Raygun: ИИ уменьшает белки без потери функций — исследование Duke в Nature

Duke University School of Medicine scientists published a generative model called Raygun in Nature on July 29, 2026. The model shrinks natural proteins by 10–25% — and in some cases by more than half — while preserving their three-dimensional structure and functional sites. Outlets have already dubbed the tool a "molecular shrink ray."

How the "shrink ray" works

Raygun is a template-based model: it takes an existing natural protein as a template and introduces substitutions and insertions-deletions of amino acids (indels) into it, rather than designing a molecule from scratch. At its core are protein language models trained on millions of natural sequences: they translate an amino acid chain into a standardized mathematical representation that the model then works with.

Control comes down to two dials. The first sets how strongly to alter the sequence, the second sets which direction to move the size in: making the protein shorter or longer. Architecturally, Raygun is built as an autoencoder, where a variable-length input is compressed into a fixed representation.

• Developer — Rohit Singh's lab, Duke University School of Medicine
• Publication — Nature, July 29, 2026 (DOI 10.1038/s41586-026-10842-8)
• Size reduction — 10–25%, sometimes more than 50%
• Model input — the source sequence, a "noise" parameter, and a target length
• Can also grow proteins beyond their natural size

What the experiments showed

Raygun was tested on three classes of proteins: fluorescent proteins (for imaging), cellular enzymes, and variants of epidermal growth factor (EGF). For EGF — a protein involved in wound healing and in signaling pathways linked to cancer — the team ran receptor-binding experiments and confirmed that the shortened variants retain function.

The key finding is that the miniaturized versions don't lose their working sites. The model introduces significant diversity into the sequence but preserves functional sites and the predicted structural integrity. Unlike models that design proteins from scratch, Raygun relies on an existing natural template — which increases the odds that the shortened molecule will turn out to be functional.

"There's a language that determines how a protein's amino acid sequence gives rise to its shape and function, but scientists don't yet fully understand that language," —

Rohit Singh, assistant professor of biostatistics, bioinformatics, and cell biology at Duke University.

Why shrink proteins

Compact proteins are needed above all for gene therapy: many therapeutic proteins are too large to fit into viral vectors with limited capacity. Shrinking a molecule without losing function removes this constraint and expands the range of accessible targets.

Beyond gene therapy, the authors cite the design of protein sensors and other biotech tools. According to Singh, protein language models work like a translator:

"Protein language models act as a kind of translator: they learn patterns from millions of sequences and link those patterns to biological structure and function," —

Rohit Singh, Duke University.

What this means

Raygun turns protein redesign from slow manual engineering into a controllable process: you set a target length and degree of change — and get a working variant. This brings the "guided evolution" of proteins closer to practical tasks in medicine and biotechnology.

Frequently Asked Questions

What is Raygun?

Raygun is a generative AI model for protein redesign from Duke University, published in Nature on July 29, 2026. It takes a natural protein as a template and creates shortened, lengthened, or heavily altered versions of it while preserving function.

How much does Raygun shrink proteins?

Raygun shortens proteins by 10–25%, and in some cases by more than 50%, while preserving predicted structure and functional sites. The model can also grow proteins beyond their natural size.

Where is Raygun used?

The main application is gene therapy, where compact proteins are easier to package into viral vectors. The authors also cite the design of protein biosensors and other biotechnology tools.