AI and Protein Engineering · global
AI “Resizes” Proteins: Raygun Deletes and Adds Amino Acids, but Retaining Function Is the Key Test
This protein generation model does not design molecules from scratch. Instead, it builds on the scaffolds of natural proteins, adjusting their length and sequence. Cell-based and binding experiments show that some products can still function, while also clearly revealing the functional costs of substantial size reduction.
Protein size is not merely a matter of appearance. Molecules that are too large may not fit into viral vectors commonly used in gene therapy; when used as cellular markers or research tools, they may also interfere with the biological processes being observed. A study published in *Nature* presents Raygun, a generative artificial intelligence framework that attempts to add or remove amino acids directly from existing proteins while preserving their folding patterns and key functions as their length changes.
Raygun is built on ESM-2, a protein language model with 650 million parameters. Conventional models represent each amino acid as a set of numerical values, so the overall representation changes in size when protein length changes. The research team instead compresses proteins into fixed-dimensional probability distributions, allowing sequences of different lengths to be compared and sampled within the same space. Users need only adjust the degree of sequence variation and the target length, after which the model can coordinate substitutions, insertions, and deletions simultaneously rather than modifying residues one at a time.
In computational analyses, Raygun could typically shorten proteins by 10% to 25%, with some designs exceeding 50%, and could also enlarge molecules. Many candidate sequences retained their predicted overall structures and functional sites. Each generation took about 0.3 seconds on an NVIDIA A100 chip. However, most of these results remain judgments based on structural models and sequence scores and cannot be equated with proteins folding correctly and functioning in biological systems.
The team therefore selected several types of proteins for experimental testing. Eight shortened fluorescent proteins were expressed in human embryonic kidney cells, six of which produced fluorescence above background levels. The shortest eGFP and mCherry variants contained 199 and 206 amino acids, respectively, making them shorter than 96% of known fluorescent proteins in the database. However, these variants were less bright than the original proteins and would still require further directed evolution and optimization to become practical imaging tools or biosensors.
The enzyme-reduction results more clearly demonstrated the technology’s limits. The researchers first generated 500,000 candidate TurboID biotin ligases and selected 11 for testing after screening for sequence, stability, and structure. Six could be expressed in cells, but only two showed significant biotin ligase activity, and they were only 1% and 6% shorter than the original protein. One version shortened by about half could be expressed but showed no significant enzymatic activity, indicating that preserving shape does not guarantee that complex catalytic functions will also be retained.
Raygun was also used to enlarge epidermal growth factor, EGF, which contains only 53 amino acids. The team screened 10,000 candidates and submitted four for experimental testing. Two bound more strongly to the EGFR receptor than natural EGF, with dissociation constants of 0.274 and 0.561 micromolar, respectively, compared with 0.759 micromolar for natural EGF. This demonstrates that lengthening a protein does not necessarily dilute its function, but stronger binding does not yet indicate better therapeutic efficacy. Cellular signaling, toxicity, and in vivo behavior were not addressed in this experiment.
The model described in the paper was trained on approximately 80,000 UniRef50 protein sequences. Public software records from 2024 describe another pretrained version trained on 95,000 sequences and approximately 2.9 GB in size, suggesting that the public model may differ from the version used in the formal study. At this stage, Raygun is better suited as a starting point for rapidly proposing candidate sequences, followed by task-specific computational screening, protein preparation, cell experiments, and animal validation. If this workflow can be reproduced reliably in the future, shrinking therapeutic proteins, antibodies, or molecular tools could reduce vector and manufacturing constraints. For now, the results demonstrate feasibility while also underscoring that design proceeds far faster than experimental validation.