Apple Unveils SimpleDesign for Joint Protein Sequence and 3D Structure Generation

Apple researchers have released SimpleDesign, an end-to-end generative AI model that simultaneously designs protein amino acid sequences and their 3D structures using flow matching. Unlike previous joint protein design approaches that rely on discrete latent representations trained via autoencoders, SimpleDesign trains directly on raw amino acid sequences and continuous 3D coordinates. By simplifying the training architecture and eliminating intermediate conversion steps, SimpleDesign demonstrates that streamlined models can achieve competitive performance in protein folding, inverse folding, and co-design tasks. This advances AI for computational biology and could potentially streamline custom protein design for biological and therapeutic applications. SimpleDesign was trained on over 2 million protein sequence-structure pairs from the AFESM dataset by applying random sequence masking and structural noise to unifiedly master protein folding, inverse folding, and joint generation tasks. However, the current evaluation remains strictly in silico (computational), with no wet-lab experimental validation yet to confirm physical folding or biological function.