Google DeepMind and Isomorphic Labs unveiled the newest version of AlphaFold, taking its flagship AI model far beyond predicting the structure of a single protein.
AlphaFold 3 can now predict a range of complex biological structures that include virtually any biomolecules, including proteins, DNA and RNA strands, and small molecules. A new Nature research paper, published Wednesday, shows AlphaFold 3 producing more accurate predictions than both traditional and AI methods for structures showing how proteins interact with ligands, nucleic acids and other proteins.
The update includes some major changes from AlphaFold 2, which was published in 2021 in a Nature paper that has been cited over 20,000 times.
While Isomorphic summarized some results with AlphaFold 3 in a blog post six months ago, the new paper provides more details, including an algorithmic approach that now incorporates a diffusion model to generate these complex structures.
“AlphaFold 3 expands on the transformative advance that was AlphaFold 2,” Julien Bergeron, a structural biologist at King’s College London, said in a statement. “In short, it does it bigger, faster, better.”
AlphaFold 3 allows scientists to put a range of biomolecules — amino acid sequences for proteins, DNA or RNA sequences, and SMILES strings for small molecules — into the model and generate a prediction about how those components interact.
“You can put in three different proteins with DNA with a small molecule with an ion and see that joint structure,” Isomorphic’s chief AI officer Max Jaderberg said in an advance interview with Endpoints News. “That’s really powerful for understanding more and more about the context of these molecules in the cell.”
DeepMind also debuted AlphaFold Server on Wednesday. It’s a way to provide free access to most of the model’s capabilities for non-commercial academic users. Isomorphic, an Alphabet-backed biotech founded in 2021, has been using the latest model in its own drug research, both for its internal pipeline of programs as well as partnerships with Novartis and Eli Lilly.
Isomorphic is one of a handful of richly funded, wildly ambitious startups bringing generative AI into drug R&D, alongside companies like Flagship’s Generate:Biomedicines and the recent billion-dollar debut of Xaira Therapeutics, which was co-founded by the AI protein researcher David Baker from the University of Washington.
Jaderberg said the newest AlphaFold will be particularly helpful in two areas: powering rational, structure-based drug design and learning more biology.
Max JaderbergHe referenced the 1990s, when a wave of startups like Vertex first started trying to design better drugs by knowing the structure of the target. That approach requires crystallized structures of how proteins, small molecules and other molecules fold and interact. Many times, those crystal structures can take years to get in the lab, if they can be done at all.
“With AlphaFold 3, we don’t need to do that experimental structure resolving,” Jaderberg said. “We can just use the model. Instead of months or years, we get a structure out in a matter of seconds.”
That can make it easier and faster to rationally design molecules, according to Jaderberg, who said the model typically takes 10 seconds to 60 seconds to generate a prediction, with the time depending on the complexity.
The predictions aren’t just faster but more accurate. AlphaFold 3 was 76% successful in predicting protein and small molecule interactions compared to 52% or lower with the next best models, which includes some from Baker’s lab, according to the Nature paper.
The AlphaFold update can help researchers learn more about proteins involved in diseases, he said, giving clues about how these proteins work in the body.
“We believe the result of understanding a lot more about that means we’ll be targeting things in a better way and ultimately [that will] lead to a reduced failure rate in the clinic,” Jaderberg said.
The update is a significant but small step in the vision set out by Demis Hassabis, the AI leader overseeing DeepMind and Isomorphic. He has laid out the long-term ambition of increasing the complexity of its AI predictions from single proteins to complexes to pathways to cells.
But there’s still a long way to go to achieve that vision of upending biology. Jaderberg declined to share any additional details on Isomorphic’s own pipeline beyond a focus on oncology and immunology. Structure predictions are “super-useful” but also “just part of the story,” Jaderberg said. His team is also working on using AI to better understand the function of biomolecules and predicting binding affinity — or how strongly molecules bind together.