Every reporter Cai Ding Every editor Lan Suying
Predicting the three-dimensional structure of proteins from amino acid sequences has always been the most challenging problem in structural bioinformatics. But a few years ago, AlphaFold, an artificial intelligence sequencing model based on deep learning, created by DeepMind, a Google company, solved this problem.
On the evening of May 8 (Wednesday), Beijing time, Nature published a paper jointly signed by DeepMind's AlphaFold team and London drug research and development company Isomorphic Labs, introducing AlphaFold 3, which is the third generation version of AlphaFold. The new protein structure prediction system can predict the "Protein Data Bank" with unprecedented accuracy Almost all molecular types in the complex structure.
According to the paper, AlphaFold 3 is a revolutionary system. For the interaction between proteins and other molecular types, the accuracy of AlphaFold 3 in benchmark test is 50% higher than that of the best traditional method available, and no structural information needs to be input, which makes AlphaFold 3 the first artificial intelligence system that surpasses the method based on physical tools in the prediction of biological molecular structure.
In fact, since the first generation of AlphaFold came out, structural biologists have never stopped discussing its capability boundary. Previous studies have confirmed that AlphaFold cannot predict the impact of new mutations on proteins. However, this cannot hide AlphaFold's unprecedented ability to predict biological structure. Thomas C., a researcher of Los Alamos National Laboratory and a senior scientist of the New Mexico State Union. Terwilliger said in a paper published in Nature last November that although AlphaFold's predictions were not all accurate, it provided a credible hypothesis that could be used as a hint mechanism. All these capabilities are probably just the beginning of the growing application of AI methods in structural biology.
AlphaFold 3 is 50% more accurate than the best available method
According to the paper, based on the improvement of AlphaFold 2's ability, AlphaFold 3 can now predict the interactions between proteins, nucleic acids, small molecules, ions, complexes that modify protein residues, and antibody antigen interactions, and its prediction accuracy significantly exceeds that of current prediction tools, including AlphaFold Multimedia. The DeepMind team said that this means that AlphaFold 3 has brought humans to a wider range of biomolecular fields beyond protein. This leap may open up more revolutionary science, from developing bio renewable materials and more flexible materials to accelerating drug design and genomics research.
AlphaFold 2 was launched in 2020, and its 3D structure can be predicted according to the sequence of amino acids (basic components of protein) of the protein. John Jumper, the first author of the paper and senior research scientist of DeepMind, and colleagues said that so far, millions of researchers around the world have made progress in malaria vaccine, cancer treatment, enzyme design and other fields using AlphaFold 2. At the same time, AlphaFold 2 is said to have been used to predict hundreds of millions of structures, which will take hundreds of millions of years of research according to the current speed of global structural biology experiments.
It is reported that the core of AlphaFold 3 is an improved version of the in-depth learning module Evoformer, which is the infrastructure of AlphaFold 2. According to the paper, as long as the molecular input list is given, AlphaFold 3 will use a fusion network similar to the artificial intelligence image generator to combine the prediction results, which can not only generate their joint three-dimensional structure, but also reveal how the molecules are combined together.
According to the paper, AlphaFold 3 has achieved unprecedented accuracy in predicting the interaction of similar drugs (including the binding of proteins with ligands and the binding of antibodies with target proteins). In the benchmark test, the accuracy rate of AlphaFold 3 is 50% higher than the best traditional method available, and no structural information needs to be input, which makes AlphaFold 3 the first artificial intelligence system that surpasses the method based on physical tools in the prediction of biological molecular structure.
DeepMind team believes that AlphaFold 3 has the ability to bring the biological world to an unprecedented height. This system enables scientists to see all the complexities of cell systems, including structure, interaction and modification, and reveals how they are interconnected, and helps to understand how these connections affect biological functions - such as the role of drugs, hormone production and the health protection process of DNA repair.
AlphaFold 3's performance shows that developing the correct deep learning framework can significantly reduce the amount of data required to obtain biological related performance in these tasks, and expand the impact of the data already collected. DeepMind expects that structural modeling will continue to improve, not only because of the progress of deep learning, but also because of the continuous progress of experimental structure determination methods, such as the huge improvement of cryogenic electron microscopy and tomography, which will provide rich new training data to further improve the generalization ability of such models. The parallel development of experimental and computational methods is expected to push people into an era of better understanding of molecular structure and disease treatment.
The paper also introduces that DeepMind's new AlphaFold server is the most accurate tool to predict how proteins interact with other molecules in cells globally.
AlphaFold server is a free platform. Biologists can use the powerful function of AlphaFold 3 to simulate the structure composed of protein, DNA, RNA and a series of ligands, ions and chemical modifiers. "The impact of AlphaFold servers will be realized by how they enable scientists to accelerate the discovery of open issues in biology and new research fields. We have just begun to tap the potential of AlphaFold 3, and can't wait to see what will happen in the future." DeepMind team wrote in the paper.
Moreover, the AlphaFold server can help scientists put forward new assumptions and test them in the laboratory, so as to speed up work progress and achieve further innovation. If traditional methods are used, protein structure prediction will not only require doctoral level knowledge, but also cost hundreds of thousands of dollars to complete. Google cloud platform platform also provides researchers with a convenient way to generate predictions, even if researchers do not have computing resources or do not master machine learning expertise. It is reported that Isomorphic Labs is combining AlphaFold 3 with a set of complementary internal AI models to design drugs for internal projects and pharmaceutical partners in order to speed up and improve the success rate of drug design.
AlphaFold is not completely accurate, but can be used as a prompt mechanism
AlphaFold has shown strong performance in continuous iterations, and the scientific community has always discussed the boundaries of this prediction system.
Thomas C., a researcher of Los Alamos National Laboratory and a senior scientist of the New Mexico State Union. Terwilliger said in a paper published in Nature last November that AlphaFold's prediction is a valuable hypothesis. Although it can accelerate the discovery of drugs, it cannot replace the determination of experimental structure. Terwilliger's team's research shows that although AlphaFold's prediction is usually surprisingly accurate, they found that many parts of AlphaFold's prediction are incompatible with the experimental data of the corresponding crystal structure.
In addition, some researchers have tried to apply AlphaFold to all kinds of mutations that will destroy the natural structure of proteins, including a mutation related to early breast cancer. However, it was found that AlphaFold could not predict the impact of new mutations on proteins, because there is no evolutionarily related sequence that can be used for research.
However, it should be noted that the Terwilliger team still gave a very positive evaluation of AlphaFold's ability in the above papers. The team wrote that although there are limitations, AlphaFold prediction has been changing the way of generating and testing protein structure assumptions. Although AlphaFold predictions are not completely accurate, they provide credible hypotheses that can serve as a prompt mechanism and allow the design of experiments with specific expected results.
"All these capabilities are probably just the beginning of the increasingly extensive application of AI methods in structural biology. AI methods will certainly expand from proteins to nucleic acids, ligands, covalent modifications, environmental conditions, as well as the interactions between all these entities and a variety of structural states. With the addition of more factors and the expansion of the database of sequence and structural information, the accuracy of these predictions and the uncertainties associated with them are likely to continue to improve. The predicted results will become more and more useful structural hypotheses, laying a solid foundation for the experimental and theoretical analysis of biological systems. " The Terwilliger team added.
Editor in charge: Liu Debin
