On May 8, 2024, the DeepMind team published an article in the journal Nature, "Precise Structure Prediction of Biomolecular Interactions Using AlphaFold 3". This is the latest AI released by DeepMind in the journal Nature to solve a specific class of problems, following the release of AlphaGeometry, an artificial intelligence that "can solve plane geometry problems at the Olympiad level" on January 17, 2024.
Proteins are composed of protein primary structure, and proteins will spontaneously fold to form protein tertiary structures during protein folding. Protein structure is critical to protein biological function. However, understanding how amino acid sequences determine the tertiary structure of proteins is extremely challenging, which is known as the "protein folding problem".
Protein folding
AlphaFold 3 is the latest version of the AlphaFold series of artificial intelligence programs released by DeepMind. As the name of AlphaFold "Alpha Fold" suggests, the AlphaFold series was designed as an artificial intelligence for predicting protein structures.
在2020年11月的第14届CASP（蛋白质结构预测技术的关键测试，Critical Assessment of protein Structure Prediction）竞赛中，AlphaFold 2的中位分数为92.4（满分100分）。 其准确度远远高于其他任何程序。
Subsequently, AlphaFold 2, with its database of predicted structures of 200 million proteins of all known DNA sequences, was made available to scientists as a free and open source. According to John Juper, senior researcher and head of DeepMind's protein structure team, "AlphaFold 2 has been cited more than 20,000 times in other published scientific papers and is being used to study drugs to treat malaria, cancer and many other diseases. ”
In the biopharmaceutical field, AlphaFold greatly reduces the time and cost of obtaining protein structures, accelerating protein structure-based drug development. Fiona Marshall, now director of the Novartis Institute for Biomedical Research, exaggerated: "AlphaFold makes everyone a structural biologist. ”
Transcription factors and ribosomal RNA molecular structure models. Visual China|Figure
Despite the amazing performance of AlphaFold 2, there is still a long way to go before the data from AlphaFold 2 can be used in real-world applications such as drug design. This is because AlphaFold 2 has its own limitations.
As mentioned earlier, the function of AlphaFold 2 is to predict the structure of proteins and how they are folded. While knowing the shape and structure of a protein is often a key part of understanding its function. But in living organisms, including the human body, proteins do not operate in isolation. In addition to the three-dimensional structure of the protein, there are also the interactions between the protein and various other molecules. However, AlphaFold 2 cannot predict the interaction of proteins with other types of molecules, such as DNA, RNA, ligands, and ions in living organisms. It is also unable to predict the interactions between these other molecules. In addition, AlphaFold 2 cannot determine the conformational state of the predicted protein structure. There is also uncertainty about the accuracy of its prediction structure, even if it is possible for different regions of the same protein to be predicted. These limit the usefulness of the AlphaFold 2.
Huge boost
DeepMind's newly released AlphaFold 3 has been greatly improved and improved in these areas.
According to DeepMind and Isomorphic Lab, AlphaFold 3 has a 62% prediction accuracy for protein-protein interactions, which is twice as accurate as AlphaFold 2. Not only that, but AlphaFold 3 can predict the interaction of proteins with small molecules such as DNA, RNA strands, ligands, and ions, as well as the interactions between these small molecules, compared to AlphaFold 2, which can only predict interactions between proteins. The paper states that the model can accurately predict "complexes containing almost all molecular types in the protein database."
In terms of accuracy, AlphaFold 3 achieved 76% accuracy in predicting protein-small molecule interactions, compared to 52% for the previous best prediction software. AlphaFold 3 was 65% accurate in predicting interactions between DNA, compared to 28% for the previous best prediction software.
The huge improvement of AlphaFold 3 is a typical example of the various amazing advances in the field of artificial intelligence in recent years. Its performance also reflects the latest developments and achievements in the field of artificial intelligence in recent years.
相较于之前的AlphaFold 2，AlphaFold 3的主要改进包括，大大简化了之前的多重序列比对（Multiple sequence alignment，MSA）流程；将之前的Evoformer替换成了Pairformer；将结构模块（Structure Module）替换成了扩散模块（Diffusion Module）。
These improvements have greatly improved the efficiency of AlphaFold 3. More importantly, it changes the way AlphaFold 3 works. On AlphaFold 2, protein structure is predicted by simulating the way proteins fold based on their physical and chemical properties. As a result, AlphaFold 2 is often inefficient, inaccurate, and can only be used for protein prediction.
The diffusion module used in AlphaFold 3 is based on a completely different logic to achieve prediction.
The diffusion model is the most advanced of today's deep generative models. This is the so-called SOTA model: state-of-the-art model. The diffusion model has excellent performance in many fields such as computer vision, natural language processing, waveform signal processing, multimodal modeling, molecular graph modeling, time series modeling, and adversarial purification. Earlier, OpenAI released SORA, an artificial intelligence model that can generate videos from text descriptions, is based on the diffusion model.
