Taking a cue from the automotive industry, the biopharma sector has been increasingly focused on autonomous capabilities as a means towards improved accuracy, reproducibility, and efficiency in research and development (R&D).
Alongside breakthroughs in artificial intelligence (AI) have come significant strides in developing robotic arms and other advanced machinery. AI agents that are able to coordinate and even plan experiments promise to free up human researchers for more abstract, conceptual scientific work while expanding the scale of R&D possible.
Go deeper with GlobalData
Discover B2B Marketing That Performs
Combine business intelligence and editorial excellence to reach engaged professionals across 36 leading media platforms.
Ginkgo Bioworks is a leading developer of autonomous labs working in the biopharma space. The company’s CEO, Jason Kelly, spoke with Pharmaceutical Technology to explain how these systems work and the potential impact they may have on the industry. He likens Ginkgo’s work in biopharma to what self-driving car companies like Waymo have done for the transport industry, and says “AI scientists” are fast becoming a reality.
This interview has been edited for length and clarity.
Frankie Fattorini (FF): What is the principle behind an autonomous biopharma lab?
Jason Kelly (JK): The analogy I like to give is from transportation. If you think of low variability and high automation, that’s a train. If you go down in automation and up in variability, you’re at the car; you’re in the loop to control it, but where you want to take that thing is totally variable. That’s what the transportation industry has looked like for the last 100 years.
I don’t know if you’ve been to San Francisco recently and seen Waymos [self-driving cars] driving around. It’s like being on a subway, but it will take you right to your house. We call that autonomous.
We have labs with a low amount of variability and a high amount of automation, called “workcells”, and companies like Thermo Fisher, HighRes Biosolutions, and Biosero sell them. [In a workcell] there is an arm in the middle of 10 pieces of equipment in a big glass box. It’s great, and fully automated, but it has to run the same experiment it did yesterday. You can’t just make it do something new.
A lab bench would involve low automation and have a lot of variability. You’re mostly going to see lab benches and scientists in white lab coats [at academic or pharmaceutical labs] standing in front of them doing things by hand.
What we’re trying to build is the Waymo [version]. We want to replace the couple billion dollars that are going to the lab benches, and we want the scientists to order whatever experiment they want; [like] they’re sitting in the back of a Waymo telling it where to go.
FF: What does a Ginkgo autonomous lab for biopharma research look like?
JK: In Boston, we have an 18,000 square foot lab with 70 robots that have, collectively, about 90 different lab devices all in one setup. Ginkgo scientists regularly submit new protocols into that system, so on a busy day, we might have 30 unique protocols and 80 or 100 total protocols, because people will ask for replicates.
For example, 40 or 50 scientists have access to 50 or 100 different kinds of lab devices: centrifuges, heat blocks, liquid chromatography-mass spectrometry (LC-MS) devices. They’re wandering around that lab, moving samples to those devices, and then programming the devices.
Every device is connected to every other device. That’s because in a manual lab, you can easily move samples between equipment. In that lab, you’re not limited to just what you can reach with the arm around you, as you would be in a traditional work cell.
We have connected all our arms with a MagneMotion track, which is like an industrial transport automation system, and it delivers 96-, 384-, or 1,536-well plates. It’ll deliver those to any device on the system. Then each device has a six-axis robotic arm. It picks up the plate and it puts it, with very high reliability, onto the device.
Then, importantly, every device is connected to our software. You can program the device using our software. Increasingly, you can talk to the software in English and translate the protocol into the code that moves it through that 70-robot system here at Ginkgo.
FF: How close are autonomous labs to full, independent autonomy?
JK: There is a key distinction to make as people think about this technology becoming a big part of the industry. There are two different technologies being developed: one is an autonomous lab, and the second is an AI scientist.
We’re very focused on the autonomous lab part. We want to make it a beautiful experience, whether it’s the AI model or a human saying, “this is the protocol I want done, and I want it done now.” Companies like Edison Scientific, Potato.ai, and others are trying to make reasoning models into better scientists.
We did an experiment with a project we did with OpenAI [where] GPT-5 operated as an AI scientist running Ginkgo’s autonomous lab. We worked on a project for cell-free protein synthesis. Academic groups and industry groups have worked on dropping the cost of the protein you get from a sample in cell-free synthesis because it’s expensive.
A paper from Stanford and Prof. Michael Jewett’s lab benchmarked all the best mixes with different reagents and so forth to say which is the cheapest for making protein per dollar. We tried to beat that.
So we did rounds of 100 384-well plate experiments. GPT-5 designed the experiment, we’d run it and give the data back, and GPT-5 would design the next experiment. After four rounds, it beat the Stanford paper. After six rounds, it had beaten it by 40%.
