AlphaFold3, a protein structure prediction model that has made huge waves in the field of biological sciences, has triggered widespread attempts to reproduce it because it only published the paper but did not provide code. DeepMind's strategy allows many scientists to only use AlphaFold3 a limited number of times on specific servers, which undoubtedly whets everyone's appetite. However, the Ligo team composed of three undergraduates from the University of Oxford successfully reproduced AlphaFold3 in just four months and plans to open source it, bringing exciting news to the scientific community. The editor of Downcodes will give you an in-depth understanding of this outstanding achievement of the Ligo team.
At the intersection of biological science and computer science, AlphaFold3 has been like a superstar since its release, attracting countless attention. It is a pity that Google DeepMind only gave us one paper, but did not provide any code or model weights. It is like a delicious cake, but it only allows everyone to look at the appearance without being able to taste it. Faced with this "behind closed door" approach, many teams are scrambling to carry out reproduction work.
In this heated atmosphere, a start-up company called Ligo stood out and became the first team to reproduce AlphaFold3. The three founders of this team were all undergraduates at Oxford University. They achieved this feat in just four months, which is a great gift to the scientific community.
AlphaFold3 is regarded as a milestone in the field of biological sciences, especially in protein structure prediction, and its application potential is huge. However, DeepMind’s strategy is quite disappointing. Their works are only available to scientists on specific servers and have a limited number of calls per day, which seems to be paving the way for future commercial interests. Even so, researchers are excited about this achievement because it has the potential to completely change the rules of the drug discovery game.
Just when many scientists felt frustrated, the Ligo team bravely took the first step. Not only did they reproduce the AlphaFold3 model, they also planned to open source it so that more people could benefit. The Ligo team says their model is currently effective at predicting protein structures, and other capabilities will follow shortly.
The process of reproduction is not simple. The team completely converted the model architecture in the DeepMind paper into PyTorch code. In the process, they discovered some problems in the original paper, such as the formula error of the loss function, which may affect the training effect. In addition, they also optimized the original model, such as introducing a residual layer to improve gradient flow.
What is exciting is that the Ligo team not only followed the ideas of the original model in this work, but also innovated and tried a more efficient implementation method. They even used only 8 A100GPUs during the training process to generate the corresponding model, and the efficiency is eye-catching.
Although DeepMind temporarily closed the results due to commercial reasons, Ligo's successful reproduction gave people hope and triggered more teams to follow up. In addition to Ligo, Columbia University's OpenFold team and independent developer Phil Wang are also actively participating in this open source movement, forming a vivid scientific research ecosystem.
Project address: https://github.com/Ligo-Biosciences/AlphaFold3
The successful reproduction of the Ligo team not only broke DeepMind's closed strategy, but also provided more convenient research tools for scientists around the world. This is not only a victory for AlphaFold3, but also a victory for the open source spirit, which heralds the vigorous development of the field of protein structure prediction in the future. We look forward to more teams joining in to jointly promote the progress of biological sciences!