Apple researchers detail SimpleDesign protein sequence-structure co-design model

Apple researchers published a study describing SimpleDesign, a model that jointly generates protein sequences and 3D structures in a single end-to-end training process. The work extends the lab's earlier SimpleFold approach, which used flow matching with general-purpose Transformer blocks to predict structure from sequence. SimpleDesign trains directly on paired amino acid sequences and 3D coordinates from more than 2 million sequence-and-structure pairs, skipping the separate tokenization stage many co-design models use. The researchers report competitive results across co-design, structure generation, and sequence generation benchmarks, and say generated sequences were generally as good as or better than most competing multimodal models. The study states results remain limited to computer-based evaluations, with no experimental confirmation that generated proteins fold, function, or behave safely in biological systems.