OpenAI Partners with Retro Biosciences to Engineer Stem Cell Proteins with AI

OpenAI has developed its first biology-focused AI model in collaboration with Retro Biosciences, marking the company's entry into scientific discovery and longevity research. The new model, GPT-4b micro, shows promising results in redesigning proteins that could improve stem cell production efficiency by up to 50 times, according to preliminary data.

Key Points:

• GPT-4b micro specifically targets protein engineering for cellular reprogramming
• Early tests indicate significantly improved efficiency in converting adult cells to stem cells
• The collaboration connects to Sam Altman's broader investments in longevity science

At the heart of this project lies a fascinating scientific challenge: transforming adult skin cells into stem cells, a process that has traditionally been highly inefficient. The current success rate hovers below 1%, taking several weeks to complete. OpenAI's new model tackles this problem by redesigning what scientists call Yamanaka factors—specific proteins that trigger this cellular transformation.

Retro Biosciences was founded in 2021 by Joe Betts-LaCroix with a $180 million investment from Sam Altman. The company employs over 60 scientists and runs five discovery programs, including autophagy and in vivo reprogramming. In 2024, Retro partnered with Multiply Labs in an $85 million agreement to automate its cell therapy manufacturing processes, a critical step in scaling treatments for age-related diseases.

OpenAI's approach with GPT-4b micro, differs significantly from Google DeepMind's AlphaFold, focusing instead on protein interactions rather than structural predictions. By training on protein sequences and interaction data, the model proposed modifications to the Yamanaka factors, yielding proteins that preliminary tests show are 50 times more effective at creating stem cells. According to Betts-LaCroix, the model delivered results faster and better than human-led efforts.

Google DeepMind Makes AlphaFold 3 Open Source for Academia

Now, academic researchers can download all necessary code for AlphaFold 3 inference and request access to the model parameters

MaginativeChris McKay

The technical achievement here is notable. John Hallman, an OpenAI researcher, notes that "Just across the board, the proteins seem better than what the scientists were able to produce by themselves." The AI model suggests bold modifications, sometimes altering up to a third of a protein's amino acids – a scale of change that would be practically impossible to test through traditional laboratory methods.

The implications for medicine could be significant. Harvard University aging researcher Vadim Gladyshev, who consults with Retro, explains the potential impact: "For us, it would be extremely useful. [Skin cells] are easy to reprogram, but other cells are not." This advancement could pave the way for more efficient organ development and cell replacement therapies.

However, the scientific community will need to wait for peer review and publication of the results before drawing definitive conclusions. OpenAI and Retro have committed to releasing their research findings, though the timeline remains unclear. For now, the model itself won't be immediately available for wider use, remaining a demonstration of capability rather than a commercial product.