Raygun AI Shrinks Proteins by 10%-25%, Preserving Function Across 701 Million Parameters
Updated
Updated · Nature.com · Jul 29
Raygun AI Shrinks Proteins by 10%-25%, Preserving Function Across 701 Million Parameters
3 articles · Updated · Nature.com · Jul 29
Summary
Raygun encodes proteins as fixed-dimensional probability distributions, letting it generate shorter, longer or heavily edited variants while preserving predicted structure and functional sites; the model can cut length by 10%-25% and sometimes by more than 50%.
Using just noise and target-length controls, the 701 million-parameter framework produces candidates in 0.3 seconds each—about 100 times faster than diffusion-based approaches—and handles substitutions, insertions and deletions in one step.
Cell tests backed the design claims: 6 of 8 miniaturized fluorescent proteins still fluoresced, two shortened TurboID variants retained biotin-ligase activity, and two expanded EGF variants bound EGFR more tightly than wild-type EGF.
The study argues that template-guided AI design can complement de novo protein generation, especially when researchers need to resize existing proteins for uses such as biosensors, proteomics tools or gene-therapy payload limits.
How does Raygun's single-step AI bypass the slow, iterative denoising that plagues current diffusion-based protein design models?
Could shrinking proteins by half using AI solve the biggest size bottlenecks in modern gene therapy delivery?
If an AI can rewrite and miniaturize natural proteins, what hidden side effects might these synthetic variants trigger inside human cells?
Raygun’s 2026 Revolution: AI Miniaturization of Proteins for Gene Therapy and the New Frontier of Biosecurity
Overview
Raygun, introduced in a 2026 Nature publication, is a breakthrough protein design tool that solves gene therapy delivery limits by shrinking proteins while preserving their function. Using a novel probabilistic encoding, Raygun makes insertions and deletions easy and generates new protein candidates in just 0.3 seconds—about 100 times faster than previous methods. It distributes deletions across the protein to maintain structural balance and can even identify and remove redundant domains, as shown with TurboID. However, while Raygun excels at structural design, maintaining complex enzyme activity often requires further lab optimization. The rise of such AI-driven tools also raises biosecurity concerns, prompting new regulatory efforts to ensure safe and responsible use.