Apple Releases 'SimpleDesign', a Protein Design Model That Skips Multi-Stage Training

Apple researchers have released SimpleDesign, a new model that simultaneously designs protein sequences and structures. While existing generative models underwent a multi-stage process involving autoencoder training to convert data into a latent space and generative model training within that space, SimpleDesign adopts a single-stage (end-to-end) approach that trains directly in the data space without these steps.

Key Technologies and Performance

• Single-Stage Training: Uses a single objective function combining discrete cross-entropy for sequences and regression targets for structures.
• Multimodal Backbone: Employs a Transformer-based backbone to effectively handle the differences between sequence and structure modalities, maintaining global self-attention for both modalities alongside modality-specific processing.
• Training Data: Trained on over 2 million sequence-structure pairs.

The model demonstrates competitive performance on co-design and unconditional sequence/structure generation benchmarks, showing that high-quality protein design is possible without complex multi-stage training pipelines.