Apple researchers have introduced SimpleDesign, a novel AI model capable of generating both protein sequences and their corresponding 3D structures in a single end-to-end system. This new development builds upon Apple’s prior effort, called SimpleFold, which focused solely on folding — predicting a protein’s 3D shape from its amino acid sequence. SimpleDesign seeks to go further by tackling protein design as a joint problem.
What Sets SimpleDesign Apart
Many existing protein design systems divide the task into separate stages. First they compress structural data into latent representations using autoencoders, then they train generative models on those representations to produce new proteins. SimpleDesign skips this multi-stage, tokenization-heavy pipeline. Instead, it works directly in the data space: training on paired amino acid sequences and 3D structural coordinates. This lets the model simultaneously generate viable sequences and coherent folds without an intermediary representation.
How It Works & Training Process
The training set for SimpleDesign includes over two million pairs of protein sequences and their predicted 3D structures, most sourced from a combined dataset that pulls in AlphaFold predictions and additional sampling. During training, both sequence and structure data are corrupted randomly: amino acids in the sequence are masked, while the 3D structure has noise added. The level of corruption varies. In some cases, sequences are mostly intact and structures heavily corrupted — akin to the folding problem. In others, the structure remains true and the sequence is heavily masked, which resembles inverse folding. When both are partially degraded, the model learns to co-design.
Despite its simpler architecture and training pipeline, SimpleDesign performs competitively across benchmarks measuring structure generation, sequence generation, and co-design. The protein structures it generates are plausible, and the sequences it proposes often match or outperform those from more complex multimodal models. However, all results so far are computational — no lab-based experiments have yet confirmed actual folding, functionality, or biophysical behavior.
Background & Why It Matters
SimpleDesign follows SimpleFold, Apple’s earlier model that applied flow-matching techniques combined with Transformer-based components for protein folding. Flow matching starts from a noisy initial state and gradually transforms it into a desired structure; it’s an alternative to diffusion models that typically remove noise iteratively. These approaches, while common in image generation or language tasks, are now being adapted for proteins. SimpleFold simplified structure prediction; SimpleDesign aims to simplify full protein design.
Progress in computational protein design has implications in multiple fields — from therapeutic development and enzyme engineering to materials science. Models that can reliably design both sequence and structure may shorten development cycles, reduce costs, and open the door to entirely new molecules. But computational promise must still be validated in real-world lab settings.
SimpleDesign represents a promising shift toward more streamlined AI architectures in protein design. Its ability to bypass latent space tokenization and operate directly suggests potential for more efficient training and broader applications. What to watch next: whether experimentally validated proteins confirm the computational predictions, how well the model handles novel folds outside its training set, and whether this approach scales to applications where safety, functionality, and durability are critical.
