Google releases blockbuster AI model! Predicting all biomolecules on earth will greatly accelerate research into treating diseases such as cancer

DeepMind, owned by Google, caused a stir in the academic world overnight.

On May 8, DeepMind officially announced a new AI model: AlphaFold 3.

Relevant research papers were published in the authoritative "Nature" magazine and occupied the front page as soon as they were published.

After ChatGPT, there are countless AI models, but the one that is most qualified to claim to change the world may only be AlphaFold 3.

AlphaFold’s super evolution makes the biological world more “high-definition”

We have learned in middle school biology classes that proteins are long-chain molecules formed by connecting amino acids through peptide bonds and folded into complex three-dimensional structures in space.

The three-dimensional structure determines the function of the protein and directly affects drug design and disease treatment.

It can be said that protein structure prediction is one of the most important propositions in biology.

However, predicting the three-dimensional structure of proteins is a difficult task and often requires complex experiments. It has even been described as "a problem that has troubled biologists for 50 years."

In 2016, DeepMind's AlphaGo defeated a professional nine-dan player and rewrote the ancient skill of Go.

DeepMind's AlphaFold wants to decipher the codes of biology and peer into the mysteries of life itself.

In 2018, AlphaFold 1 was released.

In 2020, AlphaFold 2 was launched, which can already accurately predict the shape of proteins on a large scale and down to the atomic level in minutes.

Now, we have AlphaFold 3, an AI with even greater ambitions: going beyond proteins to explore all biological molecules.

Biomolecules are the molecules that make up living organisms, including proteins, DNA, RNA, etc.

DeepMind believes that only by understanding how biomolecules interact in millions of combinations can we begin to truly understand the processes of life.

In a word, AlphaFold 3 covers a wider range than its predecessor and can accurately predict the structure of biological molecules such as proteins, DNA, RNA, ligands, and how they interact.

Let’s first look at some prediction results of AlphaFold 3.

7PNM is the spike protein of a common cold virus.

As shown in the figure, AlphaFold 3 predicts the structure of 7PNM (blue part) when it interacts with antibodies (green part) and monosaccharides (yellow part), which is consistent with the real structure (gray part).

There is meaning behind the predictions. By studying these proteins, scientists can better understand the immune system and coronaviruses including COVID-19, and even come up with better treatment options.

In addition to protein structures, AlphaFold can predict molecular complexes, complex structures composed of multiple molecules.

The enzyme pictured below comes from a soil fungus that is harmful to plants.

AlphaFold's prediction results, including an enzyme protein (blue part), an ion (yellow sphere) and some monosaccharides (yellow part), fit the real structure (grey part).

A deeper understanding of how this enzyme interacts with plant cells could help researchers develop healthier and more resistant crops, bringing practical benefits to agricultural production.

Similarly, AlphaFold 3 predicts a molecular complex composed of a protein (blue part), an RNA strand (purple part), and two ions (yellow part), which also closely matches the real structure (gray part).

This complex is involved in protein synthesis, one of the basic processes of cellular life activities and health, and its research significance is equally profound.

By showing the accuracy of the prediction results and emphasizing related uses, DeepMind wants to tell the world that AlphaFold 3 is a "revolutionary model."

On the one hand, the research scope is broader. Expanding the field of view beyond proteins, especially small molecules such as ligands, can cover more drugs.

On the other hand, accuracy is also improved. For protein interactions with other molecular types, AlphaFold 3 improves accuracy by at least 50% compared to existing prediction methods. Some of the important interactions have even been improved by 100%.

In this way, AlphaFold 3 can benefit more research, accelerate drug design, promote genomics, develop healthier crops, develop biorenewable materials…

Speaking of technical principles, AlphaFold 3 is based on the improvement of AlphaFold 2. The core is the Evoformer deep learning architecture and uses a diffusion network similar to Midjourney.

The process of using AlphaFold 3 is a bit like chatting with a large language model. Input a description of a biomolecule, and AlphaFold 3 generates a three-dimensional structure of these molecules and studies how they interact.

The process of AlphaFold 3 giving prediction results is similar to the AI ​​Vincentian graph diffusion model that gradually denoises, starting from a fuzzy atomic cloud and gradually converging into an accurate molecular structure.

Generative AI that speaks human language enhances the productivity of cubicle workers. AlphaFold means nothing more to scientists than this.

Predicting protein structure in the laboratory can cost a PhD in time and hundreds of thousands of dollars. There are hundreds of millions of predictions, and even millions of people may not be able to complete them in their lifetimes.

