Apple unveils simpledesign protein design model for simultaneous sequence and structure generation

Apple researchers have introduced SimpleDesign, a protein design model capable of generating both amino acid sequences and three-dimensional structures simultaneously. The model, published as a preprint on arXiv, bypasses traditional multi-stage training by directly training on raw data. It uses over 2 million protein sequence-structure pairs from the AFESM dataset, which combines AlphaFold predictions with other samples. The model randomly masks parts of sequences and adds noise to structures during training, enabling it to handle tasks like protein folding and inverse folding. SimpleDesign builds on Apple's earlier SimpleFold project, which predicted protein structures from sequences. This new model extends the approach to design tasks, generating sequences that can form specified structures. While the model shows competitive performance in benchmarks, researchers caution that computer simulations alone cannot confirm if the generated proteins can fold, function, or operate safely in biological systems.