On May 8th, local time, Google's DeepMind released the latest version of its biological prediction tool, AlphaFold 3. This significant achievement has caused a major stir in the scientific community. According to experts in protein prediction, AlphaFold 3 is described as "revolutionary" and "impressive." They state that this work can be used to sequence protein, DNA, RNA, small molecules, and nearly all biological molecular structures and interactions. The model for AlphaFold3 was jointly developed by DeepMind and the AI pharmaceutical company Isomorphic Labs, and the research has been published in the journal Nature. However, the lack of transparency due to not disclosing the underlying code and only providing restricted access has sparked widespread criticism in the scientific community. As of May 14th, over 650 researchers have signed an open letter expressing disappointment in the lack of code availability in the paper and criticizing the journal for violating its guidelines on code accessibility.
Image: Multiple scholars co-signed letter to provide open-source model for academia within 6 months
AlphaFold 3 is currently only accessible through a web server, with a daily request limit of 20 times (initially limited to 10 requests upon release). Additionally, users face restrictions in analyzing molecules. For instance, it cannot predict interactions between proteins and new drugs, possibly to avoid competition with Isomorphic Labs. The code for AlphaFold 3 was not made public during the peer review process at Nature magazine. Roland Dunbrack, a computational structural biologist at Fox Chase Cancer Center, one of the organizers of this letter, mentioned that he was unable to test the program upon receiving the manuscript. Dunbrack stated, "After contacting the journal, I was granted access to an early version of the web server, but despite multiple requests before publication, I was not provided with the code. I don't understand why the editors sent it for review under these circumstances." James Fraser, a structural biologist at the University of California, San Francisco, and another organizer of the letter, commented, "The paper provided no rationale, just stated 'code not provided,' which seems to violate Nature's policy." Nature's editorial policy mandates that authors promptly provide code to readers without inappropriate restrictions. This apparent contradiction has sparked anger among researchers. Erik Lindahl, a biophysicist at Stockholm University and a signatory of the letter, remarked, "In my view, much of this work does not meet the requirements of scientific research; in fact, it is essentially a commercial advertisement." In response to criticism, Magdalena Skipper, the editor-in-chief of Nature, defended the journal's handling of the paper in a statement. She noted that while Nature strives for transparency at every opportunity, there may be instances where research data or code cannot be disclosed. Editors consider various factors, including potential impacts on biosecurity and the ethical challenges they pose. In such cases, the editorial team collaborates with authors to provide alternative reproducible support. Since the publication of the open letter, DeepMind researchers have indicated that more information about AlphaFold 3 will be released soon. Pushmeet Kohli, a DeepMind researcher and one of the authors of the AlphaFold 3 paper, tweeted in response to criticisms and dissatisfaction from the academic community, announcing plans to provide an open-source model for academia within 6 months. "AlphaFold 3 testing falls short"
Since the release of AlphaFold 2 in 2020, scientists have widely utilized this tool to predict various protein structures, discover drugs, and map numerous known proteins.
The advancement of AlphaFold 3 lies in its ability to not only handle proteins but also simultaneously input substances such as nucleic acids, small molecules, and metal ions, predicting how they will interact with proteins. Because proteins do not act in isolation but require interaction with other substances, this process is of utmost concern to scientists. On the evening of May 9th, Yan Ning, an academician of the Chinese Academy of Sciences and director of the Shenzhen Bay Laboratory, posted a lengthy article on her personal Weibo account, stating, "My attitude towards AI can be summed up in two words, 'awe-inspiring,' its development speed is beyond imagination. Every time, I point out the issues that the current AlphaFold version cannot solve, such as glycoproteins that I have recently been obsessed with, those 'dark matter' entities, in situ structures, and so on. But ultimately, it's just a matter of time." She emphasized, "I think this server version strikes a balance between speed and accuracy, but the accuracy is not the best. I have three rather peculiar proteins on hand now. Previously, my self-built AF2 multimer could find one or two correct conformations at very low ranking positions, but this time the server version's testing fell short." She concluded, "AI will undoubtedly become more powerful, and how to embrace new technologies and ask more interesting questions is what researchers in the field are currently focusing on." Reference links: https://www.isomorphiclabs.com/articles/a-glimpse-of-the-next-generation-of-alphafold
(Original Title: Disappointed! Over 650 Scholars Criticize AlphaFold 3 for Lack of Transparency)
