A team of researchers from the University of Washington has developed a new method for protein sequence design that overcomes common challenges. The new method, called LigandMPNN, uses deep learning to model interactions between enzymes and small molecule binders and sensors. Unlike previous methods, LigandMPNN explicitly considers non-protein atoms and molecules, making it more accurate in designing protein sequences that interact with small molecules, nucleotides, and metals.
LigandMPNN uses a graph-based approach to represent protein-ligand interactions, allowing for more accurate modeling of non-protein atomic contexts. The model has been demonstrated to outperform existing methods in terms of sequence recovery and side-chain packing accuracy, as well as speed and efficiency.
Overall, LigandMPNN fills a critical gap in current protein sequence design methods and has the potential to greatly aid in protein engineering.
Check out the paper for more information. And don’t forget to follow the researchers on Twitter, Google News, and join their ML SubReddit, Facebook Community, Discord Channel, and LinkedIn Group. Sign up for their newsletter and join their Telegram Channel if you like their work.