Apple researchers have introduced SimpleDesign, an AI model for SimpleDesign protein design that generates amino-acid sequences and three-dimensional structures together. The model was trained on more than 2 million sequence–structure pairs and performed competitively on computer-based protein design benchmarks, although the generated proteins have not been tested in a lab.
SimpleDesign protein design skips a training stage
SimpleDesign takes a different route from protein-design systems that first compress structures into learned tokens and then train a separate generative model. Apple’s researchers trained it directly on paired sequences and 3D coordinates in one end-to-end process.
The model uses a multimodal Transformer, with separate processing for amino-acid sequences and structural coordinates while allowing the two streams to share attention. Its training objective combines cross-entropy for masked sequence tokens with a regression objective for noisy structural coordinates.
That lets the same system handle several related tasks. With a known sequence and damaged structure, it can work towards protein folding; with a known structure and masked sequence, it can generate a compatible sequence. When both inputs are partly corrupted, it performs sequence-and-structure co-design.
Apple’s model is promising, but not a lab result
The researchers report competitive results across co-design, structure-generation and sequence-generation benchmarks. The paper also says the approach can produce plausible structures and sequences that compare favourably with most of the multimodal models tested.
Those results are still computational. The study does not show that a protein generated by SimpleDesign folds as intended, performs a useful biological function or is safe in an experimental setting. The work is a model and a set of benchmark results, not a drug-discovery product.
The research follows Apple’s earlier SimpleFold work on protein-structure prediction. It also gives the company’s machine-learning group another example of favouring general-purpose Transformer components over more specialised architecture, at least for an area where the underlying biology remains stubbornly specialised.
The full paper is available through Apple’s machine-learning research site, with the technical version published on arXiv. The original report and background are from 9to5Mac.
What is Apple’s SimpleDesign model?
SimpleDesign is an AI model that generates protein amino-acid sequences and three-dimensional structures together rather than treating the two tasks as separate stages.
How was SimpleDesign trained?
The model was trained directly on more than 2 million paired protein sequences and structures using a single-stage, end-to-end objective.
Has SimpleDesign designed a working drug or protein?
No. The published results are computer-based benchmarks, and the generated proteins have not been experimentally tested for folding, function or safety.
