A paper published in the journal Nature on the 11th describes a new breakthrough in structural biology: a deep learning method called RoseTTAFold Diffusion (RFdiffusion) that can design new proteins. It can generate various functional proteins, including topological structures that have never been seen in natural proteins. ▲ Research diagram (part) Source: The in-depth learning of Nature website has promoted the prediction and design of Protein structure, but a general framework is still needed to overcome various challenges in protein design. Diffusion model is a generative simulation method, which has been proved to be very useful in image and text generation, and seems to be also applicable to protein design. However, the current success rate of such models is not high, and the resulting sequences cannot be folded into the target structure. Research by scientists at the University of Washington in the United States shows that by carefully tuning the previously reported structure prediction network of RoseTTAFold and integrating it into a noise reduction Diffusion model, we can generate a protein skeleton with practical significance, which determines the shape and function of the protein. The RFdiffusion model can test design combinations with different structural elements and generate proteins from scratch. This model can also perform different tasks, such as designing monomers (the basic building blocks of proteins), oligomers (multi subunit polymers), and complex structures with therapeutic or industrial application prospects. The team conducted experimental characterization of the structures and functions of hundreds of designed symmetric polymers, metal binding proteins, and binding proteins, demonstrating the practicality of this method. They generated a complex of a binding protein designed by RFdiffusion and its substrate (here is influenza hemagglutinin, a protein found on the surface of influenza virus) and analyzed its structure, and found that the results were almost the same as the designed model, thus proving the accuracy of the method. Researchers have stated that RFdiffusion is a comprehensive improvement on the current protein design method, which can generate a structure with a total length of 600 residues, with higher complexity and accuracy than before. Further improvements to this method in the future will enable the design of new proteins with higher complexity. (New News Agency)