Stanford and Microsoft teamed up to use diffusion models to generate protein structures, which has been open source

Proteins are vital to life and play a role in nearly every biological process. On the one hand, they can transmit signals between neurons, identify microscopic invaders, and activate immune responses. On the other hand, proteins have been extensively studied as a therapeutic mediator as part of the treatment of diseases. Thus, by generating new, physically foldable protein structures, the door is opened to new ways of exploiting cellular pathways to treat disease. ​

In this article, researchers from Stanford University, Microsoft Research and other institutions, inspired by the protein folding process in vivo, introduced a folding diffusion (folding diffusion, FoldingDiff) ) model, which designs protein backbone structures by mirroring the natural folding process of proteins.

• ##Paper address: https://arxiv .org/pdf/2209.15611.pdf
• Code address: https://github.com/microsoft/foldingdiff​

Specifically, they describe the protein backbone structure as a series of consecutive angles that capture the relative orientation of the constituent amino acid residues, which The inherent displacement and rotation invariance of this representation greatly alleviates the need for complex equivariant networks.

This study trained a denoised diffusion probabilistic model based on the transformer backbone and demonstrated that our model can unconditionally generate highly realistic protein structures with complexity and structural patterns similar to those of natural protein.

Some netizens said: I wonder if this model will bring some competition to AlphaFold.

Methods and results​

We can understand proteins as chains of amino acid residues of variable length. Typical amino acids are 20 species, sharing the same three-atom N-C_α-C backbone, but with different side chains connected to the C_α atoms (usually denoted as R, see Figure 1).

These residues assemble to form polymer chains that fold into a 3D structure, the shape of which largely determines the protein's function. These folded structures can be described using four levels:

• Primary structure, which simply captures the linear sequence of amino acids;
• Secondary structure describes the local arrangement of amino acids;
• tertiary structure describes the complete spatial arrangement of all residues;
• quaternary structure , describes how multiple different amino acid chains come together to form larger complexes.

The study proposes a simplified protein backbone framework that follows the biological process of protein folding while eliminating the need for complex equivariant networks. Rather than viewing a protein backbone, N amino acids in length, as a three-dimensional coordinate, they viewed it as a sequence of six internal, consecutive angles. That is, given the position of the current residue, the vector of six interior angles describes the relative positions of all backbone atoms in the next residue. These interior angles can be easily calculated using trigonometric functions, iteratively adding atoms to the protein backbone and then converting back to 3D Cartesian coordinates.

The picture below shows the results of an experiment. The Ramachandran diagram of the natural structure (figure a) contains three regions corresponding to the LH α-helix, RH α-helix, and β-sheet. All three regions are fully reproduced in the structure generated here (Fig. 3b). In other words, FoldingDiff is able to generate secondary structure elements within the protein backbone. Additionally, experiments show that the FoldingDiff model correctly learns that RH α-helices are more common than LH α-helices. Previous work using equivariant networks was unable to distinguish between these two types of spirals.

The following figure shows the secondary structure in the test main chain (4a) and the generated main chain (4b) Two-dimensional histogram, the results show that the generated structure reflects the true structure of the protein, with multiple α-helices, multiple β-sheets, and a mixture of the two.

The figure below shows that 111 of the 780 generated structures (14.2%) are designable, and their scTM scores ≥0.5 (Fig. 5a), which is higher than the value of 11.8% reported by Trippe et al. We also see that the generated main chains are more similar to the training examples and tend to have better designability (5b).

For more information, please read the original paper.

The above is the detailed content of Stanford and Microsoft teamed up to use diffusion models to generate protein structures, which has been open source. For more information, please follow other related articles on the PHP Chinese website!