Apple researchers have published an AI model for protein design called SimpleDesign. What sets it apart is that it can generate a protein's amino acid sequence and its three-dimensional structure at the same time.
The work appeared on the Apple Machine Learning Research page, and it questions an established habit in the field.
Why both have to be generated together
A protein's function does not depend on the order of its amino acids alone. The shape the chain takes in space is at least as decisive as that order.
So when designing a new protein, the relationship between sequence and structure has to be handled jointly. Generating each well on its own does not mean you have produced a coherent whole.
The step that was removed
Some existing generative methods work in two stages. First an autoencoder turns the protein's structure and sequence into latent tokens, then a generative model learns in that compressed space.
The Apple team argues this multi-stage process is not necessary. Rather than compressing the data through a separate encoder first, SimpleDesign is trained end to end directly on sequences and structures.
| Approach | Flow |
|---|---|
| Established method | Latent representation via autoencoder, then generation |
| SimpleDesign | Direct end-to-end training on sequence and structure |
| Result | One stage, and the information it loses, disappears |
The two-stage setup carries a cost: information is lost during compression, and the generative model no longer sees the raw data but its compressed form. Removing the intermediate layer removes that loss.
How the model works inside
SimpleDesign processes the two data types within the same model but uses different computational paths for each. A discrete cross-entropy objective handles sequences, a regression objective handles structures.
A global self-attention layer then establishes the links between sequence and structure in a shared context. The two separately processed representations meet in a single attention layer.
Results and limits
The researchers trained the model on more than 2 million sequence-structure pairs. Competitive results are reported on co-design and unconditional generation tests.
Caution is needed here: this is a research model. It does not amount to a new drug or a lab-validated protein product. A design that scores well computationally still has to fold and function, and that is a separate validation process.
What Apple is doing here
Apple appearing on the biology side looks surprising at first. The company's machine learning research arm has for some time published work independent of its product line, most of it as open papers.
The contribution here is not a product but a claim of simplification: showing that the established two-stage setup is not required. If the claim holds, the gain is not only in speed but in the information lost at the intermediate stage.