1770-Programmable Design of Functional Proteins

From Paper Talk · 00:24:37

Show by 淼淼Elva

Pinal is a 16-billion-parameter generative foundation model that revolutionizes protein engineering by designing bespoke proteins directly from natural-language descriptions. Trained on an unprecedented 1.7 billion synthetic text-protein pairs, the model utilizes an innovative two-stage architecture that first translates textual prompts into Foldseek 3Di structural tokens and subsequently generates compatible amino acid sequences. This structure-mediated approach overcomes the bottlenecks of traditional design by bypassing complex geometric constraints and leveraging a global probability re-ranking strategy. Extensive benchmarking demonstrates that Pinal significantly outperforms existing tools like ProteinDT, Chroma, and ESM3 in both structural foldability and textual alignment. Furthermore, experimental validation across four distinct design tasks confirms its ability to successfully produce functional enzymes and proteins, including catalytically active enzymes and an H-protein that improved metabolic pathway performance.References:Dai F, You S, Zhu Y, et al. Programmable design of functional proteins from natural language[J]. Nature, 2026: 1-10.前往小宇宙评论区与主播互动

More episodes of Paper Talk

More episodes

PodSearch is a free, independent podcast search engine. Read about PodSearch, the privacy policy or contact us.