Apple's SimpleDesign model can generate protein sequences and structures at once

Apple researchers published a protein design model called SimpleDesign earlier this month, with the paper released as a preprint on arXiv. The model can jointly generate amino acid sequences and 3D structures of proteins, using direct training on raw data and skipping the intermediate representation steps in traditional multi-stage training.

Many existing joint protein design models first train an autoencoder to convert protein structures into discrete latent representations, then train a generative model on those representations. SimpleDesign instead uses amino acid sequences and continuous 3D coordinates for end-to-end training. The work follows Apple's earlier SimpleFold project.

The team used over 2 million protein sequence-structure pairs, mainly from the AFESM dataset. Results show SimpleDesign achieved competitive performance in joint design, structure generation, and sequence generation benchmarks.

However, the paper only reports computer evaluation results. The team has not yet experimentally verified whether the generated proteins can actually fold, function, or run safely in biological systems, so these results cannot directly prove real-world applicability.