Michael Polansky is training an AI model on skin that’s still alive

Other rivals are pursuing various flavors of organ-on-a-chip and microphysiological systems for preclinical testing.

Polansky doesn’t seem especially preoccupied with any of them — less, it seems, out of arrogance than because he seems to have his hands full. Besides, there’s plenty of room for everyone in this moment. Unlike AI companies training on scraping the internet, there is no “biology internet” to scrape.

And Outer Biosciences has two other reasons to focus on its own knitting. First, the data it generates doesn’t exist anywhere else, which, conceivably, makes the company’s position more defensible, if a lot slower, to build than “traditional” software-based AI startups. It’s also cheaper to run, with modest compute demands compared with training a large language model. In fact, all of the company’s AI work currently runs on-premise, not in the cloud, because “we don’t want the data in the cloud,” Polansky says.

Whether over time, Outer Biosciences becomes a standalone commercial-ingredients business, licenses its discoveries, or eventually reorganizes around a single breakout compound, Polansky says he hasn’t settled on – and the team doesn’t need to. The more important goal, he says, is a predictive model that’s accurate enough that the company can spot promising directions in skin biology without having to run every experiment physically first, opening up a rate of discovery in dermatology that doesn’t currently exist.

For now, instead, the work of turning a promising compound into a real product — the formulation, the manufacturing scale-up, the supply chain, the safety testing — is still done manually by the same scientists who discover the compounds in the first place. Building out a product-development team, with people who’ve done this kind of work before, is next on the roadmap.

“I think it’s going to be fun,” he says, “to have people know that this is what we’ve been doing.”