In the pharmaceutical world, AI drug discovery is a topic that has been discussed endlessly with everyone, zeroing in on the belief that algorithms can magically trim timelines and drastically bring down the costs of research and development. It’s as if AI represents the “Easy Button” for the industry.
That is wildly overhyped, according to Diogo Rau, Eli Lilly’s chief information and digital officer.
“So even though I am a huge AI proponent, I’m also the first one to say, ‘No, AI doesn’t mean that we’re going to cut drug discovery times down from 10-year trial development times … down to two years,’ which is something that I swear everybody wants me to say every time,” a smiling Rau declared in a recent interview.
He went on to add more detail.
“If we do everything perfectly, we’re still going to have to wait for the biology to work, for medicines to work for your body,” Rau said. “You’re not going to be able to take [away] that much past five years, even if you get every single other thing out there. So I do think that overhyping is a potential killer for this industry.”
This view may seem a bit ironic given how Lilly and its tech partner Nvidia have announced their joint AI efforts that will occur in South San Francisco, where cross-functional teams will devote their energies toward AI-propelled drug discovery. But it’s certainly more realistic given several AI drug discovery firms have not made meaningful progress —at least not yet.
Before joining the Indianapolis-based pharmaceutical company, Rau spent 10 years at Apple as head of engineering for Apple’s retail stores and online store. Hence, this is certainly not a perspective borne from technology skepticism. Rau explained that not being able to cut down development timelines as aggressively as some wish doesn’t mean that leveraging AI is a mistake or that AI drug discovery isn’t the direction everyone should be headed.
“I don’t think anybody in their right mind would argue that in the 2040s, 2050s, the majority of discovery work is going to happen with humans in lab coats, the way that they were discovering medicine 100 years ago,” Rau said. “But we’re not quite ready for that change yet. But I think we need to know that that’s the direction that it’s heading. Discovery is probably the place that has the biggest potential in my mind of every aspect, but it’s one of the hardest to crack.”
Meanwhile, he’d rather talk about other areas of the biopharmaceutical industry where AI can have a great impact — areas that people are less inclined to wax lyrical over.
Rau said Lilly has adopted AI in its manufacturing processes because manufacturing is a repeatable process particularly well suited for a technology like artificial intelligence. Take glass containers for medicines. AI is being used to monitor the product and make sure there are no defects.
“We do something like 70 or 80 photographs … per autoinjector that comes off our manufacturing lines in a few hundred milliseconds, taking them from all angles,” he said, noting that it far exceeds human capacity to detect errors.
It’s an example of AI boosting safety.
“How often do you really truly get manufacturing defects in containers for medicines anymore?” Rau challenged. “And I mean, that’s been basically eliminated with AI, so that’s a very real thing.”
Another area where Lilly is using AI is demand forecasting, which he described as being very crucial for manufacturing, especially when it comes to shifts in supply chain.
“AI can think a lot more deeply into your supply chain, can figure out patterns a lot more, can predict demand signals much better, and that’s definitely outperformed humans as well, and is a very real opportunity that we’ve captured,” he said.
Rau added that digital twin technology — something that has also been hyped on the clinical trial side and has yet to pan out — is something that Lilly has leveraged in its manufacturing, especially in the production of one GLP-1 drug. Lilly had what it thought was an optimal manufacturing process built and tested by engineers involving a particular piece of equipment that was on the “critical path of the manufacturing process,” Rau recalled. The company layered on AI to duplicate the manufacturing process virtually.
“We modeled the device, we modeled the machine, we modeled the inputs and everything else around it, we modeled the steps in the process. We were able to replicate it with very high fidelity so that the digital twin very accurately predicted how everything was going to behave in terms of performance, temperatures, all kinds of things like that,” Rau explained. “We then used the digital twin to run a huge number of simulations of different configurations, different process steps … and what was probably surprising to all of us was the optimal solution that it came up with, which was quite a bit better, really turned out to be true in the physical world.”
So the model spat out a process that worked well virtually, but was replicated in actual manufacturing.
He declined to say how many additional units of that GLP-1 medicine Lilly was able to manufacture with the AI-reconfigured manufacturing process, but that’s precisely what occurred.
“I don’t know if I’m going to be able to be able to disclose how much we produced in addition to it, but literally our revenue numbers would be materially different and the number of patients, more importantly, the number of patients we would’ve reached would’ve been materially different this past year if we had not used AI to digital twin a critical step in our manufacturing process,” he said.
This use of a digital twin can be applied to many different manufacturing processes and not just in creating GLP-1 drugs though the revenue gains will not be as large.
“The bigger point that I’m getting at is sometimes you have revenue opportunities in all kinds of places besides drug discovery if you do things right,” he said.
Photo: claudenakagawa, Getty Images
