Apple's SimpleDesign model trains on 2M protein pairs, skips multi-stage

Protein function depends on the interplay between amino acid sequence and three-dimensional structure. Generative models for protein design have often used a multi-stage approach: first training autoencoders to tokenize data into latent representations, then training a generative model in that latent space. Apple's researchers hypothesized that this multi-stage training is not necessary for performant co-design models. They present SimpleDesign, which is trained directly in data space with a single-stage end-to-end objective. This comes after protein folding models like AlphaFold2 achieved groundbreaking results through architecture and training pipeline designs, and after flow matching models emerged as a powerful method for generative modeling on protein structures, typically trained in two stages. SimpleDesign's single-stage approach may simplify the training process, but its broader impact on drug discovery and protein engineering hinges on whether it can match multi-stage models across diverse tasks.