Apple researchers have built a protein model that learns amino-acid sequences and three-dimensional structures together with a single end-to-end training objective. SimpleDesign avoids the usual multi-stage process of first training separate encoders and then teaching another model to generate their compressed representations.
Proteins work through the relationship between their sequence and folded shape, so useful design systems need to model both. SimpleDesign combines a classification-style loss for discrete sequence tokens with a regression loss for structure. Its Transformer backbone gives each modality some specialized processing while retaining shared attention across sequence and structural information.
The team trained the system on more than two million sequence-structure pairs. It reports competitive results on benchmarks for co-design, where both sequence and shape are generated, and for unconditional sequence or structure generation. The finding supports the researchers’ hypothesis that strong co-design does not necessarily require a separately trained latent representation.
This is a research result rather than a claim that generated proteins are ready for medical or industrial use. Benchmark performance does not establish that a proposed molecule will fold, function or remain safe in the laboratory. The practical contribution is a simpler training route that researchers can compare with more elaborate protein-generation pipelines.