*For medical professionals only
AlphaFold 3 is simply omniscient.
In 2018, the AlphaFold shocked the world as soon as it appeared, and its evolutionary AlphaFold 2 achieved a spectacular high score of 92.4 in the 2020 CASP.
To date, millions of researchers have used AlphaFold 2 to make progress in the fields of malaria vaccines, cancer treatments, and enzyme design, according to data published by its parent company, DeepMind. AlphaFold has been cited more than 20,000 times and has won several scientific awards, including the Breakthrough Prize in Life Sciences.
However, the function of AlphaFold 2 is limited to the prediction of protein structure, but we all know that proteins are not static, they will change their function due to structural modifications, and they will also interact with DNA, RNA, and other substances. These are things that AlphaFold 2 can't see through.
This issue has been resolved today.
DeepMind's new prediction model, AlphaFold 3, predicts the structure and interactions of all living molecules with unprecedented accuracy, with at least 50% more accuracy and twice the accuracy of individual categories compared to existing technologies. The research paper was published in the journal Nature.
Frank Uhlmann, a biochemist at the Francis Crick Institute who has already tried AlphaFold 3 in depth, praised it and said that AlphaFold 3 will revolutionize structural biology research and greatly lower the research barrier to entry.
Diagram of the title of the dissertation
DeepMind is simply very confident
AlphaFold 3 is still based on AlphaFold 2, but with two important updates. One is to reduce the number of multiple sequence alignment (MSA) processes by replacing the Evoformer model with a simpler Pairformer model; The second is the application of the Diffusion Module, which has recently been widely used in image generation AI, to replace the structural model based on specific amino acid frameworks and side chain twist angles in AlphaFold 2, which greatly increases the chemical structure that AlphaFold can "see".
AlphaFold 3 model structure
AlphaFold 3 is quite "smart" and learns quickly, as you can see from the graph below, 20k steps can boost prediction performance to 97% of the maximum for all metrics.
As for the instability of some indicators with the increase of the number of training steps, the researchers believe that the overfitting is mainly due to the insufficient sample size.
Model training curves
Compared with other prediction models, AlphaFold 3 significantly outperforms classic docking tools such as Vina without inputting any structural data, and compared with the blind docking tool RoseTTAFold All-Atom.
RoseTTAFold All-Atom的研发者华盛顿大学计算生物物理学大牛David Baker也坦然承认，AlphaFold 3的预测性能要更优秀。
AlphaFold 3 can be used to predict protein structure and interactions with other substances
However, AlphaFold 3 is not flawless, for example, it is still a bit confused in terms of chiral structure, and it still has an error rate of 4.4% after adjustment.
Another key limitation is that what AlphaFold 3 can think of is mainly limited to the static structures in the protein structure library (PDB), and the simulation of protein conformational changes is still insufficient.
For now, at least, AlphaFold 3 has demonstrated unprecedented predictive accuracy and is capable of being an industry-changing heavyweight.
Unlike AlphaFold 2, AlphaFold 3 will not be fully open source, scientists can use the AlphaFold 3 server to make ten predictions a day, and according to Uhlmann's feedback, the experience is quite good now, and it is desirable to upload data in ten minutes, which is really good and fast.
However, the predictions do not provide a protein structure that could be druggable, and considering that DeepMind has already established a new subsidiary, Isomorphic Labs, to develop drugs using AlphaFold 3, this part of the functionality should be visible for a fee.
Resources:
[1]https://www.nature.com/articles/s41586-024-07487-w
[2]https://www.nature.com/articles/d41586-024-01383-z
[3]https://www.isomorphiclabs.com/articles/alphafold-3-predicts-the-structure-and-interactions-of-all-of-lifes-molecules
The author of this article丨 Dai Siyu