Apple's New AI Model Designs Proteins Without Complex Preprocessing

Apple researchers have published a new study detailing SimpleDesign, an artificial intelligence model designed to create both the amino acid sequence and the three-dimensional structure of proteins in a single step. According to 9to5Mac, this approach simplifies the complex workflow typically required for protein design, aiming to make the process more efficient and accessible.

The model builds on Apple’s earlier work with SimpleFold, which focused on predicting protein shapes from known sequences. SimpleDesign extends this logic to the broader challenge of designing new proteins from scratch. By handling both sequence and structure generation simultaneously, the system attempts to reduce the computational overhead associated with traditional methods that rely on multiple distinct stages of training.

Simplifying the protein design workflow

Most current protein design models operate in two separate phases. First, they train a component to convert complex protein structures into simplified digital tokens. Then, a second component learns to generate new proteins using those tokens. This multi-stage process adds layers of complexity and potential error. SimpleDesign skips this intermediate translation step entirely, learning directly from raw amino acid sequences and structural coordinates.

This direct approach means the model does not need to compress data into a separate latent space before generating output. Instead, it works within the data space itself. The researchers argue that this end-to-end training process is not only simpler but also sufficient to produce high-quality co-design results, removing the need for the intricate architectural components found in previous systems.

Training through varied corruption levels

The training process for SimpleDesign is notably flexible. The researchers used more than two million pairs of protein sequences and structures, primarily sourced from the AFESM dataset. During training, they intentionally corrupted both elements of each pair. Some amino acids in the sequence were hidden behind masked tokens, while random noise was added to the corresponding three-dimensional structure.

The extent of this corruption was varied to teach the model different skills. If the sequence remained mostly intact but the structure was heavily distorted, the model practiced predicting shapes from known sequences, a task known as protein folding. Conversely, if the structure was clear but the sequence was masked, the model practiced inverse folding, generating sequences that fit a specific shape. When both were partially scrambled, the model learned to design both simultaneously.

Balancing simplicity and capability

While this streamlined architecture offers efficiency gains, it represents a trade-off in how the model handles data representation. By avoiding the use of discrete tokens for structures, the model relies heavily on its ability to interpret continuous coordinates directly. This requires a robust underlying architecture, which SimpleDesign achieves by pairing flow-matching techniques with general-purpose Transformer blocks.

The result is a system that avoids the computationally expensive techniques used by previous leaders in the field, such as AlphaFold. However, the success of this approach depends on the model’s ability to generalize across different levels of data corruption. For researchers, this means a tool that is easier to deploy and train, but one that still requires careful validation to ensure it can handle the full spectrum of protein design challenges without the safety net of staged processing.