Google DeepMind’s latest medical breakthrough borrows a trick from AI image generators

Much of the recent AI hype train has centered around the mesmerizing digital content generated by simple prompts, alongside concerns about its ability to destroy the workforce and make malicious propaganda that much more persuasive. (Fun!) However, some of the most promising — and potentially far less sinister — AI jobs lie in medicine. A new update to Google’s AlphaFold software could lead to new research and breakthroughs in treating diseases.

AlphaFold software, from Google DeepMind and (also owned by Alphabet) Isomorphic Labs, has already shown it can predict how proteins fold with shocking accuracy. It has cataloged a staggering 200 million known proteins, and Google says millions of researchers have used previous versions to make discoveries in areas such as malaria vaccines, cancer treatments and enzyme design.

Knowing the shape and structure of a protein determines how it interacts with the human body, allowing scientists to create new drugs or improve existing ones. But the new version, AlphaFold 3, can model other important molecules, including DNA. It can also map drug-disease interactions, which could open exciting new doors for researchers. And Google says it does so with 50 percent better accuracy than existing models.

“AlphaFold 3 Takes Us Beyond Proteins to a Wide Range of Biomolecules,” Google DeepMind Research Team wrote in a blog post. “This leap could unlock more transformative science, from developing biorenewable materials and more sustainable crops to accelerating drug design and genomic research.”

“How Proteins Respond to DNA Damage; how do they find it, repair it?” Google DeepMind project manager John Jumper said With cable. “We can begin to answer these questions.”

Before AI, scientists could only study protein structures through electron microscopes and develop methods such as X-ray crystallography. Machine learning streamlines much of this process, using patterns recognized by its training (often invisible to humans and our standard tools) to predict the shapes of proteins based on their amino acids.

Google says part of AlphaFold 3’s progress comes from applying diffusion models to its molecular predictions. Diffusion models are central parts of AI image generators such as Midjourney, Google’s Gemini, and OpenAI’s DALL-E 3. Incorporating these algorithms into AlphaFold “sharpens the molecular structures that the software generates,” such as With cable explains. In other words, it takes a formation that seems vague or unclear and makes highly educated guesses based on patterns from its training data to clarify it.

“This is a big advance for us,” said Google DeepMind CEO Demis Hassabis With cable. “That’s exactly what you need for drug discovery: you need to see how a small molecule will bind to the drug, how strongly, and also what else it might bind to.”

AlphaFold 3 uses a color-coded scale to indicate the level of confidence in its predictions, allowing researchers to exercise due care on results that are less likely to be accurate. Blue means high confidence; red means it’s less secure.

Google makes AlphaFold 3 free for use by researchers for non-commercial research. However, unlike previous versions, the company is not offering the project as open source. One prominent researcher who makes such software, University of Washington professor David Baker, expressed his frustration with the With cable that Google has chosen this route. However, he was also amazed by the software’s capabilities. “The structure prediction performance of AlphaFold 3 is very impressive,” he said.

As for what’s next, Google says that “Isomorphic Labs is already collaborating with pharmaceutical companies to apply it to real-world drug design challenges and ultimately develop new, life-changing treatments for patients. “