OpenAI gets into longevity science

GPT puts on a lab coat: OpenAI CEO Sam Altman announced a new model called GPT-4b micro, designed for engineering proteins.

• The model designs proteins that can turn regular cells into stem cells—and Altman says it does so faster than human scientists can.
• This project started in 2023, with Altman’s funding of longevity startup Retro Biosciences, which aims to extend the human lifespan by 10 years. Given their relationship with its CEO, Retro approached OpenAI about working together on AI-enabled protein engineering.

Altman also said he is “confident” that OpenAI is on the way to building an “artificial general intelligence” (AGI), a superintelligent form of AI which can handle any intellectual task a human can. When it comes to scientific research—especially in biomedicine—this would be a game-changer.

Does this sound familiar?: When you think of AI-enabled protein research, you probably think of Google’s Nobel-winning AlphaFold. 

• But while AlphaFold predicts the shape of proteins, GPT-4b micro re-engineers the proteins themselves.
• OpenAI trained the model on samples of cross-species protein sequences and information about how proteins interact. The resulting model is a small language model, given this boatload of scientific data is still nothing compared to how much data the company’s popular GPT large language models (LLMs) work with.

How this connects to longevity: Retro studies a set of proteins called Yamanaka factors. 

• When these proteins are added to human skin cells, they cause them to morph into stem cells, which may then be the starting point for rejuvenating interventions for animals and even building new human organs.
• However, Yamanaka factors aren’t very efficient in this transformation process. Over several weeks, less than 1% make it all the way.
• The goal of applying super-charged protein engineering to this problem is increasing Yamanaka factors’ function, which preliminary demonstrations from OpenAI suggest seems to be possible.

Medical innovation stays hot for big tech players: With the launch of DeepSeek, the battle amongst competitive LLMs has never been hotter. As this market gets more and more competitive, AI applications in biomedicine may be turning into a new way for AI players to stay competitive. 

• The OpenAI protein engineering model won’t be available for outside scientists to test and replicate the results of for some time. The company plans an official launch sometime in the future. For now, the announcement serves as a marker of OpenAI’s focus as the AI race continues to heat up.
• We’re also watching: Meta’s AI for neurotech project, which it seems we’ll be waiting a while for.