Advertisement

If you have an ACS member number, please enter it here so we can link this account to your membership. (optional)

ACS values your privacy. By submitting your information, you are gaining access to C&EN and subscribing to our weekly newsletter. We use the information you provide to make your reading experience better, and we will never sell your data to third party members.

ENJOY UNLIMITED ACCES TO C&EN

Structural Biology

AlphaFold 3 to offer structure prediction via web browser

Tool can predict structures of proteins and nucleic acids bound to ions, ligands, and one another

by Laurel Oldach
May 16, 2024 | A version of this story appeared in Volume 102, Issue 15

 

Credit: Google DeepMind
The predicted structure of this enzyme (blue) binding to a calcium ion and several monosaccharides (yellow) matches closely with the experimental structure (gray).

Google DeepMind, the artificial intelligence company behind the popular AlphaFold tool, has published the next generation of the application’s protein structure prediction software. The new tool can predict the structures and interactions of an array of molecules, including nucleic acids, small-molecule ligands, and macromolecule modifications (Nature 2024, DOI: 10.1038/s41586-024-07487-w). “Every time I gave an AlphaFold 2 talk . . . people would always ask me, That’s really good, but I have a DNA-binding protein. Could you just tell me how it binds the DNA?” John Jumper, a director at Google DeepMind, said during a press conference. Unlike past versions, AlphaFold 3, which was codeveloped with drug company Isomorphic Labs and first announced in October, can tackle such interactions.

AlphaFold 3 shares many capabilities with the recently published RoseTTAFold All-Atom model. Though it lacks RoseTTAFold’s ability to design new proteins, AlphaFold 3 has one big advantage: a new web server with a simple interface that requires no coding by the user. It enables anyone with a Google account to input the name of a protein or nucleic acid and develop predictions for the structures of complexes it may form with other molecules.

During the press conference, Julien Bergeron, a biologist at King’s College London who helped test AlphaFold 3, predicted, “Every structural biology and protein biochemistry research group in the world will immediately adopt this.”

But not everyone in the field is so convinced. In an open letter submitted to Nature’s editors, a group of structural and computational biologists object to what they call “several deviations from our community’s standards.” DeepMind and Isomorphic Labs did not publish the code behind the model or make it available to peer reviewers—which the letter’s authors say also represents a failure by Nature to enforce its own policies. The letter had amassed more than 670 signatories by May 14 (DOI: 10.5281/zenodo.11192369). In response, Max Jaderberg, chief AI officer at Isomorphic Labs, and DeepMind vice president of research Pushmeet Kohli posted simultaneously to X that the companies planned to release the code within 6 months for academic use.

While DeepMind has released the code for predecessor model AlphaFold 2, it has not released training data or procedures—a critical part of a machine learning model. In mid-May, researchers led by Mohammed AlQuraishi at Columbia University published OpenFold, an open-source equivalent to AlphaFold 2 that includes that information (Nat. Methods 2024, DOI: 10.1038/s41592-024-02272-z). In principle, a laboratory at a company or university could train its own variant of OpenFold—for example, using proprietary structure data—and generate predictions that resemble what AlphaFold 2 might predict, without using Google’s servers. A lab might also optimize the model to solve a different type of problem. In an email to C&EN, AlQuraishi writes that his team has already started an effort to reproduce AlphaFold 3.

Casual users beware: AlphaFold 3 will predict the most likely structure of any combination of input molecules, whether or not they genuinely interact. According to Jaderberg, the model assigns low confidence scores to complexes that are not thought to occur in the real world. His team is working on ways to predict binding affinity, which would help identify specious interactions—and assist in Isomorphic Labs’ drug design efforts.

UPDATE:

This article was updated on May 16, 2024, to add information about a letter to Nature by structural and computational biologists as well as a paper on an open-source program called OpenFold. The update also added a space to correct the name of AlphaFold 3.

Article:

This article has been sent to the following recipient:

0 /1 FREE ARTICLES LEFT THIS MONTH Remaining
Chemistry matters. Join us to get the news you need.