Apple researchers have introduced SimpleDesign, an AI model capable of simultaneously generating amino acid sequences and 3D protein structures. The architecture builds upon the company's previous SimpleFold project, which utilized flow matching techniques to predict protein structures from known sequences.
Unlike many existing systems that convert protein structures into discrete tokens before generation, SimpleDesign operates directly on pairs of amino acid sequences and 3D coordinates. This streamlined approach eliminates the tokenization stage, simplifying the overall design process.
The model was trained on more than 2 billion protein sequence-structure pairs, primarily sourced from the AFESM dataset. During training, researchers applied varying levels of noise and masking to both sequences and structures, enabling the model to learn both reconstruction and design tasks.
In performance tests, SimpleDesign demonstrated competitive results in protein design and structure recovery. While the generated sequences and structures appear plausible, the researchers noted that these findings are currently limited to computational analysis and have not yet undergone experimental verification.
