Source: Quantum

When it comes to the contribution of artificial intelligence to science, you may think of AlphaFold, which is Google DeepMind’s protein folding program, whose creators won a Nobel Prize last year.

Now, OpenAI says it’s getting into science, too—by building a model for protein engineering.

The company says it has developed a language model that can conceive of proteins that can transform ordinary cells into stem cells—and it has easily beaten humans.

This research is OpenAI’s first model focused on biological data, and the first time the company has publicly claimed that its model can provide unexpected scientific results. As such, it's a step toward determining whether AI can make real discoveries, which some consider to be a major test toward "artificial general intelligence."

Last week, OpenAI CEO Sam Altman said he was “confident” that his company knew how to build general artificial intelligence, adding that “superintelligent tools could dramatically accelerate scientific discovery and Innovation is far beyond what we humans can do on our own.”

The protein engineering project began a year ago, when San Francisco-based longevity research company Retro Biosciences approached OpenAI about a partnership.

This collaboration is no accident. According to reports, OpenAI CEO Sam Altman personally provided $180 million (approximately 1.318 billion yuan) in funding to Retro.

(Source: OpenAI)

Retro's goal is to extend normal human lifespan by 10 years. To do this, the company studied so-called Yamanaka factors, or induced pluripotent stem cells. This is a group of proteins that, when added to human skin cells, turn them into young-looking stem cells, the type of cells that can give rise to any other tissue in the body.

Researchers at Retro and well-funded companies such as Altos Labs see the phenomenon as a possible starting point for rejuvenating animals, creating human organs or providing replacement cells.

But this kind of cell "reprogramming" is not very efficient. It takes weeks, and less than 1% of the cells treated in a lab dish complete the regeneration journey.

OpenAI’s new model, called GPT-4b micro, can be trained to suggest ways to redesign protein factors to enhance their function. According to OpenAI, researchers used the model’sThe proposed two Yamanaka factors increase the efficiency by more than 50 times - at least according to some preliminary measurements.

"Overall, these proteins appear to be better than those produced by scientists themselves," said OpenAI researcher John Holman.

Holman was the lead developer of the model, along with Aaron Jack of OpenAI and Rico Meni of Retro.

Outside scientists cannot tell whether the results are true until they are published, which the two companies say they plan to do. The model has yet to become more widely available, either - it's still a custom demo rather than a formal product launch.

“This project is about showing that we are serious about our contribution to science,” said Jack. "But whether these capabilities will be available as separate models or integrated into our main inference models - that remains to be determined."

The model works differently from Google's AlphaFold, which can Predict protein shape. OpenAI says that because Yamanaka factors are unusually soft and unstructured proteins, they required a different approach, for which its large language models are well suited.

The model is trained on protein sequence samples from many species, as well as information on which proteins tend to interact with each other. While the data is large, it's only a fraction of the training data for OpenAI's flagship chatbot, so GPT-4b is an example of a "small language model" using a centralized dataset.

Once Retro scientists had the model, they tried to guide it toward possible redesigns of Yamanaka's protein. The prompting strategy used is similar to a "few samples" approach, where the user poses a question to the chatbot by providing a series of examples with answers, and then provides an example for the bot to respond to.

Although genetic engineers have the means to direct molecular evolution in the laboratory, they can usually only test a limited number of possibilities. Moreover, even proteins of average length can be altered in nearly infinite ways (because they are made of hundreds of amino acids, each of which has 20 possible variations).

However, OpenAI's models often suggest changing one-third of the amino acids in a protein.

“We immediately put this model into the lab and got real-world results,” said Retro CEO Joe Bates-Lacroix. He also said that the concept of the model is very good and has improved over the original Yamanaka Shinya factor in quite a few cases.

We need better ways to make stem cells, said Vadim Gradyshev, an expert on aging at Harvard University and a consultant to Retro. "For us, this would be very useful. [Skin cells] are easily reprogrammed, but theirThat's not the case with other cells," he said. "And reprogramming it in a new species - it's usually very different, and you get nothing. "

Exactly how GPT-4b arrived at its guess is unclear - as is often the case with artificial intelligence models. "It's like AlphaGo beating the best human player in Go, but It took a long time to figure out why," Bates-Lacroix said. "We're still trying to figure out what it does, and we think the way we're applying it is just scratching the surface. ”

OpenAI CEO Sam Altman (Source: TechCrunch, CC BY 2.0, via Wikimedia Commons)< /p>

OpenAI said no money was involved in the collaboration, but since the research may benefit Retro (its largest investor is Altman), the news may raise concerns about OpenAI. CEO's side hustle raises more questions

Last year, the Wall Street Journal said Altman's extensive investments in private technology startups amounted to an "opaque investment empire" and was "generating." "There are more and more potential conflicts" because some of these companies also have business dealings with OpenAI.

For Retro, as long as it has business relationships with Ultraman, OpenAI Being associated with the AGI race could boost its profile and its ability to hire employees and raise money. Bates-Lacroix did not respond to questions about whether the early-stage company is currently in the funding stage.

OpenAI said Altman was not directly involved in the work and the company never made decisions based on Ultraman's other investments.