Google DeepMind's new AI predicts how the molecules of life interact

• The interactions AlphaFold 3 predicts are key for many crucial processes in cells. The interplay between and changes to proteins, DNA, RNA, ions and other small molecules dictate their function — and disfunction from disease.
• For example, when a protein on the surface of a cell binds to another protein on a virus, the molecules change shape, setting off a process that fuses the virus and the cell so the virus can invade. The details of those interactions could help develop precise vaccines or antiviral drugs.

"Biology clearly is a dynamic system, so we need to understand interactions between different structures, proteins and other things to really understand what they do," Google DeepMind CEO Demis Hassabis told Axios.

• "AlphaFold 3 is a big step in that direction."

Driving the news: AlphaFold 3 is the next iteration of AlphaFold models that took on and solved one of biology's toughest problems: predicting the structure of proteins from their amino acid sequence.

• The new AI model handles a larger number of chemicals using a different approach.
• It leverages a generative AI technique called diffusion, which is similar to those that drive image and video generators, like DALL-E.

How it works: AlphaFold 3 takes a cloud of atoms and then refines it, step by step, until the model converges on the most accurate molecular structure it can predict.

• The number of inputs it can handle is "dramatically expanded" from earlier AlphaFold models, John Jumper, director at Google DeepMind, told Axios.
• The reported accuracy ranges from 40% to 80%, depending on the interaction AlphaFold 3 is trying to model, and the program provides a measure of how confident it is in its result. These test results were published in the journal Nature.
• AlphaFold 3 performs better than existing tools for almost all categories of interactions they looked at.

"These problems we're doing, we wouldn't consider them solved," Jumper said. "We're still at an accuracy we'd like to improve."

• The tool at this point helps to more quickly — and cheaply — home in on possible structures that can be jumping off points for more detailed studies.
• DeepMind also launched a server for researchers to access AlphaFold 3, but it has some restrictions about what can be modeled, particularly for drug candidate molecules.

• Some scientists say these restrictions could limit its impact.

Between the lines: The diffusion technique comes with a risk.

• In what's known as disordered regions, or flexible parts of a protein that can take on many shapes, the model will produce a plausible-looking structure but one that couldn't exist — a biological form of the hallucinations that plague other AI models.
• AlphaFold3 does report its low confidence in these results, Jumper said, adding that the team has reduced that risk by adding more data to the regions where these hallucinations typically occur.

The big picture: "My dream is to build a model of a virtual cell," Hassabis told Axios, but "the challenges become almost sort of exponentially more difficult."

• "The system is going to have to learn some fundamentals about how biophysics works. We think we can do that," he said.
• "The question is getting the right amount and the right quality of data."

Experimental tools to image what's happening inside cells without killing them are being developed.

• If those tools arrive, "that will be huge for AI to then learn from that," Hassabis said. Or researchers may have to build physics simulations that can provide synthetic data.

While today's AI can already help science and medicine, as AI gets applied to solve more types of problems, the AI models themselves will improve too, Hassabis said.