Last week’s announcement of AlphaFold 3, a new artificial intelligence–powered program from Google DeepMind, sparked excitement in the scientific community for its promise to vastly improve predictions of the structure and interactions of proteins and aid discovery of new drugs. But DeepMind and Nature, which published the research, have come under fire for offering only restricted access to the program and failing to release the computational code underlying it.
In an open letter that had received more than 650 signatures as of 14 May, researchers write that they’re “disappointed” by the lack of resources accompanying the publication, and accuse the journal of undermining its own rules on code availability. “While companies have the right to capitalize on their innovations, using the imprimatur of academic publications without the possibility of reproducing the results, far less building on them, subverts the enterprise,” the researchers write in the letter, posted on 11 May. “When journals fail to enforce their written policies about making code available … they demonstrate how these policies are applied inequitably and how editorial decisions do not align with the needs of the scientific community.”
Experts in the field of protein prediction have described AlphaFold 3 as “transformative” and “very impressive.” They say the work marks a significant advance on earlier programs, such as AlphaFold 2 and RoseTTAFold All-Atom, and will help generate more accurate predictions of proteins’ interactions with biomolecules such as DNA and RNA.
However, whereas its predecessors were released with downloadable code, AlphaFold 3 is currently only accessible through a web server. On its launch, each user could run just 10 requests per day, although that number has since risen to 20. Users also face limitations on the molecules they can analyze. It isn’t possible to predict interactions between proteins and novel drugs, for example, reportedly to avoid competition with drug discovery efforts of DeepMind spinoff Isomorphic Labs.
AlphaFold 3’s code wasn’t available during the review process for the Nature paper, either. Roland Dunbrack, a computational structural biologist at the Fox Chase Cancer Center and letter co-author, says he received the manuscript without any way to test the program. After contacting the journal, he got access to an early version of the web server, but repeated requests for code in the leadup to publication went unanswered, he says. “I don’t understand why [Nature’s editors] sent it out for review under those conditions.”
The paper doesn’t provide a justification, simply noting, “Code is not provided”—an omission that appears to “flagrantly violate” Nature’s policies, says James Fraser, a structural biologist at the University of California San Francisco (UCSF) and one of the letter’s organizers. Nature’s submission guidelines state that custom code supporting a paper’s main claims must be made available to referees upon request, and its editorial policies specify that “authors are required to make [code] promptly available to readers without undue qualifications.”
The apparent contradiction has prompted ire from researchers. “In my opinion, large parts of this work [do] not fulfill the requirements of scientific studies,” Erik Lindahl, a biophysicist at Stockholm University and signatory on the letter, tells ScienceInsider. “[I]t is effectively an ad for commercial services.”
Nature has defended its handling of the paper. “While seeking to enhance transparency at every opportunity, Nature accepts that there may be circumstances under which research data or code are not openly available,” Editor-in-Chief Magdalena Skipper says in a statement. Editors “reflect on many different factors, including the potential implications for biosecurity and the ethical challenges this presents. In such cases we work with the authors to provide alternatives that will support reproducibility.” The paper includes “pseudocode”—a description of the steps run by the program—she adds.
DeepMind has restricted access to its products on biosecurity grounds before, although study co-author John Jumper, a senior researcher at the company, reportedly told Fortune magazine last week that biosecurity experts advised the company that risk from AlphaFold 3 was marginal, and outweighed by potential benefits. A news story in Nature, meanwhile, quoted Pushmeet Kohli, vice president of Research at DeepMind, as suggesting restrictions were implemented so as not to compromise the ability of Isomorphic Labs to pursue commercial drug discovery plans.
Regardless, since the open letter was posted, DeepMind researchers have indicated that more information on AlphaFold 3 is on the way. A DeepMind media representative pointed Science to a 13 May social media post in which Kohli, another co-author, announced the increase in the web server’s daily request limit to 20. The team is also “working on releasing the AF3 model” for academic use within 6 months, Kohli wrote, a move welcomed by researchers who spoke to Science.
Stephanie Wankowicz, a computational structural biologist at UCSF and another of the letter’s organizers, says she hopes the episode will encourage the computational biology community to set concrete standards for research communication, particularly given the increasing influence of for-profit companies in this field.
It’s also an opportunity for journals to reflect on their role in upholding scientific standards, Fraser says. If they apply standards selectively, “it’s like they’re the traffic cops that are letting some people speed and pulling other people over for rolling through a stop sign. And that’s not fair.”
