Most coverage of AI in biology focuses on one capability: creating proteins from scratch. A new tool published in Nature does something different, and in some ways more useful. It takes proteins that already exist and resizes them, making them smaller or larger, without breaking what makes them work in the first place.
The tool is called Raygun.
Why Editing Proteins Is Harder Than Designing Them
To understand why Raygun matters, it helps to understand the problem it solves. Over the past several years, AI-based protein language models have become remarkably capable. These systems are trained on millions of protein sequences, learning the underlying patterns of how proteins are structured, much like a language model learns grammar from text. They can now generate entirely new proteins, including antibodies with clinical potential.
But designing a new protein and modifying an existing one are two different challenges. When researchers tried to use existing AI systems to resize or alter natural proteins, a persistent problem emerged: the modifications tended to disrupt either the protein’s structure or its function. The shape changed. The behavior changed. The protein stopped being what it was supposed to be.
This is not a minor technical inconvenience. For many real-world applications in biotechnology, researchers do not want a completely new protein. They want a version of a trusted, well-understood protein that fits a different context. As Rohit Singh, a computational biologist at Duke University in Durham, North Carolina and a co-author of the study, explains: sometimes a smaller protein can do things or fit into places where a larger protein cannot.
How Raygun Approaches the Problem Differently
Most protein language models represent proteins as sequences of amino acids of varying lengths. The challenge with this approach is that changing the length of a sequence tends to cascade into structural and functional changes. The sequence is not just a list; it encodes relationships between parts of the protein that determine how it folds and behaves.
Raygun, which is built on top of a large language model for protein design called ESM-2, takes a more mathematical route. Rather than working directly with amino acid sequences, it breaks each protein into segments and translates the information in each segment into numerical patterns. From those patterns, it constructs a standardized representation of the protein that follows a defined set of rules. The model learns those rules, and that learning is what allows it to generate the protein at different sizes while preserving its structural integrity.
The analogy to natural evolution is instructive. Raygun uses some of the same basic operations that evolution uses to modify proteins over time: adding or removing individual protein subunits, or substituting one subunit for another. Evolution does not usually redesign proteins from scratch. It tinkers. Raygun is, in a sense, a controlled version of that tinkering.
Fajie Yuan, a computational biologist who works on protein language models at Westlake University in Hangzhou, China, and who was not involved in the research, describes Raygun as a meaningful and creative first step toward editing existing proteins. The framing is deliberate: this is a beginning, not a finished solution. But it points in a direction that existing tools have not been able to reach.
What This Means Beyond the Lab
Here is what most coverage of protein AI misses: the distinction between creation and modification is not just technical. It reflects a deeper question about how science actually works in practice.
Researchers and biotechnology developers often build on what they already know. A protein with a well-documented safety profile, a known mechanism of action, or an established manufacturing process represents years of accumulated knowledge. The ability to adjust that protein, to make it smaller so it can cross a biological barrier, or larger so it stays in circulation longer, without starting over from scratch, is not a minor convenience. It is a different kind of capability entirely.
This is where AI tools like Raygun augment scientific teams in a concrete way. The volume of possible protein modifications that could be explored manually is enormous. Computational tools handle that space systematically, freeing researchers to focus on the modifications that are most promising and to evaluate results rather than generate candidates by hand.
The study is published in Nature, and the research represents an early proof of concept. The path from a tool that can resize haemoglobin in a controlled setting to a tool routinely used in drug development is long. But the direction is clear: AI in biology is moving from pure generation toward something more like editing, and that shift opens a different set of possibilities.
In Short
Raygun is an AI tool that can make natural proteins smaller or larger without disrupting their structure or function. It does this by translating protein information into numerical patterns and learning the rules that govern protein architecture, rather than working directly with amino acid sequences. The significance is not just technical: the ability to modify trusted, existing proteins rather than design new ones from scratch addresses a gap that previous AI systems could not fill, and opens practical possibilities for biotechnology applications where size and fit matter.
Based on reporting from Nature: Machine Learning.