But with AlphaFold, scientists can ask bold questions, innovative hypotheses, and then test them in the laboratory to accelerate the research process.

A sentence from DeepMind is enough to summarize the significance of AlphaFold 3 to ordinary people:

AlphaFold 3 brings the world of biology into HD.

Hand AlphaFold to the world, awaiting a new renaissance of scientific discovery

In order to mock OpenAI's closed source, Musk gave it a nickname: CloseAI.

Google, which has contributed to several OpenAI papers, has a more open source spirit in some aspects.

In July 2021, AlphaFold 2 published a paper in Nature and also made the code open source.

To date, AlphaFold 2 has been used to predict hundreds of millions of structures. Millions of researchers around the world use AlphaFold 2 in areas such as malaria vaccines, cancer treatments and enzyme design.

Also in July 2021, DeepMind collaborated with the European Bioinformatics Institute (EMBL-EBI) to release the AlphaFold protein structure database, providing the most complete and accurate picture of the human proteome to date.

DeepMind mentioned in its official blog that this is one of the most important data sets since the mapping of the human genome. Now they will put the power of AlphaFold into the hands of scientific researchers around the world for free.

Within a year, more than 500,000 researchers used the AlphaFold database to view more than 2 million structures, accelerating solutions to real-world problems such as plastic pollution and antibiotic resistance.

Since then, the database has continued to expand.

In July 2022, DeepMind released the predicted structures of almost all proteins known to science, totaling more than 200 million, which in addition to humans also includes predicted structures of plants, bacteria, animals and other organisms.

The AlphaFold database is like a "Google search" for protein structures, and also like a starry protein universe. The three-dimensional structure of protein is the cornerstone of life. It looks exquisite and beautiful, making people marvel at the magic of nature's creation.

However, the AlphaFold 3 released this time has a slightly more conservative attitude than AlphaFold 2, which has attracted some criticism.

AlphaFold 3 is currently not open source and cannot be deployed locally. Researchers can only access most functions for free through DeepMind’s latest research platform AlphaFold Server, and the use is non-commercial.

What hinders scientific progress most is the number of accesses to the service: only 10 predictions can be made per day.

Behind the stingy behavior, DeepMind may have its own business considerations – its subsidiary Isomorphic Labs has cooperated with pharmaceutical companies to apply AlphaFold 3 to drug design.

Demis Hassabis, co-founder and CEO of Google DeepMind, expressed optimism that the first AI-designed drugs may be ready for testing within the next few years.

Of course, AlphaFold still has limitations.

Chinese structural biologist Yan Ning once answered the question in 2022 about whether AlphaFold 2 will replace scientists.

In terms of Nav/Cav, AlphaFold 2 is still stuck at their 2017 level, and when testing the interaction between new small molecules and proteins, none of the AI ​​predictions were correct.

Yan Ning explained that biostructural science is not only about folding, but also about understanding the dynamic changes of proteins, understanding interactions with other biological macromolecules or regulatory small molecules, and understanding the state of cells in situ. Because there is insufficient data, these are all This is an area where AI is still powerless.

Today, AlphaFold 3 takes a big step forward where AlphaFold 2 fell short, allowing us to see the possibility of predicting the interactions of different biomolecules, but it is still focused on static predictions of molecular structures, which can sometimes produce hallucinations. .

In a previous interview, Demis Hassabis criticized the hype surrounding AI.

He believes that AI should be used as the "ultimate tool of science," such as the AlphaFold model for predicting protein structure. Humanity is about to usher in a new renaissance of scientific discovery.

The hero sees the same thing. Nvidia’s Huang Jenxun is also very optimistic about the AI ​​track in medical and biotechnology. He introduced many AI medical services at the 2024 GTC AI Conference and reached agreements with companies such as Johnson & Johnson in the fields of surgery and medical imaging. cooperate.

Sora simulates the physical world, while AlphaFold 3 allows us to understand the biological world and returns to our original expectations for AI – to accelerate scientific discovery, promote human progress, and understand life itself.

Although AGI is still far away, text, pictures, videos, and protein molecules are different from each other, but they echo each other.

AI has indeed become so powerful and increasingly relevant to daily life, and we can expect more innovations to emerge and more mysteries to be solved every day.

It is as sharp as autumn frost and can ward off evil disasters. Work email: [email protected]

# Welcome to follow the official WeChat public account of Aifaner: Aifaner (WeChat ID: ifanr). More exciting content will be provided to you as soon as possible.

Ai Faner | Original link · View comments · Sina Weibo