Just as SORA can be trained to produce videos that conform to the laws of real-world physics in the vast majority of cases. AlphaFold 3, which uses the diffusion module, can also be trained to directly generate the structure of biomolecules, including proteins. In other words, AlphaFold 3 no longer simulates the process of protein folding, but directly determines the position coordinates of each atom in the biomolecule at the atomic level, and then generates the structure of the entire molecule.
With this change, AlphaFold 3 is no longer limited to predicting the structure and interactions of proteins, but can be used to simulate and predict the structure of biomolecules including DNA, RNA, ligands, and small molecules, including ions, and the interactions between them.
As Jim Fan, a senior scientist at NVIDIA, said on his personal social media: "AlphaFold 3 is the latest and biggest breakthrough in the field of biology iterated by artificial intelligence. What's new is that AlphaFold 3 uses diffusion to 'render' the molecular structure, and then denoising the specific structure of the molecule from a cloud of atoms over time. ”
Commenting on the huge boost to AlphaFold 3, DeepMind's CEO Demis Hassabis said at a launch event announcing the breakthrough on May 7: "Today's announcement of AlphaFold 3 is a significant milestone for us. Biology is a dynamic system, and you have to understand how various properties in biology arise through the interactions between different molecules in the cell. You can see AlphaFold 3 as our first big step in this direction. ”
Hassabis is also optimistic that the first AI-designed drugs will be in clinical use in the next few years.
It also needs to be improved
It's important to note that while AlphaFold 3 has been greatly improved and improved over its predecessors, it still has some areas for improvement.
Among them, one of the most influential problems is the so-called "illusion" of artificial intelligence. At this stage, large-scale artificial intelligence models trained based on deep learning, including GPT and SORA, have more or less "hallucination" phenomena. That is, AI generates nonsensical or erroneous content. With the generative diffusion model, it is inherently hallucinatory. This may be just a "harmless" problem for GPT and SORA, which are mainly used for everyday conversations and generating videos. However, for an AI like AlphaFold 3, whose real use is for biological research and pharmaceuticals, it would be unacceptable to make a mistake like generating a "structure that seems reasonable".
For this, the DeepMind team took what is known as cross-distillation. That is, add the structure of the previous version of AlphaFold-Multimer v2.3 prediction to the pre-trained dataset and let AlphaFold 3 learn it. This in turn reduces AlphaFold 3's hallucinatory behavior. Like AlphaFold 2, DeepMind has added a confidence component to AlphaFold 3. AlphaFold 3 will mark the confidence of different parts in the given prediction structure for the user to identify.
In addition, for macromolecules such as proteins, there is a very special property in structural chemistry called chirality. It refers to the fact that certain molecular structures, like the left and right hands, appear to be mirror-symmetrical. In some reactions, isomers with chiral symmetry will exhibit different reaction results. As a result, it is necessary to make a strict distinction between chirality in the prediction of molecular structures such as proteins. However, the DeepMind paper notes that despite various approaches, there is still a 4.4% chance that AlphaFold 3 will violate chirality in its predictions.
In addition to this, the DeepMind paper also points out that in some predictions, AlphaFold 3 also shows that the atomic positions in the product overlap with each other.
These problems show that AlphaFold 3 is still a long way from being truly an artificial intelligence that can solve human diseases. The problems that DeepMind explicitly mentioned in the paper also show that they have a clear understanding of the improvements and limitations of AlphaFold 3.
As Yan Ning, an academician of the Chinese Academy of Sciences and a structural biologist, said on his personal social media: "My attitude towards AI can be summed up in two words: 'awe', and its development speed is beyond imagination. …… I think this server version is a balance of speed and accuracy, and the accuracy rate is not the best. …… But again, AI will become more and more powerful, and how to embrace new technologies and ask more interesting questions is what relevant researchers are more concerned about now. ”
Unlike GPT and SORA, which are so versatile, artificial intelligence seems to be very close to our daily lives. From AlphaGeometry at the beginning of the year, to AlphaFold 3 more recently, and back to AlphaGo, the DeepMind team has launched artificial intelligence that specializes in solving a specific problem. These specialized types of AI, like AlphaGo and AlphaGeometry, approach or exceed the highest levels of human intelligence in some areas of pure intelligence. Or like the AlphaFold series, which really makes a difference in the world of scientific research.
Perhaps, this is the right way for artificial intelligence to turn on. After all, we are human beings, and artificial intelligence is a tool for our creation.
Contributing writer for Southern Weekly, Zuo Li
Editor-in-charge: Zhu Liyuan