Since they are involved in practically every biological activity, proteins are essential for life. The capacity to create innovative, physically foldable protein structures using computation may result in new biological understandings and therapeutic developments for diseases without cures. In this study, inter-residue angles in protein backbones are used rather than cartesian atom coordinates. Training a denoising diffusion probabilistic model with a basic transformer backbone shows that the model produces very realistic protein structures with complexity and structural patterns that are similar to those of naturally occurring proteins.

Example Use Case: Generating 3D protein backbone structures from the pre-trained model.

Technology: Diffusion Model

Limitations:

• Compared to natural proteins, the lengths of the produced structures are still quite small
• When a protein is formulated as a series of angles, errors that occur early in the chain dramatically change the overall generated structure
• The tool is incapable of handling multi-chain complexes or ligand interactions
• Only capable of producing static structures
• Evaluating and visualizing (PyMOL) part is not included in this version. Please see the code page.

Metrics: Some metrics related to TM-Scores can be found in the article.