The initial withholding of the code for AlphaFold3 has been criticized for not aligning with the principles of scientific progress, which rely on the ability to evaluate, use, and build upon existing work. DeepMind has since promised to make the AlphaFold3 code and model weights available for academic use within six months. However, questions remain about whether this version will have the full range of capabilities, particularly the ability to predict the structure of proteins in conjunction with potential drug molecules.
Key takeaways:
- Google DeepMind's decision not to initially release the computer code for its latest protein-structure-prediction AI, AlphaFold3, has led to a global race among researchers to develop their own open-source versions of the model.
- DeepMind later reversed its decision and promised to release the code by the end of the year, but questions remain about whether the released version will have the full range of capabilities, especially the ability to predict the structure of proteins in conjunction with potential drug molecules.
- Several scientists and teams, including Mohammed AlQuraishi's 'OpenFold' team and independent software engineer Phil Wang, are working on creating open-source versions of AlphaFold3, with some aiming to remove limitations and improve the model's performance.
- There are concerns about the high cost of training the models on experimentally determined protein structures and other data sets, with estimates suggesting it could cost upwards of US$1 million in cloud-computing resources to train AlphaFold3 in the same way that DeepMind did.