Institute researchers use AI to find manufacturing pathways to treat EB, and beyond

By Christopher Vaughan

May 23, 2025

Sometimes called the “Butterfly Disease” because the slightest touch can cause patients’ skin to peel away and blister, Epidermolysis Bullosa (EB) has long been a singular focus for institute member Anthony Oro, MD. Over the last 15 years, he and other Stanford researchers have made major gains in genetically engineering people’s own cells to create healthy skin that can transplanted and treat the disease. And yet there remained some major roadblocks to treatment that have frustrated Oro and others trying to find a complete cure for EB.

Now, researchers at Stanford and collaborating institutions have created an AI model that has helped them overcome some of those roadblocks. In a recent issue of Development Cell, Oro and his colleagues describe their work using artificial intelligence to guide their discovery of the key factors that determine how healthy cells can be manufactured from stem cells to treat EB. The AI application they created has broad implications for creating cell-based treatments not only for EB, but also for many other diseases.

“The questions was ‘what is the best way to come up with a definitive, cell-based therapy that can cure this disease?’” said Oro.

Because EB patients have missing or altered versions of some of the proteins that hold skin cells together, EB patients suffer burn-like injuries and blistering whenever their skin is rubbed. Because of this, they are subject to excruciating pain from their earliest years, and the continual cycle of injury and healing skin makes them much more likely to get cancer. Historically, most EB patients did not live past the age of 30.

In the last 15 years, Oro and his colleagues, often with the collaboration of ISCBRM member Marius Wernig, MD, PhD, have created induced pluripotent stem cells out of patients’ skin cells, then genetically engineered those cells to replace the defective proteins, and grow normal skin that can be transplanted back onto the patient.

This approach has been highly successful, but still did not address on major problem that afflicts EB patients: the “skin” lining the esophagus—the tube leading from mouth to stomach—is also affected in EB patients. Normal swallowing can lead to the same blistering and scarring seen on exterior skin. Furthermore, the cells in the esophagus develop differently than skin, so researchers needed to understand the molecular “recipe” that guides stem cells to become esophageal cells.

Finding the right instructions to guide cells development “is kind of like America’s Test Kitchens,” said Oro. “You could have all the same ingredients—flour, eggs, milk, baking soda, etc.—but how you put them together determines whether you create pancakes, souffleé, or pastry.”

The problems is that there are two dozen different genetic inducers involved in this developmental pathway in the human embryo, and countless ways to combine all these “ingredients.” But there is only one right recipe to turn stem cells into esophageal cells.

“The question was, do you just use trial and error to find the right answer, of is there a better way to do it?” Oro said.

As with many things these days, artificial intelligence seemed like it might offer a shortcut. Oro and his Stanford colleagues collaborated with researchers at UC Santa Cruz and the University of Washington to feed published genomics, proteomics and other “-omics” data into Manatee, an artificial intelligence program developed by Joshua Stuart at UCSC.

“Manatee was super helpful in predicting the inducer pathways that would be useful in getting us to the cellular destination we wanted,” Oro said. “A lot of those predictions went against published knowledge,” he added.

Their approach was to look at every cell signal and see how—and when—those signals affect the target genes. “A lot of the time it turns out that, in nature, the key component acts earlier or later than you expected,” Oro said. Manatee’s recommendations were tested and confirmed first in cell culture, and then ground-truthed in developing embryos.

As is often the case, AI results are good but not perfect, Oro said. “One of the signals turned out to be quite a bit stronger in the in-vitro assays than predicted by Manatee because one of the data sets we were using was incomplete. Machine learning still follows the rule ‘Garbage In, Garbage Out.’”

But for the most part, Manatee offers a powerful platform not only creating esophageal cells, but for understanding and producing cells for almost any tissue. “This can tell you the order and timing of cell signals, help you define a manufacturing pathway, and help you automate it into a closed system for actual manufacturing,” he said.

The researchers are already using their artificial intelligence platform to collaborate with other Stanford researchers to come up with a viable manufacturing pathway for corneal cells, Oro said. He also points out that the method could also be useful to investigate esophageal and other cancers, as well as answer questions about how developmental pathways evolved.

“The work was funded by NIH, Department of Defense, and Stanford Innovative Medicine Accelerator.”