Amgen’s AI Futures: Digital Twins, Unstructured Data, Human Review

By Allison Proffitt 

September 3, 2024 | Jerry Murry, senior vice president, process and development, at Amgen, gave the plenary presentation at CHI’s Bioprocessing Summit last month. With an eighteen-year tenure at Amgen—and experience at Merck and Pfizer before that—Murry has seen nearly two decades of advances in the world of bioprocessing and manufacturing.  

He spent a good deal of time talking about the advances in bioprocessing and bioreactor design, but he also touched on Amgen’s AI achievements and next steps.  

An early use case has been using automated visual inspection—computer vision—to confirm quality for syringes and vials. All product receives visual inspection. Initially this was done manually, but the percentage of visual inspection that has been done automatically using AI algorithms and computer vision has been growing. “We found [visual inspection] to be very limiting,” Murray said. “In some cases, we get rejection rates that are in the double digit. And you can't be throwing away 10% of your product at the very last step during inspection!”  

Amgen worked on refining AI and machine learning computer vision algorithms to increase the percentage of product that can accurately inspected, but it was a long process, that “caused a lot of concern in the overall regulated industry,” Murry said. “Validation and learning are two things that need to be balanced. If you have a validated process and you’re allowing the machine to learn, how do you keep that in validation?” Answering that question took several years and conversations with FDA, but today, Murry reports, 95% of syringes and vials are released through automated visual inspection.  

Digital Twins in Manufacturing

The most important hallmark of next-gen manufacturing, Murry said, is applying advanced predictive technologies and AI. Here Amgen’s aspirations are quite high. “We want [predictive technologies] to live in our manufacturing network, which is where we get most of the efficiencies.” Amgen wants to have predictive in silico models for every unit of operation and visual twins for every deliverable—every pre-filled syringe, every bioreactor, every filling or inspection line.  

“All of those have digital twins that allow us to digitally find defects and optimize the running of the line before we every actually run anything,” Murry said. “That allows us to think about automation and robotics and allows us—now that we have usable artificial intelligence—to be able to apply that to the running of those lines.”  

Those aren’t the only AI goals Amgen has. Murry says the company is exploring how AI might write Biologics License Applications. “If you think about a BLA—especially the module 3 CMC section [Quality: Chemistry, Manufacturing, and Controls] of the BLA—it’s very templated. All the same data goes to the same spot every single time” he said, making the BLA a good use case for AI. The Quality Overall Summary, in turn, comes directly from the BLA, Murry said. Today many scientists “sitting in the basement somewhere” pull together such regulator documents, Murry said. Instead, Amgen envisions AI creating a first draft from an Amgen semantic layer that is then reviewed by regulatory teams. AI can pull various summaries, redact where necessary, and more.  

“Early days,” Murry conceded, “but it works, believe it or not!” Amgen pooled thousands of the requests for information and quotes the company has received over the past 20 years and is using Microsoft Copilot over those data to fill out new RFPs and RFQs based on prior answers. Amgen still uses human review for accuracy, of course, Murry said, but he called the new workflow, “a game changer.”  

AI Vision Casting 

It’s a point he made again as he shifted his plenary session to a fireside chat with Ran Zheng, current CEO at Landmark Bio and a former colleague from Amgen. The conversation focused mostly on AI challenges and opportunities.  

Zheng started with Amgen’s approach to AI hallucinations. “At least right now you still need humans in the process. You still need people that have worked on the product and understand the data,” Murry said. But he didn’t rule out verification by another AI model. “We work with Microsoft, and Microsoft has built into [their AI] some checks and balances. It’s not perfect yet… but I think there is a time when we will be able to have checks and balances.” 

The space is moving really rapidly, Murry noted. “I would guess that at this time next year, we’ll hear use cases of people who have done this completely digitally. And hopefully the FDA is building up their AI so their AI can review our AI!”  

Amgen has several other AI implementations in the proof-of-concept stage, Murry said. AI has been used to monitor pH, glucose, and other parameters in a successful proof of concept study, though the pilots have not been used in commercial or clinical environments. The team has also used AI to write tech transfer documents, and Murry believes that AI could automate some testing.  

“The amount of time that we have people spending taking data from one area to another is just too much!  In fact, I was talking to a group of scientists just the other week, and one of them said, ‘It’s to the point where I would just rather rerun the experiment in my own lab than try to get the data out of that system.’”   

The anecdote carries with it a warning. Scientists don’t want to use clunky, outdated systems. And the next generation of researchers simply won’t. “I’m blown away by the emerging talent,” Murry said. “They don’t want to work in some of these older systems we have that are clunky and they’re not intuitive. The other technologies they utilize in other parts of their life are way more advanced than they technology that we’re providing them to do the most progressive scientific research.”  

There are certainly still challenges, Murry acknowledges, and he highlighted to areas for increased research and attention.  

Protection of data and security is a top concern. “I know for us it was a very big bar to get to the security that Microsoft offers us by using Copilot. For the most part we partner with Microsoft; Microsoft has a very robust data integrity package,” he said. “That was a very big concern of ours! An AI platform that could access 40 years of Amgen data could be a competitive advantage to somebody.”  

The other area of opportunity Murry highlighted was agility in dealing with unstructured data. The future is not models that need very clean, very structured data, he argued. “Before, when we were trying to use these other AI tools, we would have to spend months curating the data.” Over Amgen’s history naming conventions were irregular and legacy data were fragmented and poorly labeled. Prior knowledge is messy and spread between handwritten notebooks and old and outdated systems.  

“But the amazing thing with some of these large language models… are able to [actually use] all this crazy data that will never be used for anything else.”  

Curation is a huge barrier for many companies, Murry said. “If the tools require super clean information, that’s going to be a barrier to adoption,” he said bluntly.