. And by the way, do only that and not other things that are untoward. And so this is sort of a frontier of drug discovery that's been tough and very empirical. That's a hot area for this kind of technology because these strange arrangements of atoms don't look like other drugs that have come before, but they do follow the principles of organic chemistry and seem to engage these targets effectively. I don't know of one that's come through the machine-driven discovery process, which is really machine plus human, that's made it to the clinic yet, but they're coming. And I think that's exciting because those have been structures that don't exist in nature and yet the machines are alien and they can predict these interactions.

**Patrick** (2:45)
I'm always struck by Derrick Lowe's arguments, where he's always sounding this note of caution, I guess, about the optimism and maybe what he might view as boosterism around AI and biomedicine, where as I see at least his two claims are, one, it's really hard to select the targets and AI doesn't help you that much there, and then so much fails at human toxicity. And again, at least so far, AI has not been all that helpful at that step. Do you agree with him or is he overrating these particular challenges and maybe underrating the challenges that AI does help solve or thoughts in that argument?

**Dave Ricks** (3:26)
Probably we need to create the equivalent of what got created with human language, which is a more complete repository of biological knowledge to train against before the machines get a lot better. And today, I don't know, I would estimate we might know 10 to 15 percent of human biology. So the machine is not going to be good at all until we get way above 50 percent. That probably requires robotic 24-7 experiments just to create training data sets. You know, sort of this kind of big lift effort, the kind of thing actually NIH should be doing right now, I would think. But that effort is not ongoing, at least in our country. But I think if that gets going, I think we'll know more and the machines get better at the harder big problems, system prediction.

**John** (4:14)
Patrick and I just didn't finish any college, so not only don't have formal training in computer science, but don't do formal training in anything. You did not come up through the science side of Eli Lilly, but you seem extremely comfortable with the science. What has been your method for ingesting all this stuff? And especially as you're essentially making science decisions at the end of the day with the top level of capital allocation decisions, just how do you learn?