Apple researchers unveil SimpleDesign, a new AI model for protein design

Apple researchers have unveiled a new AI model called SimpleDesign, designed to jointly generate protein sequences and structures. The study, titled "SimpleDesign: A Joint Model for Protein Sequence and Structure Codesign," builds upon their previous work with SimpleFold, which streamlined the process of predicting protein 3D structures from amino acid sequences.

SimpleFold utilized a flow-matching model, which differs from the more common diffusion models used in image generation. The researchers found that pairing flow matching with general-purpose Transformer blocks enabled SimpleFold to avoid computationally expensive techniques typically used in protein folding models like DeepMind's AlphaFold.

In their new study, Apple researchers expanded SimpleFold's approach to protein design, presenting SimpleDesign as an effective multi-modal protein design model trained directly in data space. Unlike existing models that rely on a multi-stage process, SimpleDesign learns to generate amino acid sequences and 3D structures in a single end-to-end training process.

To train the model, the researchers used over 2 million protein sequence-and-structure pairs, introducing noise to both amino acid sequences and 3D structures during training. This corrupted data allowed SimpleDesign to learn both protein folding (from known sequences) and inverse folding tasks, enabling it to generate plausible protein structures and sequences.

The results were promising, with SimpleDesign performing competitively across various benchmarks for protein co-design, structure generation, and sequence generation. However, the generated proteins were not experimentally tested to confirm their safety and functionality in biological systems. Nonetheless, the study demonstrates a significant step forward in AI-driven protein design, with the full research paper providing an in-depth look at SimpleDesign's architecture, training process, benchmarks, and results.