Better late than never: Google DeepMind has today released the computer code underlying its latest AI protein prediction software to an eager research community. Many scientists are pleased by the move, though some remain upset it took 6 months for the company to reach this point.
When DeepMind announced AlphaFold3 in a Nature paper on May 8, researchers lauded the technology’s ability to predict not only proteins’ structures, but also how they interact with DNA, RNA, and other proteins, a boon for drug discovery and other fields. But they criticized the announcement itself: Despite Nature’s editorial guidelines stating computational code must be made available alongside published studies, the new paper contained only ‘pseudocode’—a list of steps a program runs—and a link to an online portal that allowed scientists to do a limited number of predictions daily.
The approach contrasted with DeepMind’s publication of AlphaFold2, complete with code, in Nature in 2021, and ran counter to accepted standards of openness, reproducibility, and peer-review, researchers argued in an open letter that garnered hundreds of signatures. Following the backlash, DeepMind committed to releasing the full code for noncommercial use within six months of the paper’s publication.
Now, it has made good on that promise. The computational model itself was made public today on the code repository GitHub with a noncommercial license, while the ‘weights’—numbers that help tune how an AI model works—are available to academics who complete a short application form.
“We want to thank the community for the patience,” says Pushmeet Kohli, vice president of science at DeepMind. Although he and his colleagues stand by how they released the program, Kohli says, they recognized the community’s desire to work with the code directly. It has taken months to prep and test the model for today’s public release, he adds.
Researchers applaud the move. “I’m delighted that [the DeepMind team] is keeping their promise to release the code, because this means the actual in-depth review of an important paper can finally start,” says Erik Lindahl, a biophysicist at Stockholm University and signatory on the open letter. “The model and weights being released are of huge importance” for efforts to evaluate and build on the work, adds Stephanie Wankowicz, a computational structural biologist at Vanderbilt University and an organizer of the letter. Still, she says, “the delay of six months is unacceptable.”
AlphaFold3 is the latest incarnation of AlphaFold, the AI that revolutionized the prediction of protein structures based solely on their amino acid sequence and won two DeepMind researchers, John Jumper and Demis Hassabis, a share of the Nobel Prize in Chemistry earlier this year. Until today, however, researchers could only use the program only via DeepMind’s online portal, which permitted just 10 (and now 20) requests per day with a restricted set of molecules.
In a statement in May, Nature’s editor-in-chief Magdalena Skipper did not specify why the journal had waived its requirement to share the full code but said editors considered factors like “potential implications for biosecurity and the ethical challenges this presents.” A news story in Nature, meanwhile, quoted Kohli as suggesting the team had restricted access to AlphaFold3 to avoid compromising the ability of Isomorphic Labs, a commercial spinoff from DeepMind, to pursue drug discovery plans.
Kohli now tells Science that DeepMind’s team prioritized developing the portal rather than releasing code “to make sure that we provided the easiest interface to the most … people.” Jumper says researchers have done some “incredible work” via the portal, which remains unchanged by today’s news, and says he suspects most scientists will continue to work this way, as it’s more practical for groups with limited computing power.
The DeepMind researchers also contend that, contrary to some critics’ claims, the Nature paper was reproducible, as demonstrated by the fact that multiple groups have since made their own versions of AlphaFold3 based on the pseudocode. AI-focused companies such as Baidu, Ligo Biosciences, and Chai Discovery have already released the results of such efforts.
These alternative “implementations” will likely still be useful, even with AlphaFold3’s code now released, notes Daniel Buchan, a bioinformatics researcher at University College London. For one thing, “it’s good and important that methods can be replicated,” he says. Comparing and contrasting the models will likely lead to improvements in the future, Wankowicz adds.
Particularly important are implementations that are free from restrictive user licenses, like the one being developed by the nonprofit OpenFold consortium, researchers say. Otherwise, “if I help a colleague … with a novel ligand that might be a good lead cancer drug, and at some point they want to work with a pharmaceutical company to commercialize it, things can get very complicated,” says Roland Dunbrack, a computational structural biologist at the Fox Chase Cancer Center who was inititially asked to review DeepMind's manuscript for Nature but never received the code to do so.
A number of research teams already have plans to work with AlphaFold3’s code. The team behind a paper out today in Nature Computational Science, which describes a program called MassiveFold, says it wants to integrate AlphaFold3 into its software. MassiveFold helps users take advantage of parallel computing to reduce the time it takes to run lots of predictions in AlphaFold2—potentially from months to hours. By integrating DeepMind’s new code, “the user will be able to get the best predictions [with this approach] from either AlphaFold2 or AlphaFold3,” says MassiveFold developer Guillaume Brysbaert, a research engineer in bioinformatics at CNRS in France.
Jumper says the DeepMind team is looking forward to what today’s public release brings. “In AlphaFold2, we saw so much creativity,” he says. “I’m really excited to see what the … community discovers about how AlphaFold3 works—how can it be applied to new problems?”
