On May 8th local time, the internationally renowned academic journal Nature published the latest achievement on Google AI model Alpha Fold3, which was jointly released by Google Deep Mind and its UK subsidiary Isomorphic Labs.
It is reported that the model predicts the structure and interaction between proteins and other biomolecules with unprecedented accuracy. The research team stated that Alpha Fold 3 will help change people's understanding of the biological world and drug discovery, thereby ushering in a new era of artificial intelligence in cell biology.
(Source/Nature)
Draw approximately 6 million protein structures
As early as the 2018 CASP13 (Protein Structure Prediction and Evaluation) competition, Alpha Fold 1 attracted widespread attention due to its outstanding performance in protein structure prediction. At that time, Alpha Fold 1's performance far exceeded that of other participating teams.
In 2020, Deep Mind released Alpha Fold 2. In the CASP14 competition that year, Alpha Fold 2 demonstrated unprecedented accuracy and made breakthrough progress in predicting protein structure. According to the evaluation results at the time, the average error of Alpha Fold 2 was only about one atomic diameter (0.1 nanometers), and its prediction accuracy was comparable to traditional experimental methods, which caused a huge response in the scientific community at that time. Deep Mind subsequently collaborated with the scientific community to analyze various biologically important protein structures through the use of alpha folding 2, accelerating the progress of biomedical research.
It is understood that in 2021, Deep Mind collaborated with the European Institute of Bioinformatics to publicly release Alpha Fold DB, a database containing tens of thousands of protein structure predictions. The disclosure of these data allows researchers worldwide to access protein structure information for free. In 2022, Alpha Fold 2 also underwent a comprehensive upgrade.
According to Deep Mind, Alpha Fold 2 has been used by 1.8 million researchers in just 3 years, mapping approximately 6 million different protein structures. But these images are images of individual static proteins and do not include chemical communication occurring within cells. "Biology is a dynamic system," explained Demis Hassabis, CEO of Deep Mind. "You must understand how the interactions between different molecules in cells produce biological characteristics."
AI assists in the treatment of diseases and drug development
Today, 6 years later, Google Deep Mind has launched the latest Alpha Fold 3, aimed at helping scientists design drugs and target diseases more effectively. Max Judelberg, Chief Artificial Intelligence Officer at Isomorphic Labs, stated that the functionality of Alpha Fold 3 provides new opportunities for researchers to quickly identify potential new drug molecules. It is understood that Isomorphic Labs has previously established partnerships with pharmaceutical companies Eli Lilly and Novartis.
Max Judelberg said, "This allows our scientists and drug designers to create and test hypotheses at the atomic level, and then use alpha folding 3 to generate highly accurate structural predictions within seconds. In contrast, traditional experimental methods may take months or even years."
According to the team, the prediction accuracy of Alpha Fold 3 has significantly improved compared to many existing specialized tools, including those based on its predecessor. Compared with existing prediction methods, this version of alpha folding has improved at least 50% in predicting the structure and interactions between proteins and other molecules. In some key research areas, its prediction accuracy has even doubled, achieving a 100% improvement. Alpha Fold 3 can accurately predict the structure and interactions between proteins, DNA, RNA, and ligands. This progress provides hope for the treatment of diseases such as cancer and autoimmune diseases.
"We have seen incredible progress, and we believe these advancements will unlock many new sciences," said John Jemper, head of the Alpha Folding team at Deep Mind. He pointed out that the technology also has the potential to improve plant biology knowledge and food security. "We have started to see biologists and early testers using it to understand how cells work and to think about what kind of errors it may lead to," he said.
After reviewing the overview of his paper, the reporter found that the latest Alpha Fold 3 has several significant updates.
The latest version of alpha folding will rely less on protein information related to the target sequence. Meanwhile, Alpha Fold 3 also uses a machine learning network called a diffusion model, which can be used by image generative artificial intelligence. John Jemper said, "This is a significant change."
Meanwhile, researchers indicate that Alpha Fold 3 is far superior to existing software tools in predicting proteins and their structures. For example, scientists interested in finding new drugs typically use a software called Docking to simulate the degree of binding between chemicals and proteins, and Alpha Fold 3 has been shown to be superior to this software, as well as another AI based tool called RoseTTAFoldAll Atom4.
John Jamper and his colleagues also validated their claims with data in earlier reports. The report shows that in over 400 test cases, Alpha Fold 3 was able to accurately simulate the interactions between known proteins and drug like small molecules, with 76% of cases being able to simulate correctly, while RoseTTAFoldAll Atom's proportion was approximately 40%. For the interaction between proteins and antibodies, the accuracy of Alpha Fold 3 is 62%, while the accuracy of AlphaFoldMultimer (a software package previously used by the company to simulate the interaction between proteins and other biomolecules) is 30%.
It is understood that in order to encourage the widespread adoption of new models, Deep Mind researchers have also released an alpha folding server, which is a free online platform. For this purpose, Deep Mind researchers have released a database containing approximately 200 million protein structures.
The advantages and disadvantages of the new model
A study published this week by the Boston Consulting Group shows that drugs discovered by artificial intelligence have a higher success rate in early trials compared to drugs discovered by other methods. The study points out that these data are an early analysis of the technology's effectiveness in drug discovery, and suggests that artificial intelligence can double the productivity of pharmaceutical research and development.
According to the introduction and interpretation of this model by Science, another internationally renowned academic journal, Julian Bergeron, a biologist at King's College London, was given the opportunity to test Alpha Fold 3. He said that the software is revolutionary in accelerating research speed. Researchers do not need to spend years studying proteins in the laboratory, but can obtain results within minutes. "We can start testing hypotheses in computers," said Julian Bergeron. "I am very certain that every structural biology and protein biochemistry research group in the world will immediately adopt this system."
Brian Schoichett, a pharmaceutical chemist at the University of California, San Francisco, has been using alpha folding to search for drugs. He said that due to limitations in modeling protein drug interactions, he does not feel that alpha folding 3 has the same impact as alpha folding 2 before.
As the research team reported, Alpha Fold 3 still has limitations in many aspects and requires further iteration and improvement. Including stereochemical limitations, accuracy issues in prediction, structural and conformational limitations, lack of dynamic information, and specific target prediction limitations.
