On May 8th local time, Google DeepMind released the latest version of its biological prediction tool, AlphaFold—AlphaFold 3.

In a paper published in Nature, the company introduced AlphaFold 3 and called it a revolutionary new model that can predict the structure and interactions of all molecules of life with unprecedented accuracy.

As a single model that calculates the entire molecular complex in a holistic manner, it can not only predict the structure of proteins but also the structure of almost all life molecules, including proteins, DNA, RNA, ligands, etc., for the interactions between proteins and other types of molecules, thus playing a crucial role in drug discovery.

Compared with existing prediction methods, its prediction accuracy has doubled. Specifically, AlphaFold 3 has achieved unprecedented accuracy in predicting drug interactions, including the binding of proteins with ligands and the binding of antibodies with their target proteins.

Without the need to input any structural information, the accuracy of AlphaFold 3 is 50% higher than the best existing traditional methods on the PoseBusters benchmark, making AlphaFold 3 the first artificial intelligence system to surpass the prediction tool of physical-based biomolecular structures.DeepMind also stated that AlphaFold 3 provides a more detailed and dynamic description of molecular interactions than any existing tool.

Advertisement

AlphaFold 3 brings high-resolution into the biological world. It enables scientists to see all the complexities of cellular systems, including structures, interactions, and modifications.

This new window into the molecules of life reveals how they are connected and helps to understand how these connections affect biological functions, such as the action of drugs, the production of hormones, and the DNA repair process.

As it allows the study of a wide range of biological molecules beyond proteins, AlphaFold 3 may pave the way for more transformative science, such as the development of bio-renewable materials and more resilient crops, as well as accelerating drug design and genomics research.

DeepMind CEO Demis Hassabis told media reporters, "Biology is a dynamic system, and biological properties are manifested through the interactions between different molecules in cells. You can think of AlphaFold 3 as our first step towards (modeling) this goal."Google DeepMind Director John Jumper also told the media that this marks a "huge leap for the model, as it really simplifies the whole process of getting all these different atoms to work together."

How does AlphaFold 3 work?

It is understood that within every plant, animal, and human cell, there are billions of "molecular machines." They are composed of proteins, DNA, and other molecules, but they cannot work alone. Only by seeing how they interact in millions of combinations can we begin to truly understand the processes of life.

Taking a series of molecules as input values, AlphaFold 3 can generate their 3D structures, revealing how they bind together.It can model large biological molecules such as proteins, DNA, and RNA, as well as small molecules (also known as ligands). In addition, AlphaFold 3 can also simulate chemical modifications of these molecules. Chemical modifications control the healthy functions of cells, and once disrupted, they can lead to diseases.

Due to the higher complexity of AlphaFold 3, it is necessary to improve the underlying model architecture. To this end, DeepMind has turned to diffusion technology.

In fact, in recent years, AI practitioners have been steadily improving diffusion technology, which has driven the development of image and video generation tools such as OpenAI's DALL-E 2 and Sora.

The working principle of diffusion technology is to train a model to start from a noisy image, and then gradually reduce the noise until an accurate prediction appears. This method allows AlphaFold 3 to handle larger input data sets.

That is to say, the capabilities of AlphaFold 3 come from its new generation architecture and training, which covers all life molecules in the training data. The core of the model is an improved version of the Evoformer module, which is a deep learning architecture that supports the performance of the model.After processing the input content, AlphaFold 3 uses a diffusion network to combine its predictions, similar to the prediction process in artificial intelligence image generation tools. The diffusion process starts with a group of atoms, goes through many steps, and ultimately forms the most accurate molecular structure.

What does AlphaFold 3 do?

The ability to predict antibody-protein binding is crucial for understanding various aspects of the human immune response and the design of new antibodies, which has become an increasingly popular treatment method.

Currently, AlphaFold 3 is bringing the capability to design drugs by predicting molecules commonly used in drugs, such as ligands and antibodies. These molecules bind to proteins, which can change their interactions in human health and disease.DeepMind has stated that its drug development subsidiary, Isomorphic Labs, is combining AlphaFold 3 with a set of complementary in-house artificial intelligence models to attempt to identify crops resistant to adversity and develop new vaccines, as well as collaborating with pharmaceutical companies to try to develop new treatments for diseases.

In detail, Isomorphic Labs is using AlphaFold 3 to accelerate drug design and improve its success rate by helping to understand how to approach new disease targets and develop new methods to pursue previously unattainable targets.

