Apple researchers have introduced SimpleDesign, a new AI model that jointly generates protein sequences and 3D structures in a single end-to-end training process. Building on last September's SimpleFold study — which used flow-matching with Transformer blocks to predict protein structures more efficiently than models like Deep Mind's AlphaFold — SimpleDesign extends this streamlined approach to the broader challenge of protein design. Traditional co-design models typically require multi-stage training: first tokenizing structures into discrete representations via autoencoders, then training generative models on those latent spaces.

SimpleDesign bypasses this by learning directly from paired amino acid sequences and 3D coordinates. The model was trained on over 2 million sequence-structure pairs primarily from the AFESM dataset, which combines AlphaFold Database predictions with additional samples. During training, researchers corrupted both sequences (via masked tokens) and structures (via added noise) at varying levels, enabling the model to handle protein folding, inverse folding, and simultaneous co-design tasks.

This unified approach eliminates intermediate tokenization steps, potentially reducing computational complexity while maintaining performance.