CasGen

Genome editing draws on an enormous natural reservoir of CRISPR-associated nucleases — Cas9 orthologs alone span thousands of catalogued sequence groups — yet practical work leans on a handful of characterized variants, chiefly SpCas9. That concentration inherits a fixed set of constraints: strict protospacer adjacent motif requirements, off-target cleavage, and protein sizes that complicate delivery. Widening the toolkit means engineering the known enzymes or finding new ones, and both routes are bottlenecked by experimental screening.

CasGen approaches the problem generatively. Developed at the University of Texas at Arlington with the UNT Health Science Center, it is a transformer autoencoder whose latent space is shaped by two auxiliary objectives so that Cas and non-Cas proteins occupy separated, compact regions. Sequences are produced not by sampling a language model left to right but by Bayesian optimization over that structured latent space, with a classifier gating every vector before it is decoded. The design is deliberately narrow: where general protein generators such as ProtGPT2 span the whole sequence universe, CasGen trades breadth for a latent geometry tuned to two enzyme families. It remains a bioRxiv preprint and has not been peer reviewed.

#Key Features

• Classification-regularized latent space: An MLP head trained to separate Cas from non-Cas proteins is optimized jointly with reconstruction, pushing the encoder to make class identity apparent in the latent representation.
• Margin-based regularization: A loss term minimizes intra-class distances and maximizes the distance between Cas and non-Cas embeddings, giving the generator a well-bounded region to sample from rather than an entangled continuum.
• Bayesian latent search with a classifier gate: A surrogate model and acquisition function propose latent vectors; those the classifier calls non-Cas are returned for reproposal, and only Cas-classified vectors are decoded and folded.
• Structure-in-the-loop selection: Predicted fold quality feeds back into the search, guiding it by foldability rather than likelihood alone.
• Novelty screening before analysis: A BLAST pass against reference databases keeps only designs below 50% identity to known proteins, so the pipeline reports new sequences, not memorized ones.

#Technical Details

The architecture pairs a transformer encoder — multi-head self-attention with residual connections, layer normalization, and dropout — to a one-dimensional CNN decoder, following the ReLSO regularized-latent-space paradigm. Sequences are tokenized over the 20 standard amino acids plus start, end, unknown, and padding symbols and standardized to 1,600 positions, covering the full length range of the collected Cas proteins. The objective is a weighted sum of reconstruction cross-entropy, classification cross-entropy, and the margin term. Training used 3,021 Cas9, 287 Cas12, and 597 non-Cas sequences; Cas entries came from InterPro and were kept only after domain verification for RuvC, HNH, and PAM-interacting regions, while non-Cas negatives were drawn from endonuclease, protease, exonuclease, and helicase families. Class-weighted losses based on inverse frequency offset the imbalance.

Generated designs were assessed structurally rather than experimentally. AlphaFold 2 produced five models per sequence with per-residue pLDDT, FoldSeek scored similarity to known Cas structures, and AlphaFold 3 modeled the designs in complex with guide RNA and double-stranded DNA. TM-scores to reference structures fall between 0.70 and 0.85, RMSD below 2.00 Å against the SpCas9 and Nme1Cas9 crystal structures 5F9R and 6DJV, and AlphaFold 3 pTM and ipTM values between 0.70 and 0.90. Designed Cas9 orthologs share 28% to 55% identity with their natural counterparts and Cas12a designs up to roughly 48%. A designed Nme1Cas9 at 49.8% identity retains the catalytic residues of both the HNH and RuvC domains; a designed SpCas9 at 28% identity keeps a functional RuvC domain with two substitutions while contacting the PAM largely through backbone atoms rather than base-specific interactions.

#Applications

The intended use is expanding the pool of candidate genome editors before wet-lab work begins — proposing Cas9 and Cas12 variants that preserve catalytic architecture while diverging substantially in sequence, then triaging them by predicted fold and complex geometry. Divergence at that scale is the raw material for altered PAM preferences, reduced immunogenicity, or smaller delivery-compatible enzymes, which makes the output relevant to synthetic biology groups and therapeutic editing programs assembling ortholog panels.

#Impact

CasGen argues that latent-space regularization can substitute for scale in generative protein design: a corpus of roughly four thousand sequences is orders of magnitude smaller than those behind general protein language models, and the classification and margin terms are what make sampling from it productive. The claims stop at computation. No generated protein was synthesized, and no cleavage, specificity, or off-target activity was measured; the authors identify in vitro and in vivo assays as the next step. In this respect the work sits well behind OpenCRISPR-1, a generated Cas9 that was expressed and shown to edit human cells. The preprint's stated code repository is not publicly accessible, and no weights have been released, so the results cannot currently be reproduced or the model reused.