What’s Hot in 2025: Protein Structure Prediction using AI

Will 2025 be the year of open source AlphaFold3 alternatives?

At the risk of sounding like a broken record, it is hard not to highlight the area of AI-driven protein structure prediction for yet another year!

As anticipated in last year’s hot topic on this subject, 2024 was indeed another year of exciting developments in this field, particularly with the release of Google DeepMind’s AlphaFold3 in May and the awarding of the Nobel Prize for Chemistry to three researchers in the field. AlphaFold3 represents a large step forward in the technology, offering the ability to predict not just isolated protein structures, but molecular complexes comprising, for example, two proteins or a ligand and a protein.

The new release was not, however, without controversy. The previous version of the program, AlphaFold2, had been made freely available to anyone along with a huge database of more than 200 million predicted protein structures. However, when the AlphaFold3 was first published, the program’s code was not made available. It has been speculated that this decision has been made because Google does not want to lose the competitive advantage for its own drug discovery arm, Isomorphic Labs. More recently, DeepMind has released the code for academic (i.e., non-commercial) use only and thus it will be fascinating to see the applications of AlphaFold3 that emerge during 2025 now that at least some researchers outside DeepMind have access to it.

We should not forget that other AI-driven protein structure prediction tools are also available. The leading alternative to AlphaFold3 is probably RoseTTAFold All-Atom from David Baker’s lab at the University of Washington in Seattle, and further developments of this approach can also be anticipated in 2025. While the code for RoseTTAFold All-Atom is licensed under an MIT License, the trained weights and data for the program are only made available for non-commercial use.

In view of this, and the restricted access to AlphaFold3, various fully open-source initiatives are underway aiming to produce programs with similar performance but that are freely available for use by anyone, including commercial entities. Two leading examples are OpenFold and Boltz-1. In 2025, we may expect to see other efforts in this direction and evaluations of their performance compared to AlphaFold3 and RoseTTAFold All-Atom.

The pace of developments in this field shows no sign of letting up and 2025 promises to be another exciting year in the application of AI approaches to protein structure prediction. The new generation of computational tools, allowing the prediction of various types of molecular complex, should bring further valuable insights for science in general and drug discovery in particular.

—David Clark, Senior Research Leader, UK Small Molecule Drug Discovery, Charles River