Apple Unveils SimpleDesign: An End-to-End Protein Co-Design Model Bypassing Latent Discretization

IT Home reported on September 12 that Apple researchers unveiled a protein design model named SimpleDesign earlier this month, with the research paper published as a preprint on arXiv. The model can jointly generate the amino acid sequence and 3D structure of a protein. By training directly on raw data, it eliminates the need for intermediate representation conversion steps found in traditional multi-stage training pipelines.

The researchers noted in the paper that many existing joint protein design models require training an autoencoder first to convert protein structures into discrete latent representations before training a generative model to process these representations. In contrast, SimpleDesign performs end-to-end training directly using amino acid sequences and continuous 3D coordinates.

This research builds upon the technical approach of Apple's earlier SimpleFold project. SimpleFold uses a flow matching model to predict 3D protein structures directly from amino acid sequences. SimpleDesign extends this simplified architecture to protein design tasks, not only predicting structures but also attempting to simultaneously generate amino acid sequences capable of forming the corresponding structures.

The research team utilized more than 2 million pairs of protein sequence and structure data, primarily sourced from the AFESM dataset. This dataset aggregates predicted structures from the AlphaFold database alongside other samples. During the training process, the model randomly masks portions of the amino acid sequence while adding noise to the corresponding 3D structure.

By adjusting the degree of corruption applied to the sequence and structure, the model can learn different protein tasks. When the sequence is largely intact but the structure is heavily perturbed, the task resembles protein folding—recovering the structure from a known sequence. When the structure is largely intact but the sequence is heavily masked, it resembles inverse folding—generating an amino acid sequence capable of forming a specified structure.

When both the sequence and structure are partially corrupted simultaneously, the model must jointly process both types of information, thereby learning the co-design of protein sequences and structures.

IT Home note: Protein folding refers to the process by which an amino acid chain forms a specific 3D structure, whereas inverse folding involves finding amino acid sequences that could potentially correspond to a target structure.

According to the research findings, SimpleDesign achieved competitive performance across benchmarks for joint protein design, structure generation, and sequence generation. The researchers also found that the model can generate protein structures with plausible conformations, and the overall quality of the generated amino acid sequences is comparable to or better than that of most competing multimodal models.

However, the paper currently reports only computational evaluation results. The research team has not yet experimentally validated whether the generated proteins can actually fold, function, or operate safely in biological systems. Therefore, these results do not yet directly prove that the proteins designed by the model possess practical real-world utility.