Raygun

Most generative protein design builds new molecules from nothing, conditioned on a target structure or function. Evolution rarely works that way: it edits what exists, using substitutions alongside insertions and deletions to remodel a template into something of a different size. Computational template-based design has struggled to follow: point substitution is the only edit most methods handle well, and protein language models represent n residues as n vectors, so proteins of different length occupy different-dimensional spaces and cannot be sampled between.

Raygun, from Rohit Singh's and Scott Soderling's labs at Duke University with collaborators at UC San Diego, encodes a protein not as a point but as a probability distribution of fixed dimension. Any sequence of at least 50 residues is partitioned into 50 contiguous blocks; averaging embeddings within each block yields, by a central limit theorem argument, a block-level multivariate normal, and together the blocks define a 64,000-dimensional distribution for the protein. Length therefore becomes a decoding choice, not a property of the encoding.

The designer supplies two parameters: a target output length, setting the scale of insertions and deletions, and a noise level, setting the substitution rate. Generation is single-shot rather than iterative, at roughly 0.3 seconds per sequence, so producing tens of thousands of candidates and filtering hard is practical. Published in Nature in 2026.

#Key Features

• Length-agnostic representation: A fixed-size distribution per protein makes different lengths commensurable, which is what permits large coordinated insertions and deletions rather than substitutions alone.
• Single-shot sampling: Candidates are drawn straight from the template's distribution rather than denoised over many steps, so a heavily modified design costs no more than a conservative one — around 100-fold faster than diffusion-based de novo approaches such as EvoDiff.
• Functional site retention: Miniaturization preferentially removes less-structured regions and preserves annotated active and binding sites at above-chance rates, though those sites are never labelled in training.
• Wet-lab validated across families: Designs were expressed and assayed in cells for fluorescent proteins, a biotin ligase, and a growth factor, not scored by in-silico metrics alone.

#Technical Details

Raygun is an autoencoder over ESM-2 650M embeddings. Parameter-free Reduction and Repetition layers perform the length transforms, while trained T-Block modules — an ESM transformer for global context paired with a 1D convolution for local structure — sit around them and hold most of the parameters. Training is self-supervised: compress a sequence, decompress it to its original length, penalize deviation in embedding and sequence space. Before a campaign the decoder alone can be fine-tuned on related sequences, encoder frozen. The published model has 701 million trainable parameters, trained on 94,734 UniRef50 proteins across 19 length bins from 100 to 1,000 residues; released checkpoints reach 800 million parameters trained on 8.8 million UniRef50 sequences.

Structural evaluation used AlphaFold-3-inferred structures. At length changes near 10%, generated sequences hold a median TM-score around 0.78 against their template with a median pLDDT drop of about 17%; mTOR tolerates a 25% reduction at TM-score 0.69. Across four PFAM families spanning all major SCOP classes, 48.25% of candidates retained the family's PFAM domain by HMMER at 50-200% of median family length. In the wet lab, 8 miniaturized fluorescent proteins were selected from 70,000 samples per template by pseudo-log-likelihood, hmmscan, and a brightness predictor; 6 fluoresced significantly over control, and two mCherry-derived designs at 199 and 206 residues are shorter than 96% of FPbase entries. Of 11 miniaturized TurboID variants, 6 expressed and 2 retained ligase activity; all 4 expanded EGF variants expressed, 2 binding EGFR more tightly than wild type.

#Applications

The clearest use is fitting proteins into spaces they currently do not fit. Gene therapy payloads are bounded by viral vector capacity, so a protein shortened by a quarter can become deliverable where the wild type is not. Imaging and proteomics benefit similarly: smaller fluorescent tags perturb their fusion partners less, and a compact biotin ligase widens where proximity labelling can be placed. The same machinery generates variant libraries around a template for directed evolution, and EGF expansion improved binding without a binder-specific objective.

#Impact

Raygun's durable contribution is conceptual: protein function survives a length-agnostic encoding, which reframes indels as a sampling parameter rather than a combinatorial search problem. That closes a gap between what evolution does routinely and what computational design could express, complementing structure-first de novo methods rather than competing with them. Consistent behaviour on beta-sheet and alpha-helical folds alike sets it apart from design methods that skew toward helical bundles. Limitations are real: hit rates are moderate, structural fidelity degrades as the design departs further from the template, and templates shorter than 50 residues fall outside the encoding. Terms differ by artifact. The repository is CC BY-NC-SA 4.0 with additional grants: academic, non-profit, and government users may commercialize downstream outputs, while for-profit organizations get a 60-day evaluation trial and then need a separate agreement. The checkpoints are not in the repository and do not inherit that restriction — all three Zenodo model records, plus the training-data and benchmarking deposit, are CC BY 4.0.