AI tool 'Raygun' shrinks and expands proteins while preserving function, opening new avenues for protein engineering

Scientists report recently in Nature a new AI framework called Raygun that can miniaturize, expand, and modify natural proteins while preserving their function. The work, led by senior author Rohit Singh at Duke University, with first author Kapil Devkota, introduces a generative approach that treats proteins as probability distributions rather than sequences, enabling coordinated insertions, deletions, and substitutions.

Raygun takes a template protein sequence, a noise parameter controlling substitutions, and a target length controlling insertions/deletions. It generates variants in 0.3 seconds—about 100-fold faster than diffusion-based methods. It can shrink proteins by 10–25% (sometimes over 50%), expand them, or introduce extensive sequence diversity, all while preserving predicted structural integrity and functional sites.

The team validated Raygun on fluorescent proteins. Starting from eGFP (238 amino acids) and mCherry (236 amino acids), they generated miniaturized variants as short as 199 and 206 amino acids—shorter than 96% of known fluorescent proteins. Six out of eight tested variants showed fluorescence in human cells. One candidate even carried a non-canonical chromophore sequence, demonstrating exploration beyond typical constraints.

They also applied Raygun to TurboID, a widely used biotin ligase for proximity labeling. Two moderately miniaturized variants (317 and 304 amino acids) retained biotinylation activity. An extremely miniaturized 165-amino-acid variant (roughly half the original size) was expressed but lacked activity, suggesting that highly engineered functions may require additional optimization, such as directed evolution.

In a protein design competition, Raygun expanded the 53-amino-acid epidermal growth factor (EGF) to 55–57 amino acids. Two variants showed stronger binding to the EGFR receptor than wild-type EGF (Kd = 0.274 µM and 0.561 µM vs 0.759 µM), outperforming all other EGF-based approaches in the competition.

Raygun complements de novo protein design by starting from existing proteins and exploring outward, making large-scale insertions, deletions, and substitutions as tractable as point mutations. This could enable redesign of sensors, reporters, and gene therapy payloads where protein size is a constraint—for example, fitting within adeno-associated virus (AAV) packaging limits.

The study used a model trained on only 80,000 proteins, and longer proteins challenged zero-shot reconstruction. The authors note that for highly engineered functions, closing the loop with experimental feedback may be needed. They also highlight biosafety considerations and endorse responsible AI for biodesign principles.

Raygun demonstrates that protein function can be captured in a length-agnostic representation, enabling the kind of coordinated sequence modifications seen in natural evolution. The approach may extend to RNA and other biological sequences.

Reference: Devkota, K., Shonai, D., Mao, J. et al. Miniaturizing and modifying natural proteins with Raygun. Nature (2026). DOI: 10.1038/s41586-026-10842-8