Summary

Apple researchers describe SimpleDesign, a single-stage multimodal model for jointly generating protein sequences and three-dimensional structures. It was trained on more than 2 million sequence–structure pairs, but the supplied evidence does not show that its generated proteins have been experimentally tested.

Contents

What SimpleDesign changes

Apple Machine Learning Research has presented SimpleDesign, a Transformer-based multimodal generative model designed to handle protein sequences and three-dimensional structures together. The work is listed as a paper published in Transactions on Machine Learning Research in September 2026.

A protein sequence is an ordered chain of amino acids. Its structure describes how that chain folds and arranges itself in three dimensions. These are different forms of information, but the source describes their complex interplay as central to protein function.

Many generative approaches described in the research area use multiple stages. They may first compress sequences or structures into latent representations—internal, lower-dimensional encodings—and then train a generative model on those representations. Apple presents SimpleDesign as an alternative that trains directly on sequence and structure data with a single-stage, end-to-end objective.

The model is intended for sequence-and-structure “codesign”, meaning that it treats the two modalities as parts of one modelling problem rather than as entirely separate generation tasks.

How the model is trained

SimpleDesign combines two types of learning objective.

For protein sequences, it uses discrete cross-entropy. Cross-entropy is commonly used when a model predicts tokens or categories, such as the next amino acid in an ordered sequence. For structures, it uses a regression objective, which is appropriate for continuous-valued targets.

The architecture uses Transformer-based multimodal backbones. It applies modality-specific processing so that sequence and structure information can be handled according to their different forms. At the same time, global self-attention operates across both modalities. Self-attention allows a model to calculate relationships between elements in its input; in this case, the design is intended to help it relate sequence information to structural information.

Apple says SimpleDesign was trained on more than 2 million sequence–structure pairs. The supplied research description does not specify how those pairs were assembled or filtered, how the data were divided between training and evaluation, or the exact representation used for structures.

The direct, single-stage approach could simplify the training pipeline by avoiding a separately trained autoencoder or other data compressor. However, whether that produces a practical advantage over multi-stage systems depends on the experiments. The supplied evidence does not quantify such an advantage.

What the reported results establish

Apple reports that SimpleDesign achieved competitive performance on three broad groups of benchmarks:

  • co-design of protein sequences and structures;
  • unconditional sequence generation;
  • unconditional structure generation.

This is a claim about benchmark performance, not a claim that the model is superior to all other systems. The supplied source does not provide benchmark names, numerical scores, comparison tables, uncertainty estimates, ablation studies or the specific baselines used.

The work therefore establishes that Apple researchers have developed and described a joint generative modelling framework, trained it at a scale of more than 2 million sequence–structure pairs, and evaluated it on generation benchmarks according to the company’s research summary. It does not, from the supplied evidence, establish that the generated proteins fold as predicted, perform a desired biological function or offer a therapeutic benefit.

Why the limitations matter

Protein design is not complete when a computer produces a plausible sequence or structure. A generated candidate must still be assessed for properties such as structural stability, folding behaviour and biological activity. Computational generation alone does not demonstrate that a protein will work in a biological system or be safe for a medical application.

The supplied material reports no wet-laboratory experiments, functional assays, animal studies, clinical studies or therapeutic candidates. It also does not provide enough information to assess possible dataset bias, the quality of the structures used for training, or how well the model generalises to novel protein families.

The work is best understood as a machine-learning and computational-biology contribution. Although the research context identifies drug discovery and protein engineering as potential areas of relevance, no medical use, patient benefit or regulatory status is supported by the evidence provided.

The source also does not state whether SimpleDesign’s code, model weights or training data are publicly available. It therefore cannot yet be assessed as a tool that researchers or protein engineers can directly use.

What to watch next

The full paper and follow-up evaluations would need to clarify several points:

  • the exact benchmark definitions and numerical results;
  • the baselines used for comparison;
  • the composition and filtering of the sequence–structure dataset;
  • the effect of removing or changing parts of the single-stage objective;
  • whether generated proteins have been experimentally produced and tested.

Independent reproduction would also help determine whether the reported performance depends on the particular dataset and evaluation setup. Experimental testing would be required to establish whether the model’s outputs correspond to proteins that fold and function as designed.

Sources