AlphaFold 3 takes an important step towards decoding molecular behavior and biological computing, a Nature sub-journal commented

Editor | KX

If we fully understood how molecules interact with each other, there would be nothing left to learn about biology, because every biological phenomenon, including how we perceive the world, ultimately originates in cells Behavior and interactions of biomolecules within.

The recently launched AlphaFold 3 can predict the 3D structure of biomolecular complexes directly from the sequences of proteins, nucleic acids and their ligands. This marks significant progress in our long-term exploration of how biomolecules interact.

AlphaFold 3 represents a breakthrough in predicting the three-dimensional structure of a complex directly from its sequence, providing insights into biomolecular interactions.

1. Code of a Biomolecule

A one-dimensional (1D) sequence of a biomolecule (such as a protein or nucleic acid) that specifies a cellular function, similar to a piece of code that specifies a program. This sequence represents code in a programming language and is "compiled" into code in machine language through a folding process, forming a unique 3D structure.

2. Execution of the program

The program is performed by the interaction between the folded biomolecules and other molecules within the cell.

3. Interactions of biomolecules

Due to their unique three-dimensional structure, biomolecules will only interact with a small number of molecules within the cell (such as DNA sites), and these interactions will trigger a series of carefully planned chemical and structural transformations that together define biochemical programs (e.g., transcription). The products of biochemical processes, such as RNA, represent the output of the executing program.

4. Encoding of Biological Sequences

Thus, in biology, a one-dimensional sequence of biomolecules encodes the program and the means to compile and execute the program; this sequence encodes the software and hardware. Predicting the three-dimensional structures formed by complexes of biomolecules based on their one-dimensional sequences is a critical step in understanding how biological programs are performed, with profound implications for our ability to understand, rationally manipulate, and design biological systems.

Illustration: "Computational Dogma" of Molecular Biology (Source: Paper)

1. AlphaFold 2

• Released in 2020, a revolutionary protein structure prediction algorithm
• Excellent median accuracy Compared with other methods
• provides predicted structures of 200 million known proteins

2. RoseTTAFold

• released in 2021, a protein prediction tool based on deep learning
• prediction accuracy is comparable to AlphaFold 2, faster, Lower computational requirements
• Utilize multi-track neural networks for high accuracy

3. AlphaFold Multimer

• Modified version of AlphaFold 2
• Trained for protein complex datasets
• Improved protein-protein complexes Prediction

4. AlphaFold 3

• Released in May 2023
• Go beyond professional tools to predict the 3D structure of protein complexes
• Significantly improve the prediction accuracy of protein-ligand and protein-nucleic acid complexes
• Predict structures containing multiple covalent modifications

5. Technology update

• Replace the structure module with the diffusion module
• Directly predict the Cartesian coordinates of individual atoms
• Expand to a wider chemical space

  

  Illustration: Illustrative example of the diffusion process that powers AlphaFold 3’s diffusion module. (Source: paper)

As a simplified illustration of AlphaFold 3:

1. Imagine taking the three-dimensional coordinates of each atom in a typical biomolecule complex.
2. Iteratively add more and more Gaussian noise to it until we get a randomly distributed cloud of atoms in space (forward diffusion).
3. Diffusion models use multi-layer neural networks to learn to reverse this process (reverse diffusion).

In this way, the diffusion module in AlphaFold 3 learns to:

• Predict the coordinates of every atom in a given complex without the need for a predefined residue framework.
• Broader chemical space including nucleic acids, ions, ligands and chemical modifications.

Other improvements:

• Replaced Evoformer with Pairformer, a newer Transformer architecture.
• Less emphasis on MSA handling.
• Update metrics to adapt to changes in network architecture.

Progress and Limitations:

• 進步：提高了預測精度，減少了對序列比對的依賴，增加了對殘基相互作用的重視。
• 限制：有時無法正確模擬分子的手性，無法預測大型蛋白質-核酸複合物的結構，生成模型可能會出現「幻覺」。

RNA 預測：

• AlphaFold 3 對 RNA 標靶的預測準確度高於其他方法，但不如頂級人類專家。

AlphaFold 伺服器：

• 提供使用者友善的介面，但原始程式碼和執行檔不公開。
• 偽代碼取代了原始碼，導致了爭論和阻礙了進一步的發展。

  

  1. 在考慮AlphaFold 3 帶來的結構預測突破時，重要的是要記住，結構生物學的目標不是預測生物分子及其複合物的3D 結構，而是預測它們的行為以及執行生物程序時會發生什麼。

2. 為了在預測分子行為方面取得進展，我們必須認識到結構預測問題並不像看起來那麼明確。生物分子及其複合物不會折疊成單一結構，而是形成數千種不同構象的集合，每種構像都有不同的機率和壽命。
3. 了解這些構象景觀以及它們在生物分子相互作用時如何變化，對於定量預測親和力和動力學速率至關重要。
4. 從各種條件下的序列預測構象集合是我們現在必須集中精力解決的問題，從而獲得對分子行為的定量和預測性理解。
5. 儘管利用 AlphaFold 3 根據生物分子序列預測其自由和相互複合的 3D 結構，是理解分子行為和生物計算的重要一步，但實驗人員不必擔心被淘汰。結構生物學領域即將變得更加充滿活力。

論文連結：https://www.nature.com/articles/s41594-024-01350-2

The above is the detailed content of AlphaFold 3 takes an important step towards decoding molecular behavior and biological computing, a Nature sub-journal commented. For more information, please follow other related articles on the PHP Chinese website!