Recently, the world has been swept by AI news again, when Google DeepMind, which developed AlphaGo and cried Ke Jie, announced their latest generation of AlphaFold 3 model on Nature.
AlphaFold, which sounds a bit like the name of a foldable phone model, is his new AI that specializes in predicting protein structures.
It can predict almost all molecular structures in living organisms. This means that biomedical research has since opened up a real business. God's perspective, any biomolecular mechanism of action will be opened from the black box and turned into perspective mode.
Many media and netizens began to cheer, The 21st century, this is really the century of living things...
To understand how awesome the newly released AlphaFold 3 is, we must first know how much shock DeepMind and its AlphaFold have given to the molecular biosphere...
We have all learned in the nine years of compulsory education that the most substance in living organisms is protein, and in order to understand the underlying principle of biomolecules, we must know what each protein looks like.
Let's put it this way, before AlphaFold, there were two main ways to predict protein structure.
One is to X-ray the protein crystal, that is, to take a film and then analyze the film, and then to understand what it looks like. The second is nuclear magnetic resonance (NMR) spectroscopy, which captures the general shape outline and then speculates on its structure.
These traditional methods are not only slow, have a small scope of application, require constant trial and error, but also cost money, and each time a film is made, it costs tens of thousands of dollars to reach a Xiaomi Su7.
That's why protein research is expensive and requires a lot of experience... Only those experienced masters, protein immortals, can guess the exact shape of the protein faster and take fewer films.
So people wondered, can this kind of work that requires experience to be solved by AI? That's where DeepMind comes in, and in order to overcome the problems of traditional filmmaking, the first generation of AlphaFold had a showdown when it chose the technical route:
Don't make a film!
Since proteins are made up of amino acids, the original AlphaFold used the method of using the known protein structures published from various places, and summarized the distance and link angle of each pair of amino acids in these proteins into a graph, and the AI used a neural network to digest them, and then let the AI make its own predictions.
The first-generation AlphaFold in 2018 was released and won the 13th Protein Structure Prediction Competition (CASP).
AI, it's amazing, isn't it?
However, the original AlphaFold had a problem in that it relied more on the features of local data for training, and it was less able to extract relationships between elements that were more distant. It's like a writer who can only write short essays, but can't learn to write long novels.
The problem is that many protein molecules are dependent on long distances, which makes the original AlphaFold a bit stretched.
Fortunately, AlphaFold 2.0, released in 2020, uses the Transformer model that later became popular on ChatGPT.
The attention mechanism of the Transformer model perfectly solves the problem of long-distance amino acids, how much progress has it made?
Version 1.0 of the 2018 Protein Structure Prediction Contest scored less than 60 points for accuracy, but version 2.0 of the 2020 Contest scored a staggering 92.4 points, covering 98% of all known proteins, and more importantly, it is completely open source.
It can be said that version 2.0 has basically solved the prediction problem of single-chain proteins, and in 2021, the AlphaFold-Multimer based on the 2.0 revision was released, which also supports multi-chains, and has also made breakthroughs in accuracy, with a prediction accuracy of more than 70% for the interaction between proteins.
So now many companies are also using them, and even helped the research and development of some new crown vaccines abroad.
However, in DeepMind's view, the victory in protein structure prediction is far from the full potential of AI, because the complex molecular structure in living organisms includes not only proteins, but also nucleic acids, small molecule ligands, etc.
It's like if you spent ten years learning the key unlocking technology, but as soon as you got out of the school, you found that everyone used fingerprint lock combination locks, and there were too few people who used traditional keys!
So this time with AlphaFold 3, they have updated a more awesome all-round model, which can not only predict various small molecules such as proteins, DNA, RNA, etc., but also reveal how they interact with each other.
So how does this work? The answer is, they used Diffusion.
Yes, it's the famous diffusion model, which you must have heard of when AI painting was on fire. It works by continuously encoding the original image, and then allowing the AI to learn to predict the generation process of these mosaics, and then in turn to generate the mosaic from the mosaic to the image.
However, just as the AI draws bad fingers and the Sora chair video wears the mold, AlphaFold 3 with the blessing of Diffusion will also make prediction errors, especially in some structures that look similar and are difficult to distinguish, such as chiral molecules that you learned in high school organic chemistry.
So in these error-prone areas, DeepMind uses an operation called cross-distillation, which is to let the second-generation version of the Transform model predict first, and then add the prediction data to the training of AlphaFold 3, which is equivalent to letting the second generation play the role of a teacher and lead the third generation to do it, so as to reduce prediction errors.
How good is the generation? Let's look directly at the official picture
AlphaFold 3's prediction of 7BBV-enzyme (present in a soil fungus) in which the enzyme protein (blue), ions (yellow spheres), and monosaccharides (yellow) almost coincide with the true structure (gray).
AlphaFold 3 predicts the structure of the spike protein (blue) of cold viruses when it interacts with antibodies (turquoise) and monosaccharides (yellow), and accurately matches the real structure (gray)
AlphaFold 3's prediction of protein complexes, in which the protein (blue) binds to DNA (pink), the prediction model closely matches the experimentally determined real molecular structure (gray).
In addition to the fact that the generation quality is quite amazing, the accuracy is also far ahead of the atomic level. It is better than other products in the simulation of protein and nucleic acid ligands, and the simulation of antigen and antibody is also excellent.
And operating the AlphaFold3 is even easier. With ChatGPT, you still have to figure out how to ask a good question and write a good prompt, whereas in AlphaFold 3, you just need to enter a list of molecules and it can predict how they will fit together.
Imagine a phenomenon that used to take a lot of time, effort and money to observe, but now you can simply enter a parameter on a website and click on it, and in a few minutes you can produce a biomacromolecule model with extremely high clarity and accuracy.
Even the biochemical processes within the cellular system, how the DNA works, and how the drug and hormone reactions work can all be understood in a very short time.
These far leading data, and everyone's enthusiasm, seem to say: this release is no longer a leap forward, but a revolutionary breakthrough, and the entire traditional biomedical research method seems to be about to be changed.
However, Shichao thinks that optimism is good, but in addition to optimism, science must be pertinent and rigorous.
At a time when various media and netizens are "exploding", "subverting" and "changing the world", many bigwigs in the circle have also expressed some comments on AlphaFold 3.
For example, Professor Yan Ning's team found that version 3.0 overturned in a glycoprotein prediction, and its performance was even worse than that of the previous version.
There are also many scientists who complain that 3.0 is not open source compared to 2.0, and there is a limit to the number of times it can be used.
There are even those who question DeepMind's boss, Hassabis, who himself founded an "AI-focused drug company" that claims to "use AI to redefine drug discovery", but from 2021 to now, they have not launched any drugs today.
Of course, this is a bit embarrassing, after all, in the process of drug development, protein structure problems are only a small part of it, which does not have a decisive impact on the progress of drug development...
In short, Shichao feels that the three generations of AlphaFold's products are indeed gratifying, but it still has many problems to break through in the long road of life science. But at the end of the day, progress is always a good thing, and hopefully DeepMind can do more and do it faster.