The BRICS TimesOpenFold3 Unlocks AlphaFold3’s Secrets for Scientists
A new AI model is revealing the inner workings of the top artificial intelligence tool. This tool predicts how proteins interact with small molecules, such as various drugs.
This model, OpenFold3, was launched on October 28. It is a recreation of Google DeepMind’s AlphaFold3. A large group of researchers led by Mohammed AlQuraishi at Columbia University carefully studied AlphaFold3’s code. They then built a copy of the AI platform. This platform predicts structures of proteins combined with other molecules. These molecules include nucleic acids and drug chemicals. AlphaFold3 is only available for use by individuals, non-commercial groups, or journalists. However, companies—and anyone else—can use the open-source OpenFold3 model. It is available for commercial purposes, including developing new drugs.
Why Protein-Molecule Pairing Matters
Predicting protein-molecule pairings is vital for designing drugs. This is “because this is how biology works. Biology is not proteins in isolation. It’s biomolecules interacting with each other,” says Woody Sherman. He is the founder and chief innovation officer at Boston-based Psivant Therapeutics. Sherman also leads the OpenFold executive committee.
Proteins are among the hardest-working molecules in the human body. How these ‘workhorses’ function largely depends on their specific shape. AlphaFold2 previously solved the problem of predicting the shapes proteins adopt. The team behind that AI model won the 2024 Nobel Prize in Chemistry for this achievement. AlphaFold3 later included interactions with other proteins and molecules.
The Call for Open-Source Transparency
DeepMind did not initially open the AlphaFold3 code for other researchers. This changed only after hundreds of scientists signed a petition demanding transparency. Stephanie Wankowicz, a computational structural biologist at Vanderbilt University, coauthored the petition. “It’s hard to evaluate a computational product without seeing the raw information,” she explains. Other researchers need the code to check the accuracy and reliability of predictions. This also helps them learn what data is necessary to improve the model, Wankowicz notes.
Re-creating AlphaFold2 gave OpenFold creators crucial insight into how the AI functions, she mentions. AlphaFold2 was marketed as an AI that learned protein folding from amino acid building blocks. However, it actually memorizes protein structures it has already seen. It then uses these memories to predict how similar proteins might look, Wankowicz states. Looking into AlphaFold3 may offer similar insights into protein-drug pairs.
Other teams have tried to rebuild AlphaFold3. They “have gotten close, but not super precise,” Wankowicz says.
The Importance of Detail in AI Recreation
Reproducing subtle tricks and tweaks is hard, says Sherman. These are often in the AlphaFold3 creators’ minds. They do not appear in the code or supplemental information. Some are technical settings for specific parts of the calculation. “Nobody’s specifying that,” he states. “But details matter, especially when you’re dealing with the large models and with lots of data.” The OpenFold3 team tried hard to copy AlphaFold3. Still, some small differences remain, he confirms.
Biology is also a crucial factor, Sherman points out. Proteins exist in cells surrounded by water and ions. They constantly vibrate and move. Static images created by AI models or lab-made protein snapshots do not capture this. The OpenFold3 team plans to add water and movement into their model. This better reflects how proteins truly exist in nature, Sherman says.
Industry Adoption and Future Impact
Pharmaceutical companies embraced OpenFold3 even before its official launch. Five companies formed the Federated OpenFold3 Initiative. They aim to train the AI model on proprietary data. This will build a more powerful prediction tool while protecting company secrets. This partnership was announced on October 1 by Apheris, a Berlin-based company managing the platform.
Only about 2 percent of the protein structures in public databases used for training AlphaFold3 and OpenFold3 are paired with druglike molecules, says Robin Röhm. He is the cofounder and chief executive of Apheris. Drug companies, in contrast, have thousands of such structures privately.
Each company in the federation will train a version of OpenFold3. They will use about 4,000 to 8,000 protein-drug pairs from their own library, Röhm explains. Apheris then combines these locally trained AIs into one centralized version. This version knows how about 20,000 proteins and drugs interact. It does not contain the original proprietary data. The global version then returns to each company for more training rounds, continuing the process.
Despite the bigger datasets, do not expect immediate, dramatic changes in drug discovery, Sherman advises. OpenFold3 “is a starting point,” he says. “It’s going to be the next stage, and the next stage and the next stage that are where we’re really going to start seeing that meaningful impact on drug discovery.”
