Apple unveils SimpleDesign AI model for protein design

September 12, 2026 at 03:22 PM1 min

This digest was compiled by AI from multiple sources — links to the originals are below.

Apple researchers have introduced SimpleDesign, an AI model that simultaneously generates amino acid sequences and 3D protein structures. The model simplifies the two-stage approach used by many protein design systems by training directly on sequence-structure pairs. It builds on Apple's earlier SimpleFold model and was trained on over 2 billion sequence-structure pairs.

Key Facts

SimpleDesign was trained on over 2 billion sequence-structure pairs, primarily from the AFESM dataset.
The model builds on Apple's SimpleFold, which used flow matching to generate 3D protein structures from amino acid sequences.
SimpleDesign skips the tokenization step common in many protein design systems, training directly on sequence-structure pairs.
In tests, SimpleDesign showed competitive results on protein design, structure recovery, and sequence recovery tasks.

Model Architecture

SimpleDesign combines flow matching with transformer architecture blocks used in text generation. The model was trained by corrupting both parts of each sequence-structure pair: amino acids were randomly replaced with masking tokens, and noise was added to the 3D structure. Varying degrees of distortion allowed the model to learn structure recovery, sequence recovery, or simultaneous design tasks.

Training Data

The training set comprised over 2 billion sequence-structure pairs, mostly from the AFESM dataset, which combines AlphaFold structures with additional samples. Apple researchers introduced distortions into both sequence and structure during training to teach the model multiple tasks.