AlphaFold 3's partner, the AlphaFold Server

Previously, for AlphaFold 2, DeepMind released open-source code, allowing researchers to delve into its code to better understand how it works.Hassabis stated that DeepMind currently has no plans to release the complete code for AlphaFold 3. Instead, the company has released a public interface for the model called AlphaFold Server, which imposes restrictions on the molecules that can be tried and can only be used for non-commercial purposes.

DeepMind said that the interface will lower the technical barriers and expand the use of the tool to a group of biologists who have less knowledge of the technology.

In summary, AlphaFold Server is a tool for predicting how proteins interact with other molecules throughout the cell.

This is a free platform that scientists around the world can use for non-commercial research. With just a few clicks, biologists can harness the power of AlphaFold 3 to model structures composed of proteins, DNA, RNA, and a range of ligands, ions, and chemical modifications.

According to the introduction, AlphaFold Server can help people propose novel hypotheses for testing in the laboratory, accelerate workflows, and achieve further innovation.Moreover, the AlphaFold Server also provides users with an easily accessible way to generate predictions, allowing them to use this tool regardless of whether they have sufficient computational resources or expertise in machine learning.

At the same time, DeepMind will also expand its free AlphaFold online educational courses with EMBL-EBI and establish partnerships with more organizations to provide scientists with the tools needed to accelerate adoption and research, including in underfunded areas such as neglected diseases and food security.

What are the risks of AlphaFold 3?

Mohammed AlQuraishi, an assistant professor of systems biology at Columbia University who did not participate in DeepMind's work, believes that the latest version of the model will be more suitable for drug discovery."The AlphaFold 2 system only understands amino acids, so its utility for biopharmaceuticals is very limited," he said, "But now, the latest AlphaFold 3 system can, in principle, predict where drugs bind to proteins."

"It makes the system more versatile, especially for drug discovery (which is still in the early stages of research), it is now much more useful than AlphaFold 2," he said.

Kulexhi indicated that the release of this new tool marks a huge leap, but also comes with some flaws.

Kulexhi said that not open-sourcing the code means that AlphaFold 3's main capability to predict the interaction between proteins and small molecules is essentially not available for public use. He believes: "Currently, AlphaFold 3 mainly serves as a publicity tool."

And like most models, the impact of AlphaFold will depend on the accuracy of its predictions. For some uses, the success rate of AlphaFold 3 is twice that of similar existing advanced models like RoseTTAFold. But for other aspects, such as protein-RNA interactions, Kulexhi said it is still very inaccurate.DeepMind also stated that, depending on the interactions modeled, the accuracy of AlphaFold 3 is around 40% to 80%.

Due to the imprecise predictions, users can only use the results of AlphaFold as a starting point, and then attempt other methods.

Moreover, new technologies bring new risks. As detailed in the AlphaFold 3 paper, the use of diffusion techniques can cause the model to produce "hallucinations," generating structures that seem reasonable but are actually impossible to exist.

To address this, researchers at DeepMind have added more training data in places most prone to hallucinations, in an attempt to reduce this risk, but this does not completely eliminate the problem.

Building on the external consultation conducted by DeepMind for AlphaFold 2, DeepMind has now engaged with over 50 domain experts as well as professional third parties in the fields of biosafety, research, and industry to understand the capabilities and any potential risks of the AlphaFold 3 model. Prior to the release of AlphaFold 3, DeepMind also participated in community and forum discussions.Regardless of the range of its accuracy, it still serves as an excellent first step for researchers in the search for new structures or substances, such as answering questions like which enzymes might be capable of breaking down plastic in water bottles.

In addressing these issues, the use of tools like AlphaFold is much more effective than experimental techniques such as X-ray crystallography.

Previously, experimental protein structure prediction required a time span equivalent to obtaining a doctoral degree, along with a cost of hundreds of thousands of dollars.

The previous version of AlphaFold 2 has helped people better map the human heart, establish models of antibiotic resistance, and identify the eggs of extinct birds.

DeepMind stated in a press release: "We have only just begun to tap into the potential of AlphaFold 3, and we can't wait to see what the future holds."It seems like your message is incomplete. You've mentioned "翻译成英文" which means "translate into English," but there is no text provided to translate. Please provide the text or content you would like to have translated into English, and I'd be happy to assist you.