The University of Washington's Institute for Protein Design (IPD, the team of 2024 Nobel chemistry laureate David Baker) and NVIDIA have jointly released RFD4-Proteina, a new unified protein design model, open-sourced under the BSD-3 license. It fuses two technical lines — IPD's RosettaFold Diffusion and NVIDIA's Proteina-Complexa — optimizing a protein's amino-acid sequence, 3D structure, and the molecules it interacts with simultaneously, and can devote more computation to difficult designs much like a language model's "thinking." Benchmarks show computational design success rates 5 to 10 times higher than previous models, with candidate proteins generated 5 to 10 times faster.
Even more convincing is the wet-lab validation: a team at the University of Manchester's Manchester Institute of Biotechnology (MIB) used the model to design brand-new enzymes for synthesizing lenacapavir, an HIV prevention drug, targeting a key amide-bond-forming step. The reaction is extremely challenging because the enzyme must promote the desired amide bond formation while suppressing a competing reaction with water, present in vast excess. Of 67 computational designs, four showed clear activity; the best converted 79% of a bulky chiral amine into the target product, with high selectivity for the required S-enantiomer. Baker says the model will open the door to new solutions in medicine, sustainability, materials, and biotechnology.
Source: Read the original article
Comments