Designing Protein Structures from Natural Language (arxiv.org)

A groundbreaking advancement in protein design has been made with the introduction of Prompteina, the first all-atom structure generation model capable of translating natural language descriptions into protein structures. Traditional protein design methods often required specific input formats for different structural and functional properties, limiting their versatility. In contrast, Prompteina utilizes a unique paired dataset known as Prompteina-Data, which includes 1.8 million protein structures linked to over 50 million descriptive captions. This innovation allows researchers to specify diverse design goals in plain language, significantly enhancing the model's usability. Prompteina demonstrates impressive performance, surpassing existing models on 15 of 18 structural properties and significantly improving binder designs. Notably, it achieved a 14% improvement in interface confidence over leading baselines and identified nine binders for the previously elusive GM2A target, with the strongest exhibiting a binding affinity of 107 nM. The success of Prompteina underscores the potential of using natural language as an effective tool for guiding protein structure generation, opening new avenues for research in synthetic biology and drug development.