Artificial intelligence (AI) is learning to mimic expert scientists and thus improve the processing steps needed to produce cell therapies. Multiply Labs and Stanford University’s Laboratory for Cell and Gene Medicine (LCGM) are tackling this together, with the goal of scaling cell therapy processing more quickly than currently possible.
Today, not only are cell therapy processes artisanal, but the robotics coding is manual, Fred Parietti, PhD, co-founder and CEO of Multiply Labs, notes. The advantage imitation learning and other advanced robotics control techniques bring to cell therapy manufacturing, he says, is that, “we’re automating the automation. Engineers are teaching the robots to train themselves.”
The process isn’t as simple as telling robots to “shake a flask,” Parietti explains. “We’re seeing that, for example, when people need to resuspend cells, they shake the flasks in their own particular ways, and that’s hard to program.”
In fact, Multiply Labs engineers tried to suspend cells using simple motor robotic motions, like moving only one or two robotic arm joints at a time. “That experiment failed spectacularly. The cells [in the flask] didn’t move.” But, by mimicking the exact motions of expert cell processing scientists, the robots can learn the motions and replicate them, thus achieving cell suspension.
To do this, Parietti says, “Currently, we invite the scientist to our offices, put sensors on the flask or the tool, and record their exact motions. Our vision is for scientists to send us a cell phone video of them performing a certain task,” which the robot would learn to imitate. That phase may be possible by the end of this year, he speculates.
Best for unstructured tasks
AI imitation learning is best used for tasks that are unstructured and poorly described, like cell suspension, mixing reagents in a bag, or spinning or rotating containers in certain ways for prescribed times. Duplicating the elliptical pattern needed for magnetic beads to bond to cells, is another example.
Compared to manual processes, Parietti says robotic automation lowers costs by about 70%.
“We’re trying to find effective ways to quickly program the robots for many more types of therapies,” he says.
“The challenge,” he continues, “is that AI is sometimes unpredictable. That’s because, unlike the standard, deterministic algorithms that were traditionally employed to control robots, this new generation of machine learning is based on probabilistic algorithms.” Therefore, Parietti stresses, “It’s vital for these robots to always have a safety check to ensure predictability so that robots will never do anything outside the bounds of safety.”
AI imitation learning has broad applications, and, Parietti says, “I believe the pharmaceutical industry is one of the best.” The industry’s ability to reduce error rates and increase process throughput for life-saving therapies make it an attractive candidate for this type of automation. “The value of pharmaceutical applications makes cutting-edge automation very justifiable.”
