A model designs proteins by learning sequence and structure together in one step.

Proteins do essential jobs in living things. A protein's role depends on both its amino acid sequence and its three-dimensional structure. Designing proteins usually requires complex multi-stage training. In those setups, models first compress protein data into hidden codes, and then another model learns to generate those codes. The researchers hypothesized that this multi-stage process is unnecessary. They created SimpleDesign, a multi-modal model trained directly on the raw data in one single stage. It works by combining two targets at once: predicting sequence steps and calculating structure shapes. A special Mixture-of-TransformerMixture-of-TransformerA type of neural network that combines multiple transformer modules to handle different types of data while sharing some global attention. setup handles the differences between sequences and structures while letting them communicate globally. When tested, SimpleDesign achieved strong performance across co-design tasks and unconditional sequence and structure generation.

Why it matters

This could simplify drug discovery and protein engineering by providing a more efficient way to design new proteins. Scientists often spend years tweaking proteins manually; a unified model might reduce time and resources.

Questions answered