"Better late than never"! Google DeepMind today unveiled its latest technology to an eager academic communitySource code of artificial intelligence protein prediction software. Although some scientists expressed dissatisfaction with the fact that DeepMind took six months to release the code, the vast majority of researchers welcomed it.
On May 8, DeepMind published an article in Nature magazine announcing the launch of AlphaFold3, a technologyAbility to predict not only protein structure but also interactions with DNA, RNA and other proteins, which has great significance for fields such as drug discovery.
However, they were critical of how the technology was published: despite Nature's editorial guidelines requiring published research to provide computational code, the paperOnly "pseudocode" is provided— a description of the steps to run the program — and a link to an online portal that allows only a limited number of predictions per day. This is in sharp contrast to AlphaFold2, which DeepMind published in Nature in 2021 and provided complete code.Violates accepted standards of openness, reproducibility, and peer review. In an open letter of support signed by hundreds of people, the researchers made this clear.
Faced with resistance, DeepMind pledges to release paper within six months of publicationExpose all code to non-commercial users. Today, it delivers on that promise.
The computational model is available in the code repositoryPublicly released on GitHub under a non-commercial license, and the “weights” (numeric values) used to adjust the AI model are also available to scholars who complete a short application form.
"We appreciate the community's patience," said Pushmeet Kohli, DeepMind's vice president of science. Although he and his team firmly believe that the program they released is correct, Kohli acknowledged that the community wants to work directly with the code. He added that they spent months preparing and testing the model for today's public launch.
Researchers welcome this. Erik Lindahl, a biophysicist at Stockholm University and one of the signatories of the open letter, said: "I am very pleased to see the DeepMind team fulfilling its promise to make the code public, which means that an in-depth review of an important paper can finally begin." Vander Stephanie Vankowitz, a computational structural biologist at the University of Pennsylvania and one of the organizers of the open letter, added: "The models and weights published are important both for evaluation and for further research based on them." But she also noted that "a six-month delay is unacceptable."
AlphaFold3 is the latest AlphaFold version, it is aProtein structure prediction AI based on amino acid sequence, which earlier this year earned two DeepMind researchers, John Zimmer and Demis Hassabis, nominations for the Nobel Prize in Chemistry for this technology. Until today, however, researchers could only use the program through DeepMind's online portal, which was limited to 10 (now 20) requests per day and could only handle a limited collection of molecules.
In a statement in May, Nature editor-in-chief Magdalena Skipper did not specify why she dropped her request to share the full code, but she said editors considered "the potential implications of biosafety and the ethical challenges that arise." Meanwhile, a news report in Nature quoted Koch as saying that the teamAccess to AlphaFold3 has been restricted to avoid impacting the drug development programs of DeepMind’s commercial subsidiary IsomorphicLabs. Colley told Science magazine that the DeepMind team prioritized developing the portal rather than releasing code "to make sure we have the simplest interface for the most people." Zimmer said researchers have done some "incredible work" through the portal, and that work has not changed after today's news. He suspects most scientists will continue to work this way because it's more practical for teams with limited computing power.
DeepMind researchers also disputed claims by some critics that the Nature paper wasCan be repeated, because multiple teams have developed their own versions of AlphaFold3 based on pseudocode. AI-focused companies such as Baidu, LigoBiosciences and ChaiDiscovery have published results from these efforts. Daniel Buchan, a bioinformatics researcher at University College London, points out that these alternative "implementations" may still be useful now that the code for AlphaFold3 is now public. "It's good and important to be able to reproduce the method," he said. Vankowitz added that comparing and contrasting these models may lead to improvements in the future. Of particular importance, researchers say, are implementations of unrestricted user licenses such as those being developed by the nonprofit OpenFold Alliance. Otherwise, "if I help a colleague design a completely new ligand that could be a potential anti-cancer drug, and one day they want to work with a pharmaceutical company to commercialize it, things get very complicated," says Roland Dunbrack, a computational structural biologist at Fox Chase Cancer Center. He was originally asked to review a DeepMind manuscript for Nature but never received the code to review. There are already some research teams planning to work with AlphaFold3’s code.
The authors of a paper published today in the journal Nature Computational Science say they hope to integrate AlphaFold3 into their own software. The program, called MassiveFold, helps users leverage parallel computing to reduce the time required to run large numbers of predictions on AlphaFold2—from months to hours.
Guillaume Brysbaert, a bioinformatics researcher at the French National Center for Scientific Research and developer of MassiveFold, said that by integrating DeepMind's new code, "users can get the best prediction results from AlphaFold2 or AlphaFold3." Zhu Mo said that the DeepMind team is looking forward to the results being released publicly today. "In AlphaFold2, we saw a lot of creativity," he said. "I'm really excited to see how the community finds AlphaFold3 works - how can it be applied to new problems?"
References:https://www.science.org/content/article/google-deepmind-releases-code-behind-its-most-advanced-protein-prediction-program